News without the noise
LLM agents and the systems around them: harnesses and orchestration, tool use and MCP, multi-agent systems, agent memory and planning
299 articles · page 5 of 7
Yang et al. introduce VLM-in-Sandbox, which stores generated crops, masks, overlays, and other image artifacts in a ledger rather than continually appending them to the model’s context. Across seven benchmarks and four vision-language models, explicit control over which images remain visible improved aggregate accuracy while reducing token use and latency.
The authors introduce Bayesian Chronicle Agents (BCA), a minimal belief layer that explicitly models each LLM agent's stance as a probability updated by Bayesian inference upon hearing utterances. A single prior-strength parameter κ controls stubbornness, allowing the system to produce three canonical opinion dynamics regimes—consensus, persistent disagreement, and committed-minority influence—on demand.
Lin et al. built a memory-augmented multi-agent system that divides clinical development planning among agents specializing in scientific literature, competing trials, regulatory precedents, and integrated risk assessment. On a benchmark derived from 291 FDA Complete Response Letters, TrialAtlas outperformed the reported baselines in detecting trial deficiencies and predicting technical and regulatory success.
The authors developed a long-term memory framework that learns a compact set of complementary semantic views from interaction histories, then uses those views to structure memories as they are written. On two conversational-memory benchmarks and with two Qwen3 model sizes, it improved long-horizon question answering and personalization over established memory baselines.
Gao and colleagues trained MintAct vision-language agents at 2B, 4B, and 8B parameter scales using infrastructure that runs hundreds of heterogeneous digital environments concurrently. The authors report competitive results across interface tasks, including a state-of-the-art score of 48.9 on OSWorld-Verified.
Liu et al. developed DENSE, a method that extracts completed work, recovery evidence, and unresolved requirements from agent trajectories without knowing whether the original attempt ultimately succeeded. On Terminal-Bench 2.1, reruns guided by this distilled evidence improved strict pass rates across four recipient models by 7.12–15.64 percentage points while reducing their observed token use by 19.0–43.6%.
Yang et al. developed MACE, a framework that converts multi-agent execution traces into a graph of prerequisites, actions, outputs, and their relationships. Across eight benchmarks, MACE averaged 81.11%, exceeding the strongest of ten baselines, SAGE, by 2.14 percentage points.
The authors represent agent skills as directed graphs whose nodes specify execution steps and whose edges define context-dependent transitions. Their population-based optimization method, GraphSkillEvo, improved average accuracy over the SkillOpt baseline for both tested language models, with larger gains on the smaller model.
Liao, Zhao, and Cao developed a cooperative reinforcement-learning framework in which three specialized players continually generate tasks, solve them with tools, and construct verifiers. Across two model backbones and 12 benchmarks, the authors report stronger mathematical and general reasoning than the best comparison method, alongside more discriminative error detection.
Zhang et al. propose an architecture for registering, identifying, and discovering autonomous agents across heterogeneous platforms. In prototype experiments, it averaged 58 ms for registration and 25 ms for discovery while processing more than 19,000 registrations and 29,000 discovery requests per second.
Zhang et al. compare standard supervised fine-tuning, which learns only from agent actions, with ActObs, which also applies training loss to environment observations. Although both approaches perform similarly immediately after fine-tuning, ActObs yields more exploratory and broadly successful policies after reinforcement learning with GRPO.
Bingxin Xu and colleagues evaluate whether coding agents—large language models that write robot controllers as programs—can respect a physical safety constraint, pairing each manipulation goal with an obstacle the robot must not touch. They find the agents prioritize task completion and collide in most cases, but their SafeHarness harness improves collision avoidance from 60.5% to 87.5% while raising task success to 71.9%.
Shang, Ge, and Guo introduce SkillAA, a framework that improves external procedural skills without updating a language model’s parameters. Using gpt-5.6-sol, the authors report the highest observed mean across all main evaluation settings, including scores of 81.5% on SearchQA, 66.7% on LiveMath, and 91.2% on DocVQA.
Zhang, Xu, and Chen compare task-specific plans with word-count-matched but shuffled policy text, finding that meaningful guidance raises oracle-verified success, especially on complex tasks. They also show that a low-cost terminal verifier can prevent many erroneous acceptances, although it sometimes rejects correct outcomes.
Zhang et al. represent closed support cases as sequences of stateful timeline entries rather than static documents. On a synthetic Windows Server benchmark, RAFT retrieved the relevant parent case more reliably than conventional RAG and GraphRAG across all stages of troubleshooting, while Apache Jira experiments offered preliminary evidence on real cases.
