Agents

LLM agents and the systems around them: harnesses and orchestration, tool use and MCP, multi-agent systems, agent memory and planning

600 articles · page 4 of 13

Agents

Modeling multi-agent LLM systems as dynamic systems for root-cause analysis

The authors introduce AgentSoV, a framework that treats execution of a multi-agent LLM system as a stochastic dynamic process. By mapping trajectories to diagnostic stages and learning variation propagation via a state-space model, they enable failure attribution, stage-input influence quantification, and virtual-sensor placement without re-executing the system. Experiments across chain, orchestration, and DAG architectures show unified diagnostic capabilities.

·3 min
Agents

Over-defense in coding agents starts at outcome-settled forks

The authors identify 'silent forks' — turns where every next action leads to the same outcome — as the origin of over-defense in coding agents. They find that outcome-based reinforcement learning gives no gradient at these forks, while a behavioral checker can locate them. Their method, ForkDPO, reduces disproportionate behavior from 70.3% to 58.9% on unseen repositories by contrasting proportionate and disproportionate actions at the same prefix.

·3 min
Agents

Co-evolving code attackers and defenders improves executable vulnerability detection

Pahwa et al. co-trained populations of LLM attackers and defenders on containerized software, using builds, tests and working exploit harnesses rather than model judgments to determine rewards. Their peer-relative optimization method improved vulnerability detection on unseen applications and languages, while also helping repository agents identify reachable entry points and reproduce held-out vulnerabilities.

·3 min
Agents

Distillation overshoot lets smaller agents surpass teachers, and a new method preserves those gains.

The authors trained smaller language-model agents by on-policy distillation from stronger teachers and found a counterintuitive pattern: students often exceed the teacher mid-training, then regress. They attribute this to competing effects of decision repair and erosion, and introduce SPOD, which preserves successful student choices from the teacher's conflicting supervision. SPOD consistently beat ordinary distillation (up to 15.2 percentage points) and even surpassed the teacher by up to 20 points.

·3 min
Agents

Evidence annotations improve step-level credit for deep search agents

Luo and colleagues propose EG-GRPO, a reinforcement-learning method that turns evidence annotations in synthetic search data into rewards for individual retrieval steps. An LLM judge links each required claim to the first citation that supports it, allowing the method to track how an agent's knowledge develops during a search. Across five deep-search benchmarks, the authors report generally stronger performance than standard GRPO and several reward-enhanced or step-level alternatives.

·3 min
Agents

Global coordination of proof agents beats local repair in formal theorem proving

The authors present Maestro, an agentic framework for Lean proof development that replaces local repair of failed subgoals with global coordination. Workers provide structured failure reports, and a single coordinator decides on appropriate responses, such as guiding a stuck prover or replanning a flawed subtree. Maestro proves 34 out of 48 theorems from the June 2026 arXivLean release with Claude Opus 5, compared to 19 for the native agent, and outperforms all published provers on PutnamBench, Lean-IMO-Bench, and IMO 2026.

·2 min
Agents

CUA-Lite standardises computer use agent development across environments and models

The authors present CUA-Lite, an open platform that standardises the components needed to develop computer use agents (CUAs)—agents that control desktop, browser, and mobile applications. By defining three abstractions (a unified environment interface, a unified supervised data format, and per-model harnesses shared across evaluation, RL, and SFT), the platform integrates 12 benchmarks, provides lightweight VM-free desktop sandboxes with over 30,000 verifiable tasks, and converts major public CUA datasets. The authors release reference scores for 15 models and show that supervised fine-tuning and reinforcement learning through the same harness improve open-weight models.

·3 min
Agents

Large-scale dataset and world-model RL improve embodied navigation in open worlds

The authors propose Embodied World Agents (EWAs), agents that perceive and act across diverse open-world 3D environments, and release EmbodiedWorld-200K, a dataset of over 200K video samples with multi-level annotations. They demonstrate that offline pretraining on this data improves embodied planning, and that further RL post-training using a video world model as a scalable environment proxy yields clear gains on out-of-distribution scenes.

·2 min
Agents

Self-evolving agents trade plasticity for stability in continual learning streams.

