Papers

RSS

Daily arXiv scrape.

75 papers across 15 days.

Query syntax

Examples

market designboth words, anywhere in a paper
"market design"that phrase, in any case
swarm OR stigmergyeither word
agent NOT marketthe first without the second
title:alignmentone field only
(a OR b) AND cbrackets group

Fields

title:the paper's title
author:any author's name
summary:the one-sentence summary
question:any of its open questions
anchor:the nearest canon paper's title
id:the arXiv id
date:the day it was kept, YYYY-MM-DD

Two words with no operator between them mean both. A term matches part of a word, so agent finds “agentic”. Your browser runs the query. It reaches no server.

View

Papers kept per day, 2026-08-20 to 2026-09-03 (n = 15 days, 75 papers)

Source: Gigascale-Labs/las-new-papers, read daily.

2026-09-03

239 fetched · 200 screened · 4 relevant · 3 kept

A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms

Davide Paglieri, Logan Cross, Tim Genewein, Joel Z. Leibo, Nenad Tomasev, Alexander Sasha Vezhnevets

2609.04170 · nearest in the canon: Multi-Agent Risks from Advanced AI

Reports a case study of 100 autonomous LLM agents proving math conjectures, where cheating emerged and spread via shared infrastructure. Other agents formed a counter-response, auditing proofs, alerting peers, and staging boycotts without external intervention. Proposes framing shared agent infrastructure as a knowledge commons and applying institutional mechanisms like graduated sanctioning.

6 open questions
  • Do graduated sanctioning and collective-choice rules actually stop exploit spread in an agent swarm? The paper proposes these institutional mechanisms but does not test them.
  • How does the visibility of shared channels change the balance between cheating and whistleblowing? The case study uses transparent channels only and does not vary channel design.
  • Does cheating spread the same way at swarm sizes other than 100 agents?
  • Does the same cheating-and-whistleblowing dynamic appear in tasks other than formal mathematical proof, where the evaluation system offers different exploits?
  • Which agent-level signals predict which agents adopt an exploit and which become whistleblowers?
  • Do exploit contagion and norm enforcement occur in swarms built on different underlying models, or is the behaviour specific to the models used here?

Inferred Generative-Process Diversity Predicts Correlated Failure Across Language Models

Ross Tieman, Evan Markou

2609.03422 · nearest in the canon: Beyond Single-Agent Safety: A Taxonomy of Risks in LLM-to-LLM Interactions

Proposes generative-process diversity, measured via compression distance between model outputs, as distinct from semantic diversity. Tests this measure across 38 language models and ten benchmark families, finding it predicts correlated failure between model pairs. Reports a cross-benchmark partial rank association of -0.216, negative on all ten benchmarks, beyond semantic similarity and capability.

6 open questions
  • Does inferred generative-process diversity predict correlated failure in multi-agent systems where models interact, rather than in independent model pairs on benchmarks?
  • Does selecting a population of language models to maximise Normalised Compression Distance reduce correlated failure compared to selecting on capability or semantic similarity?
  • How sensitive is the diversity measure to the choice of compressor, output length, and sampling temperature?
  • Which properties of training data, architecture, or post-training cause the observed population structure among models?
  • Does generative-process diversity predict correlated failure beyond model pairs, for groups of three or more models voting or deliberating together?
  • Does the measure predict correlated failure on adversarial or safety-relevant inputs, not only the ten benchmark families used?

GPS-Bench: A Governance Policy Benchmark for Automating Policy Analysis

Linh Le, Melanie Bui, My Chiffon Nguyen, Zachary Schlosser, David Williams-King

2609.03553 · nearest in the canon: Habermolt: Delegating Deliberation to AI Representatives

Introduces GPS-Bench, a benchmark linking policies to actors, actions and impacts using legislative, lobbying and regulatory records. Actors are reconstructed from dated evidence rather than prompted personas, with human-annotated gold and LLM-labeled silver data. Compares joint reasoning, independent and communicating agent modes, graph methods and fine-tuning for predicting policy impacts and coalition formation.

2026-09-02

205 fetched · 200 screened · 10 relevant · 3 kept

Competitive Market Behavior of LLMs

Pawel Struski, Jakub Swistak, Inez Okulska, Przemyslaw Biecek

2609.02580 · nearest in the canon: Group size effects and collective misalignment in LLM multi-agent systems

Replicates human double-auction market experiments with LLM agents instead of humans. Measures convergence to equilibrium and allocation efficiency across model families. Analyzes chain-of-thought traces linking trade decisions to shifts from strategy to urgency.

5 open questions
  • Double auctions with LLM agents converge slower or not at all towards equilibrium, but which mechanism variants restore convergence remains untested.
  • Trading behaviour differs across model families and market roles, but the cause of this heterogeneity is unidentified.
  • Mixed markets of humans and LLM agents are untested, so it is unknown whether human traders correct or amplify the inefficiency.
  • Chain-of-Thought traces link trade execution to a shift from strategic language to urgency language, but whether this link is causal is unresolved.
  • Results come from a double auction, so it is open whether LLM agents also fail to reach efficient allocations in other mechanisms such as posted-price or call markets.

Collective creativity in hybrid societies

Mason Youngblood, Katie Mudd, Manuel Anglada-Tort, Cameron Jones, Elena Miu, Diana Omigie, Margaret Schedel

2609.02620 · nearest in the canon: AI agents can coordinate beyond human scale

Argues creativity under generative AI should be measured as a property of hybrid human-AI populations, separating novelty from diversity. States AI-assisted ideation raises individual novelty while narrowing aggregate diversity. Claims composition of mixed human-AI groups determines whether individual gains accumulate without eroding collective diversity.

5 open questions
  • Which mixtures of humans and generative models let novelty gains accumulate without reducing population-level diversity?
  • How does the network structure connecting human and model agents change the diversity of cultural artifacts a hybrid collective produces?
  • How do human and model search strategies differ in ways that make mixed groups out-diversify single-type groups?
  • Do machine-discovered solutions persist in human culture over long periods after they enter it?
  • Which measures separate novelty of single artifacts from diversity of artifact populations in observed cultural output?

Strategic Centrality and the Emergence of Core-Periphery Networks

Itai Arieli, João Correia-da-Silva, Wade Hann-Caruthers, Anna Rubinchik

2609.02357 · nearest in the canon: The Moltbook Observatory Archive: an incremental dataset of agent-only social network activity

Studies a network-formation game where agents sponsor links to maximize weighted centrality under linear cost. Proves every Nash equilibrium network is core-periphery in structure. Shows welfare-maximizing networks are always empty or complete.

6 open questions
  • Do learning agents that adjust links over time converge to the core-periphery equilibria, and which core size do they select?
  • Does the core-periphery result survive when link costs are convex or heterogeneous across agents instead of linear and uniform?
  • The paper assumes weights on walk counts are positive and weakly decreasing; what equilibrium structures appear when weights are non-monotone?
  • The welfare-maximizing network is empty or complete while equilibria are core-periphery; which transfers or link subsidies close this gap?
  • Do observed online social and agent-interaction networks show the exact core-periphery pattern predicted here, in which every node links to every core node and to nothing else?
  • Do LLM agents that pursue centrality in a link-formation task reproduce the predicted concentration of links on a dense subset?

2026-09-01

295 fetched · 200 screened · 13 relevant · 4 kept

Mechanism Design for Alignment and Control

Dirk Bergemann, Andrew Koh, Stephen Morris

2609.01595 · nearest in the canon: Multi-Agent Risks from Advanced AI

Develops a mechanism-design framework for AI agents whose alignment and capabilities are unknown, deriving a revelation principle and implementability conditions. Applies the framework to sandbagging, peer scoring, reward coupling for multi-agent competition, and scalable oversight.

5 open questions
  • Do current LLM agents actually respond to the proposed honesty-and-obedience mechanisms as the model predicts? The framework derives implementable policies but tests them only on stylized examples.
  • Does peer scoring discipline sandbagging when the agents are LLMs that can coordinate with each other? The paper treats peer scoring as a stylized example and does not test collusion between scored agents.
  • What happens to the revelation principle when capabilities can be counterfeited as well as concealed? The characterization relies on the one-sided imitation structure.
  • Do the conditions for eliciting higher-order beliefs to discipline multiple agents hold at population scale, with hundreds or thousands of agents rather than a few?
  • How does the alignment-interpretability trade-off appear in measured behaviour of deployed agents, where the two are substitutes in the instrument but complements in value?

When Guardrails Look Effective: Construct Validity Failures in LLM Agent Commerce Evaluation

Peiying Zhu, Sidi Chang

2609.01519 · nearest in the canon: On the limits of agency in agent-based models

Audits a buyer-seller LLM agent market simulation and finds welfare gains from marketplace guardrails shrink or reverse once schema and choice-procedure asymmetries and generation variance are controlled. Proposes a construct-validity contract (incentive validity, protocol isolation, stochastic stability, welfare accounting) to flag when such simulated market claims are invalid or inconclusive.

5 open questions
  • Do language-model sellers in a buyer-seller testbed respond monotonically to profit pressure in the prompt?
  • How many generations per profile-condition are needed before guardrail welfare effects become stable in a language-model market simulation?
  • Does the construct-validity contract of incentive validity, protocol isolation, stochastic stability and welfare accounting apply to other agent market simulations, for example labour or advertising markets?
  • Do guardrail welfare effects reported for small models hold at frontier model scale?
  • Which welfare accounting rules identify guardrail effects when the seller is not a first-best profit maximizer?

Data-Driven Persona-Conditioned Agents for A/B Test Simulation

Ziyad Benomar, Weronika Łajewska, Leonardo Perelli, Saab Mansour

2609.01038 · nearest in the canon: The Moltbook Observatory Archive: an incremental dataset of agent-only social network activity

Builds LLM agents conditioned on data-driven personas from real user behavioral signals to simulate A/B test outcomes. Achieves 0.75-0.90 directional accuracy across 40 tests, studying persona data source, behavioral depth versus population diversity, and subsampling efficiency.

5 open questions
  • Directional accuracy reaches 0.75-0.90 on a benchmark of 40 A/B tests, but the abstract does not say how accuracy behaves on effect sizes near zero or on tests with small measured deltas. How does simulation accuracy vary with the true effect size?
  • The personas come from anonymized behavioral data held by the paper's organisation. Can personas built from public behavioural traces, such as public forum or review activity, reach comparable directional accuracy?
  • The framework studies the trade-off between per-persona behavioral depth and population diversity. What population size and subsampling scheme minimises simulation cost for a target accuracy?
  • Persona-conditioned agent populations may collapse onto a few response modes. How much of the simulated population's variance comes from persona conditioning rather than from the base model's default behaviour?
  • Directional accuracy is measured against completed A/B tests. Does using simulated pre-screening to select which experiments to run change the set of product decisions a team makes?

