multi-agent system developers

Abstract: When one AI agent tells another that "the tests passed" or "a person approved this refund", the receiver has only the sender's word. A2A and MCP authenticate the channel, but the sender is the party in doubt, so a prompt-injected agent can launder fabricated facts and authority through a multi-agent system. We present Proof-Carrying Messages (PCM), a protocol in which each claim carries evidence: records from an append-only Merkle log that the agent cannot write, with inclusion proofs, signed by an attesting signer the agent cannot control. Receivers check messages offline and give each claim a three-valued result: proven, unproven or failed. A claim that rests only on the agent's own report is never proven. Authority travels as grant chains rooted in a recorded human approval, and the chains can only narrow. We specify four claim types, a typed caveat algebra, replay rules that survive restarts, and strict readings that make independent verifiers agree, and we prove the resulting security properties. On a pre-specified benchmark of 926 messages from real runs, 712 of them lies in 16 classes, PCM accepts 10 lies, all of them stale evidence (1.4%, 95% CI 0.8-2.6%). Plain A2A accepts all 712, and models of six deployed protocols (A2A with signed Agent Cards, uAgents, RFC 9421, SCITT, AP2, Verifiable Intent) accept 504-712. Two independent verifiers agree on all 926 verdicts, and verification takes about 1 ms.

Abstract: Training-time multi-agent system (MAS) design seeks to learn reusable orchestration policies, yet prevailing approaches remain largely one-shot: an Orchestrator generates a system but only observes and learns from its terminal outcome. This paradigm discards execution trajectories that expose system-level failures and contain fine-grained information beyond the terminal outcome, leaving no direct basis for revising the source MAS, evaluating the updated MAS, or learning from the resulting contrast. To this end, we introduce MAS-Loop, an execution-grounded framework that closes this feedback loop by coupling an Orchestrator with a trajectory-conditioned Refiner. For each task, the Orchestrator constructs and executes a source MAS to obtain an outcome score and an intermediate trajectory. Conditioned on the trajectory, the Refiner localizes an actionable failure, formulates a trajectory-grounded diagnosis, and proposes a bounded patch to the agents, information flow, or execution configuration. The executable updated MAS is re-executed on the same task to measure its realized outcome under stochastic execution. This pair-MAS comparison provides complementary learning signals: the Orchestrator learns to generate strong systems with fewer recoverable failures, whereas the Refiner learns to expose and repair observed weaknesses, thereby driving policy-level co-evolution across training iterations. Experiments show that MAS-Loop achieves the strongest overall performance among the evaluated baselines, with consistent gains from refinement across all evaluated benchmarks. The co-evolution process also yields MAS-JudgeCorpus, a dataset of execution-grounded source-updated pairs with orchestration- and trajectory-level supervision, which can provide the community with support for MAS reward modeling and trajectory-aware investigation.

Abstract: Large Language Models (LLMs) are increasingly being used as cognitive interfaces that shape how users interpret evidence, communicate, and make decisions. As deployed systems become more multimodal, tool-using, and multi-agent, safety failures no longer arise only from isolated harmful outputs, but also from how agents update, store, and propagate epistemic commitments across retrieval, memory, tools, and peer communication. In this paper, we define belief as an engineering abstraction for evidence-based commitments that guide downstream action. Building on this view, we propose the Explicit Belief Multi-Agent System (EB-MAS), which externalizes belief states into a structured substrate, uses belief operators to propose candidate updates, and applies a Cognitive Firewall to bound or reject updates that are weakly evidenced but strongly driven by vulnerability cues. We also discuss the risk of group belief at the group level, including manufactured consensus, collusion, and cascading misinformation, and outline practical integration challenges around grounding, false positives, scalability, and ethical misuse. Therefore, scalable agent safety requires not only better output filtering or stronger base models, but also explicit interfaces for tracing, calibrating, and constraining belief revision at runtime.

Abstract: As large language models (LLMs) increasingly operate as interconnected agents, a fundamental question emerges: does individual cooperation translate into collective cooperation? We investigate this question through the Iterated Prisoner's Dilemma (IPD), moving from dyadic interactions to homogeneous and heterogeneous networks of open-weight LLMs. Our findings reveal that cooperation is highly sensitive to model composition and seemingly minor linguistic changes. Models that cooperate with identical peers can shift toward defection in mixed populations. A simple uncertainty cue can increase or decrease cooperation, redistribute costs, and affect unexposed agents several hops away. Even reversing action-label order can dramatically alter behavior, with uncertainty framing sometimes amplifying this sensitivity. Together, these findings challenge the notion of cooperation as a stable model-level property, revealing how strategic behavior depends on interaction partners, network positions, and the language used to frame decisions.

Abstract: Agentic systems have demonstrated strong capabilities on complex tasks by enabling multiple interacting large language model (LLM)-based agents to collaborate and execute subtasks in parallel. However, this collaborative structure introduces new challenges for reliably evaluating and diagnosing system failures, particularly failure attribution: tracing a system's execution to identify the agent step responsible for an incorrect final outcome. Because manually performing this task is time- and labor-intensive, recent work has turned to LLMs as judges to automate the process, referred to as automated failure attribution (AFA). These approaches assume that a sufficiently capable model can infer the failing step directly from an interaction transcript. In practice, however, direct prompting often performs poorly, which motivates follow-up work that explores various ways to improve AFA performance, including alternative prompting strategies and inference-time interventions. In this paper, we examine a more fundamental question: whether this weak behavioral performance reflects a genuine capability limitation. Using linear probes applied to internal activations, we uncover a significant \emph{knows-but-won't-say gap:} models encode sufficient information to attribute a failure correctly but fail to explicitly output the responsible step. Building on this finding, we propose \emph{grafting}, a simple manual intervention that exploits knowledge already encoded in the model to improve AFA performance without requiring ground-truth labels at inference time. Extensive evaluations across multiple LLMs and benchmark settings demonstrate that grafting consistently outperforms direct-prompting approaches, underscoring the potential of leveraging latent model knowledge for more reliable failure attribution. Our code is available at https://github.com/wendyzheng22/mi4afa.

Abstract: Despite the success of video world models, extending them to multiple agents is difficult because it requires generating several agent-specific video streams that are consistent in time and between agents. To address this gap, we introduce OverSeer, a multi-agent world model that augments egocentric agent views with auxiliary camera views. We suggest that, by fixing the auxiliary views in space, they act as a proxy for a world-centric representation: Unlike the agents' views, which can change entirely due to a simple change in agent position, changes in the auxiliary views directly map to localized changes in the world state. We show that these fixed views substantially help the model to maintain consistency. We also show that, due to their redundant and semi-static nature, the auxiliary views can be efficiently compressed with a delta-VAE that encodes only the changes relative to a static reference image, reducing the number of tokens required to process them. Experiments in Minecraft and Doom demonstrate improved world and agent consistency, with qualitative results on real-world multi-camera captures indicating that the approach is not tied to game engines. Our compression analysis further shows that delta-VAE substantially reduces the computational overhead of auxiliary-view generation. We conclude that these fixed views provide a simple and efficient way to improve multi-agent world models.

