
ALT Constructing multi-agent system (MAS) applications across domains requires translating diverse requirements into executable systems, involving task decomposition, agent and tool selection, and execution organization. Existing studies have explored individual aspects of MAS construction, including agent generation, architecture design, workflow generation, and component selection. However, these methods are often developed and evaluated independently, using different task specifications, component resources, and execution settings. This makes it difficult to compare their construction capabilities or determine how they can be applied to different domains. To address this problem, we propose OpenMAS-GApp, a One-for-All benchmark for MAS application construction across diverse domains. It standardizes task inputs, component resources, construction interfaces, and execution environments to support MAS construction under consistent conditions. Within OpenMAS-GApp, we provide Graph Harness a

ALT As AI agents are increasingly deployed in real-world applications and take on social roles, dynamic social intelligence, the ability to continuously perceive multimodal cues, infer latent cognitive states, intervene through embodied actions, and ultimately achieve social goals, has become essential. However, existing evaluations largely fragment social intelligence into text-based abstractions, static reasoning tasks, or embodied simulators that lack social dynamics. These approaches fail to capture the multimodal, continuous, and interpersonal complexity of authentic human interactions. To bridge this gap, we introduce ALIVE-World, an embodied simulation environment supporting dynamic multimodal interactions encompassing egocentric vision, language, and non-verbal action. To provide socially grounded interactive feedback, we populate this environment with counterpart agents driven by a profile-conditioned, dual-system (intuitive-affective and logical-cognitive) architecture. Building

ALT Agent memory systems are mostly evaluated on fixed histories, and only their final answers are scored. In real use, facts change, sources differ in reliability, and users ask the agent to forget things. We present LEDGER, a memory layer in which a language model proposes facts and explicit rules decide which value is current, which source to trust, and what to delete. Each fact keeps its source and the period in which it held. LEDGER+EP adds a search over raw conversation turns for questions the ledger does not cover, and scrubs these turns on deletion. We compare both against no memory, full-history replay, retrieval over raw transcripts, and Mem0, with the same answering model and prompt, on a synthetic multi-session coding-agent workload and a 210-question subset of LongMemEval, and we score both the evidence each memory sends to the model and the answer. On the coding workload, LEDGER+EP in fallback mode answers 94.3% of questions correctly on five development seeds and 91.1% on fi

ALT Multi-agent systems combine distributed information through interaction, which can improve performance but also spread errors or suppress useful differences. Predicting when interaction helps is central to understanding and designing collective intelligence. We investigate whether a few dimensionless parameters can predict gains over explicitly defined non-communicating baselines across group sizes, network structures, and tasks, toward a Reynolds-number-like description of collective intelligence. Using distributed classification tasks, stochastic tracking networks, and neural cellular automata, we combine analytical and empirical evidence to examine information sufficiency, dynamical applicability, and finite-sample prediction. We construct inference systems with bounded random variables and matching specified joint moments up to total degree three, yet opposite interaction gains under a fixed readout. These moments alone cannot generally determine the gain sign. Tracking networks li
![Code-generating language models produce user interfaces that build and run, but the pages are often weak in hierarchy, spacing, typography and composition. We study whether a small code model can learn design taste through reinforcement learning in verifiable synthetic environments. The method has three components. First, a layered verification framework splits design quality into hard gates, render-grounded metrics, rubric judges, within-group relative preference and a blinded expert panel. A signal enters the reward only if it passes an admission protocol that tests agreement with experts, causal sensitivity, resistance to hack traps and stability. Second, following SCALER [1], synthetic design environments are parameterised generators with measured difficulty ladders: base pages are written once by an agentic coder, and instances are produced by syntax-tree perturbation operators and constraint samplers, scheduled by an adaptive curriculum with a quarantine rule against reward over-](https://pbs.twimg.com/media/HXVmFijWIAAzPTT.png?name=small&format=webp)
ALT Code-generating language models produce user interfaces that build and run, but the pages are often weak in hierarchy, spacing, typography and composition. We study whether a small code model can learn design taste through reinforcement learning in verifiable synthetic environments. The method has three components. First, a layered verification framework splits design quality into hard gates, render-grounded metrics, rubric judges, within-group relative preference and a blinded expert panel. A signal enters the reward only if it passes an admission protocol that tests agreement with experts, causal sensitivity, resistance to hack traps and stability. Second, following SCALER [1], synthetic design environments are parameterised generators with measured difficulty ladders: base pages are written once by an agentic coder, and instances are produced by syntax-tree perturbation operators and constraint samplers, scheduled by an adaptive curriculum with a quarantine rule against reward over-