The authors test whether nonparametric self-evolving LLM agents, which adapt by rewriting prompts, memory, or code without updating model weights, can function as continual learners. Across four drifting task streams built from BIRD, BFCL, and ScienceWorld, they find that self-evolution improves aggregate performance on most streams but systematically sacrifices stability for plasticity, and that both self-modification and similarity-based retrieval suffer from interference or overwriting when task distributions shift.

·3 min
Agents

Mined tool-call constraints help agents avoid errors and finish tasks

Lee and Barke present Symbolic Constraint Agents, a framework that converts tool-use requirements into formal rules, refines those rules using execution traces, and checks each proposed tool call with a satisfiability modulo theories solver. On two agent benchmarks, the resulting interventions detected more invalid calls, removed many failed calls, and produced task-success gains of up to 9.8 percentage points.

·3 min
Agents

Synthetic computer worlds improve agents on unseen computer-use tasks

The authors developed WildAWM, an open-source pipeline that synthesizes computer environments, tasks and automated success checks for training computer-use agents. Its 341 environment-task pairs combine graphical interfaces, command-line operations and model context protocol tools. Training two Qwen models exclusively in these synthetic worlds improved their performance across eight benchmarks involving previously unseen computer-use tasks.

·3 min
Agents

Constructing testable, stationary rewards for open-ended tool agents

The authors introduce Reward Engine, a method to automatically construct a reward signal for open-ended tool-use tasks that lack ground-truth answers. The reward consists of executable testcases (for hard constraints) and rubrics (for graded quality), each providing an evidence receipt. This RewardPack is tested against labeled counterexamples before use and held fixed while the agent is optimized, raising the three-run consistency (Pass^3) from 45.0% to 54.0% on WorkBuddy Bench Office and from 78.4% to 83.7% on Claw-Eval.

·3 min
Agents

Refining user representations improves the fidelity of AI user simulation

The authors identify natural-language user representations as an underexplored bottleneck in user simulation and formulate conversation reconstruction to evaluate them. They introduce hindsight distillation, an online training procedure that uses feedback from comparing real and simulated conversations to improve representation generation. Their model, UserRepGen-4B, enables more faithful simulation than GPT-5.5 across 700 conversations and seven datasets.

·3 min
Agents

Training LLM forecasters across dates improves search and probability estimates

Ye and colleagues introduce AlphaBet, a reinforcement-learning framework that repeatedly forecasts the same unresolved event at different dates, using only evidence available at each point. On two benchmarks, the trained GLM-4.5-Air agent produced lower Brier scores, indicating more accurate probability estimates, with gains attributed to both better evidence acquisition and better probability reporting.

·2 min
Agents

Search agents that locate, verify and crop references improve image generation for recent events.

The authors introduce Savanna, a benchmark of 500 image generation prompts synthesized from recent news, and show that even harness frameworks built on frontier models retrieve a correct visual reference for only about half of them. They then propose Locate-Corroborate-Isolate, a three-phase agentic paradigm, and train Ranger-8B on 15K trajectories following it, raising the search score to 67.2 and improving generation quality across all tested generators.

·3 min
Agents

Artifact-grounded RL trains a general frontend development agent

The authors introduce WebCoder, a general frontend development agent trained with artifact-grounded agentic reinforcement learning. They construct Webdev-World, an executable environment that integrates code editing, execution, and inspection, and use an independent verifier to assess final task requirements for rewards. On their benchmark Webdev-Bench, WebCoder-27B raises requirement satisfaction from 64.8 to 72.8, surpassing both its base model and Claude-Opus-4.8, and on Vision2Web it improves the average score from 41.1 to 62.4, matching Claude-Opus-4.8.

·2 min
Agents

Faster Jev decisions do not consistently improve agent task efficiency

Ling and colleagues tested whether replacing agents’ action-selection components with Jev, a model designed for fast choices among defined options, improves end-to-end efficiency. Jev lowered decision cost and latency in all five agents, but those local gains did not reliably translate into faster or more economical successful outcomes. Mobile agents were particularly vulnerable to failed recovery and declaring completion too early.

