AI Agent Papers2026

Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses

Proposes Recuris, a recursive Experiential-Working Memory architecture that separates Working Memory (task progress) from Experiential Memory (skills) and uses a Meta-Agent to validation-gate localized skill updates, enabling bounded recursive skill evolution for long-horizon agents.

Visit Website

Introduction

Long-horizon interactions accumulate noisy histories that obscure task state and misalign skill use; Recuris flips the problem by keeping a compact Working Memory to guide skill selection from an evidence-backed Experiential Memory and using localized, validation-gated updates to evolve skills over time.

Key Findings

Structured coupling of Working Memory and Experiential Memory localizes execution failures to specific memory components, making updates more targeted and reliable.
A fixed Meta-Agent converts execution evidence into gated updates to Skill Memory, producing a bounded recursive loop that reshapes future execution.
Broad empirical gains: improved success in 35 of 37 model–benchmark pairs; +17.8 points on GPT-5.6 Sol and +15.6 on Claude Opus 5 (taking Opus 5 to 87.9%); +16.6/+13.5 on Qwen3.6-27B/35B on SkillFlow; up to +32.2 points on the longest tasks and up to 80% reduction in common long-horizon failures.
Code and artifacts released for reproducibility and follow-up experimentation.

Who it's for and trade-offs

Great fit if you build or evaluate LLM-driven agents on long-horizon tasks and need a systematic way to convert episodic experience into reusable, validated skills. The approach scales with interaction horizon and is model-agnostic across modern LLM backbones. Look elsewhere if your tasks are strictly single-step or if your primary constraint is removing all external orchestration (Recuris assumes an external harness and a Meta-Agent for policy updates).

Back

Information

Websitearxiv.org
OrganizationsNUS, Princeton University, Stanford University, University of Oxford
AuthorsZhaochen Yu, Yingcheng Wu, Zhenfei Yin, Kaiyuan Chen, Zhe Zhao, Mengdi Wang, Shuicheng Yan, Ling Yang
Published date2026/08/25

More Items

Computer Vision Papers2026

On-Policy Self-Distillation in Diffusion Models

Wei Zhou, Xiongwei Zhu +15

Converts image-level rewards into explicit intermediate targets for diffusion-model denoising via an on-policy self-distillation loop. Constructs bounded positive/negative targets around anchors from reward gradients, fits those targets with finite updates, and refreshes a behavior policy by EMA—improving aligned performance across backbones while reducing GPU hours.

distillation RL vision image evaluation+2

AI Agent Papers2026

AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces

Sungho Park, Wonjoong Kim +11Affiliation: KAIST, Affiliation: Southern University of Science and Technology +1

Automatically optimizes runtime harnesses for LLM agents by diagnosing failure traces and iteratively applying structured, generalizable patches. Combines batch-based failure diagnosis, code-like patch generation across prompts/tools/middleware, and validation-aware selection to raise long-horizon task success on multiple benchmarks.

LLM ai-agent agent-skills long-horizon benchmarks+4

AI Video Papers2026

Annotations as Rollouts: Efficient and Scalable Reinforcement Learning for Video MLLMs

Yunheng Li, Guohong Mu +5VCIP, School of Computer Science, Nankai University, Brain and Artificial Intelligence Lab, Northwestern Polytechnical University +2

Treats human annotations as oracle rollouts and separates them from on-policy baselines to improve reinforcement learning for video multimodal LLMs. Key features include a decoupled advantage estimator, sign-balanced pruning, and scalable gains across model sizes and data budgets.

video multimodal rl LLM vision+3