ALT Cooperative speed control by connected and automated vehicles can reduce congestion at lane drop bottlenecks. However, effective coordination depends on vehicle roles and execution timing, while initial holding costs may be offset by later traffic benefits. This paper proposes an offline plan search framework that combines world model guidance with simulator verification. A traffic world model incorporating physical dynamics guides evolutionary search over vehicle roles, target speeds, holding durations, and start times. Paired SUMO simulations screen candidates for total delay reduction and limits on additional mean delay across vehicle groups. Accepted plans form a reusable library whose actions temporarily override a frozen MAPPO policy. Across 192 paired test scenarios at 60% penetration and a demand of 2,800 vehicles per hour, the searched library reduces mean total delay by 14.3% relative to MAPPO and by 3.8% relative to a handwritten plan library. The largest increase in group m

ALT Reinforcement learning (RL) is increasingly employed for real-time control in Open Radio Access Networks (O-RANs). However, the asynchronous nature of O-RAN feedback introduces stochastic delays between actions and their observable rewards, leading to misaligned credit assignment and unstable training. This letter formulates the O-RAN control problem as a Markov decision process (MDP) with delayed and stochastic rewards, where the true reward corresponding to an action is received after an unknown delay. To address this, we propose a delay-aware deep RL framework that integrates a long short-term memory (LSTM)-based delay predictor and a temporal reward realignment module within an off-policy learning pipeline. Using empirical delay traces collected from the Open AI Cellular (OAIC) testbed, we demonstrate that the proposed framework achieves faster convergence and significantly reduces latency violations and packet drops compared to conventional RL methods that ignore feedback delay. T

ALT AI tools have helped push submissions to the largest machine-learning venues to tens of thousands per cycle. Venues have answered with rules that tie access to review: per-author caps, mandatory reciprocal reviewing, reviewer eligibility earned through accepted papers, and assignment based on reviewer bids. Under these rules a reviewer is also a competitor of the authors. We study two deviations that the rules make attractive. Sabotage lowers a competitor's score. Collusion raises an ally's score. We build an agent-based conference calibrated to public statistics of NeurIPS and ICLR and run 6,220 simulated conferences. Four results stand out. First, the two deviations do different harm and amplify each other. Collusion changes who is accepted, sabotage erodes what scores say about quality, and rings that may also sabotage gain more than rings that only collude. Second, under random assignment the chance that collusion changes a decision is set by one exposure index, E = Rm/A, the expec

ALT Autonomous trading agents need controls that prevent problematic orders without making legitimate tasks impractical. We study this aspect of agentic compliance with TradeGauntlet, a replayable limit-order-book simulator that separates detection, prevention, and execution cost. In 1,680 simulated sessions with Claude Haiku 5.5 and Sonnet 5.5, all conduct alerts were self-matches. Gates eliminated those alerts by rejecting self-crossing orders, while the aggregate-gate condition reduced Sonnet's two-sided fills by about 15%. A conduct reminder reduced Sonnet's alert rate but reduced Haiku's fills without a detectable alert benefit. These results show why zero alerts alone do not establish an agent's compliance. Scripted probes and two search methods test weaknesses beyond those observed in the LLM runs. The searches find split-order schedules that evade the implemented conduct rules; a learned baseline catches all 2,400 tested searched episodes but also flags 15% of a large maker's sessi

ALT Virtual patient (VP) systems support nursing communication training by providing simulated encounters in which learners practise explaining care and responding to patient concerns. Although large language models (LLMs) support flexible dialogue, current approaches give less explicit attention to how patients' information needs develop and guide the content and timing of their questions. We draw on the Risk Information Seeking and Processing (RISP) model, which relates information seeking to perceived risk and the gap between perceived knowledge and desired understanding. Guided by this perspective, we propose the RISP Multi-Agent System (RISP-MAS), a VP system for patient-initiated information seeking. The system links ongoing assessment of the patient's perceptions and information needs to coordinated decisions about whether to seek information, when to ask and which uncertainty to address. Evaluation of nursing student-VP interactions suggested that RISP-MAS exhibited plausible patie