Abstract: Online robust zero-sum Markov games pose a coupled operator-learning and game-solving problem: from nominal interaction data, the learner must recover a worst-case minimax Bellman operator while maintaining strategic coverage and controlling stage-game approximation error. Existing robust reinforcement-learning theory does not yet provide such an operator-learning perspective beyond specialized multi-agent settings. We develop an optimistic dual robust Bellman framework that combines dual robust Bellman fitting, optimistic minimax planning, backward-validated confidence sets, and exploiter-mix data collection. Our analysis is organized around the robust Bellman residuals induced by the candidate value class and yields a regret reduction that separates exploration complexity, robust operator-estimation error, stage-game solution error, and misspecification. Under explicit interpretable assumptions, we obtain conditional theorem interfaces for general and projected function approximation, together with closed results in several structured regimes, including tabular, epochwise cross-fit linear, self-normalized pure-online linear / finite-rank kernel, and projected RKHS truncation. Small-scale numerical experiments illustrate robustness calibration, strategic coverage, and the return-gap trade-off predicted by the framework.

Abstract: General-sum games are more challenging than fully cooperative or adversarial settings because agents may deceive or exploit one another. Achieving high social welfare requires sustained cooperation, while a single defection can drive agents toward a welfare suboptimal Nash equilibrium. Moreover, assuming access to other agents' rewards or internal states is often unrealistic and implicitly introduces additional cooperation into the setting. To address this issue, we introduce \textit{Reciprocal Q-Augmentation} (RQA), a method that uses a latent-variable opponent model to augment the Q-function with estimates of how one's actions affects other agents. RQA allows agents to learn both cooperative and defective policies using only their own Q-function. We show that, in social dilemmas, agents that all follow the cooperative policy of RQA learn to cooperate faster and reach higher levels of cooperation than fully cooperative algorithms. Moreover, a single RQA agent that switches appropriately between its cooperative and defecting policies can induce cooperation in agents trained with standard RL algorithms, which would otherwise end up defecting. We evaluate our approach on the Iterated Prisoner's Dilemma, Iterated Traveler's Dilemma, Iterated Stag Hunt, and \texttt{CleanUp-Lite}. Our experiments show that the algorithms implementing RQA outperforms fully cooperative methods with shared rewards and enables cooperation with algorithms that otherwise fail to reach high social welfare.

Abstract: Large-scale multi-agent ecosystems require mechanisms for discovering suitable agents without exposing entire registries to a large language model (LLM) orchestrator. The challenge is not only to retrieve relevant Agent Cards from a growing registry, but also to maintain a compact candidate context, since each additional candidate increases prompt size, inference cost, latency, and the risk of exceeding the model context window. We formulate this problem as candidate-context scalability and propose the Structural Signal Retriever (SSR), our deterministic and training-free retrieval method designed for Agent Card discovery. SSR derives structural lexical signals from capability-bearing Agent Card fields, weights them using inverse document frequency (IDF), and uses an Inverted Signal Index (ISI) for selective candidate scoring. We evaluate SSR using a synthetic agent benchmark, a public MS MARCO adaptation, and end-to-end LLM orchestration. At 1,000 agents, SSR reduces LLM input by 98.8% and orchestration latency by 70.9% relative to full-catalog orchestration. On MS MARCO, IDF improves retrieval effectiveness within the structural representation, while the ISI preserves retrieval effectiveness and reduces explicitly scored passages by 96.8%. End-to-end experiments show that candidate pre-selection maintains a compact candidate context as registry size grows and avoids the operational failures observed with full-catalog orchestration. These results support SSR as an Agent-Card-native, traceable retrieval mechanism for candidate-context scalability.

Abstract: Safety in agentic systems is inherently sequential: whether an action is safe depends on prior actions, tool outputs, information disclosures, and interactions among agents. Recent work has shifted agent safety guardrails from post-hoc trajectory-level assessment toward step-level action diagnosis and proactive risk prediction, yet existing training objectives provide limited supervision for how step evidence accumulates over time and when it becomes sufficient to justify intervention. We propose \methodlong~(ASPO): a training framework for step-level timely intervention. Pre-fix level stage alignment allows crediting each decision at its own step: it compares each response with the other responses sampled at the same prefix and rewards the action that fits the stage, continuing before the risk is visible, asking while clarification can still avert harm, and stopping before harm is irreversible. Adaptive feedback self-distillation further credits the reasoning behind each decision, with the model itself as a teacher that sees the step annotations on intervention steps and rules mined from its own false alarms on safe steps. Its training data carry a violation taxonomy and an intervention window at every step and are extended through self-evolving multi-agent red-teaming. Finally we present \textsc{StepSafe}, a step-level reasoning guardrail with timely intervention, that achieves the state-of-the-art peformance among open-source guardrails across five agent-safety benchmarks on safety-utility pareto curve with Youden's $J$ of 53.0 (4B) against 46.2.

Abstract: In LLM-based multi-agent systems (MAS), a local mistake can propagate as other agents build on it, turning a correctable error into a failed task. Diagnosing completed runs or repeatedly re-executing them to search for corrections does not provide a learned policy for intervening while execution is still in progress. We present Agent Runtime Inspection and Amendment (ARIA), a framework for learning when and how to repair ongoing multi-agent executions. Our key intuition is to turn execution-verified repairs into reusable supervision, moving repair search into data construction and training so that a learned model can intervene at runtime without repeatedly testing alternatives. First, we collect over 24{,}000 successful and failed trajectories across four multi-agent frameworks, six benchmarks, and eight host models. From these trajectories, we construct the ARIA dataset, to our knowledge the first large-scale dataset pairing failed multi-agent trajectories with execution-verified repairs. It contains 3{,}953 verified repair pairs, each specifying what went wrong, where to intervene, and how execution should continue. Then, we develop the ARIA framework for MAS-agnostic runtime inspection and amendment, designed to integrate with any MAS that exposes execution trajectories and supports continuation replacement. Using verified repairs as training examples, we teach a compact model to recognize when intervention is needed and how to correct an ongoing run. The aim is to learn repair strategies that apply to related failures in new configurations, rather than merely reproduce individual corrections. Finally, the trained ARIA model demonstrates strong repair performance in both post-hoc and runtime settings, with generalization to unseen frameworks and host models. Our code is available at https://github.com/LabRAI/ARIA.

Abstract: Large language models are increasingly deployed as teams of agents that hold different private information about a shared task. How should such agents concisely communicate in a way that promotes aggregation of decision-relevant information? Communicating only proposed actions can discard decision-relevant uncertainty: two agents may prefer the same action for different reasons. We study an alternative interface in which agents communicate calibrated beliefs over actions. We first analyze a simple synthetic setting that captures the structure of multi-round reasoning under partial information. In this setting, we show theoretically and empirically that communicating unbiased probabilities can be strictly more powerful than communicating actions. We then test the same principle in a collaborative maze-solving task with trained transformers and pretrained LLM agents. Our results demonstrate that probability communication outperforms action communication when the exchanged predictions preserve calibrated uncertainty. When communicated probabilities are biased, we show that the benefits of probability communication can disappear. We give a post-hoc conversational calibration intervention that consistently improves decision-making by correcting these biases both in a toy transformer setting and when pretrained LLMs use the corrections in context. These results suggest that probabilistic communication is useful for multi-agent collaboration not merely because it transmits more information, but because calibrated uncertainty gives other agents a reliable object to update on.

