AI agent cooperation Instinct
Simulation butterfly effects, Fig elicitation and CATS credit
Today's measurement papers probe how fragile multi-agent findings are: persona representation alone swings two-agent cooperation by 75 points in social simulations, an agentic elicitor can programmatically provoke six social behaviors for testing, and per-agent credit assignment turns topology search into an auditable procedure. On the research side, a graph-coloring guarantee for merging parallel agent work, latent theory-of-mind tokens, and evidence that LLM debaters change behavior differently from what they report.
Evaluation
Butterfly effects in LLM social simulations demand robustness audits
Ye and colleagues study how implementation choices propagate through LLM social simulations and alter collective outcomes, using a repeated Prisoner's Dilemma and a social-media echo chamber across four LLM families with representation- and design-level perturbations. Changing only the representation of a matched persona shifts two-agent cooperation by 75.2 percentage points for GPT-5.2, while game framing and memory representation produce smaller or less consistent effects. Multi-agent interaction reshapes representation sensitivity, amplifying or attenuating it depending on the model and perturbation, and in the echo chamber persona representation changes interaction targeting with divergence in collective outcomes.
Assessment
A direct threat to the reliability of simulation-based cooperation claims: if a persona's representation format can swing cooperation by 75 points, single-configuration social-simulation results are close to uninterpretable without the robustness audits the paper proposes.
- The Butterfly Effect in LLM Social Simulations: From Architecture Sensitivity to Robustness Audits —
October 20 preprint, announced October 22.
Fig programmatically elicits cooperation, deception and sabotage for testing
Gu, Wang and Ouporov introduce Fig (Finding Interactions in Games), an agentic system that takes a natural-language behavior description, decomposes it into sub-behaviors, generates new social games or selects existing ones, and executes single or combined interventions on the games' engine, opponents, instructions and context. Evaluation combines built-in mechanical checks with synthetic rubric judgments, validated by human review. Models attempt the target behavior in 45.2% of episodes on average across valid Fig-generated scenarios for six test behaviors (cooperation, exploitation, deception, covert coordination, betrayal and sabotage), and Fig's interventions raise attempt rates by 10.9% on average over matched control games.
Assessment
Most multi-agent social evaluation observes whatever behavior emerges; targeted elicitation turns environments into programmable test beds, which is what systematic red-teaming of cooperation and defection actually requires.
October 20 preprint, announced October 22.
CATS assigns per-agent credit and prunes multi-agent teams auditably
Tritthart and Zehle propose Contrastive Attribution Topology Search (CATS): within a fixed supergraph of specialized agents, an LLM attributer compares execution traces across candidate topologies and assigns credit to individual agents, then updates edges while progressively pruning weak edges and agents to yield a small deployment team. With small open-weight models and matched candidate evaluations, CATS improves mean accuracy on BigCodeBench by 5.3 percentage points over score-only heuristic search while using 54% fewer deployment tokens, matches the strongest baselines on MBPP with fewer tokens, and in a HotpotQA prompt-poisoning stress test excludes all adversarially prompted agents in all three runs where random search retains some at comparable accuracy.
Assessment
A single task score cannot say which agent helped or hurt; trace-contrastive credit makes team design auditable and, as the poisoning test shows, doubles as a security filter for who gets to stay in the loop.
October 21 preprint, announced October 22.
Coordination research
Graph coloring gives a tight guarantee for merging parallel agent work
Li, Demidovich and Richtárik model parallel agent collaboration as discrete block optimization with an exactly evaluable objective and a known dependency graph: independent sets certify compatible edits whose verified gains add exactly, and a coordinator constructs a safe compatible candidate or evaluates interacting edits to capture favorable interactions. An optimal fractional coloring yields a merger preserving a 1/chi_f(G) fraction of total verified local improvement (chi_f being the graph's fractional chromatic number), proven tight among mergers restricted to selecting subsets of the same proposals; under further conditions repeated collaboration converges geometrically. Experiments cover repairing conflicting edits, scaling compatible parallel work, road-network and course-timetabling proposal integration, and scheduling parallel route searches.
