EviSD: Evidence-Conditioned Self-Distillation for Search Agents
_reachsumit · x · 2026-08-04
The paper proposes EviSD, an evidence-conditioned self-distillation framework designed to address the limitation of outcome-based RL (e.g., GRPO) in distinguishing the contributions of individual actions within multi-turn search processes.
Core Mechanism:
- During training, the model samples actions as a student from the original context. Simultaneously, the same model acts as a "privileged teacher," re-scoring these actions using action-aligned context injected with supporting evidence and golden answers.
- The teacher-student gap is converted into a bounded correction applied only to generated action spans, providing localized guidance without altering the inference process.
Results: Across 7 QA benchmarks and 3 backbones, EviSD achieved the highest macro-average Exact Match in all settings, outperforming the strongest baselines by 1.3–2.3 points while modulating only 6.7%–15.1% of response tokens.
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