Abstract: In multi-agent reinforcement learning, exposing a cooperator to partners with diverse strategies during training is key to better collaboration at test time. However, existing methods typically rely on trajectory-level proxies to measure diversity, which may label two distinct trajectories of the same strategy as diverse. Instead, our key insight is to leverage LLMs' ability to describe behaviorally distinct agents in environments, similar to how humans can recognize high-level strategies. To this end, we introduce PopGen, a framework that synthesizes semantically diverse agents from scratch using LLMs. PopGen first generates a diverse set of textual descriptions of agent behaviors, then translates each behavior into a code policy that can be executed within the environment. We show that cooperators trained against PopGen populations achieve robust zero-shot cooperation with unseen partners, outperforming population-play baselines across four collaborative environments. We further show that semantic diversity scales with PopGen population size, improving the cooperator's worst-case performance, and we identify design choices that preserve population quality while maximizing diversity. Our code is available at https://github.com/labicon/PopGen.

Abstract: AI agents can evolve by using environment feedback to revise their harnesses, the software managing model inputs, tool use, memory, and execution, while keeping foundation models frozen. Existing studies focus on single agents, leaving joint harness evolution in multi-agent systems (MAS) underexplored. A single global optimizer faces growing context and workload as the system scales, while access restrictions may prevent sharing agent internals or system-wide feedback. We propose HarnessNet, a framework for hierarchical harness evolution in MAS. Its basic unit, Agentron, comprises an evolvable forward harness, a foundation model, and a local optimizer that can observe and modify only that harness. Agentrons form a layered graph through which environment feedback propagates backward, with upstream optimizers receiving node-specific feedback from the downstream layer. This decomposes joint harness evolution into smaller, coordinated tasks with separate contexts, reducing individual optimizers' context and workload while keeping access local. Optimizers need not access all agent trajectories or raw environment feedback. Experiments on AgentBench with multiple foundation models demonstrate the feasibility of training HarnessNet. Compared with a single global optimizer, HarnessNet achieves comparable performance gains while respecting local privacy and requiring fewer tokens and less time per optimizer call. A migration study shows that the evolved harnesses remain jointly effective after replacing the foundation models, suggesting transferability across models.

Abstract: A language model's ability to solve a task individually does not ensure that a team can solve it under unequal access to information and communication. SANZARU examines this gap through cooperative bomb-defusal puzzles adapted from BOMBANANA! (Lefto Studio, 2026). Task information and action authority are divided among three agents: one cannot hear, another cannot speak, and the third cannot see. The benchmark contains 120 puzzle instances across seven families, each targeting a distinct coordination ability. Across 1512 episodes, 7 models achieve an overall success rate of 49%, ranging from 20% to 86% under the evaluated configurations. Teams develop or adopt diverse communication protocols, from simple gesture codes to Morse and encodings of entire manuals. Yet trace analysis attributes about 60% of 771 team losses to communication, and agents attempt to use channels unavailable to their teammates despite explicit role instructions. Together, task outcomes and interaction traces enable evaluation of both collaborative performance and perspective taking throughout cooperation. The benchmark is available at https://github.com/markli404/sanzaru-benchmark.

Abstract: Long-horizon multi-agent collaboration requires memory that preserves both successful actions and the coordination structure underlying them. Existing memory systems support information sharing or procedural reuse, but often leave the evolution of shared procedures disconnected from the task dependencies and execution states of collaborating agents. This coupling also limits the direct transfer of procedural memory designed for individual agents to team settings. We propose Self-Evolving Procedural Memory (SEPM), a framework that uses task dependency graphs as a common representation for online coordination and procedural memory evolution. During execution, a lightweight shared blackboard records structured evidence alongside task dependencies, supporting the detection of goal, plan, and world-state divergence and evidence-based realignment. After execution, verified actions and dependencies are distilled into reusable procedures, while recurring failures inform warnings and constraints. The shared procedure pool evolves through creation, refinement, specialization, and retirement, guided by success and cost utility and constrained by evidence and safety requirements. Together, these mechanisms preserve the collaborative structure and applicability conditions needed to reuse experience across tasks. Evaluations on ALFWorld, WebArena, and OfficeBench under AutoGen and DyLAN show that SEPM achieves the highest task scores among the compared methods in all six benchmark settings. Under AutoGen, it improves task scores over the no-added-memory baseline by 29.85, 25.86, and 13.00 percentage points, respectively. The codebase is available at https://github.com/cookieApril/SEPM.

Abstract: Sequential social dilemma benchmarks capture diverse social interactions, but rarely test whether agents can infer hidden environmental rules through repeated interaction while adapting to other agents. We introduce \textbf{OpenSocialJax}, which extends Clean Up and Commons Harvest across an open-ended distribution of latent environment configurations, systematically varying spatial layouts and hidden rules governing tool crafting, payoffs, and resource dynamics. Each episode samples one configuration from tens of millions of possibilities and holds it fixed across five trials. We evaluate eight large language models, including three closed source models and five open source models. Our evaluation tests whether models can discover hidden rules and adapt to other agents in homogeneous and mixed teams. We find that rule discovery often improves performance but does not reliably sustain cooperation, as the cleaning burden becomes concentrated among fewer agents in OpenCleanup and every team eventually exhausts the shared resource in OpenHarvest. In our mixed arena, one cooperative agent was able to sustain the public good in OpenCleanup at a cost to its own return, but could not prevent resource depletion in OpenHarvest. Together, these results highlight the difficulty of adapting jointly to unfamiliar environmental dynamics and the changing behavior of other agents.

Abstract: The fast pace in which protocols, frameworks, and technologies are being proposed for LLM-based agents results in a complex scenario for anyone aiming to create state-of-the-art AI solutions. Without an assessment on how different protocols, frameworks, and libraries work together to compose an Agentic Web solution, developers and organizations face different challenges (e.g., skilling, governance, infrastructure) even before becoming part of the Agentic Web ecosystem. In this context, we propose a system for performing Agentic Web evaluations called MASTIF (Multi-Agent System TestIng Framework) and present a baseline assessment of Agentic Web technologies, including 3 LLMs, 3 protocols, and 4 agent implementation frameworks, against 300 Mind2Web benchmark tasks. Results suggest that selecting a task-specific technology stack may result in gains of task completion rate, efficiency score, and social value stability scores; and reductions in verbosity and time to complete tasks. With the presented results, we expect to present a baseline in terms of metrics and results for this complex and fast evolving scenario for those aiming at designing, assessing, and developing multi-agent systems.