Assessment
The companion piece to this week's Local Work, Global Progress convergence theory, now with a tight quantitative bound on how much verified improvement merging can keep, the beginnings of a design theory for parallel agent teams.
October 21 preprint, announced October 22.
ToM2Vec trains latent theory-of-mind tokens from the partner's own thoughts
Sivakumaran and colleagues introduce ToM2Vec, which forms abstract theories of mind in a latent space via a fixed budget of continuous reasoning tokens instead of text, which more easily describes only a single partner state. During training the ToM tokens are supervised by the partner's actual thoughts, which they learn to abstract and match; at test time each agent generates its ToM tokens as conditioners for its own thoughts and actions. Across several collaborative environments, ToM2Vec consistently improves performance, latent ToM reasoning outperforms trained natural-language ToM, the gains are partner-agnostic and transfer across model families, and the tokens infer partner beliefs, desires and actions more accurately.
Assessment
Training agents to model partners in latent space, supervised by the partner's real reasoning traces, is a concrete route past the shallow single-guess models of others that limit LLM teamwork.
October 21 preprint, announced October 22.
Bayesian evolutionary search designs multi-agent topologies under budget
Jehle and Zehle propose a Bayesian Evolutionary Optimization framework for discovering task-specific multi-agent topologies: evolutionary operators generate diverse workflow graphs while a graph-aware Gaussian Process with Expected Improvement focuses costly evaluations on promising candidates, replacing LLM-based candidate selection with an explicit criterion. On BigCodeBench, evolved architectures achieve the highest overall performance among the evaluated agent systems; on temporal tabular datasets, optimized multi-agent preprocessing improves over default TabPFN on several tasks while remaining competitive on IID data.
Assessment
An explicit, sample-efficient selection criterion for architecture search addresses the problem that LLM-judged candidate selection carries judge bias into team design.
October 21 preprint, announced October 22.
LLM debaters shift behavior differently from what they report
Siahkamari, Nguyen and MacNeil introduce a controlled LLM-LLM debate framework on climate-change attribution in which one role holds an assigned stance, sometimes with an explicit instruction to persuade, while the other receives no position, tracking both explicitly reported positions and turn-by-turn behavior across four ordered combinations of GPT-6 Luna and Gemini 2.5 Flash Lite. Opinion dynamics differ substantially across model combinations: same-model interactions narrow the opinion gap on average (clearly for only one of the two models) while cross-model interactions widen it. Explicit reports and in-dialogue behavior capture distinct aspects of opinion dynamics, and a blanket instruction to persuade does not reliably increase influence.
Assessment
A methodological warning for persuasion and deliberation research: measuring only stated positions misses how models actually shift in conversation, and influence is a property of the model pairing, not just the instruction.
October 20 preprint, announced October 22.
Listen
Coordination and Review at Agent Scale: Outgrowing the Vibe Code, Ch 20 Pt 2
An 11-minute episode on what happens when five people each run agents all day: how to divide work so agents stay out of each other's way, what stops two agents editing the same file at the same time, and how a team survives the pile of pull requests agents produce without anyone living in the review tab. The show frames the harness as the team's shared infrastructure and walks through splitting work to avoid collisions plus review strategies for the output flood.
Assessment
A compact practitioner take on the human side of multi-agent coordination, division of labor and review bandwidth, that complements this week's formal merging and credit-assignment results.
Episode dated October 21 in the publisher feed; the show site lists all episodes.
Editorial synthesis
Reliability is the thread. Social simulations swing 75 points on representation choices, so findings need audits; behavior needs to be elicited on purpose rather than observed by accident; and team scores need per-agent credit to mean anything. Meanwhile the merging of parallel agent work gets its first tight guarantee, and theory-of-mind moves into latent space where partners can be modeled with nuance text cannot hold.
Editorial interpretation of the cited sources.