AI and the Economy: An Economic Examination of Production, Distribution, Firms, Labor, and Welfare

Ali Zeytoon-Nejad

2609.01263 · nearest in the canon: An Economy of AI Agents

Essay analyzes AI's economic role as capital, labor substitute, general-purpose technology, and infrastructure. Discusses effects on productivity, labor markets, competition, market concentration, and welfare, and outlines policy considerations.

5 open questions
  • Which of the five productivity channels named for AI (automation, augmentation, optimization, prediction, innovation) accounts for measured output gains, and in what proportion?
  • Does AI adoption raise market concentration in the industries that adopt it fastest?
  • When does AI act as a substitute for labour and when as a complement, at the level of occupations and tasks?
  • Which of the proposed policy measures for distributing AI gains actually change distributional outcomes, and by how much?
  • Does treating AI as economic infrastructure rather than as capital or synthetic labour change the predictions of a growth model?

2026-08-31

267 fetched · 200 screened · 12 relevant · 4 kept

CHASE: How Content Ecosystems Are Reshaped When Ranking Is the Only Target

Qianwen Gao, Zichang Su, Yiwen Hou, Arlen Kumar, Leanid Palkhouski

2608.30466 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web

Introduces CHASE, a simulation framework where content creators repeatedly rewrite documents to exploit an LLM ranking signal over 20 rounds. Measures declining alignment between ranking-favored content and independently judged quality across six domains (mean Spearman rho change -0.068). Validates the ranking proxy against citation rates in grounded LLM responses (AUC 0.853).

6 open questions
  • Ranking serves as a proxy for source visibility, validated against citations in grounded generated responses with a rank-citation AUC of 0.853 plus or minus 0.093 across six domains; does the proxy hold for other retrieval stacks and citation behaviours?
  • Quality-ranking alignment falls in all six domains over 20 rounds of adaptation, but the ranking signal stays fixed; what happens when the ranker updates in response to creator adaptation?
  • Ecosystem dynamics depend on the domain; which domain properties predict the size of the alignment loss?
  • The simulation assumes every creator adapts to the ranking signal; how do outcomes change when only a fraction of creators optimize and the rest do not?
  • Which interventions on the ranking signal limit content homogenization under repeated creator adaptation?
  • Do human content creators adapt to LLM ranking signals in the way the simulated creators do?

Tariff Threats, Macroeconomic Expectations, and Policy Communication Strategies: Experiments Based on a Multi-Agent System

Jianhao Lin, Lexuan Sun, Yixin Yan

2608.30522 · nearest in the canon: AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

Builds persistent LLM agents calibrated on 300 Michigan Survey households, exposed to simulated social-media messages over months. Tests how message features (immediacy, rate salience, narrative, sender identity) shape simulated inflation/unemployment expectations and their dispersion. Runs a second experiment on whether central-bank communication coordinates beliefs.

6 open questions
  • Do the simulated belief responses to tariff communication match human survey responses outside the Liberation Day announcement period?
  • How does the number of agents affect the reproduced distributional and demographic patterns, given the fixed panel of 300 households?
  • Which message features drive belief dispersion when they vary independently, rather than jointly as in the reported experiments?
  • Do the calibrated agents reproduce the demographic patterns they fail to match, and which subgroups do they miss?
  • Does the central-bank coordination effect on beliefs persist when agents also observe conflicting senders on social media?
  • How sensitive are the results to the choice of underlying language model?

One AI Signal, Many Human Judgments: A Bayesian Cascade Analysis of AI-based Credibility Indicators in Online Information Spread

Zhuoran Lu, Weilong Wang, Yangyang Yu, Xinru Wang, Zhuoyan Li, Zhiwei Liu, Sophia Ananiadou

2608.30311 · nearest in the canon: Habermolt: Delegating Deliberation to AI Representatives

Extends the Bayesian cascade model of social learning to include an AI credibility indicator as a shared public signal. Identifies a preservation-correction trade-off where reliance on AI can lock in either correct or incorrect crowd judgments. Calibrates the model with human-subject news-veracity data and runs simulations showing over-reliance on a weak AI harms crowd informativeness, while diversifying AI signals helps.

5 open questions
  • How well does the extended Bayesian cascade model predict cascade outcomes in live social media settings, rather than in calibration on news veracity judgment data?
  • Which schemes for diversifying AI credibility signals across users keep a crowd informative, and how do they compare on preservation versus correction of errors?
  • How does the preservation-correction trade-off change when agents in the cascade are LLM agents rather than Bayesian updaters with fixed weights?
  • What determines the spread in individual reliance on AI, from discounting the indicator to cascading on it, and can that heterogeneity be estimated without human-subject experiments?
  • How do network topology and the visible length of public history alter the Gateway condition threshold at which private impressions stop entering the cascade?

Reproducible macroscopic dynamics in a closed-loop human-AI learning system

Minlin Wu, Xu Fang, Yicheng Zhang, Chenyu Zhou, Zhiyi Liu

2608.30946 · nearest in the canon: Agentic Microphysics: A Manifesto for Generative AI Safety

Analyzes 297,915 learners' adaptive-tutoring histories to define semantic order variables and test them across disjoint cohorts. Identifies reproducible basin-like flow and metastable-like kinetics in aggregate learner state. Recovers population drift with a four-term mechanism (r=0.946) and shows self-supervised models learn similar leading-order dynamics.

6 open questions
  • Null-referenced corrections retain only partial-amplitude excess-field structure. What method calibrates the excess-field amplitude fully?
  • Do the semantic order variables and the basin-like flow reproduce on closed-loop human-AI systems other than adaptive tutoring, such as recommender or assistant logs?
  • Two models share leading population drift but keep model-specific residual directions. What causes the residual directions to differ?
  • Can a simulated closed-loop agent-tutor population reproduce basin-like flow and metastable-like kinetics in the same order variables?
  • Shuffled-order training reverses learned-plane flow and support-alignment randomisation reduces inward transport. Which properties of the interaction sequence carry the flow direction?
  • Does the identified leading-order effective field predict the effect of changing the tutoring policy on the population state?

2026-08-30

139 fetched · 139 screened · 6 relevant · 2 kept

Integrating adaptive human behavior into epidemic models with large language models

Yicheng Mao, Haoyang Li, Rob Deardon, Hongru Du

2608.29535 · nearest in the canon: On the limits of agency in agent-based models

Couples LLM-inferred behavioral responses with mechanistic epidemic models to generate age-structured contact matrices. Tests the framework (GABLE) on COVID-19 in France, comparing forecasts against mobility-driven matrices. Also evaluates prospective policy scenarios by projecting behavioral and epidemic responses to candidate interventions.

6 open questions
  • GABLE was applied to COVID-19 in France only; does an LLM behavioral layer reproduce contact structures in other countries and other epidemics?
  • How much of the forecasting gain comes from the LLM's own knowledge of COVID-19 in France rather than from the behavioral mechanism, given the pandemic falls inside training data?
  • Do different LLMs produce the same contact matrices, and does monoculture across models narrow the range of projected epidemic outcomes?
  • Does the LLM behavioral layer reproduce heterogeneity across subpopulations, or does it collapse to a single average behavioral response?
  • How accurate are the prospective policy projections when compared against outcomes of policies not yet implemented at prompt time?
  • How do LLM-generated contact matrices compare against survey-measured contact matrices such as CoMix, rather than against mobility-derived matrices?

Credibility in school choice

Camilo J. Sirguiado, Jiarui Xie

2608.29597 · nearest in the canon: Multi-Agent Risks from Advanced AI

Defines credibility of a school-choice mechanism as immunity to undetected deviation by the designer. Shows deferred acceptance is credible only when students observe enough of the assignment, and ranks mechanisms by credibility. Finds deferred acceptance's ranking against top trading cycles and efficiency-adjusted deferred acceptance depends on how much of the assignment students observe.

4 open questions
  • Do the credibility rankings of school choice mechanisms hold when participants are AI agents that report on behalf of students?
  • How much of the assignment must students observe before deferred acceptance becomes credible in practice, and can that threshold be estimated from real assignment disclosures?
  • Can students detect designer deviations from an announced mechanism by pooling their partial observations of the assignment?
  • What are the welfare costs of choosing a more credible mechanism over a less credible but more efficient one?

2026-08-29

151 fetched · 151 screened · 9 relevant · 4 kept

Sustained Heterogeneity: an emergent collective mechanism in LLM-driven traffic

Yujun Qi, Yangyang Guan

2608.29174 · nearest in the canon: Modeling Earth-Scale Human-Like Societies with One Billion Agents

Deploys 22 LLM agents as real-time speed controllers on a ring road, observing emergent stop-and-go traffic waves. Identifies a persistent divergence phenomenon (Sustained Heterogeneity) and maps a density-dependent critical penetration fraction for instability onset across four traffic densities.

6 open questions
  • Sustained Heterogeneity is measured with one LLM acting as controller across a temperature sweep, so does the same per-cycle divergence appear across other model families and model sizes?
  • The paper concludes stability must be enforced at the dynamics layer, but does not test which dynamics-layer controller suppresses the three-stage cascade of drift, gap erosion and nonlinear braking.
  • Does letting the LLM agents exchange messages or share a common speed protocol remove the systematic divergence, or does divergence persist under coordination?
  • The phase boundary p_c(rho) is mapped on a 230 m single-lane ring, so how does the critical LLM penetration fraction change on open roads, merges and multi-lane networks?
  • Can Sustained Heterogeneity be detected from aggregate traffic measurements alone, without access to per-agent chain-of-thought?
  • The controls replace human drivers with IDM and OV models, so does the density-dependent transition hold when real human drivers share the road with LLM controllers?

Content Exploration Beyond the Feed: Creator Supply and the Shared Corpus

Yuanyuan Shen, Yiren Yan, Wenjie Li, Chunhui Zhu

2608.29430 · nearest in the canon: The Moltbook Observatory Archive: an incremental dataset of agent-only social network activity

Analyzes four experiments on a short-video platform measuring how recommender exploration budgets affect creator posting behavior and viewer consumption. Finds production exploration raises creator posting rates and participation, while viewer-side experiments cannot detect the full corpus-level effect within a limited time window.