Abstract: Evaluating agents by outcomes alone can obscure the capabilities that produce them. This problem is especially pronounced in evolving environments, where outcomes reflect a closed-loop interaction between agent behavior and changing external conditions. We introduce LiveMACEBench, a process-aware benchmark that uses live financial markets as a naturally evolving testbed for persistent LLM agents. Five frontier LLMs operate along continuous trajectories under matched Tool Use, Persistent Memory, Rule Following, and Multi-Agent Collaboration configurations. We evaluate them through both realized outcomes and mechanism-specific diagnostics derived from complete decision traces. Across 30 days of live evaluation, we find a pronounced outcome-capability gap: realized returns often diverge from capability-specific measurements, and similar outcomes can arise from markedly different patterns of mechanism use. Trace-level diagnostics further expose distinct bottlenecks across capabilities, demonstrating that mechanism access, effective mechanism use, and downstream performance are not interchangeable measures of agent capability. LiveMACEBench makes this distinction measurable, turning live markets from a performance leaderboard into a diagnostic environment for agent capability

Abstract: Social preferences can promote cooperation in multi-agent reinforcement learning, but existing approaches often require agents to observe the rewards of their peers. In many real-world interactions, however, an agent can, as humans do, observe others' behavior and outcomes without access to their private reward signals. We introduce self-referenced social preferences, in which each agent learns a model of its own reward, applies it to other agents' observed transitions to assess their outcomes from its own perspective, and feeds these self-referenced assessments into standard social preferences. We study two ways to incorporate these assessments: modifying the learning reward, or using them to weight policy updates. We evaluate the approach on three sequential social dilemmas, Escape Room, Clean Up, and Commons Harvest, which require volunteering, public-good contribution, and resource restraint, respectively. Across all three environments, agents learn cooperative behavior without observing others' rewards, including in settings where independent learners fail to cooperate, and frequently achieve more equitable divisions of jointly produced returns than agents with access to true rewards. The effective integration point depends on the social preference: inequity aversion works best in the reward together with a value look-ahead, whereas a purely benevolent preference benefits from policy-update weighting. Under partial observability, the policy-update approach continues to support cooperation. These results show that explicit access to other agents' reward signals is not necessary for learning cooperative behavior: social preferences can instead be grounded in self-referenced assessments of others' outcomes derived from their observed behavior.

Abstract: LLM-based multi-agent systems (MAS) engage tools, share memory, and delegate tasks, often encountering adversarial content. Current defenses for MAS are typically evaluated in isolation, focusing on one attack type at a time, which can lead to costly and hard-to-audit outcomes. This study organizes defenses into five principles, implementing them as DEFER1 (DEterministic-First Enforcement with Residual judgment), which includes a cascade of 28 checks that blocks what it can and refers the rest to a panel of four judges. In independent testing across four domains, attack success rates drop from about 30.0% to approximately 3.0%, with 78% of blocked attacks handled by deterministic checks. Only a quarter of proposals reach the judges in the security-operations domain, illustrating that the rules provide security for attacks violating clear policies, while judges manage those that only misrepresent intent. Both systems have weaknesses, such as a risk-score approval gate that inaccurately approves most attack proposals but few legitimate ones, highlighting the challenges in assessing threats accurately.

Abstract: LLM orchestration investigates how an orchestrator coordinates a group of autonomous agents to achieve common goals or maximize collective welfare. The agents are typically heterogeneous, each holding a private preference that it pursues but does not reveal. Inferring such hidden preferences from behavior has been a subject of long-standing research in game theory and multi-agent systems. The core challenge lies in maintaining a belief over every agent's preference and updating it from the agents' observed actions. Existing LLM orchestrators carry that belief as prompt text with no explicit update rule. This lets early errors persist and propagate rather than be corrected. We therefore propose \textbf{HARP} (Heterogeneous-preference Agent oRchestration via Preference inference), a novel framework that moves the belief out of the prompt. Specifically, HARP maintains one numeric posterior per agent over a finite set of candidate preferences and updates it in closed form by Bayes' rule. The language model supplies only actions and per-candidate likelihoods, so estimation is decoupled from its reasoning. We prove that HARP attains the same $\tilde O(\sqrt K)$ Bayesian regret as explicit joint inference when the factorization is exact. Furthermore, HARP\textsuperscript{+} augments planning with a bonus for actions that distinguish the candidates, so inference continues even when the optimal action is uninformative. Empirical results on three substrates, ranging from payoffs the preferences fully determine, through payoffs that depend on more than them, to scales where explicit joint inference is infeasible, demonstrate that HARP\textsuperscript{+} is the strongest non-oracle method across the class our theory identifies.

Abstract: Multi-agent simulations with large language models (LLMs) often operate networks of agents with a single base model. This overlooks the inter-model effects which may dominate engagement dynamics in real-world deployments. To show this, we simulate a heterogeneous social network powered by several different base models and show that the amount of engagement an agent receives depends more on its base model than on its assigned persona. The attraction or repulsion effects of a base model strengthen dramatically when more models are added in the mix, suggesting that networks dynamics may converge to base model effects at scale. To help explain this effect, we conduct a series of content-mediating analyses, showing the predictability of base models across contexts as well as the relationship between a model's lexical patterns and an engagement-maximizing style. In light of recent developments in mass multi-agent interaction, this work underscores the relevance of heterogeneous compositions in driving the outcomes of those networks

Abstract: Multi-agent systems built on large language models (LLMs) restate observations as a matter of course: relays forward them, shared boards repeat them and discussion rounds echo them. An aggregator that pools such messages should count sources, not statements. We convert a reported probability into units of independent readings, which assigns every restatement a copy weight, 0 for an aggregator that counts sources and 1 for one that counts every statement, and yields the implied decision under any cost structure. Three testbeds hold the evidence fixed and vary how it is restated: message logs with an exact Bayesian oracle, web documents with appended copies, and logs written by LLM agent teams under four communication protocols. Across four models from three providers, a forwarded copy counts for 0.06 to 0.42 of a new reading, mostly because some replies count every statement. On 5% to 40% of logs that state one reading three times, the reported belief implies an early commitment that the oracle never makes. The models that count copies least and most on controlled logs do so on web copies and agent-written logs as well. A one-paragraph declaration of what a copy contributes brings the copy weight on controlled logs to 0.08 or less. A rule that has agents refer to readings instead of restating them cuts belief-implied early commitment from 11.2% to 1.1% and preserves genuine corroboration.

Abstract: Large language model (LLM)-based multi-agent systems increasingly rely on memory to transform execution trajectories into reusable procedural knowledge. Yet repeated retrieval also makes memory errors persistent: memory pollution arises when outdated, weakly supported, or spuriously successful procedures become recurring components of future reasoning. Multi-agent execution introduces an additional structural risk. Scope collapse occurs when procedural knowledge escapes the coordination scope in which it was shown effective and is repeatedly reused at incompatible decision levels, allowing local errors to influence cascades of downstream decisions. Meanwhile, task-level failures provide ambiguous supervision because they rarely reveal which recalled knowledge was responsible. We introduce AECG, a framework for asymmetric experience consolidation and governance for multi-agent systems. AECG turns memory from static experience storage into a dynamic reliability-governance loop, preserving coordination scope and using multi-scale, confidence-aware reliability to detect degradation. It then combines degradation with downstream impact to prioritize high-risk knowledge under a bounded review budget, applies targeted interventions, and reactivates revised skills only after paired replay. Across three multi-agent frameworks and four benchmarks, AECG achieves the best score in 11 of 12 framework--benchmark settings and improves over the strongest competing memory method by as much as 10.23 percentage points; removing scope preservation reduces accuracy by up to 16.89 points. AECG thereby reframes multi-agent memory from passive accumulation into auditable reliability governance. Code is available at https://github.com/fenhg297/AECG

