arxiv arXiv cs.AI · just now · research

APEx distills agent procedural experience for adaptive deep research

from English

Researchers propose APEx, a hierarchical experience utilization framework that organizes interaction history into instance-level trajectory memories and category-level procedural skills to improve deep research agents.

The system couples these memories through a closed-loop architecture comprising Executor, Distiller, and Planner modules.
These modules are optimized via a three-stage alternating GRPO training paradigm for reward-guided skill distillation.
At test time, distilled skills serve as procedural priors for online Planner adaptation using skill-guided test-time reinforcement learning.
Experiments on 7 benchmarks show APEx surpasses GPT-5.4 by 14.7 points and the strongest memory-augmented baseline by 3.0 points.

APEx enables ground-truth-free self-improvement with skill-alignment regularization to prevent policy drift, achieving state-of-the-art performance on complex questions.

Importance 2/3 arXiv cs.AI AI agents Research paper

Read original

arxiv arXiv cs.CL · 2h ago

APEx distills agent procedural experience for adaptive deep research

Researchers propose APEx, a hierarchical experience utilization framework designed to improve deep research agents that answer complex questions using external tools. The system organizes interaction history into instance-level trajectory memories and category-level procedural skills, coupling them through an Executor, Distiller, and Planner architecture.

arxiv arXiv cs.LG · 21h ago · 3 views

CANOPY enables Qwen3-14B to top AppWorld using outcome-only RL

The paper introduces CANOPY (Coverage-ANchored On-PolicY RL), a protocol that addresses signal starvation and policy drift in long-horizon reinforcement learning for small open models. By scaling same-task exploration and keeping updates KL-anchored, the method allows agents to learn effectively from end-of-task verification alone.

arxiv arXiv cs.CL · 2d ago · 4 views

WebWorld uses browser-as-world-model to self-improve web code

The authors present WebWorld, an interface that allows a Vision-Language Model (VLM) to interact autonomously with the browser as a deterministic world model for web code. This approach addresses the structural flaw in VLM-driven self-improvement where the model judges its own repairs by introducing a counterparty that cannot be fooled.

media MarkTechPost · 3d ago · 6 views

Google releases EnvHarness to adapt static agent environments for training

A team from Google Cloud AI Research, Washington University in St. Louis, and UNC Chapel Hill has released EnvHarness, a programmable layer that transforms static agent benchmarks into adaptive training worlds. The system wraps existing environments using plug-in components that operate strictly through the standard reset() and step() interface, allowing the environment to adapt to the policy's training without altering the underlying simulator or human-built verifiers.

arxiv arXiv cs.CL · 7d ago · 10 views

JIT-Agent enables on-the-fly harness evolution to boost agentic LLM performance

The authors present JIT-Agent, a harness intelligence model trained to synthesize task-adaptive agent harnesses for arbitrary off-the-shelf agentic LLMs. This approach formalizes the agent harness as a composable artifact governed by a fixed four-module protocol, allowing the system to customize and repair harnesses on the fly.