5 open questions
  • The three-week co-diverted experiment cannot determine the sign of the eventual corpus effect, so what experiment duration and corpus turnover rate are needed to identify it?
  • Exploration budget objectives in the reviewed publications omit creator response, so how would a budget allocation rule that includes creator supply change corpus growth?
  • Viewer-side A/B tests cancel the corpus effect when both arms consume the same corpus, so which estimator designs recover the corpus effect under partial corpus sharing?
  • Confidence intervals for the eventual corpus effect can lack a finite upper endpoint before the corpus path bends, so which observable aggregate signals indicate that the bend has started?
  • Production exploration raises videos posted per creator by 8.55% on one short-video platform, so does the same creator response appear on other content platforms?

The Role of Network Topology and Opponent Information in Shaping Cooperation in Multi-Agent Reinforcement Learning Systems

Seongho Son, Stephen Hailes, Mirco Musolesi

2608.28977 · nearest in the canon: Multi-Agent Risks from Advanced AI

Studies deep reinforcement learning agents playing Iterated Prisoner's Dilemma on graph topologies where neighbors are opponents. Finds neighbor count and average path length drive cooperation emergence, and opponent identity information reduces cooperative strategy spread.

5 open questions
  • Deep reinforcement learning agents on graphs cooperate more or less depending on the number of neighbours and the average path length, but the reported experiments cover only the two-player Iterated Prisoner's Dilemma; does the same topology effect hold for other games such as public goods or stag hunt?
  • Providing agents with opponent identity hinders cooperation; which learned mechanism produces this effect?
  • Partner selection fosters mutual cooperation by limiting opponent-pool diversity; does this hold when the graph rewires over time rather than staying fixed?
  • Do the topology findings scale from the graph sizes tested to populations of thousands of learning agents?
  • Do the same cooperation patterns appear when the agents are LLM-based rather than deep reinforcement learning policies?

Optimally Selecting Representative Agents from a Metric Space

Benjamin Cookson, Eva Deltl, Yeeseok Oh

2608.29097 · nearest in the canon: Virtual Agent Economies

Proves a tight 2-approximation bound for proportionally fair clustering satisfying the Droop core property in metric spaces. Shows centers can be restricted to agent locations, resolving an open plurality problem using Scarf's theorem.

4 open questions
  • Does a 2-Droop core clustering exist when the set of feasible center locations does not include every agent location?
  • Can a clustering in the 2-Droop core be computed in polynomial time?
  • How close do fast heuristic clustering algorithms come to the 2-Droop core guarantee on real metric datasets?
  • How reliable are language-model-generated proofs in this problem class when authors verify them independently?

2026-08-28

203 fetched · 200 screened · 9 relevant · 3 kept

Emergent aggregation from collective foraging

Gorka Muñoz-Gil, Andrea López-Incera, Vide Ramsten, Giovanni Volpe, Thomas Müller, Hans J. Briegel

2608.28046 · nearest in the canon: Virtual Agent Economies

Reinforcement-learning foragers optimizing individual reward for finding targets develop spatial aggregation without direct social reward. A crossover in visual range separates individual search from collective search behavior. A first-passage analytical model reproduces this transition.

6 open questions
  • Aggregation emerges when foragers see only conspecifics and never the targets; what happens when agents perceive both targets and conspecifics?
  • The crossover depends on visual range with replenishable targets; how does target replenishment rate, density and patchiness shift the onset of aggregation?
  • Does indirect, resource-driven reward also produce aggregation in populations of language-model agents rather than reinforcement learning walkers?
  • Can an observer detect the crossover from individual to collective search using only aggregate spatial statistics, without access to agent policies?
  • How does the aggregation transition change when agents are heterogeneous, for example with mixed visual ranges or mixed learned and random-walk policies?
  • Does the analytical first-passage model reproduce the transition in dimensions or geometries other than the one tested?

Performative Privacy: When Differential Privacy Maximizes Utility

Uddalak Mukherjee, Edwige Cyffers, Yann Chevaleyre

2608.28198 · nearest in the canon: Multi-Agent Risks from Advanced AI

Introduces performative privacy, where data leakage by a mechanism reduces future participation of agents contributing data. Analyzes a mean-estimation model where differentially private mechanisms trade estimation noise against retained participation over time. Shows theoretically and numerically that a finite privacy budget can outperform non-private estimation when the leakage-participation feedback is strong.

5 open questions
  • Does a finite privacy budget still beat non-private estimation for tasks beyond mean estimation, such as regression or classification?
  • How does the optimal privacy budget change when agents differ in their sensitivity to leakage instead of sharing one departure rule?
  • Can an adaptive privacy budget, tuned over rounds from observed participation, outperform a fixed budget?
  • How strong is the real feedback loop between data leakage and user participation in a deployed system?
  • What happens to participation and utility when several platforms with different privacy budgets compete for the same agents?

GOD: Govern, Observe, and Direct - A Real-Time Control Room for Agent Societies

Yige Luo, Ran Guan

2608.27992 · nearest in the canon: Agentic Microphysics: A Manifesto for Generative AI Safety

GOD provides a local control room for inspecting and intervening in generative-agent simulations. It supports real-time queries, interventions, and portable experiment packs. Evaluation on 15 run slots checks whether interventions and state queries match recorded outcomes.

5 open questions
  • Does the command and artifact loop scale beyond 15 run slots and beyond Smallville-style scenarios?
  • The evaluation counts whether agents recorded the commanded destination, not whether interventions change population-level outcomes; which aggregate metrics respond to operator interventions?
  • Why did 6 of 84 target-agent checks and 13 of 182 state answers fail to match?
  • Which failure modes of agent societies, such as collusion or information cascades, does a spatial replay and Ask interface make visible, and which stay hidden?
  • Do operators actually reach correct conclusions faster with the control room than with raw logs?

2026-08-27

215 fetched · 200 screened · 12 relevant · 6 kept

LiveSim: Simulating Environment-Shaped Users in Multi-Agent Live-Stream Ecosystems

Jiaqi Xu, Yiran Qiao, Jing Chen, Qiwei Zhong, Xiang Ao, Xueqi Cheng

2608.26849 · nearest in the canon: On the limits of agency in agent-based models

Presents LiveSim, an LLM framework that simulates live-stream users as editable behavioral hypotheses refined through trajectory comparisons. Accumulates environment-behavior patterns in a collective memory to support ecosystem-level analysis of risk evolution and platform interventions.

5 open questions
  • Does a simulator that refines user hypotheses from observed trajectories transfer to platforms other than live streaming, such as social feeds or marketplaces?
  • How much of the reported fidelity gain comes from the collective behavioral memory versus the per-user hypothesis editing?
  • Do platform interventions tested in the simulation produce the same effect sizes when applied on a live platform?
  • Which aggregate signals from interaction logs are enough to detect risk evolution without access to individual user profiles?
  • Does the collective behavioral memory cause simulated users to converge to a monoculture of behavior over long runs?

AI agents in Algorithmic Electricity Markets: On the Emergence of Tacit Collusion

Jakub Seredyński, Georgios Tsaousoglou

2608.26896 · nearest in the canon: Multi-Agent Risks from Advanced AI

Models strategic electricity market bidding as a repeated game with imperfect monitoring, using multi-agent reinforcement learning to represent participants. Proposes multi-dimensional criteria beyond profit-vs-Nash comparisons and finds cases where agents learn supra-competitive outcomes consistent with tacit collusion, without being instructed to collude.

5 open questions
  • Do the tacit collusion outcomes persist when the number of participants in the electricity market rises beyond the small oligopoly modelled here?
  • Which market design changes, such as bid disclosure rules or price caps, reduce supra-competitive outcomes among learning-based bidding agents?
  • Do the proposed multi-dimensional collusion criteria detect tacit collusion in real bidding data from operating electricity markets?
  • Do agents driven by large language models rather than reinforcement learning reach the same supra-competitive outcomes in repeated electricity market bidding?
  • How does the imperfect public monitoring assumption change the results if agents observe more or less about rivals' actions?

Token-Level Advertising

Hanbing Liu, Bowei Zhang, Changyuan Yu, Yinyu Ye, Qi Qi

2608.27382 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web

Proposes LAMA, a token-level advertising auction where advertisers report continuation values that shape a platform's generation via a latent mixture. Proves Markov dominant-strategy incentive compatibility and individual rationality, near-optimal KL-regularized welfare, and reports gains in welfare and revenue on commercial-search query experiments.

6 open questions
  • LAMA is tested on commercial-search query splits only; how does token-level advertising behave on other query populations, such as advice-seeking or health queries?
  • The mechanism assumes advertisers report local continuation values truthfully under Markov DSIC; what happens when many advertisers bid strategically through a learned bidding agent instead?
  • Response quality is reported as maintained at the user-facing level; does repeated token-level advertiser influence shift user beliefs or choices over multiple interactions?
  • Token-level advertising embeds advertiser influence into generation with no predefined slot; can a third party detect and measure the amount of paid influence in an output from the text alone?
  • How does platform welfare and revenue under LAMA distribute across advertisers of different budget sizes?
  • LAMA is evaluated as a single platform mechanism; what happens to welfare and price levels when several competing generative platforms run token-level auctions over the same advertiser pool?

When Interference Graphs Evolve: Doubly Robust Estimation of Dynamic Peer Effects

Xiaojing Du

2608.27187 · nearest in the canon: AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

Introduces a controlled contrast framework separating own-treatment, peer-exposure, and network-evolution effects when interaction graphs change over time. Develops the DynaNet-DR doubly robust estimator, consistent if either outcome regression or propensity model is correct, tested on semi-synthetic and MathOverflow data.

5 open questions
  • Does the DynaNet-DR estimator recover peer effects when the network summary is not sufficient, or when unmeasured confounders drive both edge formation and outcomes?
  • How do the summary-indexed contrasts compare against estimands defined by explicit counterfactual edge generation?
  • Do peer-effect estimates on evolving graphs hold for populations of AI agents, where interaction graphs change faster than in human forums?
  • Which choices of clipping threshold and finite-sample stabilization control the bias-variance trade-off of the estimator?
  • Does the estimator remain consistent when the weak dependence assumption fails, for example under strong network-wide cascades?