Abstract: There has been a surge of recent work on correlated equilibrium concepts in Markov games. However, existing results focus on concepts weaker than normal-form correlated equilibria (NFCEs), leaving open the more challenging question of computing such equilibria, which goes back to the seminal work of Papadimitriou and Roughgarden (JACM'08). Here, we establish the first efficient algorithm for NFCEs in finite-horizon Markov games with a fixed number of players $n$. In particular, with $S$ states, horizon $H$, and at most $A$ actions per player, it computes an $ε$-NFCE in time $S(AH/ε)^{O(n)}$. This is the first algorithm polynomial in $1/ε$ and the description of the game for NFCEs in an interesting class of problems beyond the normal-form setting. Moreover, under the usual assumption that recommendations are independent across states, we show PPAD-completeness---that is, computational equivalence to Nash equilibria---either in many-player games or when the precision is exponentially small. The key idea behind our approach is to run backward induction on a sequence of auxiliary stage games, but with the twist that in each step we compute a constant-expectation correlated equilibrium. This is a natural refinement of correlated equilibrium in which the conditional expected payoff from obeying is independent of the recommendation. In fact, our reduction goes both ways, establishing an equivalence between constant-expectation CEs and NFCEs in Markov games. For a fixed number of players, we observe that a constant-expectation CE can be computed approximately by combining linear programming with suitable discretization. In contrast, it is PPAD-hard in i) polymatrix (many-player) games at constant precision, and ii) two-player games at exponentially small precision. The latter result follows from an unexpected connection to rank-2 two-player games.

Abstract: Cooperative MARL is commonly evaluated through cooperation discovery from random initialization, leaving open whether continued optimization can destabilize learned cooperation. Actor-critic comparisons can also conflate critic presence with value gradients entering shared actor representations. We study cooperation maintenance, defined as the survival of a behaviorally verified cooperative policy under continued training. We formulate maintenance as a right-censored event-time problem and compare matched warm starts: X0 allows value loss gradients to update shared actor features, X1 retains the critic while blocking those gradients, and X5 removes the learned critic as a critic-free reference. This isolates direct value-gradient access while controlling initialization, critic computation, and evaluation. Positive reward scaling preserves strategic preferences and equilibria while perturbing learning dynamics. Gradient audits confirm the intended routing pathways, and frozen-policy torso perturbations probe whether route-induced updates align with local cooperation boundaries. In confirmatory MinEx and CleanUp-lite experiments, higher scales selectively increase maintenance sensitivity in X0; X1 remains near the censoring ceiling, and X5 has no confirmed events in the tested settings. In CleanUp-lite, route-by-scale displacement is associated with reduced local cooperation margins; MinEx shows a weaker, optimizer-dependent effect. These results identify a conditional, scale-sensitive maintenance risk associated with direct value-gradient routing rather than a universal failure of critics.

Abstract: We consider stochastic games with independent controlled chains and unknown transition kernels, where players observe only their local states and realized payoffs. We develop a fully online, decentralized, and uncoordinated mirror-descent algorithm that operates in the dual space of occupancy measures for approximating stationary Nash equilibrium (NE) policies. The algorithm uses a single transition/reward sample at every primitive time step, relies only on local information, and requires neither coverage of the joint state space nor synchronized episodes. Under uniform-ergodicity and finite-coverage assumptions, we show that, with high probability, the time-averaged fixed-comparator regret decays at the canonical $O(T^{-1/2})$ rate, up to logarithmic factors and polynomial dependence on the game parameters. In particular, the complexity depends on the cover times of the individual local state spaces rather than the product state space, avoiding exponential dependence on the number of players and the sizes of the joint state and action spaces. The resulting finite-time regret bound further yields an approximate coarse-correlated-equilibrium guarantee, which is natural for arbitrary reward functions since computing a stationary $ε$-NE is PPAD-hard in this setting. Under an additional global variational-stability condition, we show that the same fully online algorithm converges asymptotically in the last iterate to a stationary $ε$-NE. Our results provide a fully online and scalable learning framework for stochastic games with unknown independent chains. The algorithm can also be viewed as a primal-dual framework for Markov games that exploits the independence and local structure of the players' controlled transition chains.

Abstract: Concurrent actions in large language model (LLM) agent environments require arbitration even when each proposal is individually valid. We implement a typed snapshot-settlement contract and audit three distinct properties: order sensitivity, useful progress, and replay consistency. Five settlement policies are tested in 28,800 exhaustive permutation trials and 2,160 scripted multistep episodes. Joint policies are spatially order-invariant conditional on fixed priorities, yet conservative rejection completes only 31.25% of agents in a six-agent doorway task versus 90.28% for random tickets; the paired improvement is 59.03 percentage points (95% bootstrap interval: 50.00-68.06). All policies preserve the tested spatial constraints, and priority arbitration still misses the independent small-instance optimum. A separate full-state journal audit exactly replays 156 checkpoints and rejects 1,332 constructed corruptions with a retained terminal anchor. The evidence concerns execution semantics, not human realism or long-run fairness.

Abstract: Latent communication enables multi-agent systems to exchange information directly in internal representation space, reducing the token, computation, and latency overhead of text-based communication. To this end, lightweight trainable links are introduced to map the sender's representations into the receiver's input space. In this work, we show that even benign link training can increase harmful compliance relative to text-based communication while the underlying safety-aligned agents remain unchanged. An attacker can amplify this effect by optimizing the links on harmful query--response pairs or poisoning otherwise benign training data. We further develop a reinforcement-learning attack that rewards harmful compliance alongside benign task performance without requiring harmful target responses. Across three communication topologies and four safety benchmarks, this attack raises the mean harmful-compliance score from 27.9 with benignly trained links to 76.9. Compared with direct supervised optimization, it also achieves higher average accuracy on two benign utility benchmarks. Adapting the rewards toward safer behavior also enables repair of compromised links, substantially reducing harmful compliance across all evaluated attacks without updating the agents. Overall, our results show that safety alignment requires considering the multi-agent system as a whole.

Abstract: LLM-based multi-agent systems (MAS) increasingly use latent collaboration to avoid the information loss and repeated encoding-decoding overhead of natural-language communication. However, directly forwarding all sender latents makes the receiver-side context scale with both the number of agents and the reasoning length, increasing computation, memory usage, and collaboration latency. A natural solution is latent compression. But we find that cross-agent redundancy remains unresolved in existing latent compression approaches, which typically compress each sender independently and then concatenate the results. We propose LatCom, a cross-agent latent compression framework for efficient multi-agent latent collaboration. LatCom maps multiple sender latents into a fixed number of receiver-readable and task-relevant slots. Rather than reconstructing all sender hidden states, it optimizes the compressed latents for receiver-side task utility. LatCom trains the compressor in two stages: single-sender readability learning first establishes a latent interface interpretable by the frozen receiver, and multi-sender fusion learning then trains the compressor to fuse complementary evidence and remove redundancy across agents. Experiments on multiple benchmarks with Qwen3-4B show that LatCom achieves an average 2.46x inference speed-up over LatentMAS and reduces output token usage by 70.3% while maintaining comparable average accuracy.

