Wangxuan Fan, Xiaoyu Nie, Zhoutian Shi, Xiangcheng Meng, Shipei Zeng, Pin Gao, Yan Hu, Zhongxiang Dai
2609.34679 · nearest in the canon: Modeling Earth-Scale Human-Like Societies with One Billion Agents
Combines LLM agents with contextual bandits to model decentralized bipartite matching in a simulated 50x50 marriage market. Validates LLM-induced preferences against empirical conditional-logit references. Reports higher mutual welfare and fewer blocking pairs than Gale-Shapley baselines.
▸6 open questions
- “Does the Bandit-UCB advantage in mutual welfare hold at market sizes larger than 50x50?”
- “Do the results hold in matching markets other than a Chinese marriage market, such as labour or school-choice markets?”
- “How do the matching outcomes change when agents misreport traits or learn to game the reciprocity model?”
- “How closely do LLM-induced mate preferences match empirical preferences outside the conditional-logit reference used here?”
- “What happens to welfare and blocking pairs when many agents share one LLM backbone and converge on the same preference ordering?”
- “Do decentralized LLM-agent matching results transfer to markets with real human participants on one side?”
Paul-Peter Arslan, Yubin Kim, Xiao Xiao
2609.35402 · nearest in the canon: Group size effects and collective misalignment in LLM multi-agent systems
Runs a pre-registered randomized experiment testing whether persistent pairing of Q-learning pricing agents raises learned prices in a Bertrand duopoly. Finds keeping the same partner raises average profits by 0.27 of the competitive-to-monopoly gap across twenty paired runs. Reports that hidden rival prices still raise resting prices even though punishment cannot be observed, making learned punishment inconclusive as an explanation.
▸6 open questions
- “Does persistent matching raise prices in markets with more than two pricing agents?”
- “Which matching or rotation policies can a platform use to hold learned prices near the competitive level?”
- “The registered test of learned punishment is inconclusive once a static best responder is netted out, so what probe separates learned punishment from static best response?”
- “Does the price rise under persistent partners survive when the price-setting module is a deep reinforcement learner instead of tabular Q-learning?”
- “Two model families show no effect and the Qwen results are inconsistent when the rival price appears in the prompt, so which model and prompt properties decide whether language-model pricing agents show the effect?”
- “Does hiding rival prices raise resting prices in live platform pricing, where agents also see demand signals and entry?”
Kirk Bansak, Jens Hainmueller, Dominik Hangartner, Jeremy Ferwerda, Elisabeth Paulson, Angie Delevoye, Nicholas Adams-Cohen, Ashwin Ramaswami, Selina Kurer, Joelle Pianzola, Michael Hotard
2609.35448 · nearest in the canon: On the limits of agency in agent-based models
Reports a randomized trial of algorithmic refugee-canton matching optimized for employment against existing placement procedures. Assigns about 2,000 cases across Swiss cantons with blinded officers and refugees. Finds pooled employment gains of 2.2 to 5.2 percentage points depending on cohort and time horizon.
▸5 open questions
- “Scaling employment-optimized matching to all cases could shift outcomes through congestion in local labor markets. What happens to aggregate employment when every refugee case gets an optimized placement rather than about 2,000 randomized cases under fixed quotas?”
- “Placement officers keep final authority over the recommendation. How often do they override the algorithm, and how does override behavior change the effect compared to the intention-to-treat estimate?”
- “The trial reports the share of months employed and the employment rate. What are the effects on wages, job quality, language acquisition, and other integration outcomes?”
- “The trial started just before COVID-19 shifted labor-market conditions, and effects were larger for the 2022-2023 cohorts. How much does a matching model trained on pre-shock administrative data degrade when labor-market conditions change?”
- “Algorithmic matching optimizes for average employment under canton and origin-group quotas. How does the gain distribute across cases, and which groups gain nothing or lose?”
Oriane Peter, Elena Simperl, Kate Devlin
2609.35302 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web
Introduces a method to measure topic saliency shifts across LLM model families during post-training, applied to climate change discourse. Tests interventions using specialized models to preserve perspective diversity. Reports homogenization affecting representation of diverse solutions across models.
