Each one is read from its source and summarised: what it is, the problem it tackles, and what you could use it for.
Paper2026-11-04
Agents & tool use · Safety & alignment
FITTER is a multi-agent framework that integrates clinical records and wearable data with condition-specific guidelines to generate personalized FITT-VP exercise prescriptions.
ProblemThe difficulty of translating population-level exercise guidance into safe, individualized prescriptions that account for complex clinical and wearable data at scale.
Use it forGenerating personalized exercise plans for cancer survivors; Creating safe rehabilitation prescriptions for stroke patients; Automating the integration of wearable activity data with clinical diagnoses
multi-agent-systemhealthcareexercise-prescriptionllmsafety-audit
arxiv.org ↗
Paper2026-11-04
Agents & tool use
This paper analyzes average reward Multi-Agent Reinforcement Learning (MARL) in networked settings where agents use local k-hop information.
ProblemDecentralized multi-agent learning with local information often fails to converge to global optima due to multiagency and partial observability, lacking strong theoretical guarantees for practical deployment.
Use it forDesigning decentralized control policies for large-scale networked systems; Analyzing convergence guarantees for multi-agent reinforcement learning algorithms; Developing robust MARL methods that handle partial observability and multiagency challenges
multi-agent reinforcemenconvergence analysisdecentralized learningactor-criticentropy regularization
arxiv.org ↗
Paper2026-11-04
Agents & tool use
This paper proposes a method to detect unrecoverable failures in web navigation agents early by mining closed sequential patterns from execution traces.
ProblemWeb navigation agents often fail late after spending excessive tokens on unrecoverable trajectories, and existing early termination methods lack interpretability for diagnosis and human oversight.
Use it forReducing token costs for LLM-based web agents by terminating unrecoverable trajectories early; Diagnosing specific failure modes in web navigation agents using a structured taxonomy; Providing human oversight and interpretability for autonomous agent execution
web-agentsearly-terminationsequential-pattern-minininterpretabilitycost-optimization
arxiv.org ↗
Paper2026-11-04
Agents & tool use
ATLAS is a training-free intervention that stabilizes long-horizon multi-agent systems by steering hidden states toward shared task, role, and state directions.
ProblemLong-horizon multi-agent systems suffer from coupled failures of task drift, role drift, and state staleness, which existing surface-level or independent control methods fail to resolve simultaneously.
Use it forStabilizing role consistency in long-running multi-agent conversations; Improving memory retention and state maintenance in agent swarms; Reducing task drift in open-weight LLM-based agent frameworks
multi-agent systemsLLM internalsactivation steeringlong-horizon tasksrepresentation learning
arxiv.org ↗
Paper2026-11-04
Agents & tool use
ATC is a control mechanism for multi-agent LLM systems that enforces role contracts by scoring and constraining messages before they enter the shared conversation.
ProblemAgents in multi-agent systems often drift from their assigned roles over time, performing functions they are not authorized for (e.g., a verifier starting to plan), which undermines the division of labor and system relia
Use it forPreventing role drift in multi-agent systems where agents have distinct responsibilities like planning, retrieval, and v; Enforcing strict separation of duties in black-box agent architectures where internal model states are inaccessible; Improving task performance and consistency in large, diverse agent teams under stress tests
multi-agent systemsllmrole enforcementagent safetytrajectory control
arxiv.org ↗
Paper2026-11-04
Agents & tool use
An arXiv paper studying whether long-horizon AI investment research must be regenerated every time prices move.
ProblemRunning full multi-agent AI research reports on every price update is expensive in tokens and compute, but it is unclear whether that freshness actually improves portfolio returns or whether cheaper reuse of existing tar
Use it forDeciding how to allocate a limited research/compute budget between cheap price-driven target updates and expensive full ; Designing refresh policies for AI-generated equity research pipelines (when to reprice, guard, reset, or rebuild); Benchmarking portfolio construction choices (Top 10 vs Top 30, holding-window length) against SPY after trading costs
financellm-agentsinvestment-researchportfolio-optimizationcompute-budget
arxiv.org ↗
Paper2026-11-04
Agents & tool use
This is an arXiv paper proposing the Distributed Marginal Return (DMR) algorithm, a decentralized method for the Weapon-Target Assignment (WTA) problem in distributed multi-agent systems.
ProblemExisting distributed methods such as auction-consensus assume task-value independence, which ignores the coupled effects of assigning multiple weapons and therefore misprices bids in realistic WTA settings.
Use it forDistributed weapon-target assignment across cooperating autonomous platforms under limited communication; Command-and-control task allocation where multiple assignments interact and cannot be treated as independent; Generalizing auction-consensus task allocation to settings with coupled task values
multi-agent systemsweapon-target assignmenttask allocationdistributed optimizationconsensus protocol
arxiv.org ↗
Paper2026-11-04
Agents & tool use · Science & health
An arXiv paper proposing 'goal-driven self-recursive optical research,' in which agents use physical measurement evidence to revise the hardware, software, models, and protocols used to pursue a scientific goal.
ProblemOpen-loop, model-driven optical experiments: model-generated colors can drift or disappear from the display during spectral acquisition, so measurements do not reflect the intended stimulus and the model cannot be correc
Use it forClosed-loop calibration of see-through AR/VR displays against spectroradiometer measurements; Automated color-matching experiments where an agent iteratively revises its model from spectral feedback; Hardware-software co-adaptation research where measurement evidence drives changes to device control code and protocols
opticsautonomous-experimentaticlosed-loopagenthardware-software
arxiv.org ↗
Paper2026-11-04
Agents & tool use · Multimodal
AttriSwarm is a proposed agent-swarm framework for cross-page visual attribution in multimodal document QA.
ProblemAnswers synthesized from many document pages obscure which pages and which visual regions actually support them, and existing joint or single-trajectory methods suffer cross-page context pollution or high sequential over
Use it forAuditing multi-page document QA answers by tracing them back to supporting pages and evidence regions; Building verifiable retrieval/attribution pipelines over multi-page PDFs or reports; Training or evaluating page-level and region-level attribution models on cross-page QA data
multimodal-llmvisual-attributionmulti-agentdocument-qacross-page-reasoning
arxiv.org ↗