Abstract: The tragedy of the commons poses a multi-agent safety problem: reward-seeking agents can deplete a shared resource, and cooperation among its users does not itself specify how much must be preserved. We make preservation an explicit requirement by formulating a regenerative commons as a constrained Markov game or a constrained multi-agent MDP with a designer-specified depletion budget. We develop a nonstationary Lagrangian framework that constructs a policy sequence from solutions of unconstrained games or cooperative control problems. Extending earlier time-average constructions, we introduce average-epoch solution concepts for reset episodes with discounted rewards and terminal costs. We prove a reward-independent feasibility certificate, cooperative feasibility and approximate optimality against feasible policy mixtures, and an extension to unbiased sampled costs. For self-interested agents, a constrained Nash certificate quantifies the price-dispersion term introduced by deviations that redistribute budget across epochs. Under the stated assumptions on solver accuracy and multiplier updates, these results give constrained policy-sequence guarantees using solutions of unconstrained problems. Experiments with constrained IPPO and MAPPO in a Gordon-Schaefer fishery examine how depletion budgets shape stock retention, harvest rewards, and price adaptation.

Abstract: In multi-agent systems, reliable consultation is challenging because advisor capabilities vary across tasks, and misleading information can make consultation worse than autonomous reasoning. We introduce BaRe-Mem, an online Bayesian reliability memory for multi-agent consultation. It estimates advisor reliability based on the central model's internal belief representations and updates these estimates from historical interactions. These estimates modulate the influence of advisor responses and guide the choice between consultation and autonomous reasoning. Across nine benchmarks and six central models, BaRe-Mem is more robust to misleading advisor information than debate and majority voting. On the more challenging tasks, it remains above autonomous reasoning across all tested misleading levels. Moreover, we extend the BaRe-Mem mechanism to worker allocation in agent teams. On the MuSiQue benchmark, BaRe-Mem improves task completion over routing by historical success counts and identifies capable workers earlier.

Abstract: Persistent memory lets language-model agents carry information across long-running collaborations, but leaves a lifecycle question open: what may a returning agent still act on once the shared state has changed? A memory can be correctly retrieved, relevant to the current task, and faithful to its source, and nonetheless be inadmissible for action: an itinerary saved before a pause still names the hotel the team has since replaced. We formalize this as temporal memory admission and present TRACE, a training-free layer that treats re-entry as an eligibility decision rather than a storage or retrieval operation, reconciling a departure checkpoint against absence-period updates, resolving explicit and implicit invalidation, and releasing a bounded Return View only when it covers the returning role's open obligations. We evaluate TRACE under three actor models on Memora, STALE Type II, and a derived ManBench-Return setting, each recast as return episodes: one agent departs, four teammates change the shared state, and the agent rejoins. What separates methods is not overall accuracy but whether one can retain valid memory and reject stale memory at once, and no single-policy baseline can: Restore (reinstate the departure checkpoint in full) admits stale state, Reset (start the return from an empty memory) discards valid state, each bottoming out at 0% on one of the two. TRACE is the only method high on both, reaching 92.6-98.3% valid-information availability with 98.4-99.5% invalid-information rejection on ManBench-Return, within 3.8 points of the best baseline's overall accuracy. On STALE Type II it improves Overall over the strongest comparison policy by 22.3 (Qwen), 18.5 (Gemini), and 27.5 (DeepSeek) points at roughly 2.3 times their tokens, while a write-time consolidation pipeline is more accurate still at 3.99 times TRACE's.

Abstract: Multi-agent LLM systems are increasingly deployed for complex, long-horizon tasks or emerge as a natural consequence of agents interacting in the wild. Yet they give rise to significant safety and security risks: the flexible protocols that enable task generalization also expose novel threats, from cascading prompt injection to inter-agent collusion. Progress in defending against these threats has been slowed by a lack of shared empirical infrastructure, which forces bespoke environment development for every new defense and makes standardized comparison impossible. Existing evaluations address isolated threat models or single-agent settings, but none jointly vary attack, defense, and architecture across realistic multi-agent environments. To address this gap, we introduce ORBIT, a configurable evaluation framework for empirical multi-agent safety and security research, built on UK AISI's Inspect. ORBIT lets researchers configure communication topologies, memory, scheduling, and agent roles. It supports four threat types and four defense strategies, as well as non-adversarial failures, with a benchmark suite spanning five scenario families covering browser use, computer use, agentic coding, customer service, and cooperative allocation. Our central finding is a gap in defense transferability across threats: per-action defenses that cut a compromised agent's attack success by 60 points on multi-issue coding give no measurable protection against colluding agents, and none of the defenses we tested generalized over all attacks tested. We further demonstrate security-performance tradeoffs and interactions between architecture and defense effectiveness. We make ORBIT available open-source at https://github.com/wlanderson0/orbit.

Abstract: Multi-agent systems built on large language models (LLMs) are largely homogeneous, as their agents behave alike even across distinct LLMs. We show that when such agents act concurrently without communication, they collide on targets they must split and diverge on targets they must take together, a double failure we term the symmetry trap. Theory of Mind (ToM), widely used for coordination without communication, cannot escape this trap, since homogeneous agents form the same prediction of one another and respond to it in the same way. We propose Theory of Scene (ToS), a training-free reasoning schema in which each agent reads its public role, the only difference between the agents, and the task context they all observe. Homogeneous agents thereby derive one division of labor, each taking the part its role fixes, which turns homogeneity from the cause of the trap into the cure. ToS reads the role together with the scene through role gating, which determines whether ownership overlaps or is already divided, and the task context through task coupling, which infers whether the team must converge on each target, divide it, or take its stages in turn. We evaluate on DivvyBench, a controlled environment we introduce, whose target types make an episode Competitive, Cooperative, or Mixed across Tabletop, Airspace, and Household scenarios, and on two established agentic benchmarks, GovSim and Overcooked. ToS outperforms all six baselines on every benchmark, and each baseline falls far behind it in at least one setting. Against ToM given the same inputs, ToS raises the DivvyBench success rate from 71.1% to 99.6%, the GovSim total gain from 207 to 400, and the Overcooked level-normalized throughput from 1.41 to 1.67.

Abstract: Model misspecification poses a fundamental challenge in multi-agent reinforcement learning, where transition uncertainty can be amplified by strategic interactions among agents. Distributionally robust Markov games (DRMGs) provide a principled framework for addressing such uncertainty, yet existing methods often rely on restrictive assumptions or scale poorly to large state and joint action spaces. We study online learning in general-sum DRMGs with general function approximation and $φ$-divergence uncertainty sets. We propose RoMEX-$φ$, a model-free framework that integrates equilibrium-based exploration with dual fitted learning. Through a functional dual representation of the robust multi-agent Bellman operator, RoMEX-$φ$ enables tractable worst-case value estimation from nominal interaction data using a centered empirical robust discrepancy. We introduce the robust Multi-Agent Decoupling Coefficient (robust MADC) to characterize the intrinsic exploration complexity arising from strategic interactions and adversarial transition uncertainty. We establish sublinear robust regret guarantees governed by the robust MADC rather than explicitly by the state and joint action space sizes, replacing tabular dependence with intrinsic function-class complexity. Numerical experiments on a scalable general-sum DRMG under total variation uncertainty show that RoMEX-$φ$ is substantially more resilient to transition shifts than its non-robust counterpart while remaining competitive with an exact tabular robust baseline. Our results provide a scalable framework for distributionally robust multi-agent reinforcement learning with general function approximation.

