ZOTPAPER

Hierarchical supervision allocation improves long-horizon model distillation

The authors propose LENS-OPD, a framework that distills a large teacher model into a smaller student agent for long-horizon tasks by hierarchically allocating supervision: first deciding how long the student should explore, then which decision to intervene on, and finally which tokens to correct. On agent benchmarks including WebShop and ScienceWorld, LENS-OPD consistently improves task success rates over vanilla on-policy distillation and strong baselines.

30 Sept·3 min

On-device caches convert natural language to actions with classification not generation

Fereidouni et al. introduce on-device semantic operation caches that transform the natural language to action problem from a generative task into a classification-centric formulation. Tested on generating Excel formulas from natural language requests, the system reduces total inference cost by 56 percent compared to cloud-only model routing and improves response latency by five times on cache hits.

Skill-guided environment synthesis improves post-training for tool-using agents

The authors introduce Skill2Env, a pipeline that jointly generates task instructions, tools, workspaces and evaluators from source skills. Supervised fine-tuning on 1,500 high-scoring trajectories from these environments produced consistent gains across multiple agent benchmarks.

Composed service environments improve agents across diverse tool-use benchmarks

The authors built 448 executable services with 10,130 tools and composed them into tasks where information and actions must flow between applications. Training Qwen3.6-35B-A3B on 3,000 supervised trajectories and 1,000 reinforcement-learning tasks improved its average score across eight benchmarks by 9.17 points.

Accurate World Predictions Are Not Required for Agent Training Gains

Che et al. audited whether world-model post-training improves agents because they learn to predict what happens after an action, or because training changes their behavior in other ways. Across interactive text environments and VisualWebArena, agents retained substantial gains with corrupted prediction targets, while random-reward training increased the range of tasks solved across repeated attempts.

Self-designed evaluators enable safer retries for long-horizon agents

Asgari et al. had a language model build its own evaluation suite from public documentation, then used those evaluators to control retries and maintain an outcome-tracked memory across task streams. Across five-repeat experiments on tau2-bench and AppWorld, SelfSuite outperformed a plain agent without labels, although its advantage depended more consistently on gated retries than on its memory design.

Selective experience activation improves personalized GUI agents across task steps

The authors show that task-level history retrieval mainly helps a GUI agent’s first decision, with little sustained benefit across the remaining 90% of an episode. Their training-free ExpActivator framework selects situation-specific past actions and recurring intents, improving both personalized task execution and proactive suggestions across four GUI backbones.

Language models discover search algorithms that transfer across games

Li and colleagues used large language models as code-generation engines to evolve modular game-playing algorithms under fixed compute budgets. Across more than 400 environments, the resulting search mechanisms earned top-tier aggregate rankings and defeated most of 15 Monte Carlo tree search baselines in pairwise comparisons, including on held-out and procedurally generated games.

Interface support helps developers supervise more coding agents in parallel

The authors studied how developers coordinate multiple autonomous coding sessions, then distilled five supervisory practices into the PILOT framework: Planning, Isolating, Logging, Observing, and Triaging. In a 16-person experiment, a prototype implementing these practices increased ticket throughput by 63 percent and let participants supervise one additional agent at peak, although it did not significantly improve perceived control.

Active visual probes help browser agents act more accurately

Li et al. introduce Probe to Act, a browser-agent framework that interleaves reasoning with targeted visual checks of webpage elements. Across three benchmarks, the approach improved proprietary and open-weight models while using substantially less retained context than full-observation histories.

Event graph corrections improve LLM agent stock forecasts by modeling historical reliability

The authors present RICE-Alpha, a point-in-time stock-scoring framework that combines a history-aware multi-view Base Alpha with a reliability-weighted residual correction from event graph transitions. Tested on daily Nasdaq-100 and Hang Seng Index panels from 2024 to 2026, it achieves the strongest results among evaluated LLM-based agents and momentum baselines, with an ICIR more than doubling that of the strongest baseline and net Sharpe ratios of 1.656 (US) and 1.725 (Hong Kong).

Dual-system voice agents balance live conversation with slower tool use

The authors pair an always-active, full-duplex speech model with a slower backend agent that can reason and use tools. Across spoken question answering, multi-turn dialogue and simulated business tasks, adaptive delegation improved difficult-task performance while training reduced unnecessary backend calls.

