NHNM paper turns positives into natural hard negatives to train instruction-following retrievers
_reachsumit · x · 2026-10-29
An arXiv paper proposes Natural Hard-Negative Mining (NHNM) to address scarce instruction-conditioned supervision for instruction-following information retrieval (IFIR). Instead of fixing positives and synthesizing extra negatives, NHNM promotes convertible negatives under query-compatible instructions and repurposes original positives as natural instruction hard negatives, via an automated Candidate Cleaning, Reversal Construction, and Geometry Filtering pipeline.
- Applied to MSMARCO, it yields NHNM-MARCO, an IFIR training set of nearly 7K instances
- Models from 568M to 4B parameters fine-tuned on NHNM-MARCO achieve strong instruction-following performance on InstructIR and FollowIR, matching or beating substantially larger retrievers in several settings
- Controlled analyses attribute the gains to natural hard-negative supervision rather than scale or data volume
More from Research
- Fine-tuned Qwen3.5 4B goes from placing cubes to full 3D scenes — teortaxesTex · 2026-10-29
- Cross-lingual LLM effects paper wins Outstanding Paper award at EMNLP 2026 — anas_ant · 2026-10-29
- Lie detector compiled into Qwen3.5 weights flags 91% of lies with zero false positives — teortaxesTex · 2026-10-29
- Practitioner: No serious lab actually trains on the test set today — nrehiew_ · 2026-10-29
- Researchers slam secrecy around OpenAI-linked Yang-Mills proof materials — anshulkundaje · 2026-10-29
- Job-market paper: German engineering societies lifted regional patents 4.8% — Afinetheorem · 2026-10-29