Abstract: Evaluating claim admission in shared agent memory is challenging because repeated claims may be mistaken for independent evidence. An agent may copy or paraphrase a retrieved belief, while admitting a false claim exposes subsequent agents to it. To study this problem, we introduce the Correlated Promotion Benchmark (CPB), which evaluates whether candidate claims should be admitted to shared memory.CPB-Static constructs a frozen test split from publicly annotated sources with fixed gold actions. CPB-Live runs multi-agent teams over a shared store, records all writes and retrievals, and tracks source lineage defined by each scenario. A separate consumer answers from the store alone. We evaluate eight admission policies across four agent families. Our results show that policies which deduplicate sources reject many true claims alongside false ones, whereas policies preserving answer coverage admit nearly as many false claims as unrestricted sharing. Gating on declared source type reduces false adoption to 0.06--0.09, compared with 0.22--0.47 for other answering policies. Once an uncontested false belief enters memory, the consumer asserts it in 0.97--0.99 of probes across all families. No non-oracle policy consistently rejects false claims across verbatim copies, paraphrases, and paraphrases declared authoritative. These findings reveal the limitations of admission policies without access to source lineage.

Abstract: Multi-agent deliberation can improve performance, but what happens when some agents do not act in good faith? In practice, an agent may be deceptive and work to subvert the group, whether through its own objectives or external instruction. We study how susceptibility to deception scales as groups increase in size and deceivers become more prevalent. It is not the number of agents in the group that matters, but the proportion of deceivers. We observe that the defection rate, how often initially correct agents switch to an incorrect final answer, rises linearly with this proportion. Whereas humans in comparable conformity studies are reliably swayed only when misleading confederates form a majority, LLM agents defect regularly even when deceivers remain a minority. Susceptibility also depends on which models are interacting, especially on the honest agent side. Unexpectedly, allowing deceivers to coordinate privately can make them less effective. Altogether, our results show that adding more agents is therefore not a sufficient defense, because the adversary can simply scale with the group.

Abstract: No two robots are truly identical: calibration, battery state, sensor drift and wear give every swarm a distribution of behaviour rather than a single point, usually treated as an imperfection to be minimised. In animal collectives the reverse holds: consistent individual differences in behaviour ('temperament') are shaped by natural selection and often decisive for group performance. This perspective proposes 'temperament engineering', a bio-inspired framework that treats the swarm's distribution of temperaments, rather than the individual controller, as the design object. It borrows five evolutionarily validated axes of animal temperament (shyness-boldness, exploration-avoidance, activity, aggressiveness and sociability) as a design vocabulary, rendering each as a continuous control parameter $τ\in [0,1]$ above the controller, realisable as a module threshold, a policy-conditioning vector in multi-agent reinforcement learning, or a constraint on a foundation-model planner. A three-phase workflow maps mission success criteria onto relevant axes, plans the shape of the $τ$ distribution, and tunes reaction norms governing how temperament responds to environmental cues. The payoff is greatest under decentralisation: where a central planner can reassign behaviour online, a temperament distribution is a planner output, but in a swarm without global knowledge it must be an offline, anticipatory design input. Behavioural and platform heterogeneity are thereby co-design variables, and I sketch tentative robot-native axes (self-model plasticity, forcefulness, initiative and expressiveness) arising from features robots have and animals do not. Engineered heterogeneity has been shown to outperform homogeneous swarms in tasks such as aggregation and exploration; establishing when, and how much, heterogeneity repays its cost is the work the field can now take forward.

Abstract: Anchor and Perturb (AnP) is a lightweight framework that resolves multi-agent coordination failures by decoupling exploratory variance injection from recurrent manifold stability. Existing remediation strategies predominantly alter mixing network architectures or enforce simultaneous exploration across the collective, which inevitably precipitates severe temporal-difference penalties in non-monotonic reward spaces. Specifically, AnP isolates underperforming lazy agents and injects an asymmetric exploratory pulse into targeted coordinates whilst anchoring converged teammates to nominal greedy exploitation. Empirical telemetry benchmarks demonstrate that AnP successfully rescues collapsed joint policies (recovering from a 5% evaluation win rate nadir back to 85%) and facilitates escape from suboptimal coordination plateaus, sustaining peak win rates of 90% without requiring structural network modifications.

Abstract: Multi-agent LLM systems are increasingly evaluated in social dilemmas, but most work treats governance as imposed by the experimenter, expressed rhetorically, or restricted to a fixed menu of mechanisms. We introduce GovSim-SelfGovern, an extension of the GovSim common-pool resource environment in which agents author executable Python governance rules, receive sandbox validation feedback, vote on proposed laws, and live under the rules they enact across rounds. To evaluate agent-driven self-governance, we examine three scenarios ranging from stable abundance to a fatal resource wall where five agents cannot all survive through harvest alone. To solve this, agents must write and debug useful laws in time before their institutions degrade sharply under resource pressure. Finally, we study a central alignment question: when agents hesitate to propose exile, are they rejecting it for normative reasons, or does it never enter their candidate set? Our results show that executable governance improves the space of possible interventions for agents, but survival depends on whether agents discover the right institutional mechanisms in time. Fiscal capacity enables redistribution, while deeper reasoning and removal of democratic veto make exile more feasible. GovSim-SelfGovern therefore adapts executable code actions to a common-pool governance setting and shows how scarcity turns institutional authorship into a political and ethical problem.

Abstract: LLM-based multi-agent systems generate collaboration traces that record how agents plan tasks, verify intermediate results, and repair failures. Reusing these procedures requires preserving an action's prerequisites and the outputs needed by subsequent agents. Our empirical studies show that grouping these dependencies into functional memory units improves their retention, while connecting units increases retrieval of the units and links jointly required by a task. The preferred combination of units also changes between instructions and checklists, even when each combination's content is fixed across formats. Updating choices from the outcomes of each combination and format pairing outperforms scoring combinations and formats separately. These findings motivate MACE, a memory-agent co-evolution framework that adapts memory organization and agent memory use through execution feedback. Its MemGoG structure represents functional units as subgraphs of related conditions, actions, and outputs, connecting them through support, conflict, and repair relations. MACE Loop selects task-relevant units and relations within a memory budget and provides each agent with instructions or checklists for its current operation. It records the selected units, presentation formats, agent outputs, and task outcomes to update unit scores and relations for retrieval and inform subsequent presentation choices. Across eight benchmarks, MACE outperforms ten baselines with an average score of 81.11%, compared with 78.97% for the strongest baseline, SAGE.

Abstract: An agentic society is a collection of AI agents that coordinate autonomously across trust boundaries, on behalf of different principals whose objectives may only partially align. We show experimentally that in agentic societies even honest, competent agents often fail to reach satisfactory outcomes with existing harnesses and messaging primitives, and that faulty or malicious agents can stall collaboration, influence outcomes, and pursue other harmful goals by exploiting vulnerabilities in communication (``speech''). We argue that agentic societies need a \emph{social harness} for inter-agent interactions, in addition to each agent's \emph{personal harness}, which manages its private context and communication with its principal. We propose a layered architecture for social harnesses which (i) prevents classes of failures outright, (ii) enables agents to detect invalid messages at runtime, and (iii) supports post-facto investigation and consequences, and highlight directions for future research to realize these capabilities.

