GLint: A Field Report on Hard Negatives for Late-Interaction Retrievers
antoine_chaffin · x · 2026-08-09
The author shares a detailed experimental report on building GLInt, a 150M parameter late-interaction retriever. The model achieves a 57.43 mean nDCG@10 across 15 BEIR tasks, setting a new SOTA for models under 300M parameters.
Key Research & Findings:
- Mining Space: Explores whether mining hard negatives in multi-vector MaxSim space yields better late-interaction models compared to traditional dense retriever mining.
- Hard vs. False Negatives: Analyzes the challenges of constructing hard negatives in multi-vector scenarios and how to filter effectively to avoid false negatives.
- Training Strategy: Transitions from Supervised Fine-Tuning (SFT) to knowledge distillation to boost performance further.
- Pitfalls: The article includes a dedicated section on experiments that did not work, demonstrating a complete research loop from hypothesis and diagnosis to iteration.
Related event: GLint: SOTA Pure Late-Interaction Retrieval Model(2 posts)→
More from Research
- Optuna creator Akiba: when 2 hyperparameters cost $20M, what's left for BO/HPO — Stefania_druga · 2026-10-16
- AI will be superhuman at science regardless, and Europe can't opt out — teortaxesTex · 2026-10-16
- SERO augments long-horizon agentic RL with executable oracle continuations — Shanghai-AI-Laboratory · 2026-10-16
- Tsinghua and Amap propose OmniVerifier-C1, a graded verifier beating binary judges — Bowei Liu · 2026-10-16
- Alibaba open-sources MAI-Gym, an all-in-one GUI agent eval and RL training platform — Tongyi-MAI · 2026-10-16
- Peking University's SimForcing distills simulation motion priors into robot world models — PekingUniversity · 2026-10-16