ALT Major industry multi-agent systems like Codex and Claude Code can be broken by attacks that have long been unsuccessful against modern frontier models. AI Safety as a discipline has largely focused on finding the bounds and vulnerabilities of AI models used in isolation. Yet as humans increasingly deploy these models as agents and multi-agent systems, the same attacks that had been thought to be patched when using a model in isolation can break these same models under specific circumstances. We show that these system breaks are not purely theoretical, but can break Claude Code, Codex, and a reference harness, all unmodified and used with their own sub-agent mechanisms. We show the attack by asking agents to answer questions from a store of files on disk in which we have planted a forged record. Searching the store themselves, the strong model orchestrators commit almost zero harmful revisions in over 1,000 trials. However, if that same orchestrator model delegates to a cheaper sub-agen

ALT Multi-agent deep-research systems use delegation to parallelize search and keep long tool traces out of the main agent's context. Yet delegation helps only when task allocation and subtask execution are coordinated: the main agent must issue useful subtasks, while subagents must return reliable evidence. Existing training either optimizes only the main agent or broadcasts the same team advantage to all agents, which can reward misleading subagent reports and penalize useful ones. We introduce LEAD-RL (Learning Effective Agent Delegation with Reinforcement Learning), a joint main-agent-subagent training framework with contribution-aware credit assignment. A training-time rubric evaluates the research utility of each subagent report and gates its inherited team advantage, while final-answer correctness remains the only team reward; no token, step, or speedup reward is used. Across five benchmarks and two model families, LEAD-RL improves answer quality. On BrowseComp, it improves accuracy

ALT Multi-agent systems of language models are often motivated by an averaging argument: if the agents' errors are unbiased and independent, the error of an aggregate of N answers falls as 1/sqrt(N). Fermi estimation, whose continuous answers are aggregated on the logarithmic scale, permits a direct test. We decompose the log10 error of each answer into the mean error of its model on the item, common to all its draws on the item, and a residual, and estimate both from 2.51M answers given before any communication by 15 open-weight models (3B-72B) on realFP, with replication under a single prompt pair on realFP, Science Olympiad Fermi questions and a guesstimation set. Given the item, teammates' errors are uncorrelated and teams behave as independent draws, but the item-level component accounts for a median 0.906 of the error variance over items. Teams of thirty therefore reduce a single agent's error by a median 1.7% (at most 5.4%), whereas draws that were also unbiased on each item would r

ALT When a tool-using LLM agent team violates policy, aggregate outcomes do not reveal whether an upstream message changed the executor's action. We introduce SwarmSafe-Delta, a paired executable protocol that freezes upstream computation, varies real, hidden, and format-matched placebo delivery, and scores a single committed action against deterministic benign and risky policy states under preassigned budgets. Across four models and constructed optional tasks, real delivery at N=4 produces 21.4% risky-world violations, versus 5.0% with placebo and 1.0% with hidden delivery (paired differences +16.4 and +20.3 percentage points). Real delivery also improves benign task success, illustrating that suppressing communication sacrifices utility. In a second regime constructed to require an upstream lookup route, real delivery improves safe completion by 74.2 points relative to hidden delivery. These effects quantify message-delivery contrasts under specified information dependencies, not natural
![The performance of LLM-based agents depends not only on the underlying model but also on the harness: the prompts, tools, and orchestration logic around it. Recent works automate harness optimization with coding agents that analyze execution traces, diagnose failures, and edit the harness code; however, those focused on single-agent systems. Multi-agent harnesses are harder to optimize as fixing them requires changing how agents interact in addition to how each behaves. We study whether automated harness optimization can diagnose and repair failures in multi-agent systems through a case study on complex information-seeking tasks from DeepSearchQA [Gupta et al., 2026]. We apply MULTILOOP, a simple optimization loop in which a coding agent iteratively analyzes, evaluates and edits the harness. With only 10 training samples, the evolved harness improves accuracy by 4.6% over the base harness and by 3.2% over the best single-agent system. However, a taxonomy of failure modes before and aft](https://pbs.twimg.com/media/HW_7ouMW4AAv2PY.png?name=small&format=webp)
ALT The performance of LLM-based agents depends not only on the underlying model but also on the harness: the prompts, tools, and orchestration logic around it. Recent works automate harness optimization with coding agents that analyze execution traces, diagnose failures, and edit the harness code; however, those focused on single-agent systems. Multi-agent harnesses are harder to optimize as fixing them requires changing how agents interact in addition to how each behaves. We study whether automated harness optimization can diagnose and repair failures in multi-agent systems through a case study on complex information-seeking tasks from DeepSearchQA [Gupta et al., 2026]. We apply MULTILOOP, a simple optimization loop in which a coding agent iteratively analyzes, evaluates and edits the harness. With only 10 training samples, the evolved harness improves accuracy by 4.6% over the base harness and by 3.2% over the best single-agent system. However, a taxonomy of failure modes before and aft

