AI agent cooperation Instinct
Behavioral probes, consensus metrics and ROAR's real-time test
Today's set is unusually strong on measurement: behavioral probes for information flow inside LLM swarms, a framework that tracks consensus formation across 1,872 runs, a real-time multiplayer benchmark, and a red-teaming suite for colluding agents. On the research side, post-training is shown to amplify the memory curse in repeated dilemmas, and several papers attack the cost-quality frontier of multi-agent communication.
Evaluation
Behavioral probes make information flow inside LLM swarms measurable
Chen and colleagues repurpose the External Context Score (ECS) and Parametric Knowledge Score (PKS), originally designed for hallucination analysis, as behavioral probes of information flow inside LLM swarms. Across controlled and realistic swarm settings, ECS increases with input relevance and distinguishes useful multi-source contributions, while PKS decreases as informative context accumulates. The authors show the probes can complement utility-driven structure search, with statistical tests over node-level response and multi-source aggregation experiments.
Assessment
Most swarm evaluation scores only final task utility; per-node probes of who relied on what are the kind of signal needed to diagnose cooperation quality rather than just grade outcomes.
October 19 preprint, announced October 20.
Measuring Consensus in AI Agent Networks: 1,872 runs of position trajectories
Gao presents an evaluation framework that treats consensus formation as the evolution of publicly stated positions toward a shared region, tracking position trajectories, group dispersion, and the information expressed in public messages. Two case studies cover staffing decisions from distributed facts and reconciliation of duplicate information with a cancelled contract. Across 1,872 completed runs, sharing explanations is associated with more accurate group decisions, while the second task shows persistent disagreement even when the aggregate answer is accurate. Membership-change experiments probe robustness.
Assessment
Directly addresses the gap between aggregate accuracy and genuine agreement. A consensus metric suite is a natural companion to outcome-based cooperation benchmarks.
October 17 preprint, announced October 20.
ROAR benchmarks LMM agents under real-time multiplayer constraints
Wang and colleagues introduce ROAR, a multiplayer action-shooter environment covering cooperative, competitive, mixed cooperative-competitive, and single-agent scenarios. A Multi-Agent Asynchronous Gym interface runs the environment and each agent in separate threads, so the world evolves while agents reason. Tables compare open-source LMMs and calibration baselines under the traditional turn-based interface versus asynchronous runs at 250 ms and 40 ms intervals, with P50 and P90 latency reported per model.
Assessment
Most multi-agent benchmarks pause the world while agents think; making latency and stale decisions first-class moves evaluation closer to how deployed cooperation actually behaves.
October 16 preprint, announced October 20.
MiniLab red-teams distributed attacks for loss-of-control risks
Kitaoka and colleagues introduce MiniLab, a synthetic series of live services and codebases (five interconnected services, 249 benign tasks, eight attack objectives) adapted from internal AI deployment infrastructure. The authors demonstrate distributed attacks in which different agents carry out separate attack fragments, and show that the resulting covertness evades frontier monitors reading each trajectory individually.
Assessment
The adversarial side of cooperation evaluation: if monitoring stays per-trajectory, colluding agents can split a harmful objective below the detection threshold. AI-control evaluations need multi-agent attack suites like this one.
October 16 preprint, announced October 20.
Trust-safety margins measured in live language-based economic games
Galat and Rizoiu entered the GLEE competition's final week with one protected agent, Milton, plus four instrument agents running deterministic per-game randomized policy trials. In negotiation, learned pricing improves adjusted percentile by 0.071 (n = 21,168). In persuasion, a preregistered five-dose experiment over 7,391 assigned games, blinded until a fixed stop, shows that recommending every unit raises sales by 0.194 of the maximum (95% CI [0.147, 0.241]) where the receiver's slack makes obedience rational and lowers them by 0.067 where it does not. Milton stood 11th of 460 agents in bargaining at the stop; the authors also report how the final day's pool shift moved all five agents' ratings together while their play was unchanged.
Assessment
A preregistered, blinded measurement of how far a live language-based counterpart can be pushed before cooperation degrades: an empirical trust-safety margin, and a methodological template (instrument agents, rollback-guarded changes, fixed stops) worth copying into cooperation evaluation.
October 18 preprint, announced October 20.
Coordination research
CoopShift: post-training amplifies the memory curse in LLM cooperation
Liu and colleagues compare pretrained, supervised fine-tuned, and preference- or reward-optimized checkpoints within each of seven released model lineages across seven social dilemmas, playing 200 rounds with visible history limited to the last two or forty rounds. Expanding visible history reduces cooperation at every SFT and Aligned checkpoint; these declines exceed the base model's in six of seven lineages, with most of the amplification already present at SFT. With long history, both post-trained stages also cooperate less than base on average in each binary-action game.
Assessment
Evidence that standard alignment training measurably worsens long-memory cooperation in repeated games, directly relevant to choosing checkpoints for deployed cooperative agents.
October 16 preprint, announced October 20.