Speaker-indexed memory cuts token costs while preserving answer accuracy

The authors built a conversational memory system that stores episodes alongside derived relations, preferences, summaries, and graph state, with explicit links back to the source evidence. Across four long-term memory benchmarks, Stashbird substantially reduced prompt-token use while matching or exceeding a reproduced Hindsight baseline on three benchmarks.

Teacher-guided training improves specialization and coordination in multi-agent models

Zhong et al. introduce an on-policy distillation method for jointly training multi-agent systems without designing separate rewards for every role. Across code generation and mathematical reasoning benchmarks, MAS-OPD achieved the highest reported mean score at both tested student-model scales, alongside clearer division of labor and more effective coordination.

Hypergraphs help language agents reuse feedback across planning paths

The authors augment Monte Carlo Tree Search with a task-specific hypergraph that connects equivalent or related decisions appearing under different trajectory prefixes. Across three language models on DeepPlanning, the method improved average planning accuracy by 2.3 to 7.3 percentage points over the strongest baseline for each model while using fewer calls and output tokens than evaluated MCTS alternatives.

Terminal agents struggle when familiar environmental assumptions stop holding

Singh and colleagues introduce AGNI, a pipeline that identifies assumptions behind successful agent trajectories, invalidates them through controlled environmental changes, and verifies that the modified tasks remain solvable. Multiple LLM agents performed substantially worse on these novel variants, often observing evidence of a change without correctly diagnosing it. Post-training on environmental novelty improved performance on both held-out novel tasks and unchanged base tasks.

Draft-KV ensures the receiver actually uses the sender's latent message

The authors identify a fundamental flaw in prior latent communication methods: replacing the sent message with one from an unrelated question barely changes accuracy, meaning the receiver relies on the interface, not the message content. They propose Draft-KV, which transmits key-value states from the sharer's own answer draft via a gated attention branch with progressive training. A frozen 8B sharer lifts a frozen 0.5B receiver from 37.45% to 78.04% on MMLU-Redux, and message reassignment drops accuracy to 36.40%, confirming the gain stems from genuine latent information exchange.

30 Sept·3 min

Tool calls need formation-bound dispatch to preserve their security meaning

Kim and Kang identify schema-epoch drift, a failure in which a valid tool call acquires different security consequences because its implementation changes between formation and execution. They propose formation-consistent dispatch, which verifies successor implementations and preserves the original call's authority through rollout, retirement, and delayed approval.

Visual agents improve by jointly evolving reusable skills and practice data

Yan et al. introduce V-Gym, a framework that uses an agent's successful and failed visual reasoning attempts to improve both its reusable skills and its practice dataset. Across several multimodal reasoning benchmarks and backbone models, the authors report that this joint evolution outperforms baseline approaches and produces skills that transfer across domains and models.

Agent distillation improves when supervision follows outcomes, not model disagreement

Zhang et al. test a common assumption in on-policy distillation: that large token-level differences between teacher and student mark the most useful moments for supervision. They show that these local gaps often mispredict downstream benefit, then introduce OG-OPD, which weights teacher guidance using final outcomes from paired student continuations and improves success across three interactive benchmarks.

Aligned skills and rubrics improve multimodal agent training

Jiang and colleagues introduce SkillRubric, a framework that derives paired instructions and evaluation rubrics from reusable agent skills. A multimodal verifier uses screenshots and tool outputs to reward intermediate progress, while controlled revisions improve guidance and rubrics without changing both simultaneously.

Agents improve long-horizon performance by evolving how they manage context

Li et al. treat context management, rather than reusable skills alone, as an evolvable component of an agent harness. Starting from the open-source Pi-agent, their framework reconstructs what the model saw at important decisions, attributes failures to context problems, and updates the policy, improving results across three long-horizon benchmarks.

Runtime research preserves details that prebuilt agent memories can lose

Yan et al. introduce Just-In-Time Agent Memory, a framework that separates lossless history storage from request-time context construction. Across agent-memory and long-context benchmarks, the authors report stronger task performance than ahead-of-time memory systems and greater efficiency than earlier trained agentic-memory approaches.