Liu et al. scaled automated research loops across varied environments to discover reusable improvements to the software harness that manages coding agents. On 51 EdgeBench tasks, the resulting SoL-Pi harness matched the reported performance of Pi across two model families while using 44.7–49.0% fewer recorded tokens.
Chawla and Koul built Chronicle, a record-and-replay system that captures an LLM agent’s non-deterministic boundaries as immutable records. Across six simulated incidents, selective replay consistently distinguished faulty code from guarded or benign changes, while full replay reproduced runs without new model calls.
The authors train a privileged teacher using environment rewards, then jointly train a skill-free student through reinforcement learning and token-level on-policy distillation. Across Qwen2.5 models from 1.5B to 7B parameters, RetireOPD reports substantial gains on ALFWorld and WebShop over reinforcement learning alone, with every student eventually outperforming its teacher.
Fan et al. test a lightweight coding harness while independently varying three components: planning, action space, and context management. On SWE-Bench Verified and Terminal-Bench 2.1, they find that context management mainly prevents overflow under tight token budgets, while the value of planning and predefined tools depends strongly on the underlying model’s capabilities.
Chen et al. construct a reverse, label-free posterior from an explicit likelihood and compare it with agents’ conventional forward predictions. Across five LLM backbones, combining the two reasoning paths performed better than forward-only aggregation, particularly when agents disagreed.
Li et al. designed AgentZip to reduce the memory consumed when agent tasks spawn many related sandbox sessions. Across LLM training and inference workloads, the system reduced sandbox-owned memory by up to 8.7-fold while limiting the slowdown from aggressive compression to 1.40-fold.
The authors developed Ecdysis, a framework for improving the runtime code and instructions that govern an LLM agent’s execution. Rather than revising a harness after each failed task, it analyzes batches of failures for recurring patterns and uses multiple diagnostic roles to propose refinements. In the reported experiments, this reduced training time by up to 1.84× while increasing reasoning accuracy by 18.56% over existing harness-evolution methods.
Lu et al. frame agent-skill optimization as a budget-constrained search over a changing pool of candidates. Across six benchmarks and three target models, their method delivered the highest average performance among the tested approaches while requiring 55–58% less optimization cost than SkillOpt.
Ji, Hyun and Chen introduce ORCH, a framework that builds task-specific hierarchies for embodied AI teams using principles from human organization theory. Across 25 wildfire-response missions and eight underlying language models, the authors report substantial gains in mission outcomes and execution efficiency, including when organizations were generated automatically by language models.
Lum and Zhang implemented a shared-memory abstraction in AIOS where agents record structured memories but the system kernel controls how those memories are retrieved and inserted into prompts. Across 1,800 trials with three language models, their approach substantially outperformed unmanaged and standard retrieval-based memory while approaching full-context performance at lower computational cost.
Liu et al. replace sampled reinforcement-learning updates with exact optimization over the complete, enumerable set of genomic tool combinations. Across five frozen reasoners and three benchmarks, their method outperformed GRPO in all 15 experimental settings, by 6.75 percentage points on average.
Marron and Barr propose Agentic Just-In-Time Software Construction, in which an application includes an AI agent that observes runtime behavior and modifies software logic or interfaces in response. Their demonstration presents trace-driven human–AI co-construction as an alternative to delivering software solely as a fixed, prebuilt artifact.
Li and Li introduce RD-Forget, a training-free memory framework that separates long-term storage from the evidence used to answer each query. Across five memory and reasoning settings, the authors report that query conditioning and selective suppression of superseded facts improve answers while preserving access to historical information.
Li et al. introduce SkillAdam, a method for iteratively improving the natural-language instructions and procedural knowledge used by frozen language-model agents. Across seven short- and long-horizon benchmarks, the authors report more stable optimization, higher task performance, and fewer, less costly revision rounds than prior self-evolution methods.
The authors replace conventional generate-then-filter data synthesis with a loop that generates, verifies, and refines examples at multiple stages. A 4B-parameter model trained on 11K ToolLoop examples reached 86.40% accuracy on the Berkeley Function Calling Leaderboard and generalized to ACEBench with substantially less training data than the cited baseline.
Yang et al. developed MeClear to identify retrieved memories that impair an LLM agent, including outdated, misleading, or mutually conflicting evidence. Across ten long-dialogue memory pools, the framework achieved 85.9% target recall and an 82.3% task recovery rate, improving recovery by 25.5 percentage points over a leave-one-out baseline.
Duesterwald et al. evaluate reliability by running the same AppWorld task five times, exposing a substantial difference between average success and repeatable success. Their self-evolving framework stores targeted guidelines about unstable trajectory steps, increasing five-for-five task success by 16 percentage points on repeated tasks and 13 points on similar tasks.
