𝜋Skill: High-Density Knowledge Extraction from Single Trajectories via Circular Step-Level Analysis

ACL ARR 2026 May Submission16964 Authors

Keywords: Multimodal Agents; Training-free Continual Learning; Agent Memory; Knowledge Distillation

Abstract: Multimodal agents can solve complex reasoning tasks with external tools, but still struggle to improve from past trajectories without parameter updates. Existing training-free memory methods often rely on coarse trajectory-level summaries and static repositories, limiting fine-grained failure analysis and memory reliability. To address this, we propose \textbf{$\pi$Skill}, a lifecycle-managed continual learning framework for multimodal agents. Its Circular Step-level Knowledge Distillation (CSD) extracts high-density experience atoms from successful and failed step transitions, while Confidence-aware Lifecycle Memory Updating (CMU) tracks confidence, usage statistics, lifecycle states, and version history to refine or suppress unreliable knowledge. During inference, $\pi$Skill retrieves knowledge through semantic recall and confidence-aware refinement, with execution feedback used for online memory updating. Experiments with Qwen2.5-VL-7B-Instruct on four multimodal agent benchmarks show that $\pi$Skill improves the average Average@4 from 12.41 to 14.02 and Pass@4 from 26.53 to 29.73, demonstrating the effectiveness of fine-grained knowledge distillation and lifecycle-aware memory evolution.

Paper Type: Long

Research Area: NLP and Code Models

Research Area Keywords: multi-agent systems,tool use,retrieval-augmented language models

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: English

EMNLP 2026 AI Reviewing Experiment: no

Submission Number: 16964