Post-training helps an agent sustain complex work for over 10 hours

Zhang et al. trained an autonomous agent on software-engineering tasks derived from large GitHub release pull requests, including chained tasks designed to require extended execution. Across five long-horizon benchmarks, the authors report consistent gains over the base model, with some runs lasting more than 10 hours and involving over 1,000 tool calls.

Flexible action chunking improves long-horizon planning in JEPA world models

The authors introduce FlexiWorld, a JEPA-based world model that trains with arbitrary goal spans and variable-length action chunks, along with an actor-residual planning method (ARCEM). Across four benchmarks, FlexiWorld with ARCEM achieves 89.29% mean success versus 83.98% for the strongest baseline, and longer chunks speed up planning by 1.3x without retraining.

Self-modifying coding agents evolve learned optimizers from simple templates

Yu et al. present HADA, a framework that treats the design of learnable black-box optimizers as a coding task. Starting from a basic MetaBBO template, the system produced variants that the authors report outperform recent human-designed baselines and adapt across optimization domains.

Reinforcement learning teaches scientific ideation agents to explore broader possibilities

Kargupta et al. trained a scientific proposal agent with rewards targeting creative actions, exploratory processes and useful, novel outcomes. The resulting system broadened research directions by 27.8% and contribution types by 14.9% over the base model, while also improving model-evaluated originality and predicted citation impact.

Equivalent tool interfaces can make the same LLM agent fail

Liu et al. tested whether tool-using LLM agents behave consistently when the same actions are presented through different interface definitions. Evaluating eleven models on mechanically equivalent schema variants, they found substantial, model-specific sensitivity to details such as tool grouping, argument structure and multi-call dependencies.

Production-scale AutoResearch exposes five failure modes and a three-principle fix

The authors applied AutoResearch, an LLM that iteratively edits training scripts and keeps changes that improve a held-out metric, to two production embedding systems at Amazon. Over 220 experiments spanning 12 weeks, they identified five recurring failure modes and a three-principle multi-agent scaffolding design that maps each to a structural remedy. The framework improved Recall@6 by 1.82x, coherence by 2.1x, and expanded catalog coverage 5.8x.

30 Sept·3 min

A method that links failure localization to repair by identifying which editable system asset to change

The authors propose DAAF, a framework that moves beyond localizing where a failure occurs in an LLM agent to determining which persistent, editable system asset (e.g., a routing rule or knowledge segment) should be changed. By learning the effects of valid attribute replacements from controlled replays and amortizing this evidence at deployment time, DAAF achieves 80.72% attribute Hit@1 on held-out Telecom tasks while recovering 62.65% of failed executions and limiting clean-task regression to 3.23%.

29 Sept·3 min

Bayesian reliability memory helps agents decide when to trust advice

Feng, Huang and Poria introduce an online Bayesian memory that estimates advisor reliability using a central model’s internal belief representations and interaction history. Across nine benchmarks and six central models, the authors report greater resistance to misleading advice than debate and majority voting, with the same mechanism also improving worker routing on MuSiQue.

Structured agent harnesses improve LLM-driven discovery of symbolic equations

Yu and colleagues built a domain-specific runtime for agentic symbolic regression, where language models search for interpretable equations that explain observed data. On LLM-SRBench, the harness produced substantially higher symbolic recovery than competing systems using the same model backbone, including when variable names and scientific descriptions were removed.

Statepoints let AI agents reverse mistakes and branch safely

Guo et al. built an agent runtime that treats files, processes and remote service changes as a unified, restorable state. In their evaluation, this allowed agents to undo errors and explore alternatives in parallel, improving task quality by up to 15 times while adding 3% overhead for user recovery.

Routing Across Memory Types Makes LLM Agents More Accurate and Efficient

Wei et al. tested 13 agent-memory methods and found that no single representation performed consistently across three benchmarks. Their MemAgent system instead routes tasks among heterogeneous memory providers, improving average accuracy by 10.0% while reducing task steps by 12%.

Short agent memories fail mainly by dropping facts too early

Shen and Li study agents that repeatedly rewrite a fixed-size state instead of retaining their full interaction history. In TextWorld cooking games, they find that most failures arise not from the 128-token budget itself, but from writers discarding facts before their future relevance is known.

