The use of large language models (LLMs) to coordinate multi-agent teams raises a fundamental architectural question: whether task allocation should be centralized in a single LLM controller or distributed across autonomous LLM agents, particularly when communication is unreliable. In a deliberately simplified perimeter-defense simulation against LLM-controlled intruders, we compared a central commander, an auction-based consensus team and independent agents under five network conditions. Breach rates were 15% for the consensus team, 32% for the commander and 29% for independent agents, and of the different network conditions, only limited radio range changed a team's breach rate, raising the commander's. In consensus team, giving a low-cost LLM more authority over the bid computation raised breaches by 38 percentage points. With birds and civilian drones in airspace, we added a photograph for each to the input text state. Only the near-frontier model treated them apart from intruders. Intruders disguised as existing taxidermy bird drones were mistook as birds by both low-cost and near-frontier LLMs, going unpursued in 59-61% of near-frontier decisions (1-16% before). However, a one-sentence forewarning in the prompt lowered it to 0-25%. We find that LLM team with a dedicated external conflict resolution protocol results in fewest breaches. Code: https://github.com/the-kas-lab/consensus-within-competition-between.
Effective computer-use agents (CUAs) should seamlessly combine visual GUI interaction with code and terminal execution within a single policy. We present Mage-VL-Cowork, a 4B hybrid CUA trained from a general multimodally aligned checkpoint through an end-to-end, progressive five-stage curriculum across midtraining, SFT, and RL. To power this curriculum at scale, Mage-CUA-Forge replaces per-task agent construction with reusable, skill-conditioned generation workflows and application-specific preparers, synthesizing 293,806 desktop and web tasks across 6,915 skills while reducing recurring per-task expenditure by 82.6%. Rather than directly distilling a default teacher, we evolve the teacher harness and instantiate specialized CUA harnesses to yield high-quality trajectories alongside stage-specific supervision. Mage-VL-Cowork-4B finally climbs to 66.0% on OSWorld-Verified through different training stages with more efficient reasoning cost, achieving state-of-the-art performance at its parameter size and pushing forward the frontier of on-device intelligence.
Shikha Shrestha, William Yeoh, Ning Zhang et al. · cs.AI, cs.CL
Agents 96%
Abstract
LLM agents serve natural-language requests through tools, data sources, and other agents, and even simple requests often need several interdependent calls. Typically, the LLM picks these calls by trial and error, with no guarantee of an executable plan. We present Capability-Grounded Agentic Planning (CGAP), which instead reasons about what a request requires and what tools provide. CGAP compiles the tool catalog deterministically into a PDDL model, uses an LLM to turn the request into goals, and solves the model with a classical planner. Since tool outputs are unknown in advance, the model is abstract, tracking which inputs are available, not their values. During execution, an LLM adds entity-specific calls that abstract plans omit, and deterministic checks reject any that break the plan. Every argument is drawn from a cited source, such as a conversation span or an earlier tool response, and each requested result counts as delivered only after an LLM verifier checks it against the request. With three LLMs on ToolSandbox, ComplexFuncBench, BFCL multi-turn, and $\tau^2$-bench retail, CGAP outperforms a standard tool-calling agent by up to 28.9 points.
Parallel LLM agents tackle complex tasks by exploring multiple trajectories, yet useful discoveries remain isolated in individual execution histories. Existing sharing methods often select or summarize information before other agents' needs become clear, potentially overlooking details that become important later. We propose Ask on Demand (AOD), a communication framework that enables agents to seek targeted peer assistance when specific information needs arise. Upon an approved request, AOD retrieves relevant evidence from peer execution histories, selects a suitable peer, and generates a response tailored to the request. To determine when such assistance is worthwhile, we train a lightweight gate using paired team continuations, with supervision derived from final team outcomes and inference cost. Experiments on AppWorld and WebShop demonstrate that AOD consistently improves task performance over independent parallel execution while achieving lower inference cost than existing information-sharing baselines. These findings highlight the benefits of aligning peer communication with agents' evolving information needs to achieve more effective and efficient collaboration.
Xuanang Gao, Kepeng Lin, Xinyan Shi et al. · cs.AI
Agents 93%
Abstract
Supervised fine-tuning of Model Context Protocol (MCP) agents is typically conducted on a static data distribution, which fails to adapt to the student's evolving capabilities and leaves newly exposed weaknesses insufficiently addressed. To address these challenges, we propose \method, an agent--data mutual evolution framework that enables MCP agents to improve through a closed loop of task evolution, trajectory evolution, and supervised fine-tuning. Specifically, \method\ assesses the student's established capabilities and capability gaps from rollout feedback to identify learning targets. Task evolution then translates these targets into executable tasks by adapting the instructions and observable environment states of compatible verified seed tasks while keeping their verifiers aligned. To provide reliable and efficient demonstrations for these tasks, trajectory evolution uses reusable skills to recover failed teacher rollouts and refine successful ones. The resulting task--trajectory pairs are used to fine-tune the student, while accepted tasks seed subsequent evolution. To close the loop, we re-evaluate the updated student and use its feedback to guide the next round of data evolution. Experiments on Toolathlon, MCPMark, and MCP-Atlas demonstrate consistent improvements over the initial student, yielding absolute gains of 16.5\%, 10.2\%, and 12.6\%, respectively. Project page: https://github.com/xagao/EvoMCP.
