HindSearch: Enhancing Search Agent Training with Hindsight Critiques
_reachsumit · x · 2026-08-04
Search-augmented LM agents are typically trained with binary exact-match rewards, ignoring the reasons behind failed trajectories. HindSearch introduces a hindsight self-distillation procedure.
After each rollout, a frozen judge writes a short critique of failed trajectories using the gold answer, providing an auxiliary on-policy distillation signal for the student's search actions. On a seven-benchmark suite using Qwen2.5-3B-Instruct, HindSearch achieves a 39.4% average EM, outperforming prior search-RL baselines.
More from coding & agent
- Qwen 3.8 Coding Test: Nearly Matches K3 at Half the Price — bindureddy · 2026-08-04
- Overcoming State Loss in Long-Horizon Agents: New Framework Boosts Accuracy — Ziyu Ma · 2026-08-04
- Skip Docker: db-here Enables Zero-Risk Database Isolation for AI Agents — andersonbcdefg · 2026-08-04
- memsem: Local Semantic Memory MCP Server for AI Agents — WindSeries · 2026-08-04
- OpenAI launches ChatGPT Work agent for hours-long complex projects — emmanuelvivier · 2026-08-04
- Google Launches Managed Agents in Gemini API with MCP Support — emmanuelvivier · 2026-08-04