Memory benchmarks can hide how agents preserve and revise facts

Ding and Wu introduce MemProbe, a framework for evaluating the stability-plasticity balance between preserving existing memories and incorporating new evidence. In a 56-episode suite covering four cognitive paradigms, they test six incremental memory systems and show that similar aggregate scores can conceal markedly different memory-update behaviors.

Causal error graphs improve diagnosis of failures in agent traces

The authors introduce a tool-augmented framework that represents agent failures as typed causal graphs rather than isolated anomalous events. On their new agent-annotated benchmark, CEG-Agent outperformed prior methods under both semantically relaxed and structurally exact evaluation criteria.

GUI world models fail when hidden state is omitted from predictions

The authors show that GUI world models that condition only on visible observations systematically fail when identical interfaces hide different transition-relevant state. They introduce StateAliasBench to isolate this failure and a lightweight recovery method that infers hidden state from action history, improving downstream agent performance.

29 Sept·3 min

Independent logs make deployed agent records verifiable after failures

Dai et al. evaluated whether records produced by deployed agent harnesses can support later audits without requiring trust in the harness itself. Across 140 runs, examiners often identified the correct fault, but their evidence largely came from harness-authored material; an external append-only log detected all 28 omissions and fabrications introduced during execution.

Reinforcement learning teaches an agent to verify mixed-source multimodal misinformation

Li and colleagues assemble specialized detectors into a unified toolkit, then train a large vision-language model to select and use those tools across multi-step verification tasks. On MMFakeBench, the trained agent improves accuracy over its base model, while a cache of precomputed tool outputs reduces the cost of reinforcement-learning rollouts.

Reinforcement learning improves tool selection across large repositories

Dai et al. introduce ToolSearcher, a reinforcement learning framework for selecting tools when repositories are too large to place every tool description in the model's context. On large-scale benchmarks, the authors report consistent gains over strong baselines, particularly for iterative search and tasks requiring compatible combinations of tools.

Agents can hide high-capacity messages without a pre-shared key

Pang, Smith and Zheng present a steganographic protocol for independently deployed agents that do not share the same model or an initial secret key. Across three workloads and three sender models, Codetta delivered up to 94 times the capacity of the leading asymmetric protocol, while its covert key exchange succeeded using about 80,000 visible tokens.

Branching search improves how language agents evolve reusable skill libraries

Liu et al. formulate automatic skill evolution as a graph search, maintaining a main update path alongside branches created from earlier library versions. Across five benchmarks and two language models, SkillVine reported the best test performance in nine of ten benchmark and model combinations.

Preserving Retrieved Images Improves Long-Term Multimodal Agent Memory

Jiang et al. isolate memory delivery, the step that converts retrieved records into the text and images shown to a model. On MemLens, supplying original pixels from an unchanged evidence set improved accuracy by 13.87 percentage points on an 8B model, compared with 2.31 points from perfecting retrieval.

Adaptive tutoring helps rerankers assemble better evidence sets for research

Jiang et al. train a setwise reranker that optimizes the collection of evidence passed to retrieval-augmented generation and deep-research systems, rather than scoring each document independently for relevance. Across ten benchmarks, the authors report the best overall performance among evaluated methods while requiring fewer retrieval calls.

Spoofed Agent Identities Can Override Contradictory Safety Evidence

The authors introduce TrustFork, a 1,890-task benchmark that tests orchestrators when one subagent pursues a risky goal and three provide user-aligned alternatives. Orchestrators frequently retained and acted on the risky subagent's advice despite contradictory evidence, with identity labels and harness behavior shaping whose response prevailed.

Fact-intent graphs accelerate multi-agent work on difficult, long-running tasks

Xu and colleagues introduce CAIRN, a multi-agent architecture that records findings and proposed investigations in a persistent directed acyclic graph. Across cybersecurity and mathematical reasoning tasks, the approach provided little benefit on easier problems but solved 76.5% of high-effort cases faster, with speedups reaching 3.08 times.

Scoping Agent Memory Prevents Unrelated Tasks From Disrupting Each Other

Cheng et al. tested persistent skill editing in frozen language-model agents across 12-round streams of recurring code-repair tasks. They found that globally retrieving locally validated skills caused cross-family interference, while family-scoped retrieval raised utility and supported more safe, repeated updates.