The paper introduces the Agent-Editing World Model (AEWM), a new approach that models how reasoning and actions influence future task progress instead of simulating tool responses. AEWM includes an Action Judge that classifies decisions as Critical, Exploratory, or Noisy, and a State Revision mechanism that edits noisy reasoning–action continuations from the same observed history. The integrated system, EditAct, directly updates the underlying state during real execution, leading to significant performance gains across multiple benchmarks and agent backbones.
arXiv:2610.19951v1 Announce Type: cross Abstract: Agent skills adapt frozen models through editable external instructions. Revising several rules at once makes changes difficult to inspect, while jud...
GlobeBench is a new benchmark for evaluating text-based world models, comprising 1,632 tasks across five domains and nineteen simulation skills. It tests whether models can predict an environment’s response to an action based on the full agent trajectory, requiring understanding of state changes, environment rules, and action effects on related and unrelated objects. The benchmark compares predictions to execution-derived references using a hybrid architecture of deterministic gates and a semantic judge, and finds that GPT‑6 Astra tops the 20 tested models with a 49.57% success rate, while accuracy drops sharply when actions affect many dependent objects, leaving 631 tasks unsolved by any model.
arXiv:2608. 06197v1 Announce Type: new Abstract: Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on external simulators that are difficult to ground.
EMMA is a framework that treats the configuration of a multi‑agent system—controller routing, specialist prompts, and procedural skills—as structured episodic memory. It explicitly assigns ownership and limits updates through a reflector and backbone‑independent guards, ensuring edits are scoped, incremental, and bounded. Experiments on KramaBench, AppWorld, and BBEH show EMMA outperforms four baselines, achieving up to 32.5% higher performance and reducing training executions by 17–38%.
The paper introduces CoPE, a method that trains language agents by adding multi‑step action targets to the policy backbone. By supervising each action forecasted from preceding histories, CoPE enables acting and forecasting to co‑evolve without extra inference cost. Across five benchmarks, CoPE outperforms existing RL baselines, achieving up to twice the sample efficiency on ScienceWorld and improving overall performance on realistic applications.
The paper introduces the Pitfall Anticipation World Model (PAWM), a selective guidance system for black-box large language model agents. PAWM predicts imminent failures in a bounded multi-step future and intervenes only when a potential pitfall is detected, reducing costly and often uninformative step-by-step predictions. Trained first with supervised fine-tuning on synthetic pitfall-labeled data and then co-trained with a black-box LLM, PAWM boosts Claude Sonnet 4.5’s performance by 24.3% on ScienceWorld and 7.7% on WebArena-Lite, outperforming several baselines.
ContextPilot is a proactive context‑management framework designed to improve long‑horizon agentic reasoning with large language models. It expands the toolset to include planning, long‑term memory, and soft context offloading, and introduces a reinforcement‑learning strategy that focuses on critical editing decisions and assigns action‑level advantages. Experiments on long‑context QA and deep search tasks demonstrate that ContextPilot achieves stronger performance with a more compact working context, outperforming existing baselines across various base models and benchmarks.
The paper introduces AMBER, an append‑only memory framework for language‑model agents that interact over long horizons. AMBER lets agents jointly learn to reason, act, and write free‑form memory, guaranteeing retention by construction and enabling end‑to‑end reinforcement learning without extensive curated data. Experiments on WebArena Lite show AMBER outperforms overwrite‑based memory by 4.09 percentage points in average success and improves task completion rates in repeated runs.
Memento 3 is a frozen large‑language‑model agent that learns explicit world models by maintaining a natural‑language rulebook in external memory. The agent iteratively observes, reflects, revises the rulebook, compiles it into executable code, and verifies updates through prediction error and replay, accepting only code that the LLM deems faithful to the rulebook. In experiments, a single‑model version solved all 25 public ARC‑AGI‑3 games with perfect efficiency, and a learned feedback controller achieved perfect scores in Atari Pong without further LLM calls.
EnvLunch proposes a warm‑up training stage that uses the executable environment itself as a source of verifiable tasks, eliminating the need for human‑written or model‑generated task descriptions. By sampling valid action trajectories, executing them, and training agents to recover equivalent trajectories from final observations, EnvLunch achieves over 9‑point performance gains on downstream tasks and reduces the required annotated examples by 8 to 32 times compared to baselines. This approach also improves fine‑tuning on benchmarks beyond data analysis, offering a more efficient starting point when annotated tasks are scarce.
arXiv:2608. 03137v1 Announce Type: new Abstract: Large language model (LLM) agents must retain reusable information, control a bounded active context, and recover earlier evidence during long-horizon interaction.
arXiv:2609.38334v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed as agents for multi-step decision-making, yet transfer poorly to unseen environments. World-mode...