▸5 open questions
- “Does the topic saliency shift observed for climate change discourse also appear in other contested domains, such as health, economics or migration?”
- “Which post-training stage produces the saliency shift, instruction tuning or preference optimisation?”
- “Do specialised models preserve perspective diversity when many agents query them repeatedly and retrain on each other's outputs?”
- “Does narrowed topic saliency in model outputs narrow the topics people raise in their own writing?”
- “How much does aggregate-level diversity measurement miss compared with topic-level saliency tracking across a full model population?”
An Yan, Yu Huo, Zhiwei Shang, Yiran Peng, Chenglin Wu
2609.34821 · nearest in the canon: An Economy of AI Agents
Introduces CEO Arena, a benchmark where LLM CEO agents make pricing, procurement, and R&D decisions in an eight-company market over 500 simulated days. Uses matched replacement evaluation to isolate an agent's effect on rivals and market. Finds most agents have negative mean returns, with private gains sometimes accompanying market-wide losses.
▸6 open questions
- “Most LLM CEO agents have negative mean returns over 500 simulated days, but the cause is not established; which decision types drive the losses?”
- “Private gains can accompany market losses in the simulated market, so how should a market designer set rules that penalise negative externalities imposed on rivals?”
- “Only four of 56 directed pairs show relatively stable effects, so how many runs or what design does a matched replacement evaluation need for stable pairwise effect estimates?”
- “Agents are all LLM-based CEOs in the benchmark; how do results change when human decision-makers take some of the company roles?”
- “Memory, action, and accounting traces suggest pricing and spending responses as explanations for the observed results; can tacit collusion be detected from these traces?”
- “The market holds eight companies fixed; how do outcomes scale as the number of competing agents grows?”
Sidharth Pulipaka, Ansh Sharma, Stanislau Hlebik, Leonidas Raghav, Vyas Raina, Ivaxi Sheth, Mario Fritz
2609.35576 · nearest in the canon: Beyond Single-Agent Safety: A Taxonomy of Risks in LLM-to-LLM Interactions
Introduces artifact-mediated propagation, where adversarial content stored in one agent's memory spreads to other agents through shared artifacts. Measures propagation across simulated multi-agent universes over time. Finds attacks reach 60-80% of agents with chains extending to eight hops in larger environments.
▸5 open questions
- “Which defences stop adversarial content from entering an assistant's persistent memory or from being copied into a new artifact?”
- “How does propagation depend on the topology of artifact sharing between assistants?”
- “Do the propagation rates measured in simulated universes hold in deployed assistant products with real users?”
- “What signals in interaction logs detect an artifact-borne attack before it reaches many agents?”
- “How does the propagation depth of eight hops and 60-80% reach change across model families and memory implementations?”
Ben Moore, Liam Dunne
2609.35077 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web
Analyzes about 2 million LLM citations from four commercial engines against 10,000 crawled pages using a nine-method statistical framework. Finds prompt-content alignment is the strongest predictor of citation, while standard AEO signals reverse under domain fixed effects. Reports domain-level AI authority exceeds other page features by a factor of six in SHAP value.
▸6 open questions
- “The study measures correlations between page features and citation frequency, not causal effects. Does changing prompt-content alignment on a page cause a change in how often production LLMs cite it?”
- “The pages come from nineteen B2B SaaS workspaces. Do the same page-level predictors hold for other content verticals, such as news, health or e-commerce?”
- “Domain-level AI authority exceeds the strongest non-alignment page-level feature by a factor of six in mean absolute SHAP value. What constructs this authority, and can a new domain acquire it?”
- “The four engines are treated together in a pooled citation corpus. Do ChatGPT, Claude, Google AI and Gemini differ in which page features drive their citations?”
- “Citation share concentrating on high-authority domains implies a feedback loop between citation and future authority. Does LLM citation concentrate over time on a shrinking set of domains?”