Current LLM ensemble frameworks focus on multi-step setups with additional modules for answer ranking, often opting for token and span analysis rather than structured outputs, leading to heavyweight architectures with potential fail states along the pipeline. Faster, lighter solutions are more vulnerable to hallucination propagation and can lack output control in more complex pipelines. This paper proposes a customisable, lightweight ensemble workflow of coordinated Large Language Models that leverages JSON-structured outputs and anonymous peer-review ranking to mitigate hallucinatory outputs and single-model failure points. The pipeline is demonstrated on a relation extraction task applied to popular science articles in English, targeting four ontologically-grounded relation types (strong causation, weak causation, contrastive, and compositional), with semantic node canonicalisation and interactive, colour-coded HTML causal graphs as the final output. Performance is evaluated through an anonymous user study, achieving an average perceived accuracy of 0.778 against a human-annotated gold standard. The modular architecture supports flexible deployment across both API-based and in-house LLM setups, and the full framework is released under an open license to foster reproducibility and collaborative research.
Genliang Zhu, Xiaobo Chen, Ziyuan Wang et al. · cs.AI
Agents 91%
Abstract
Tool timeouts can leave AI agents facing ambiguous tool outcomes and partial execution: an action may have failed, committed or remained in flight while other workflow steps have already completed. We formulate evidence-guided task recovery over approved obligations, provider capabilities and states compatible with public evidence. A shared admissibility check handles key expiry, authorization timing, dependencies and residual effects; a terminal certificate separates actual goal satisfaction from publicly supported safe completion. The known-model implementation combines standard observation-contingent planning with a separately implemented effect-history adjudicator. In a finite study of 37 assignments, planning and three active controls each certify 28 completions; four saved synthetic fee units incur additional computation. A preregistered assignment-held-out comparison then evaluates planning advice against evidence prompting with two local model bundles. Across 58 paired units in 19 previously developed task families, advice certifies 30 completions versus 9, a 36.2-percentage-point difference (exact family sign-flip p=0.0078), using 272 proposals versus 399. All 116 trajectories pass independent replay. Both models exhibit a positive completion difference. The comparison measures the evaluated advice interface under a common shield; search-versus-rule advice requires a separate matched comparison. Unknown and handoff outcomes remain explicit, with guarantees tied to the declared provider capabilities rather than arbitrary third-party exactly-once execution.
Deploying 6G network slices across administrative domains requires an online decision for each arriving request: which domain hosts each virtual network function, how traffic crosses domain boundaries, and whether the slice is admitted. This joint decision problem is NP-hard under continuous arrivals. Experiments on a parametrized family of multi-domain topologies show that no fixed planning strategy dominates across all regimes: the performance ranking reverses between topology classes, and feasible placements exist for most rejected requests but remain undiscovered. These findings identify the plan layer as the decisive performance lever. We introduce ORION, a hierarchical orchestrator in which two LLM agents translate service intents into structured requests and generate abstract placement plans. A cooperative multi-agent RL coordinator selects the cross-domain partition, treating the plan as a persistent soft prior, while per-domain agents place the functions. ORION main contribution lies at the plan layer: an outcome-driven episodic memory that records, for each completed attempt, the network-condition profile, the proposed plan, and its verified feasibility outcome. The memory feeds verified successes and labelled failures back to the planner, thereby closing the adaptation loop using deployment feasibility rather than a scalar reward assigned to the planner's own proposals. To isolate the effect of the plan quality, we evaluate the planner on held-out topology families while keeping the downstream selection and execution mechanisms fixed. Under this controlled setting, the proposed memory improves admission by $4.6$--$11.5$ percentage points over the same planner without memory, while importance-based retention accounts for most of the improvement over a first-in-first-out store.
Yingzhao Jian, Zhongan Wang, Yuxuan Hou et al. · cs.AI
Multi-agent 91%Agents 36%
Abstract
Video director agents orchestrate multistage production to realize users' creative visions. Emerging research on their self-evolution advances screenplay planning effectively. However, the mutual reinforcement between directing and user interaction remains largely overlooked. Improved interaction clarifies user intent and informs workflow refinement, while more reliable execution shifts collaboration from troubleshooting to creative decisions. To this end, we introduce LIT (Learning to Interact and DirecT), a framework comprises a director agent, a simulated user, and a meta engineer. Simulated user assigns creative tasks to probe the director's weaknesses, while the meta engineer uses the task traces to revise the director's interaction and production strategy. Although this approach can address discovered interaction and production issues, it still relies on passive attribution by the meta agent. This limits its ability to adapt to broader user interaction patterns and production techniques. To solve this problem, we equip LIT with a co-evolution paradigm. The meta agent proactively learns external knowledge, thereby improving its ability to discover problems and propose solutions. The meta agent, simulated user, and director together form an evolutionary closed loop, collaboratively enhances the system's directing and interaction capabilities. Across the Test and OOD Test splits, LIT improves production and Interaction scores by 19.1% and 32.0%, respectively, relative to the initial pipeline. Controlled ablations reveal the correlation between improvements in production and interaction, where the co-evolution mechanism further encourage directing skill and interaction mode exploration.