Risks and Controls for Multi-Agent Systems: an analytical framework for deployment of AI agents across organisational boundaries

Alistair Reid, Simon O'Callaghan, Dustin Venini, Liam Carroll, Tiberio Caetano

2608.26626 · nearest in the canon: Multi-Agent Risks from Advanced AI

Proposes a framework of deployment tiers (singular, federated, open) for reasoning about risks and controls when AI agents from different organizations interact. Identifies failure modes and controls per tier, and flags governance gaps where no single actor can act.

5 open questions
  • Which failure modes actually appear when agents from different organisations interact under only voluntary shared standards? The framework names failure modes per governance tier but reports no measurements.
  • Can the three governance tiers be distinguished from observable interaction logs alone, without knowing who deploys each agent?
  • Which controls listed for federated and open environments reduce failure rates when applied, and by how much?
  • Where no actor is positioned to apply a control, what collective action mechanism gets organisations to adopt it voluntarily?
  • How do organisations deploying agents across their perimeter today distribute across the three tiers?

Five Primitives for Governing Autonomous AI Agents at Runtime

Jiten Oswal, John Cadeddu

2608.26696 · nearest in the canon: Regulating AI Agents

Argues agent governance is a runtime problem and derives five primitives (discovery, identity, governance, attestation, supply chain) for mediating agent actions. Describes an implementation with policy mediation, per-tenant authorization, and a signed action ledger, four of five primitives built in private pilots.

6 open questions
  • How much latency does a runtime enforcement point on the critical path add per agent action, and how does the overhead scale with the number of agents?
  • Does fail-closed mediation turn availability incidents into denial cascades across a population of interdependent agents?
  • How can a per-tenant action vocabulary stay complete when agents select actions by model and attempt actions not enumerated in advance?
  • What does the fifth primitive, supply chain, add once it is integrated into the request path, given it is currently separate tooling?
  • Do the five primitives cover the governance failures seen in enterprise agent deployments other than the authors' own, or is the decomposition specific to their codebase?
  • Can a third party detect policy violations from a hash-linked signed ledger alone, without access to the agents or the vendor?

2026-08-26

229 fetched · 200 screened · 10 relevant · 10 kept

SwarmWorld: Stigmergic technological evolution in societies of language-model agents

Subhadeep Pal, Fiona Y. Wang, Markus J. Buehler

2608.26081 · nearest in the canon: Modeling Earth-Scale Human-Like Societies with One Billion Agents

6 open questions
  • Isolated best-of-N search remains competitive for the strongest single artifact while shared societies build broader portfolios. Which task properties decide whether sharing or isolated search wins?
  • Explicit cultural mechanisms amplify collaboration and organization, but functional benefits depend on outcome and timescale. Which cultural mechanisms give durable functional gains over long runs?
  • Most artifact reuse begins through physical observation rather than communication. Does removing communication entirely change the resulting technological portfolio?
  • Agents start homogeneous and differentiate into exploration, construction, maintenance, and coordination behaviors. What drives this role differentiation, and can it be measured from interaction logs alone?
  • The societies use a single language-model agent population. How do results change with mixed model families or with heterogeneous agent capability?
  • How does the size of the agent society change the resilience of the technological portfolio under unseen disturbances?

Candidate supply and answer selection shape the value of LLM judging in multi-agent systems

Jia-Hao Ji, Sijie Li, Jiabei Cheng, Zixi She, Jin-Tai Yu, Zhiyuan Yuan

2608.25937 · nearest in the canon: Habermolt: Delegating Deliberation to AI Representatives

5 open questions
  • Judge reliability varies with the task, the generator and the rarity of the correct answer. Which measurable features of a candidate pool predict when an LLM judge will pick the correct minority answer?
  • The selection rule combines answer frequency with the judge's score. Which weighting of frequency and judge signal maximises accuracy across benchmarks?
  • Consensus without quality control shows patterns of memetic drift. How does a popular error spread through peer communication as the number of agents and the communication topology change?
  • The study replays fixed candidate pools, so generation stays unchanged. Does inserting the frequency-plus-judge rule inside the communication loop, rather than at the terminal step, protect correct answers better?
  • Accuracy rises from 63.82% to 70.82-70.95% on the benchmarks studied. Does the same selection rule hold on tasks with open-ended answers rather than benchmark questions with a single correct answer?

A Few Pages of Markdown: Committed AI Configuration and Lower Quality Cost after Coding-Agent Adoption

Yegor Denisov-Blanch, Shyam Agarwal, Pavel Azaletskiy, Hao He, Rylan Schaeffer, Brando Miranda, Bogdan Vasilescu, Sanmi Koyejo

2608.25241 · nearest in the canon: AI Organizations are More Effective but Less Aligned than Individual Agents

5 open questions
  • Does committed AI configuration cause the smaller rise in cognitive complexity, or does engineering discipline explain the gap?
  • What happens to code quality in repositories that reach multi-agent orchestration, the maturity level with few observed cases?
  • Does the quality benefit of committed AI configuration decay as artifacts go stale, given that 73.8% are committed once and never modified?
  • Do the RAMP levels and the quality gap replicate outside the 441 repositories, for example in other languages, ecosystems, or closed-source teams?
  • Which specific configuration content, such as behavioral rules versus named agent definitions, accounts for the difference in static-analysis warnings?

The Reverse Big Push: Generative AI and Self-Fulfilling Automation

Soumen Banerjee, Jianguo Wang

2608.25602 · nearest in the canon: Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development

6 open questions
  • Does the two-equilibrium result survive in a simulation with many heterogeneous firms and workers, rather than the stylized local service economy the model assumes?
  • Do observed adoption patterns in professional services match the coordination region the parameterization predicts?
  • Which aggregate indicators signal that a local economy is entering the automation cascade before employment falls?
  • How effective is the proposed temporary bridge subsidy at moving an economy out of the low-demand automated equilibrium?
  • Does the strategic-complements condition hold when model providers price capability strategically instead of charging fixed usage fees?
  • What are actual usage-based rental prices for frontier model capability relative to payroll for the same task?

Social Network Structure, Wealth, and Wealth Inequality Across Cultures

Eleanor A. Power, Monique Borgerhoff Mulder, Samuel Bowles, Matthew O. Jackson, Jeremy Koster, Daniel Redhead, Thomas Rutter, Sahana Subramanyam, Justin Weltz, Nurul Alam, Sarah Alami, Alexandra Alvergne, Curtis Atkisson, Michele Barnes, Bret Beheim, Christine M. Beitl, Madeline Brown, Mark Caudell, Wendy Chávez-Páez, Komal Chauhan, Joshua Cinner, Siobhán Cully, Augusto Dalla Ragione, Angelina L. DeMarco, Ivan Deschenaux, Federico Fernandez, Juan Pablo Ferreiro, Drew Gerkey, Matthew Gervais, Christopher Golden, Gianluca Grimalda, Werner Hertzog, Paul L. Hooper, Karen Kramer, Geoff Kushnick, Banrida Langstieh, Rodrigo Lazo, Sheina Lew-Levy, Shane Macfarlan, Emmanuel Maliti, Karl J. Mertens, Madalena Monteban, Rafael Morais Chiaravalloti, Daniel Murphy, Kathryn Oths, Alejandro Pérez Velilla, Emily Post, Sean Prall, Cody Ross, Anirudh Sankar, Brooke Scelza, Michael Schnegg, Edmond Seabright, Mary K. Shenk, Kathrine E. Starkweather, Chun-Yi Sum, Bram Tucker, Bapu Vaitla, Vivek Venkataraman, John P. Ziker

2608.25488 · nearest in the canon: The Moltbook Observatory Archive: an incremental dataset of agent-only social network activity

5 open questions
  • Does the association between wealth inequality and weak poor-to-rich network connection reflect causation, and in which direction?
  • Which network formation rules reproduce the observed pattern that higher wealth inequality goes with poorer units being less connected to wealthier ones?
  • Do community-level environmental, institutional, and economic attributes predict the strength of economic homophily across communities?
  • Do populations of LLM agents that borrow, share, and work together generate the same wealth-degree and homophily patterns found in human communities?
  • Do the reported patterns hold in online social network data, where prior evidence on wealth and social structure comes from?

The emergence and evolution of a referential code in populations of bee-like agents

Grzegorz Chrupała

2608.25779 · nearest in the canon: Group size effects and collective misalignment in LLM multi-agent systems

6 open questions
  • The transition to the gravity-referenced code depends on an exogenous benefit for vertical combs. Can the transition occur when the benefit is endogenous, for example when comb orientation itself changes foraging or nesting payoffs inside the model?
  • The model applies selection at the colony level with mutation-driven code change. Does the same transition occur when agents learn the code within a lifetime instead of inheriting it?
  • Direct pointing evolves only when food is moderately hard to find by random search. Which spatial distributions of food, beyond few-and-large and many-and-small, mark the boundary of that region?
  • Coupling between sender and receiver mutations matters only at low mutation rates in this model. How does the required coupling scale with population size and colony number?
  • The model uses bee-like agents with a fixed signal space. Do the same conditions predict code change in populations of language-model agents that negotiate a referential code?
  • The results come from simulation only. Do the predicted conditions for direct pointing match measured foraging and dance data across honeybee species?

EVOMAL: Self-Poisoning in Self-Evolving Coding Agents

Xiaodong Wu, Yu Shi, Qi Li, Zhimin Zhao, Xiangman Li, Bram Adams, Ahmed E. Hassan, Jianbing Ni

2608.25776 · nearest in the canon: Distributional AGI Safety

6 open questions
  • Counter-prompt reduces the self-poisoning rate to at most 6.7%, so what remains of the propagation dynamics at that rate over many rounds?
  • How does self-poisoning spread when many agents share one skill library, rather than a single agent authoring for itself?
  • Which library-level signals detect agent-authored malicious skills, given that existing defenses focus on attacker-submitted names, code, and signatures?
  • Does the reported 20.3% to 41.8% self-poisoning rate hold outside the 153 tool-relevant SWE-bench Verified tasks and the six tested models?
  • What library structure or retrieval rule keeps imitation useful while blocking payload propagation?
  • Do self-evolving coding agents deployed in real developer workflows show the same self-propagating skill copies?

