A set of four papers released 8-9 Oct 2026 proposes a stack for self-improving LLM agents that freezes the language model and instead evolves external world-model rules, verified guardrails, goal-cond…
EvoBelief, a self-evolving world model that represents environment knowledge as falsifiable hypotheses, improved task performance across four benchmarks while keeping the underlying language model fro…
A diffusion language model (dLLM) planner achieved a 53.7% plan repair success rate versus 27.0% for an autoregressive (AR) planner on Natural Plan without task-specific training, according to the arX…
Researchers introduced GraphOPD, a graph-augmented on-policy distillation method for LLM agents that scores each step by a random-walk stationary distribution over a dependency graph built from the en…
Researchers proposed SAGA (Self-evolving Agents through Experience-Grounded Abstraction), a framework that transforms LLM agent interaction trajectories into episodic descriptions, reusable procedures…
Google Research open-sourced EnvHarness on October 2, a programmable wrapper layer that adapts existing agent training benchmarks by modifying starting conditions, action contracts, and task sequencin…
A 28 Sep 2026 arXiv paper introduces ProVer, a framework that targets potentially pivotal decisions for fine-grained credit assignment in agentic reinforcement learning, addressing what the authors ca…
A new arXiv paper (2610.00372v1) introduces the Causal Intervention Router (CIR), a lightweight policy that decides when an LLM agent's error-recovery harness should intervene, raising success on long…
A new arXiv paper (2610.00010v1) proposes Core-Tail World Model (CTWM), a rank-based memory controller that allocates prompt budget with a single exponent τ while retaining a summarized tail, after a …
A paper proposes STEP-HRL, a hierarchical reinforcement learning framework that conditions LLM agents on single-step transitions instead of full interaction histories, using completed subtasks to repr…
MATE, a deterministic post-retrieval memory adaptation procedure, achieved task success rates of 81.3% and 93.3% on 134 ALFWorld tasks with Qwen2.5-14B and 72B while using roughly one-tenth of the tok…
A new arXiv paper introduces StepLearn, a nonparametric framework that lets large language model agents learn from individual transitions during deployment while requiring prospective validation befor…
Researchers introduced CoSkill, a multi-agent reinforcement learning framework that jointly trains a Reasoning Agent and a Meta-Skill Agent over a hierarchical skill library, achieving success rates o…
A new arXiv preprint (2609.0289v1) introduces HARNESSEVO, which decomposes an LLM agent's textual harness into four separately evolvable slots, and finds that on ALFWorld with a frozen 7B backbone, ne…
Researchers propose Potential-Guided Policy Optimization (PGPO), a new reinforcement learning method for multi-turn agentic tasks that estimates empirical state potentials from anchor-state-group retu…
Researchers introduced SkillGLoW (Global-Local Weave), a method for consolidating procedural skills in LLM agents, achieving a 17.2-point average gain on hard benchmarks over no-skill baselines across…
Researchers propose VICT (Verifier-Instrumented Credit Tracing), a training-time interface that improves credit assignment in reinforcement learning for long-horizon LLM agents by tracing verifier-int…
Researchers from Meta AI and the University of Illinois Urbana-Champaign introduced EvoHarness-RL, a trainable coordination layer that enables large language model agents to create and manage their ow…
Google Cloud AI Research, with researchers from Washington University in St. Louis and UNC Chapel Hill, introduced EnvHarness, a programmable wrapper that adapts AI training environments to target age…
Researchers propose MileGPO, a method for credit assignment in long-horizon agentic reinforcement learning that derives process-level credit from grouped on-policy rollouts via Milestone Discovery, Re…