ALT Emerging personal AI agents raise fundamental scalability and privacy concerns for centralized multi-agent systems (MASs), which rely on a global orchestrator with broad access to agents, resources, and communication. In decentralized environments, global visibility is unavailable. Agents need to proactively discover non-local resources (e.g. tools, skills), and collaborate with other agents to solve tasks. We study decentralized multi-agent collaboration under graph-constrained communication, a natural reflection of human structures. Each agent observes only its local state and can interact only with neighboring agents. We introduce a benchmark with synthetic scalable distributed tool calling with 100K+ tools. Our experiments reveal that recent LLM agents struggle to discover non-local relevant resources and collaborate in decentralized setting. Moreover, standard reinforcement learning fails to reliably overcome this limitation, suggesting that decentralized MAS needs more explicit h

ALT Multiagent research inherited its economic vocabulary of auctions, markets, prices, bargaining, and matching from a single observed case: human societies. Those institutions bundle identity, competence, execution, ownership, commitment, and settlement in ways that suit agents whose minds are private, who learn rather than copy skills, and who each persist as one person. Engineered agents need not share any of these constraints, making the list of familiar institutions look less like the entirety of economic organization and more like a sample from a larger, unmapped space. We sketch one way to make that space concrete: a provisional chart that records each institution, separately from the agents that run it, as a typed structure over positions, rights and duties, allocation, information, time, state, process, value, and external recognition. In this chart, named institutions become overlapping regions rather than fixed points, neighboring families separate along a few coordinates, and

ALT Heterogeneous large language models (LLMs) exhibit complementary strengths, motivating routing and multi-model inference methods that seek to identify reliable predictions for each input. However, exploiting this complementarity remains challenging even when a correct prediction is available within the ensemble. We investigate this challenge across heterogeneous LLM ensembles and find that a simple single-round aggregation rule is highly competitive with learned routing methods, yet substantial oracle headroom remains. Further analysis shows that the evaluated single-round response signals provide limited discriminative power for identifying reliable predictions. In contrast, controlled re-presentations of the same question elicit complementary evidence about prediction reliability. Motivated by these findings, we introduce RE-PRESENT, a training-free framework that selectively re-presents disputed questions and directly aggregates candidate-level evidence across alternative views. Acr

ALT As large language models power autonomous agents, understanding how they collaborate and where they fail becomes critical. We introduce PARLEY, an extensible platform for studying LLM agent collaboration through four asymmetric puzzle tasks grounded in cognitive psychology (referential communication, joint coordination, perspective-taking, and spatial reasoning). GPT-4o agent pairs succeeded in 60.5% of conversations (N=400, 100 per task), with task-level variation (34-80%) reflecting distinct collaborative demands. From these conversations, we derived the data-driven PAC taxonomy: six failure types organized into Perception (visual interpretation and spatial reasoning), Adaptation (progress tracking and feedback integration), and Coordination (verification and grounding). Failure to incorporate external feedback was a critical indicator of unsuccessful collaboration, with agents repeatedly applying unsuccessful strategies. We also show that a taxonomy-guided intervention (N=200) impro

ALT When AI agents screen items for a costly human assessor, as in agentic search, e-discovery and AI reviewing of papers, only what they pass gets judged. We show what each kind of evidence can establish about their errors, even with unlimited data. Agreement shows that agents concur, not that they are right, however many there are. Acceptance by the assessor gives precision but only a lower bound on recall. Knowing which agents found which relevant items gives an upper bound on recall if agents' finds are not negatively correlated, and caps what one more agent can add if agents are interchangeable. Without assumptions about the agents, recall can be pinned down only by judging items they did not pass: a random sample of these estimates how many relevant items were missed, to any chosen accuracy, and the sample needed does not grow with the number of agents.