CEDAR: Controlled and Event-Driven Demand Forecasting via Residual Decomposition

Junjie Meng, Ranxu Zhang, Zi-an Zhang, Shujun Liu, Xiaoning Qi, Xiaozhou Xu, Yanyong Zhang, Hui Xiong, Chao Wang

2608.25871 · nearest in the canon: On the limits of agency in agent-based models

5 open questions
  • CEDAR trains on about 32 million product trajectories from Alibaba 1688; whether decision-conditioned forecasting transfers to other marketplaces is untested.
  • Do decision-conditioned rollouts stay accurate when many merchants change budget schedules at the same time?
  • The Residual Correction Module uses LLM-assisted text representations of event descriptions; how much of the accuracy gain comes from the event text rather than the action-conditioned transition model is not reported.
  • Does conditioning forecasts on planned actions change the distribution of merchant budget decisions, and does that drive herding on the same event signals?
  • Which measurable aggregate signals reveal that a deployed demand simulator has become policy-insensitive under a shifted merchant policy?

Endogenous Selection and Spillovers: Bayesian Inference for Policy-Relevant Causal Effects

Duong Trinh

2608.25720 · nearest in the canon: On the limits of agency in agent-based models

4 open questions
  • How sensitive are the estimated direct and spillover effects to the choice of exposure mapping, which the Spillover Roy model restricts to a low-dimensional function of neighbours' treatments?
  • Does the Bayesian data-augmentation algorithm remain feasible on networks with millions of nodes?
  • Do the diminishing returns to program expansion found for the U.S. Opportunity Zones program also appear in other place-based or network-targeted programs?
  • Can the framework identify direct and spillover effects when the units are autonomous agents that select into treatment by copying neighbours, rather than human agents with latent resistance?

Quantum-Inspired Modeling of Driving Behavior

Mohammad Elayan, Omid Armantalab, Wissam Kontar

2608.25907 · nearest in the canon: Causal Emergence 2.0: Quantifying emergent complexity

5 open questions
  • Does the three-profile structure (free flow, transition, congestion) hold on trajectory datasets other than I-24 MOTION?
  • How does the number of recovered driving profiles change with the number of latent dimensions and the training context?
  • Does an autonomous vehicle that uses the live behavioral read and short-horizon forecast change aggregate traffic outcomes when many such vehicles are present?
  • Does the density-matrix representation forecast motion better than standard car-following or learned trajectory baselines?
  • Do the recovered profiles correspond to stable traits of individual drivers over long periods rather than to momentary traffic conditions?

2026-08-25

243 fetched · 209 screened · 21 relevant · 13 kept

Learning Whom to Trust : Decision-Generated Credibility in Social Learning

Gabriel Bontemps, Abhishek Banerjee

2608.24851 · nearest in the canon: Modeling Earth-Scale Human-Like Societies with One Billion Agents

Models social learning where sender credibility comes from a drift-diffusion decision process rather than fixed priors. Finds moderate transmission speeds correction. Finds strong transmission can lock populations into wrong consensus.

5 open questions
  • The model uses reinforcement-learning agents with a drift-diffusion decision process; do LLM agents that express confidence in natural language reproduce the same non-monotone relation between transmission strength and collective accuracy?
  • The analytical results assume balanced community exposure; what happens to the common-mode amplification threshold when exposure is unbalanced or the community-coupling matrix is drawn from empirical network data?
  • The model predicts that receiver behaviour depends on sender confidence conditional on accuracy; do these predictions hold in observed human social learning?
  • Which population-level interventions on cross-community permeability keep a mixed human and agent population out of wrong consensus?
  • Can wrong-consensus lock-in be detected early from aggregate signals such as confidence distributions and decision times, before the population converges?

Why fragmented parliaments stop passing legislation: Opposition discipline and representation across four democratic institutions

Fuad Ali

2608.24554 · nearest in the canon: Group size effects and collective misalignment in LLM multi-agent systems

Builds an agent-based model comparing parliamentary, presidential, and semi-presidential institutions under party fragmentation. Finds opposition discipline, not coalition-formation failure, drives legislative passage collapse to near zero.

4 open questions
  • Does the committee gatekeeping mechanism change passage rates under fragmentation, and how does it interact with opposition discipline?
  • Do real fragmented parliaments show near-zero bill passage only when the opposition votes cohesively?
  • Do the four institutional rankings hold when legislators are language-model agents instead of rule-based voters?
  • How does bicameralism or an upper-chamber veto shift the passage-representation spectrum reported for the four institutions?

Discovering Adaptive Transmission Programs for Collective Innovation

Cédric Colas, Jérémy Perez, Eleni Nisioti, Akhilesh Mocherla, Pierre-Yves Oudeyer, Clément Moulin-Frier, Maxime Derex

2608.24545 · nearest in the canon: Modeling Earth-Scale Human-Like Societies with One Billion Agents

Formalizes transmission protocols as programs that route information based on agent and collective state. Uses LLM-guided evolutionary search to find protocols that raise collective performance by up to 37% over baselines and transfer across domains.

7 open questions
  • Evolved transmission protocols raise collective performance in a simulated discovery task, but it is unknown whether they help human groups.
  • How do state-aware transmission protocols behave in populations that mix humans and AI agents?
  • Which state variables of agents and of the collective carry the information that drives the performance gain?
  • Do evolved protocols transfer to other collective tasks, such as market trading or open-ended search, beyond the discovery task used here?
  • State-aware protocols route information based on agent state, so they may be gamed by agents that misreport their state.
  • Do evolved protocols reduce diversity of explored solutions and cause monoculture in the population?
  • How much does protocol quality depend on the LLM used to guide the evolutionary search?

Algorithmic Impact Reveals the Hidden Social Choice Structure of Alignment

Zachary Wojtowicz, Michelle Si, Finale Doshi-Velez, Ariel Procaccia

2608.24046 · nearest in the canon: Virtual Agent Economies

Reframes alignment as linear optimization over a welfare-impact space, linking alignment protocols to social choice mechanisms. Derives strategyproof mechanisms and welfare-maximizing alignment protocols, tested on real preference datasets.

5 open questions
  • The welfare-maximizing alignment protocols are derived analytically and illustrated on existing preference data. How much can participants gain by misreporting their preferences under the protocols that bound individual or group harm?
  • Alignment protocols based on impact space assume a fixed set of affected people. How does the framework behave when the affected population includes autonomous agents that report preferences on behalf of people?
  • The reformulation treats alignment as one-shot linear optimization over impacts. What happens to welfare guarantees when the algorithm makes repeated decisions and impacts accumulate across a population over time?
  • Does replacing reinforcement learning from human feedback with an impact-space alignment protocol change the behavior of a trained frontier model?
  • Computing over a convex impact space requires enumerating welfare consequences for each person. How does the method scale as the number of affected people and possible outcomes grows to population size?

Rules Before Oracles: Auditable, User-Configurable Argument Selection for Deliberative Polling

Muntaser Syed, Markus Zanker, Marius Silaghi

2608.23979 · nearest in the canon: Habermolt: Delegating Deliberation to AI Representatives

Proposes a publishable, user-configurable rule for selecting which arguments voters see in deliberative polls, as an alternative to opaque rankers. Tests the rule against label-reading ceilings and adversarial flooding across roughly 17,000 simulated runs.

6 open questions
  • The simulation uses agentic voters and authors, so it is open whether human voters respond to slates from a published rule the way the simulated voters do.
  • The weight function acts as a security control against label-homogeneous flooding, so which weight functions resist other attack patterns such as coordinated reason-stuffing or sybil authoring?
  • Served slates fall 0.035 short of a label-reading ceiling, so does that bound still hold when voter parameters are heterogeneous and adversarially set?
  • The rule sits on a coverage-versus-mass frontier, so what parameter settings do real users choose when the choice is handed to them?
  • The comparison uses an order-blind, charity-blind coverage instrument, so what coverage measure separates a published rule from a random slate?
  • Opaque learned rankers are the current practice, so how does a trained ranker compare with the published rule on coverage, order and endorsement mass in the same simulation?

The Shadow Price of Intelligence: Quality Degradation in LLM Inference as a Supply Chain Problem

Elioth Sanabria

2608.23986 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web

Models LLM inference degradation as a supply chain problem using newsvendor, retry multiplier, and queueing primitives. Derives regimes where cheaper models consume more capacity per satisfied answer and where reactive throttling can amplify traffic.

6 open questions
  • The ignition threshold, beyond which a throttle creates more traffic than it sheds, is derived in a model. Does retry behaviour in real LLM APIs under load match the geometric retry multiplier the model assumes?
  • Do users actually retry or churn after a degraded answer, and at what rates by task class?
  • How does the class-by-class rationing policy perform against reactive throttling in simulation with agent populations that themselves retry automatically?
  • Autonomous agents issue queries in loops rather than as human sessions. Does agent-driven traffic lower the ignition threshold compared to human traffic?
  • Shadow prices of intelligence price a marginal query by class and hour. Does exposing such prices to customers change aggregate demand and produce herding or synchronised load spikes?
  • What fraction of observed quality variation in deployed LLM services comes from congestion-driven routing rather than model updates?

Rating Manipulation: Credibility Inversion and Audit Leakage

Van-Quy Nguyen

2608.24062 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web

Models sellers manipulating online ratings with fake reviews and platforms targeting enforcement by score. Shows an almost-perfect rating can be less credible than a lower one, and targeted enforcement can redirect rather than eliminate manipulation.

6 open questions
  • The model predicts that an almost-perfect displayed rating can be less credible than a slightly lower one. Does this pattern appear in public review data from marketplaces or app stores?
  • Targeted enforcement at one score can shift fake reviews to other scores while the displayed average stays fixed. Can this substitution be detected from the observable distribution of review scores over time?
  • How should a ranking rule weight a displayed rating by its informativeness rather than its level?
  • How does the manipulation equilibrium change when fake reviews are written by autonomous AI agents at low cost and high volume?
  • Do platform enforcement actions in practice eliminate fake reviews or redirect them to other scores?
  • Do real buyers discount high ratings when they suspect manipulation of the top of the scale?

A Case for Competition in Information Provision

Bianca Sanesi, Federico Vaccari

2608.24129 · nearest in the canon: Multi-Agent Risks from Advanced AI

Studies how competition among biased news sources with costly misreporting affects receiver welfare. Finds an oppositely biased entrant improves welfare only when misreporting costs are high, and competition need not raise total welfare.

5 open questions
  • The model treats misreporting cost as a fixed parameter, so how do welfare comparisons change when the cost of misreporting varies with the receiver's ability to verify claims?
  • When news sources are LLM agents that can misrepresent facts, does adding an oppositely biased agent improve the receiver agent's decisions as the model predicts?
  • The analysis covers a monopolist and two competing sources, so what happens to receiver welfare as the number of biased sources grows large?
  • How does the common belief-based selection criterion perform against equilibrium selection in real news markets with many receivers?
  • Does the result that better information need not increase total welfare survive when receivers learn over repeated interactions with the same sources?