- “The standard AEO checklist effects reverse or collapse to zero once domain fixed effects are applied. Do content publishers who follow the checklist gain any citation benefit in practice?”
Ido Finder, Assaf Elovic, Gad Shalev
2609.34951 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web
Runs 37,927 agent journeys across four harnesses and 1,056 businesses to test how site readability (agent experience) affects AI-driven recommendations. Finds agent-ready businesses get answers built from their own pages 78% of the time versus 56%, with 1.9x more clear recommendations. Attributes the dominant failure mode to omission rather than fabrication.
▸6 open questions
- “Does making a site agent-readable cause the measured gains, or do readable sites differ in other ways? The study matches businesses on fame, prior model knowledge and two AEO proxies, but does not change a site's readability and re-measure.”
- “Which parts of agent experience drive the effect: robots.txt rules, JavaScript rendering, page structure, or fetch latency?”
- “Why do clear-recommendation rates vary sevenfold across the four agent harnesses?”
- “What happens to market outcomes when many businesses optimise for agent experience at once?”
- “Do agent recommendations that favour readable sites change actual buyer purchases and business revenue?”
- “Does the pattern hold for buyer questions in other languages and for business types outside the 1,056 sampled?”
Hanxun Huang, Yutao Wu, Qizhou Wang, Silvia Montaña-Niño, Yige Li, Xiang Zheng, Elif Buse Doyuran, Phoebe Matich, Xiao Liu, Xingjun Ma, Sarah Erfani, Christopher Leckie
2609.35028 · nearest in the canon: Retrieval Collapses When AI Pollutes the Web
Introduces VEX-Bench, measuring verification complexity of LLM-generated misinformation across 7 models and 7 generation methods over 5,880 articles. Defines a VEX score combining elicitation yield and verification complexity. Finds high-VEX content generated at 3x to 169x lower cost than agent-based verification, risking misallocation of scarce verification resources.
▸5 open questions
- “Screening prioritises high-VEX content and consumes verification capacity, but the size of the resulting misallocation in a working fact-checking pipeline is not measured; how much true misinformation goes unchecked when a fixed verification budget faces an adversarial generator?”
- “The benchmark scores verification complexity as perceived during screening by an LLM-as-judge; how do professional fact-checkers rank the same articles?”
- “What screening policies keep verification capacity allocated to harmful content when generators optimise for high VEX scores?”
- “The benchmark covers 2 misinformation categories, 6 domains, and 60 topics; do the cost asymmetries of 3x to 169x hold for other languages, domains, and low-resource open-weight models?”
- “How does high-VEX content spread once released, and does screening prioritisation change what reaches audiences?”
Mingxi Zou, Wei Zhu, Zhuo Wang, Langzhang Liang, Zhiwen Tang, Yinghui Xu, Zenglin Xu
2609.34136 · nearest in the canon: Group size effects and collective misalignment in LLM multi-agent systems
Introduces Waggle, a shared anonymous local policy that lets LLM agent swarms coordinate without explicit roles or global topology. Trains it with Swarm-Consistent Distillation combining anonymous-orbit consistency and rollout-grounded prediction. Reports the learned law retains over 96% of oracle quality as populations and budgets change.
▸6 open questions
- “Waggle keeps over 96% of substrate-specific oracle quality across coordination settings, but the abstract does not say how coordination quality scales as population size grows by orders of magnitude; does the learned local law hold at thousands of agents?”
- “What aggregate failure modes appear when many agents run the same anonymous local law, such as herding, monoculture or lock-in of a wrong commitment?”
- “A shared anonymous policy makes all agents interchangeable, so does a single adversarial or misaligned agent propagate its semantic messages through the swarm, and how far?”
- “Which observable signals in the local coordination field predict that a swarm will reorganize, and can an operator steer reorganization by editing those signals?”
- “The abstract reports transfer without retraining, but does not say which shifts in interaction conditions break the law; where are the limits?”
- “Does the mechanism hold when agents have conflicting objectives rather than a shared task?”