Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
Targets multi-hop RAG by training the embedding retriever on its own failures.
AI Storyline
3 items · 3 sources · 3 days
Operational story trace
Retrieval-Augmented Generation
Follow in this browser to see new updates on your Live feed.
Current stateDeveloping
Latest change
A Nov 6 paper compares RAG over a raw transcript against a two-tier memory that compacts each turn into atomic facts, testing whether conversational agents should keep the transcript at all.
Earlier contextThe story so far
Late-October research targeted the retriever: a paper trained embedding models on their own mistakes so multi-hop questions find all the evidence they need. A November 2 paper, TRIO, then argued that current RAG approaches are still far from satisfactory.
editor-curated · source-linked
Arc
Oct 29Nov 6 · now
RETRIEVER · Oct 29Training embedding models on their own retrieval mistakes for multi-hop RAG1 source · show source ▾
PIPELINE · Nov 2TRIO proposes task refinement with individual optimization for RAG1 source · show source ▾
TRIO: Task Refinement with Individual Optimization for Retrieval-Augmented Generation
Frames current RAG as unsatisfactory and proposes TRIO for task refinement.
NOW · Nov 6Two-tier conversational memory evaluated head-to-head against RAG1 source · show source ▾
Compaction, decay and reinforcement in a two-tier conversational memory: a paired evaluation against retrieval-augmented generation
Moves the question from retrieval to memory: transcript search versus compacted atomic facts.
Day 1 Thursday, Oct 29, 2026
Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
Targets multi-hop RAG by training the embedding retriever on its own failures.
Day 2 Monday, Nov 2, 2026
TRIO: Task Refinement with Individual Optimization for Retrieval-Augmented Generation
Frames current RAG as unsatisfactory and proposes TRIO for task refinement.
Day 3 Friday, Nov 6, 2026
Compaction, decay and reinforcement in a two-tier conversational memory: a paired evaluation against retrieval-augmented generation
Moves the question from retrieval to memory: transcript search versus compacted atomic facts.
What to watch — open questions
- Do the multi-hop retriever gains hold on benchmarks other than those the paper reports?
- Does compacted fact memory beat transcript RAG outside the paper's paired evaluation?
- Will any of these methods ship in a widely used RAG framework?
How this thread was built
editor wrote the arc · 3 beatswatcher 1 status change
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.