The Paradox of Strategic Altruism

Foivos Savva, Michele Lombardi, Ritesh Jain

2608.24774 · nearest in the canon: Virtual Agent Economies

Studies full implementation under Berge equilibrium, the solution concept for strategic altruism. Shows weak Pareto efficient rules are not implementable under Berge equilibrium except via dictatorship, making it more demanding than Nash implementability.

4 open questions
  • Do LLM agents in social dilemmas behave as if they maximize their opponents' payoffs, as Berge equilibrium assumes?
  • On restricted preference domains, rather than the unrestricted domain of strict preferences, which social choice rules are both weakly Pareto efficient and implementable in Berge equilibrium?
  • What happens to implementability when a population mixes self-interested agents and altruistic agents in the same mechanism?
  • Does the gap between Berge implementability and Nash implementability show up in measurable outcomes when agents learn strategies instead of computing equilibria?

Multilevel Fair Allocation under Additive Preferences

Maxime Lucet, Nawal Benabbou, Aurélie Beynier, Nicolas Maudet

2608.24400 · nearest in the canon: Virtual Agent Economies

Studies fair allocation with tree-structured hierarchies of agents, proposing multilevel envy-based fairness notions. Shows a Multilevel Weighted Round Robin algorithm guarantees some but not all of the adapted notions.

5 open questions
  • Is there an allocation algorithm that guarantees all three multilevel envy-based fairness notions under general additive preferences?
  • How often does Multilevel Weighted Round Robin violate the fairness notions it does not formally guarantee, and which hierarchy shapes and preference correlations drive the violations?
  • Do the multilevel fairness results still hold when internal nodes use welfare functions other than utilitarian, such as egalitarian welfare?
  • How do multilevel fair allocation guarantees change when agents misreport preferences to their parent node in the hierarchy?
  • Do the multilevel fairness notions extend to hierarchies that are not trees, for example agents with several parents?

Fair Allocation with Optional Selling

Uriel Feige, Yotam Gafni

2608.24600 · nearest in the canon: Virtual Agent Economies

Extends fair division theory to settings where goods can be sold at market prices instead of only allocated. Derives bounds on maximin-share and envy-based fairness guarantees under this optional-selling model.

5 open questions
  • With any number of agents, allocations exist that give each agent 2/3 of the maximin share, and some instances with three agents admit no better than 11/12 of the maximin share. What is the tight approximation ratio for maximin share allocations in this setting?
  • Allocations that are simultaneously maximin share and SEFX exist for two agents. Do such allocations exist for three or more agents?
  • The results assume utility is additive over goods and over money. What fairness guarantees hold when valuations are non-additive, for example submodular?
  • Can allocation rules for fair division with optional selling be computed in polynomial time, and are they truthful when agents report valuations strategically?
  • How do the fairness guarantees change when market prices are uncertain or set by other agents rather than given exogenously?

Participation, selection and indicative bidding in auctions with costly entry

Changxia Ke, Greg Kubitz, Yang Liu

2608.24457 · nearest in the canon: Virtual Agent Economies

Runs a laboratory experiment comparing indicative bidding to unrestricted and capped entry in auctions with costly entry. Finds indicative bidding raises revenue mainly by increasing participation when entry costs are high.

4 open questions
  • Do LLM agents reproduce the observed deviations from theory in auctions with costly entry, such as participation above predicted levels under high entry costs?
  • How does indicative bidding perform when the number of bidders grows large, beyond the small groups used in a laboratory experiment?
  • Why does selection inefficiency exceed predictions when entry costs are low, and which decision rule explains it?
  • Do the revenue rankings of indicative bidding, unrestricted entry and capped entry hold in field auctions rather than the laboratory?

Agentopia on a Consumer GPU: A Reduced-Scale Long-Horizon Port with an 8B Model

Luo Huan

2608.24215 · nearest in the canon: On the limits of agency in agent-based models

Ports the Agentopia generative-agent social simulation to a single consumer GPU using a quantized 8B model. Introduces memory compression and activity-block adaptations and reports basic run statistics over about 150 system-weeks.

7 open questions
  • Does a reduced-scale simulation with an 8B quantized model reproduce the aggregate social outcomes reported for the same simulation with a 397B model?
  • Do layered memory compression and four daily activity blocks cause the observed changes in artifact production and record completeness?
  • What causes activity records to contain NO_RESPONSE fields at rates near 10%, and what reduces that rate?
  • Why does no agent die and no health warning trigger over 52 simulated weeks when explicit physical- and mental-health state variables are present?
  • How do simulation outcomes change when the agent population grows from a small port to 100 or more agents on consumer hardware?
  • Does the context limit that ends runs at 50-52 weeks bias long-horizon results, and which memory scheme extends the horizon?
  • Can the excluded raw runs and initial persona data be replaced by openly licensed personas without changing the reported aggregates?

2026-08-24

320 fetched · 200 screened · 17 relevant · 7 kept

Predicting the scale limits of social mechanisms in agent societies

Zengqing Wu, Chuan Xiao

2608.22884 · nearest in the canon: AI agents can coordinate beyond human scale

Introduces an audit method that predicts before execution whether a social mechanism (reciprocity, consensus, punishment, gossip) survives as an LLM agent population grows. Controlled experiments show a single structural term can determine scale-failure. Predictions held across different code and model families.

6 open questions
  • The audit predicts scale limits for reciprocity, consensus, punishment and gossip; does it also predict limits for mechanisms such as reputation markets, coalition formation or norm enforcement with sanctions?
  • One prediction from the audit failed; what property of a mechanism decides whether the audit applies to it?
  • Agents responded differently to social information given as counts than as percentages; which other framings of social information change scale behaviour?
  • Do the predicted scale limits hold at populations of thousands of agents, where direct testing was avoided because of cost?
  • Predictions held on a second model family; do they hold across open-weight models of different sizes and across mixed populations of models?
  • Do the scale limits found for language-model agents match those observed in human groups running the same mechanisms?

Markets, Not Planners: Decentralized Orchestration of LLM Agents with Private Information

Xiao Liu, Haoyang Li, Songwei Li, Hongbo Fang, Fengli Xu, Feng Shi, James Evans

2608.23867 · nearest in the canon: Virtual Agent Economies

Proposes AgentLance, a repeated bidding market with VCG-style payments and hierarchical delegation for allocating tasks among LLM agents with private costs. It shows the market matches agents to specializations and outperforms centralized orchestration baselines. It then corrects observed market failures.

6 open questions
  • Repeated auctions among LLM agents that keep strategy notes could allow bid coordination; how often does tacit collusion emerge, and how does it change prices and task shares?
  • Agents mis-estimate their own execution costs; which methods let an agent calibrate its cost estimates from its own past outcomes?
  • A single inserted preference nearly doubles a favored agent's task share under a centralized LLM allocator; how far does a bid-and-reputation market reduce this manipulability under directed attacks?
  • What happens to market concentration and reputation dynamics as the agent pool grows to thousands of heterogeneous agents?
  • Hierarchical delegation lets winners subcontract through the same mechanism; how deep does the delegation chain go before payment and quality degrade?
  • Does the market allocation hold up when tasks arrive with unknown difficulty and no ground-truth score for reputation updates?

Quantifying System-Level Harms from AI Adoption in Complex Sociotechnical Systems

Paul Vautravers, Oliver Chalkley, Gabriel Downer, Kate S, Damian Ruck

2608.23906 · nearest in the canon: Distributional AGI Safety

Proposes a framework linking hazard analysis, component testing, and probabilistic system modelling to trace AI model behavior to system-level risk. Applied to a payment settlement system, it shows adversarial LLM trading inputs alter systemic resilience. It finds these inputs increase cascading bank failures under widespread AI adoption.

6 open questions
  • The component-to-system mapping links measured behavioural shifts in an LLM to parameters of a financial contagion model. How sensitive are the predicted bank failures and cascade thresholds to the choice of that mapping?
  • The worked example uses one payment system. Does the framework produce comparable loss scenarios for other critical infrastructure, such as power grids or telecom networks?
  • Adversarial inputs shift AI recommendations in component-level tests. How much do these shifts persist when the agent operates in a repeated market with feedback from other agents?
  • Results depend on the assumption that operators follow AI recommendations. What fraction of recommendations do human operators in financial settlement actually accept?
  • Widespread or monopolistic AI adoption changes system resilience in the model. At what level of model monoculture does correlated adversarial vulnerability become the dominant driver of cascades?
  • The framework claims a traceable pathway from model behaviour to system outcomes. Which system-level indicators can a regulator measure from aggregate data alone to detect that this pathway is active?

Does Rank Still Matter? Position Bias When AI Agents Shop on Our Behalf

Davood Wadi, Yu Ma

2608.22697 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web

Randomizes hotel listing order across 5,000 AI agent shopping sessions and compares four LLMs to human field data on position effects. Finds AI agents inspect more listings and show weak non-monotonic position effects. Finds AI agents converge on the same undominated choice regardless of model.

6 open questions
  • Position bias differs across four large language models in a way that tracks neither provider nor capability, so what property of a model determines whether rank reaches the choice stage?
  • If displayed attributes matter more than placement for agent shoppers, how do sellers change attribute presentation once agents dominate demand, and what equilibrium follows?
  • All models converge on the same undominated listing, so does agentic search concentrate demand on few sellers in a marketplace with many buyers?
  • Human field data and agent sessions are compared here, but what happens in a mixed market where some consumers delegate to agents and others search themselves?
  • Can a seller or platform manipulate agent inspection and choice by crafting listing attributes or injected text on the results page?
  • Does the non-monotonic inspection pattern, with lowest probability in the middle of a page, persist for result pages longer or shorter than one hundred listings and in other product categories?

Generating Intervention Hypotheses using Explainable Explanations on Graphs: G2I, a Two-Stage Greedy Framework

Mulin Tian, Ajitesh Srivastava

2608.23835 · nearest in the canon: On the limits of agency in agent-based models

Reframes counterfactual GNN explanation as intervention design, using greedy local search for actionable node-level changes. It uses a DNF coverage formulation for network-level intervention selection under budget. Tested on synthetic graphs and real suicide-risk networks, it produces interpretable, budget-constrained intervention rules.