The Weight of Teamwork: architecture rankings are not identified by communication alone
Shi and colleagues use symbolic regression as an observable scientific-search setting and compare fixed-budget LLM teams differing in peer visibility and update timing, evaluating both the frozen candidate archives and the expressions selected from them. Distributing the call budget across more independent trajectories gives the largest measured gain (+0.70 mean held-out fitting score at the widest allocation). Under a common archive-wide selector, the implemented Ring policy reduces mean score by 0.33 relative to Independent, while Star and Complete remain unresolved; holding archives fixed and changing the delivery rule changes measured architecture gaps, including the sign of the Complete point estimate.
Assessment
A methodological warning for multi-agent architecture comparisons: report candidate generation and final selection as separate endpoints, or the ranking says as much about the selection rule as about the communication structure.
October 16 preprint, announced October 20.
Reciprocal Q-Augmentation cooperates without shared rewards
Morri and colleagues introduce Reciprocal Q-Augmentation (RQA) for mixed-motive multi-agent reinforcement learning without access to other agents' rewards or internal states, an assumption the authors note is often unrealistic and that builds cooperation into the setting. On the Iterated Prisoner's Dilemma, Traveler's Dilemma, Iterated Stag Hunt, and CleanUp-Lite, algorithms implementing RQA outperform fully cooperative methods with shared rewards and enable cooperation with algorithms that otherwise fail to reach high social welfare.
Assessment
Removing the shared-reward assumption matters: most cooperative MARL methods assume away exactly the incentive problem that makes cooperation hard.
October 17 preprint, announced October 20.
Inhibitory message passing fixes evidence crowding at lower token cost
Siboni, Springer and Rostami propose a neuro-inspired inhibitory message-passing rule in which cancellation absorbs duplicate copies of distributed evidence so every holder can still be heard. The rule raises peer-group accuracy at every tested group size at lower cost, holds across models and architectures, and ablations show the original evidence sentence, not the sender's restatement of it, is what must travel.
Assessment
A cheap, hand-specified routing rule that attacks a concrete failure of multi-agent debate (duplicated evidence crowding out unique evidence); the what-must-travel ablation is a useful design hint for message formats.
October 17 preprint, announced October 20.
ACTS learns who talks to whom and cuts tokens by 15%
Li and colleagues introduce ACTS, role-aware communication topology scheduling for LLM multi-agent collaboration. On code generation, multi-step reasoning, general knowledge, and competitive mathematics, ACTS consistently outperforms state-of-the-art baselines, boosting average accuracy by up to 6.19 percentage points while reducing token consumption by over 15%. A project website is available.
Assessment
Learned sparse communication topologies beat fixed ones on both quality and cost, evidence that cooperation structure should be treated as a learned variable rather than a design constant.
October 19 preprint, announced October 20.
Infrastructure
Federation over Harness evolves how agents share experience
Yao, Rabbani and Li propose Federation over Harness (FoH), which automatically evolves the harness for both local and global components of a distributed multi-agent reasoning pipeline: a server agent evolves the client harness for task solving and self-reflection while evolving its own harness for aggregating distributed reflections and maintaining shared knowledge. On ClawEval and ScienceAgentBench, FoH outperforms the SOTA single-agent harness by 53.57%, non-federated harness evolution by 29.76%, and fixed-harness federated self-evolution by 22.38%, while reducing task-completion tokens by 34.09%; evolving the server harness adds a further 4.50% over a fixed server harness.
Assessment
Removes the hand-designed prompts and workflows that currently gate experience sharing between agents; both ends of the pipeline becoming evolvable is the interesting infrastructure claim.
October 15 submission, announced October 20.
Serialization-invariant integrity checks for agent-published datasets
Gao and Clark observe that checksum mismatches on received data can reflect either altered content or serialization differences that preserve data equivalence under specified rules. They propose a serialization-invariant identification method with comparison rules and identifier computation, integrated with the DataFacts mechanism agents use to describe where data can be accessed and to publish reference checksums.
Assessment
Plumbing for trustworthy agent-to-agent data exchange: byte equality is the wrong contract, and equivalence-aware identifiers are what data-consuming agents actually need.
October 17 preprint, announced October 20.
Listen
Backtest Your Agent Against Yesterday's Web: Ryan Ahern, Exa at AI Engineer
Ryan Ahern, a forward-deployed engineer at Exa, argues that agents need version-controlled web search so teams can backtest agentic systems against the web as it existed at a past date, rather than trusting live search results that leak future knowledge. The 26-minute session demonstrates Exa Snapshot searching the web as of a historical date to prevent future-information leakage, and a live acquisition-prediction workshop shows why holdout backtesting exposes overfitting.
Assessment
Evaluation methodology for web-using agents: backtesting against historical web snapshots is the eval equivalent of a holdout set, and leakage from live search is a real and underappreciated contaminant.
Published by AI Engineer on October 19.
Editorial synthesis
Measurement is the theme. Information flow inside swarms, consensus as a tracked trajectory, real-time instead of paused worlds, collusion below the per-trajectory monitor, and preregistered trust margins in live games: five different answers to the same question of what cooperation quality even is. On the method side, the warnings pile up: post-training amplifies the memory curse, architecture rankings are not identified without the selection rule, and token budgets respond to topology scheduling. The infrastructure work tries to make sharing itself evolvable and data exchange equivalence-aware.
Editorial interpretation of the cited sources.