ISA-Distill: On-Policy Distillation for Search Agents Adapts Teacher Feedback to Student Knowledge
_reachsumit · x · 2026-11-04
A new arXiv paper introduces Information-State-Aware Distillation (ISA-Distill), a method for training search agents.
- Existing on-policy distillation gives the teacher privileged context (evidence, reference answers, hindsight), but the teacher may then rely on facts the student cannot reliably recover, undervaluing retrieval that would help establish them.
- ISA-Distill estimates how reliably the student can recover action-relevant knowledge and produces two complementary feedback views for the same sampled action: a fact-use view that supplies the relevant fact, and an evidence-seeking view that presents the information need.
- The signals are weighted toward fact use when recoverability is high and toward evidence seeking when low.
Experiments across diverse QA benchmarks and model scales show improvements.
More from Research
- Noetra engineer's first project: building an image encoder from scratch for multimodal LLMs — HirokatuKataoka · 2026-11-04
- Japanese team trains image encoder from scratch to match SigLIP2, with better CJK performance — HirokatuKataoka · 2026-11-04
- SpecJudge author argues LLM-as-judge tools should never ask models for self-reported confidence — jokiruiz · 2026-11-04
- Tsinghua's BEMULA benchmark: no LLM agent memory system passes over 1.5% of forget requests — rohanpaul_ai · 2026-11-04
- KAIST AI's FactEmbed decouples query understanding to improve composed image retrieval — kaist-ai · 2026-11-04
- Intology's Locus auto-designs molecular models, beating GPT-6 and Claude agents — rohanpaul_ai · 2026-11-04