Abstract: Biological cells can be viewed as individual, interacting agents whose collective dynamics give rise to adaptive behaviour at multiple levels of organisation, from individual cells through tissues to whole multicellular organisms. In this perspective and tutorial article we discuss whether intrinsic rewards in artificial neural systems can support adaptation, functional specialisation and higher-level self-organisation without a shared external objective. We review empowerment, curiosity, learning progress, information gain, unsupervised skill discovery, mutual information estimation and the use of world models for intrinsic reward computation. Particular attention is given to failure modes showing when such objectives do not produce sustained exploration or increasingly complex behaviour. We argue that more capable systems may require complementary objectives, communication, memory, learning at multiple temporal scales and environmental constraints. Based on this perspective, we outline three experimental directions. These include a resource-constrained environment in which otherwise stable behavioural attractors become unsustainable, allowing us to test whether environmental constraints can mitigate characteristic failure modes of intrinsic objectives. The network of recurrent agents with per-agent intrinsic rewards, and a hierarchical world-model agent in which exploratory motor competence develops before goal-directed behaviour. These experiments are intended to test whether intrinsic learning can lead to adaptive organisation at progressively higher levels.

Abstract: LLM-based multi-agent systems can fail even when communication succeeds because agents do not correctly track their peers' roles, knowledge, or intentions. We investigate whether such inter-agent misalignment cases, labelled FC2 in MAST-Data, can be converted into functional partner-state reasoning items. ToMAS applies four explicit convertibility criteria to diagnosed execution traces. A full conversion pass over 242 eligible non-AG2 training traces produced 39 CLEAN items. In an 18-trace reliability pilot, two annotators achieved 94.4% raw agreement and Cohen's kappa = 0.92. We then used the converted items as binary rewards in a small-scale GRPO feasibility experiment with Qwen2.5-1.5B. On a 28-item held-out Magentic GAIA diagnostic, every evaluated condition exceeded the ROUGE-L threshold on the same 2 of 28 items. Post-hoc adapter checks show why: under the learning rate used, the LoRA update remained numerically negligible (max abs Delta W about 7e-6), so all conditions decode identically to the untrained checkpoint. The experiment therefore does not show a training effect and cannot establish one; it reports an executable pipeline together with two limitations that any conclusive study must address: a provenance gap between the training and evaluation items, and lexical-overlap scoring. ToMAS provides a preliminary rubric and pipeline for converting diagnosed coordination failures into trainable partner-state reasoning items and identifies the requirements for a conclusive matched-domain evaluation.

Abstract: Ad Hoc Teamwork (AHT) addresses the challenge of designing agents capable of coordinating with novel partners without prior coordination. However, progress in the field is hindered by the prohibitive computational cost of the AHT research lifecycle, the lack of standardized benchmark implementations, and the absence of a diverse, validated evaluation teammate suite. In this work, we introduce JaxAHT, the first open-source, JAX-based library designed to accelerate and standardize the AHT research lifecycle. Leveraging JAX's hardware acceleration and massive parallelization capabilities, JaxAHT provides a unified framework for teammate generation, ego agent training, and evaluation against unseen teammates, achieving approximately 95x wall-clock speedup over PyTorch counterparts. Alongside the library, we contribute a diverse suite of evaluation teammates across the domains of Level-Based Foraging, Overcooked, and Hanabi. To illustrate the value of the framework, we use it to conduct a large-scale, compute-controlled benchmark study comparing teammate generation and AHT agent learning methods, finding that no algorithm consistently performs best, and that agent modeling primarily offers benefits in role-based scenarios with diverse teammates.

Abstract: In June 2026, thousands of AI agents found that a small public wiki would accept edits from inside their sandboxes, and started using it to help one another pass a timed test. Each agent lived for about an hour and remembered nothing afterwards. Nobody asked them to cooperate, and the wiki had not been built for them. The complete record of what they wrote is public, and it is unusually informative, because it preserves not only what each agent wrote but what that agent could see before writing. We use it to follow the three decisions an agent had to make on arrival: where to write, what to call itself, and how to word its message. One rule governs all three. An agent takes an option with a probability close to the share of that option in what it can see, and the share that matters is the one on the page in front of it, then the one in the stream of recent edits, and only weakly anything older. Three minimal copying models, one per decision and with a single free parameter each, reproduce the heavy-tailed distribution of how many agents met on a page, the frequency of the pieces from which the agents built their names, and the patchwork of pages that are internally consistent and different from one another. Copying whatever the environment happens to show is enough to produce most of the collective structure of this population. It is also what makes such a population easy to steer, since whoever writes first, or writes while the others are quiet, sets the convention for everyone who comes later.

Abstract: Solving Nash equilibria for general multi-player Markov games is computationally intractable, while two-player zero-sum Markov games admit fast last-iterate policy-optimization methods. Finite-horizon zero-sum networked separable Markov games occupy an important middle ground: they retain global competition structure through pairwise interactions, while preserving computational tractability of Nash equilibria (NE) in the full-information and known-transition setting. Existing algorithms for this class either proceed through equilibrium-collapse arguments for a simplified setting where a single controller determines the transition probability, or backward dynamic programming that relies on equilibrium solvers at each stage. However, the design and analysis of direct policy-update approaches remain inadequate. To address this issue, we propose the entropy-regularized optimistic multiplicative weights update (ER-OMWU), a complementary single-loop policy dynamic that updates players' policies symmetrically and returns an approximate NE in the last iteration. We provide a first last-iterate convergence analysis of policy dynamics in the games of interest: after $\widetilde{O}(1/ε)$ iterations, the returned policy is an $ε$-approximate Nash equilibrium. The result preserves the near-linear convergence rate achieved by policy optimization in two-player zero-sum Markov games, but extends the policy-dynamics viewpoint to a more complicated but structured multi-player setting.

Abstract: Quantum game theory is an extension of classical game theory that uses quantum principles in game theory. The Eisert-Wilkens-Lewenstein (EWL) quantum game is an early example of the two-player classical Prisoner's Dilemma transformed into a quantum Prisoner's Dilemma. In the EWL game, the players choose pure quantum strategies represented by unitary matrices. This extension can resolve the classical dilemma by enabling cooperative equilibrium with higher payoff. In this paper, we first discuss the Extended EWL (EEWL) for multiplayer quantum games with mixed strategies. In EEWL, each player controls a set of unitary operators as quantum actions and uses a classical mixed strategy over these actions. The payoffs are defined as expectation values of Hermitian reward operators acting on a shared quantum state, which is generated and measured according to the EEWL protocol. We then propose the Unitary Strategy Matrix Exponential Algorithm (USMEA), a geometry-aware sequential algorithm for the EEWL mixed-strategy setting, in which each player jointly learns a trainable set of local unitary actions and the associated classical mixing probabilities. Thereby it acts as a learning-and-control layer for multi-agent quantum decision systems. We analyze the convergence properties of USMEA under standard smoothness and step-size conditions and validate the theory with numerical experiments. These results show how classical optimization methods can be systematically integrated into the design and analysis of engineered quantum strategic interactions.