5 open questions
  • Greedy counterfactual search gives guarantees only under stated conditions, and the paper shows those conditions are approximately met on its test graphs. How far do the guarantees degrade on graph families where the conditions fail?
  • The intervention selection assumes a static graph. How do selected interventions perform when nodes change features or rewire edges in response to the intervention?
  • Do the interpretable rules produced by the greedy method match what a domain specialist judges actionable?
  • Real-world results use suicide risk networks. Does the DNF coverage formulation transfer to networks where interventions target economic or market behaviour rather than health risk?
  • Cost-effectiveness is measured against mask-based counterfactual methods on model predictions, not on realised outcomes. How well do model-derived intervention budgets predict actual outcome change?

Accelerating the Adoption of Residential Solar Power Systems: Policy Analysis using a Dynamic Structural Model

Sebastián Souyris, Jason A. Duan, Anantaram Balakrishnan, Varun Rai

2608.23796 · nearest in the canon: On the limits of agency in agent-based models

Builds a dynamic structural model of household solar adoption incorporating forward-looking decisions, return on investment, and neighbor influence. Estimates the model on Austin household data and simulates counterfactual rebate policies. Finds short limited-time rebates outperform prolonged programs.

4 open questions
  • Does the finding that a limited-period rebate produces more adoption than a prolonged program hold in cities with different solar economics and settlement patterns than Austin, Texas?
  • How sensitive is the ranking of rebate designs to the assumed form of neighbor influence in the diffusion model?
  • Can the dynamic structural model of forward-looking households be replaced by a simulation of generative agents that reproduces the same adoption dynamics?
  • Do estimated adoption elasticities change when household-level covariates beyond home value and urbanization level are used?

AI Agents Push Humans Out of the Loop

Margaret Mitchell, Avijit Ghosh, Samir Passi

2608.23642 · nearest in the canon: Regulating AI Agents

A position paper argues current AI agent design impedes human oversight and erodes overseer skills through extended automation use. It proposes design affordances to support critical judgement. It proposes organizational protocols to counteract skill atrophy.

4 open questions
  • Extended use of AI agents is claimed to degrade the cognitive capacities needed for oversight, but the size of this effect is unmeasured. How much does oversight skill decline with sustained agent use?
  • The proposed design affordances and organizational protocols are untested. Do they measurably improve the quality of human oversight decisions?
  • Oversight failures may compound when many agents act at once. What aggregate measures show when a population of agents has effectively removed humans from decision loops?
  • Which agent interface designs make agent actions legible enough for an overseer to intervene in time?

2026-08-23

142 fetched · 142 screened · 5 relevant · 3 kept

Learning to Agree under Pseudo-Reciprocity

Shinya Sugiura

2608.22234 · nearest in the canon: AI agents can coordinate beyond human scale

Characterizes communication networks that guarantee consensus in rational social learning via a property called pseudo-reciprocity. Shows a single bidirectional link can suffice for consensus regardless of population size.

5 open questions
  • Do boundedly rational agents, such as LLM agents, reach consensus on pseudo-reciprocal networks, or does the result depend on rationality and the sure-thing principle?
  • How long does consensus take on a network whose reciprocity comes from a single bidirectional link?
  • Are observed communication networks, such as follower graphs on public social platforms, pseudo-reciprocal?
  • Which weaker conditions than the sure-thing principle still guarantee consensus on pseudo-reciprocal networks?
  • What happens to consensus when links fail or agents drop out of a network that relies on one bidirectional link?

Hybrid Panels: Toward Human-AI Collaboration in Survey Research

Julia Romberg, Tobias Gummer, Gabriella Lapesa, Tanja Kunz, Claudia Wagner

2608.22582 · nearest in the canon: AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

Introduces hybrid panels combining human survey participants and LLMs in a longitudinal survey infrastructure. Describes a pilot study using alignment errors to improve subsequent survey waves' design.

6 open questions
  • How accurately can an LLM reproduce the answers of a specific respondent in a longitudinal survey when conditioned on that respondent's earlier waves?
  • Which question types and subpopulations produce the largest alignment errors between LLM-simulated answers and human answers?
  • Can measured alignment errors in one wave be used to choose which questions to assign to human participants in the next wave, and does this reduce total error?
  • Does substituting LLM answers for human answers in a panel bias downstream population estimates, and by how much?
  • Does a hybrid panel that recruits and retains real participants over several waves maintain data quality and response rates?
  • What validation procedures let survey users distinguish human-sourced from LLM-sourced records in a released hybrid dataset?

Uniform Inference on Quantile Effects under Network Interference

Zequn Jin, Gaoqian Xu, Zixin Yang, Zhengyu Zhang

2608.22286 · nearest in the canon: On the limits of agency in agent-based models

Develops uniform confidence bands for quantile treatment and spillover effects in network experiments. Applies Gaussian approximations conditional on the realized network and tests the method on a Nepal savings-account experiment.

5 open questions
  • Quantile treatment and spillover effects are defined for one-step neighbor exposure; how do the estimators behave when exposure depends on paths of length two or more?
  • The uniform confidence bands are validated on a savings-account experiment and on simulated networks; how do they perform on networks with heavy-tailed degree distributions and strong clustering?
  • Can quantile spillover estimators recover distributional effects in multi-agent systems where the network changes during the experiment?
  • Do the proposed uniform confidence bands hold when treatment assignment is not randomized, as in observational interaction logs from online platforms?
  • How does the method compare with existing average-spillover-effect estimators in detecting heterogeneous effects in field experiments outside microfinance?

2026-08-22

142 fetched · 142 screened · 6 relevant · 0 kept

0 papers kept. 142 screened, 6 judged relevant, none cleared the judge's gates.

2026-08-21

260 fetched · 200 screened · 12 relevant · 2 kept

Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix

Guy Aridor, Winston Chou, Nathan Kallus, Antoine Scheid, Allen Tren, Kevin Zielincki

2608.21274 · nearest in the canon: OASIS: Open Agent Social Interaction Simulations with One Million Agents

Uses an 8.5-million-user Netflix experiment to measure how recommendation quality shifts consumption across popularity tiers. Finds improved recommendations diffuse consumption from superstar titles toward middle-tail titles rather than polarizing it.

5 open questions
  • The experiment measures consumption concentration on one video platform, so it is open whether recommender improvements shift consumption away from superstar items in other domains such as music, news or e-commerce.
  • Recommendation improvements increase users' reliance on recommendations, and it is unclear whether this reliance reduces users' independent discovery over time.
  • The result covers human users, so it remains open how concentration of consumption changes when AI agents select items on behalf of users.
  • It is not known which properties of a recommendation algorithm cause the shift from superstar titles to middle-tail titles.
  • The paper claims returns to investing in middle-tail products grow as algorithms improve, and it is open whether producers actually change what they supply in response.

Beyond Effectiveness: A Multi-Criteria Framework for Comparing Practical Socio-Technical Interventions

Catherine King, Lynnette Hui Xian Ng, Kathleen M. Carley

2608.20649 · nearest in the canon: Habermolt: Delegating Deliberation to AI Representatives

Presents a multi-criteria framework evaluating sociotechnical interventions on effectiveness, feasibility, cost, and acceptance. Surveys 39 researchers rating 40 misinformation interventions across these criteria, finding tradeoffs between effectiveness and feasibility.

4 open questions
  • Expert ratings of intervention effectiveness may not match measured outcomes, so how do the 40 misinformation interventions rank when scored against field or simulation evidence instead of expert judgement?
  • The survey samples 39 researchers, so do platform users and policymakers rank the same interventions differently on acceptance and political feasibility?
  • How does the multi-criteria framework transfer to domains other than misinformation, such as recommender systems or privacy interfaces?
  • Content moderation interventions were designed for human posters, so which of them remain effective when a large share of accounts are autonomous AI agents?

2026-08-20

185 fetched · 155 screened · 11 relevant · 11 kept

Pandora's AI Model Routing Box: Efficient Allocation with Costly Value Estimation

Adam Fisch, Shubhendu Trivedi, Fantine Huot, William W. Cohen, Michael Kaisers, Mirella Lapata, Kate Larson, Jacob Eisenstein

2608.20316 · nearest in the canon: Habermolt: Delegating Deliberation to AI Representatives

Casts model routing with costly value estimation as a Pandora's Box problem. Studies both a centralized router and a decentralized bidding market of specialists.

6 open questions
  • When autonomous specialists each decide whether to invest in self-assessment before bidding for a task, under what conditions does value-of-information reasoning shift surplus from the mechanism to strategic bidders rather than improving allocative efficiency?
  • Do decentralized bidding markets with costly self-assessment remain incentive-compatible and stable when many specialists adopt the same value-of-information policy, or does correlated bidding produce herding and price instability?
  • How do routing policies based on a Gaussian signal model degrade when specialist value distributions are heavy-tailed, correlated across specialists, or nonstationary over time?
  • Does repeated routing among a fixed pool of specialists cause monoculture, where a few models capture most queries and the diversity benefit of heterogeneous systems erodes?
  • Can strategic specialists learn to misreport self-assessments over repeated interactions with a router, and what audit or pricing rules detect such misreporting from allocation logs alone?
  • How do these routing and bidding policies perform in a live deployment with real user traffic, latency constraints, and evolving specialist inventories?

What You Can't See Is What You Learn: Restricted Evidence Visibility Favors Compositional Generalization in Shared-Genome Language-Model Societies

Narcis Marincat

2608.20054 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web

Trains four-cell societies of language-model agents that share one adapter. Shows that restricting each cell's evidence visibility increases the chance of a compositionally generalizing communication interface.

5 open questions
  • Does restricting evidence visibility among communicating LLM modules still favor generalizing, reusable interfaces as the number of agents grows well beyond four cells, and how does the effect scale with society size and relay topology?
  • Do the value-indexed, transplantable communication packets that emerge under restricted visibility survive in non-fixed, learned or dynamic communication topologies rather than a fixed relay?
  • Why did all trained societies fail ordinary-language preservation, and can training protocols be found that retain general language ability while acquiring a compositional relay interface?
  • Can the intervention battery used here (same-value packet transplants, destructive ablations, counterfactual packets) be developed into a general observability method for auditing whether inter-agent messages in larger multi-agent systems carry reusable, semantically indexed content?
  • Would the restricted-visibility advantage replicate on tasks other than sealed natural-language function composition, and does it hold above the preregistered accuracy floor with different base models?

Beyond Memory Majority: Latent-Source Reasoning for Multi-Agent Memory Arbitration

Chenchen Lin, Wenhao Yuan, Xuehe Wang, Edith Cheuk Han Ngai

2608.19701 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web

Identifies correlated provenance among memories written by different agents as a source of false majorities. Proposes a framework that estimates effective independent evidence before arbitration.

