On-policy training teaches multi-hop RAG retrievers to recover from mistakes
_reachsumit · x · 2026-10-30
A new paper proposes on-policy state training for chain embedding retrievers in multi-hop RAG.
- Baseline finding: a small chain retriever trained only on gold histories beats graph-based systems and rivals search agents, but by step three over half of its test-time states contain wrong passages, hurting recall.
- The method has the encoder retrieve its own chains each epoch, labels each visited state with the earliest missing gold passage (masking the rest), and trains on these self-generated states.
- Result: the retriever learns to recover from its own retrieval errors, improving evidence completeness for complex questions.
More from Research
- UChicago seminar: hippocampal ripples, visual imagery BCIs, minimally invasive neural interfaces — plopesresearch · 2026-10-30
- Certificate-Guided Visual Routing paper makes diffusion VLMs look only when needed — _akhaliq · 2026-10-30
- Oopsie-data: Building the first open dataset of robot failure trajectories — chris_j_paxton · 2026-10-30
- AI Agents Audit 138 Economics Papers, Flag 2,035 Potential Problems — haider1 · 2026-10-30
- Schmidhuber's 'Miraculous Year 1990-91': claiming roots of GANs, Transformers, pre-training — SchmidhuberAI · 2026-10-30
- NBER opens applications for Stanford Digital Economics and AI tutorial, due Nov 3 — avicgoldfarb · 2026-10-30