GLInt: Geometry-matched hard negatives for late-interaction retrieval
beirmug · x · 2026-08-27
The article introduces GLInt (Geometry-Matched Hard Negatives), a new method for late-interaction retrieval models (like ColBERT). The key finding is that hard negatives mined by a multi-vector retriever yield better model performance than those mined by dense retrievers.
Key Data & Conclusions:
- Performance: Achieved a mean nDCG@10 of 57.43 on 15 BEIR tasks, surpassing the LateOn baseline (57.22), making it one of the strongest retrievers under 300M parameters.
- Specific Tasks: Achieved the highest scores on Quora and HotPotQA among models under 7B parameters.
- Zero-shot Eval: Reached 0.8815 nDCG@10 (acc@1=0.807) on a specific benchmark.
Technical Details:
- The method focuses on mining data in MaxSim space, using geometry-matched hard negatives to boost OOD (Out-of-Distribution) performance.
- Utilized PyLate and FastPLAID tools in practice.
- Training data (GLInt-data) and the model are open-sourced.
More from Research
- V-Rubrics: Improving Visual Faithfulness via Rubric-Based Reinforcement Learning — liuziwei7 · 2026-08-27
- FlashKDA ref impl lower bound set to -5 months ago — YouJiacheng · 2026-08-27
- Interpretability research lacks a breakthrough moment; solving 'superintelligent psyche' may take much longer — ericjmichaud_ · 2026-08-27
- Mixed SFT Outperforms Next-Chunk RL Without CoT Data — Yinhao Tang · 2026-08-27
- Real-TurnTurk: Multimodal Turkish Corpus for Turn-Taking Prediction — Ahmet Tuğrul Bayrak · 2026-08-27
- OpenMythos: 14.8k-star open reconstruction of the "Claude Mythos" architecture — tom_doerr · 2026-08-27