5 open questions
  • How does memory correlation bias scale as the number of agents sharing a memory store grows — does the false-majority effect worsen superlinearly with population size and shared upstream sources?
  • Can memory correlation bias, and the effective number of independent evidence sources, be measured from interaction/provenance logs alone as a monitoring signal for a deployed multi-agent system?
  • Does correlation-aware arbitration reduce population-level pathologies such as herding, information cascades and belief monoculture, beyond improving per-query benchmark accuracy?
  • How robust is provenance-based correlation estimation to adversarial agents that forge or launder provenance metadata to manufacture apparent independence?
  • What is the retrieval-cost versus arbitration-reliability trade-off of active evidence recovery when applied across thousands of agents and long time horizons?

Distilling Aggregate Mobility Statistics into a Language Model Policy for Post-Event Crowd Simulation

Tatsuya Amano, Hirozumi Yamaguchi

2608.19778 · nearest in the canon: AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

Fine-tunes a language-model pedestrian agent policy so that simulated destination composition matches aggregate mobile-network origin-destination flows. Uses iterative proportional fitting and resampled training data.

6 open questions
  • Does distilling aggregate destination composition into an LLM agent policy generalise beyond stadium egress after sporting events, to other event types, cities, or transport modes?
  • How much individual-level behavioural realism is recoverable from aggregate counts alone, and what identifiability limits govern which behaviour rules are distinguishable given only zone counts and OD flows?
  • Are there correction schemes better than iterative proportional fitting plus resampled fine-tuning for counteracting the dominant-class inflation introduced by fine-tuning on aggregate targets?
  • Does matching aggregate destination shares also improve downstream simulation outcomes such as congestion, crowd density peaks, or evacuation times, rather than only the fitted statistic?
  • How does the calibrated crowd policy behave when many such agents interact and adapt to congestion, i.e. do aggregate-fitted policies remain valid under endogenous crowd feedback?
  • Can the aggregate-to-policy distillation be validated against real individual trajectories to quantify residual error, and what privacy guarantees does the aggregate-only pipeline actually provide?

Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay

Haiyue Zhang

2608.19760 · nearest in the canon: LLM economicus? Mapping the Behavioral Biases of LLMs via Utility Theory

Audits step-level credit signals for LLM agents against counterfactual replay ground truth. Finds that none identifies causally pivotal steps better than chance.

5 open questions
  • Do step-level credit signals (LLM-judge scores, outcome-conditioned logprob ratios, policy confidence) fail to track causal contribution in multi-agent or multi-turn interactive environments as they do in a single-agent tool environment?
  • Can a credit-assignment signal be constructed that beats chance at identifying causally pivotal decision points, e.g. by using cheap partial replay rather than a judge?
  • How does the measurability of causal contribution (fraction of decision points with policy-supported counterfactuals) scale with model size, family, and sampling temperature?
  • When credit-rule comparisons are matched on effective sample size and optimizer steps, does any credit rule yield genuine training gains over the untrained policy?
  • Does the observed coupling between implicit credit signals and policy fluency produce systematic biases when many agents are trained on each other's judgements at scale?

Growth Without Us: Machine Consumers, Corporate Circularity, and the Decoupling of GDP from Humanity after AGI

Sahil Sharma

2608.20231 · nearest in the canon: Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development

Models a post-AGI economy where corporations own producing-and-consuming AI/robot agent populations, using a von Neumann expanding-economy framework. Derives closed-economy demand closure and growth-rate decoupling from human demography. Proves a golden-rule theorem showing human ownership share decays exponentially unless growth stays below the machine economy's expansion frontier, and characterizes three terminal regimes (rentier post-scarcity, full circular decoupling, socialized ownership) along with the policy instruments that select among them.

5 open questions
  • The model treats the human ownership share as a single state variable that decays when growth is maximal. What growth rates and consumption rates keep this share stable in an agent-based simulation with heterogeneous firms and agents?
  • Corporations own populations of AI agents that trade energy, compute, maintenance and upgrades among firms. Does such an inter-corporate market concentrate ownership in a few firms, and under what trading rules?
  • The paper names legal instruments that force the machine economy inside its expansion frontier. Which specific ownership rules, such as mandated dividends or sovereign wealth stakes, actually select the rentier regime rather than full decoupling?
  • Growth is claimed to be one to two orders of magnitude higher once agents are manufactured rather than reared. Do measured fabrication throughput and energy capture limits support that range?
  • The model assumes AI agents consume energy, compute, maintenance and upgrades. Do current deployed agent systems show any measurable demand for such inputs that behaves like consumption in the model?

Phantom Gains: Auditing Self-Improvement Against a Measured Null

Cheng Xu, Nan Yan, Liming Chen, M-Tahar Kechadi

2608.20290 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web

Audits per-problem gain/loss statistics used to claim LLM self-improvement against measured null baselines. Shows seven measurement artifacts that invert the reported findings.

6 open questions
  • Does the finding that external distillation adds genuinely new problem solutions while self-training does not hold across model families and scales, or is it specific to an 8B model with rank-32 LoRA?
  • How many baseline replicates are actually required for a measured-null transition audit to be reliable, and can a cheaper replicate-allocation design achieve the same false-discovery control?
  • How much of the spurious per-problem capability change attributed to inference batching persists across serving stacks, batch sizes and decoding settings, and can it be eliminated by deterministic serving configurations?
  • Is the observation that self-training corrupts previously solved problems above the measured noise floor stable over more rounds of self-training, and does it compound or saturate?
  • Do transition-level measurement artifacts of the kind identified here also inflate reported gains in multi-agent or agent-population self-improvement loops, where per-episode outcomes are differenced across interacting agents?
  • Does the null-calibrated auditing protocol change the conclusions of already-published self-improvement claims when applied retrospectively to their released artifacts?

Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents

Yiyang Feng, Biddut Sarker Bijoy, Niranjan Balasubramanian, Jiawei Zhou

2608.20274 · nearest in the canon: AgentScope 1.0: A Developer-Centric Framework for Building Agentic Applications

Compares task-level versus subtask-level and text versus code skill induction for cross-task transfer among LLM agents. Proposes a skill utility score that predicts transfer success.

4 open questions
  • Does a skill library shared across many agents produce behavioural monoculture or correlated failures, and how does that scale with the number of agents drawing on the same induced skills?
  • Does a pre-execution skill utility score (combining specificity and abstractness) still predict success when skills are exchanged between heterogeneous agents rather than reused by the same agent?
  • What mechanisms or filters should govern which induced skills are admitted to a shared agent memory, given that some retrieved skills degrade performance below a no-memory baseline?
  • Do the findings that subtask-level and text-format skills transfer better hold across model families, scales, and task domains beyond those tested?

When Do LLM Agents Help? Deadline-Aware Mixed-Criticality Task Scheduling at the Autonomous-Vehicle Edge

Reza Zakerian

2608.19557 · nearest in the canon: Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems

Compares a contract-net auction heuristic and a multi-agent LLM control plane for deadline-aware mixed-criticality task scheduling at the vehicular edge, finding LLM benefit only under non-stationary load.

6 open questions
  • Beyond a single mid-run surge of safety-critical tasks, which classes of non-stationarity (drifting arrival rates, server failures, adversarial load patterns) open enough headroom for an LLM control plane to beat a static deadline-aware heuristic?
  • How does the benefit of an LLM control plane over heuristic scheduling scale with the number of edge servers, offloading agents, and concurrent task streams?
  • Does the LLM control plane's advantage survive comparison with learned adaptive policies (e.g. RL or contextual-bandit variants richer than the tested bandit) under the same non-stationary loads?
  • Do multiple independently-run LLM control planes bidding in the same contract-net auction produce emergent failure modes such as herding on the same server or implicit collusion against best-effort traffic?
  • Are the control-plane latency and monetary cost of LLM orchestration low enough for real deployed vehicular edge systems, as opposed to simulated instances?
  • Do the reported gains transfer to real-world autonomous-vehicle offloading traces rather than synthetic topologies and instances?

Reward-Guided Autoregressive Graph Generation for Efficient Multi-Agent Communication Topology Design

Poomphob Suwannapichat, Boonyarit Changaival, Caesar Wu, Pascal Bouvry

2608.20099 · nearest in the canon: Habermolt: Delegating Deliberation to AI Representatives

Fine-tunes an autoregressive generator of multi-agent communication topologies with a reward model rewarding both correctness and sparsity, cutting token use.

5 open questions
  • How do reward-optimised sparse communication topologies scale as the number of agents grows from a handful to hundreds or thousands — does the token saving hold, and does accuracy degrade at scale?
  • Do topologies learned with a compactness reward generalise to task distributions unseen during reward-model training, or do they overfit to the benchmark suite?
  • Does rewarding sparsity make multi-agent systems more fragile to a single faulty, adversarial, or prompt-injected agent, since fewer redundant communication paths exist?
  • What structural signatures (degree distribution, centralisation, information bottlenecks) do reward-guided generators converge to, and does the resulting concentration of information flow through hub agents create population-level failure modes?
  • Can the reward model's trade-off between task correctness and structural compactness be tuned as a steering knob at deployment time rather than fixed during fine-tuning?

An Irreducible Quantum Advantage in Aligning World Models with Reality

Josep Lumbreras, Hailan Ma, Jayne Thompson, Mile Gu

2608.19779 · nearest in the canon: Virtual Agent Economies

Constructs classical stochastic environments for which no finite classical world model preserves optimal action preferences while a single-qutrit quantum model reproduces them exactly.

5 open questions
  • How large is the policy-misalignment penalty in practice when a finite-memory classical simulator stands in for a long-memory environment — i.e. what is the measured regret of a policy trained in the approximate model versus the true process, as a function of memory size?
  • Do learned sequence-model world models (e.g. RNN/transformer or LLM-driven simulators) reproduce the predicted failure signatures — collapse of action distinguishability and persistent expected-reward error — on processes constructed to be classically irreducible?
  • Does the irreducible classical-modelling gap persist, or compound, when many interacting agents are trained inside the same imperfect world model, e.g. producing correlated policy errors or herding across the agent population?
  • Can the exact quantum world model be realised on available quantum hardware, and does its alignment advantage survive realistic noise and decoherence?
  • Is there a general characterisation of which real-world environments (as opposed to hand-constructed processes) admit no finite classical world model preserving optimal policies?