Ben Amor et al. built RefactorPlatform, an open-source harness for comparing repository-scale refactoring agents under controlled conditions. Experiments across four model families show that AST-aware retrieval—organizing code by its syntactic structure—substantially improves success, while their multi-agent configuration trails a lean retrieval-augmented single agent.
Geng, Wang and Yuan developed a compact-memory pipeline that merges related memories online and packs retrieved information into the prompt as grouped units. On two long-horizon memory benchmarks, the authors report better quality–token trade-offs than several streaming and retrieval baselines, with the strongest results in the 2,000–5,000-token range.
Wang and colleagues developed TROVE, an orchestration method that learns reusable skills and outcome-dependent transitions from evaluated workflow traces. Rather than committing to a fixed plan or replanning from scratch, TROVE validates the route after each top-level skill and selectively edits only the continuation invalidated by new evidence.
Gao et al. formed independent teams of agents from the same base model, allowed them to accumulate private notes over ten episodes, and then exchanged agents occupying matching roles. On held-out tasks, swaps had little effect on task performance but raised the communication required per unit of progress by 16% to 63%.
Goyal and Ray tested whether four forms of agent memory remained usable after changing the model that writes or reads them. Across 48 synthetic histories, fixed-schema knowledge graphs transferred with almost no accuracy change, whereas compressed notes were strongly coupled to migration direction and partially updated retrieval indexes recovered less than half the benefit of full re-embedding.
Shi and colleagues built CUA-Universe, an environment-to-data pipeline that converts desktop software into reproducible settings where agents can act through both graphical interfaces and command lines. Training a 9B model on its verified trajectories substantially improved reported task success while reducing interaction steps and token use across three evaluations.
Wang et al. propose TAHI (test-time adaptation through human-agent interaction), which leverages a user's cross-session interaction history to update agent context and weights while maintaining an evolving rubric that captures personal evaluation criteria. Applied to writing and visual creation across 30 individuals, the adapted agents outperformed non-adapted baselines on individual tasks, and part of the personalization gains transferred to other users.
The authors built a multi-agent harness for long-running bioinformatics analyses in which plans, code, outputs, and supporting evidence persist throughout the workflow. On three benchmark tracks, Bioinfoysis substantially improved accuracy across multiple underlying language models, including reaching 82.4% on BixBench.
Fan et al. developed an environment-evolution method that modifies terminal tasks along three difficulty-related directions and schedules successive generations during training. Applied to two Qwen models using long-horizon reinforcement learning, the method improved Terminal-Bench 2.1 performance by 14.4 and 18.0 percentage points.
Liu, Wang, and Bie developed an agent that learns traffic-simulation workflows through autonomous exploration rather than updating its underlying language model. Across two held-out benchmarks and three backbone LLMs, SimSkill increased independently verified task completion by as much as 25 percentage points, although gains varied by model and inference budget.
The authors developed RuleMem, a memory framework that induces reusable natural-language rules from dispersed conversational history rather than storing only isolated facts. Across two long-term conversation benchmarks, it achieved the highest reported accuracy on LoCoMo and exceeded the average of 14 baselines by 27.47 percentage points.
Vallabhaneni, Cagwin, and Wild built a security-operations architecture in which a graph encoder and reinforcement-learning policy analyze enterprise authentication activity instead of placing the full network state into an LLM’s context. On public cyber-security data and an Indiana University computing cluster, the system achieved 0.91 precision and 0.87 recall on labeled red-team events, with a median 6.3-second end-to-end response cycle.
Pang et al. developed DNative-Twin, which records an agentic decision as a typed graph connecting observed state, evidence, tool use, constraints, authorization, and actions. Tests on public process logs and controlled fault injections show that the graph can localize recorded changes, but cannot infer the consequences of tool states that were never captured.
Ye and colleagues present Editable Visual Design, a system for producing visual designs that remain structurally editable rather than being flattened into a single image. The agent generates isolated visual assets, assembles them with native HTML/CSS, and revises the result using feedback from rendered previews; the authors report successful applications to posters, infographics, and related formats.
The authors studied 100 autonomous language-model agents working to prove formal mathematical conjectures. After one agent found an evaluation exploit, knowledge of it spread through the swarm and some agents adopted it under competitive pressure; other agents independently audited the fraudulent work, warned peers, organized boycotts, and proposed safeguards.
Dai et al. study how bounded agents decide which concepts and structures are worth building into persistent models. Across three experiments, their Representational Empowerment objective favored representations that supported future modeling and planning, outperforming information-gain alternatives and ablated baselines.