跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 2026-09-29AI 評分41

Video-RSI: Recursive Self-Improvement of Video Understanding Agents via Harness Evolution

Video-RSI: Recursive Self-Improvement of Video Understanding Agents via Harness Evolution

AI 導讀

Video-RSI 透過執行框架演進,讓影片理解智能體在不更新模型權重的情況下遞迴自我改進。智能體可重新檢視訓練影片、收集視覺、轉錄及 OCR 證據,並據此修改提示詞、工具、記憶和執行邏輯。評測基準結果顯示,演進後的執行框架提升準確度,同時處理較少影格;團隊亦釋出推理與評測程式碼及相關執行框架。

正文

We introduce Video-RSI, a framework for recursive self-improvement of video understanding agents through harness evolution. Instead of updating model weights, Video-RSI enables the agent to revisit the original training videos used for evolution, gather additional visual, transcript, and OCR evidence, and test competing explanations for its failures. Evidence-supported diagnoses are then translated into reusable changes to the agent’s prompts, tools, memory, and execution logic.

Video-RSI uses a cost-aware selection process to retain revisions based on both answer accuracy and visual processing cost. Across the evaluated video understanding benchmarks, the evolved harness improves accuracy while processing fewer frames, yielding competitive accuracy–efficiency trade-offs.

We release the inference and evaluation code, the initial and final harnesses, and the implementation of the harness-evolution workflow. We welcome feedback and discussion from the community.

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co