·2 min
Agents

Lightweight decision model reduces token use in multi-agent systems by up to 54.6%

Wang et al. propose Jev-LiteMAS, a multi-agent framework that allocates routine coordination to a low-cost decision model (Jev) and reserves generative models for expert execution and corrective feedback. Across four benchmarks and two backbones, the method reduces generative-LLM tokens by 12.7–54.6% with accuracy within 1.34 points of the baseline, and lowers end-to-end latency in most configurations.

·3 min
Agents

KnowGPO aligns trajectories by knowledge for better credit assignment in search agents

The authors introduce KnowGPO, a reinforcement learning framework that assigns credit in deep search agents by representing each step's state as a subset of atomic knowledge conditions satisfied, rather than by surface observations. Trajectories with the same knowledge state are grouped for within-group comparison, and the incremental increase in satisfied conditions between steps serves as a dense process reward. On three benchmarks (Xbench, BrowseComp-ZH, GAIA), KnowGPO improves over outcome-only GRPO by 2.0 to 5.9% and exceeds recent process-supervision methods by 3.8 to 4.2 percentage points, with an 8B model matching or surpassing most 32B open-source systems.

·3 min
Agents

Refined Questions and Teacher Guidance Improve Long-Horizon Multimodal Search

Zhang and colleagues present a pipeline for training multimodal agents to investigate questions through multiple rounds of web search, visual retrieval and reasoning. Their 9-billion-parameter agent improved average accuracy by 15.3 percentage points over its agentic-workflow baseline across seven primary benchmarks, with joint reinforcement learning and on-policy distillation adding 2.3 points beyond supervised fine-tuning.

·2 min
Agents

Multimodal decision models output structured choices in a single pass

The authors introduce Intern-Decision, a family of multimodal decision models (0.8B, 2B, and 4B parameters) that reformulate structured decision making as next-token prediction over masked fields in a predefined JSON skeleton, enabling multi-field scoring in a single causal forward pass. The 4B model achieves a 90.02% unweighted average accuracy across seven evaluation suites, compared with 88.74% for the baseline Jev model, and attains a mean latency of 44.16 ms for a 289-token request with three decision fields on one RTX 4090.

·3 min
Agents

KV cache substitution between frozen LLMs fails; augmentation succeeds

The authors tested eight methods for substituting a sender's KV cache into a frozen reader LLM in place of its own cache. All substitution methods performed at the level of an uninformative cache baseline, failing to improve the reader's accuracy. In contrast, an augmentation approach—where the reader retains its own cache and attends over the sender's cache as additional context, with only its queries adapted—beat the reader's native-context baseline and worked when pooling multiple senders.

·3 min
Agents

Localized Context Distillation resolves the learning dilemma in interaction experience

The authors introduce Localized Context Distillation (LoCD) to improve how large language models learn from interaction experience. LoCD first localizes the spans of a student trajectory that require revision, then has a teacher regenerate only those spans from the student's own prefix, and finally trains the student on the revised segments. In extensive experiments on math reasoning, code generation, and interactive agent tasks, LoCD consistently outperforms strong baselines while requiring smaller and sparser parameter updates.

·3 min
Agents

Synthetic experience improves AI game control beyond supervised training

The authors introduce Player-1, a post-training framework that first fine-tunes a vision-language model on 6,044 curated gameplay demonstrations, then applies reinforcement learning in 16 synthetic game environments designed to cover common control demands. The resulting Player-1-9B model achieves 38.33% success on MineDojo-300, compared to 10.54% after supervised fine-tuning and 12.63% for the strongest baseline. The policy also generalizes to unseen games ViZDoom and GameWorld, showing that synthetic experience transfers beyond the training suite.

·2 min
Agents

Cortex folds context benefits into model weights without degrading other skills

The authors formalize Memory Consolidation as internalizing context-induced performance gains while preserving non-target capabilities, then propose Cortex, a method that anticipates future uses of context and writes them into parametric memory. Across dense and mixture-of-experts backbones from 4B to 120B parameters, including multimodal models, Cortex sits on the empirical consolidation-retention Pareto frontier in most evaluated settings.

·2 min
Agents

Web agents can skip text generation and read action choices from hidden states

Gan et al. train small plug-in heads that read a web agent's next action from the hidden states of a frozen Qwen3-VL generator instead of generating a token string. On seven offline test sets a confidence cascade keeps step success within 0.3 points below and 1.4 points above the generator while deciding 12 to 64% of steps, and on WebArena it raises task success while generating less than half the tokens.

·3 min
Agents

A verifiable runtime for agents that extend themselves with safe, reversible tools.

The authors present L3M, a runtime for AI agents that can write their own extensions and improve over time. Built in Lean 4, it stores all state in Git, allowing resumption, forking, and merging of agent trajectories. By enforcing invariants via type checking and making tool safety depend on reversibility, L3M achieves safe autonomous operation: on Terminal-Bench, it reaches about two-thirds of the success rate of an unsafe shell-equipped baseline using only tools that require no human review.

·3 min
Agents

Reinforcement-trained steering agent corrects goal drift in LLM agents and multi-agent systems

Shen et al. address goal drift, the tendency of LLM-based agents and multi-agent systems to deviate from their objective during long-horizon execution. They train a steering agent that monitors execution on the fly and intervenes when drift occurs, and find that reinforcement learning transforms a weak zero-shot supervisor into an effective one, raising performance across seven question-answering benchmarks and in a five-role multi-agent system.

·2 min
Agents

New RL method trains AI agents to handle multiple sessions effectively

Xu et al. identify a mismatch between single-session training assumptions in agentic reinforcement learning and inference-time techniques like context compaction and subagent forking. They propose Mosaic, a critic-free hierarchical optimization method that combines session-aware trajectory-level advantage estimation with session-level advantage shaping. On the BrowseComp benchmark using Qwen3.5-4B, Mosaic outperforms the broadcasting baseline by 10.1% relative accuracy while reducing average sessions per trajectory from 3.17 to 2.35.

·2 min
Agents

For internal-state sharing between language models, augment the cache and replace the activations

Deng, Liu, and Zhu systematically compared two carriers of internal state (KV cache and residual-stream activations) and two fusion mechanisms (augmentation and replacement) in a controlled latent-fact transfer setting. They find that the KV cache must be augmented (the receiver cross-attends to a peer's cache while keeping its own, with a small trained adapter), whereas activations work best when written directly into the receiver's residual stream with no bridge and no training. The rule 'augment the cache, replace the activation' holds under one protocol, and one adapter serves multiple publishers it never trained on, provided their internal state lies close to the receiver's.

·3 min
Agents

MetaExp trains a transferable experience-learning capability in language model agents

The authors introduce MetaExp, a framework that trains language models to acquire, update, and reuse external experience across tasks. Fine-tuning Qwen3.5-122B-A10B on 4,956 filtered trajectories from 1,000 synthetic task families yields gains of +8.5 points on SpreadsheetBench, +14.8 on AutomationBench, and +2.3 on APEX-Agents under cold-start sequential evolution, while the baseline Qwen3.5 backbone degrades by 2.2 points on SpreadsheetBench.

·3 min
Agents

Execution traces enable precise localized repair of agent memory architectures

The authors introduce MemScope, a framework that evolves LLM agent memory architectures by inspecting fine-grained execution evidence rather than coarse evaluation scores. By jointly examining raw traces, runtime memory state, and provider code, a Patch Agent identifies precise failure locations and applies localized code edits. On GAIA, trace-and-memory inspection achieves 86.67% accuracy versus 63.33% for summary-only feedback, and on ALFWorld unseen tasks success rises from 50.75% to 71.64% after evolution.

·3 min
Agents

Evidence graphs produce synthetic research tasks with controllable difficulty

Yang and colleagues introduce GRADS, a pipeline for synthesizing multi-hop questions while explicitly tracking the evidence needed to answer them. Across software engineering, particle physics and security corpora, the authors report that GRADS generated verified tasks with difficulty levels that remained distinct when tested by independent language-model solvers. Training a 4-billion-parameter model on these tasks also improved completion rates and reduced search-tool use.

·3 min