Research map / Learning theory and algorithms / Combinatorial and search optimization
HOBIT: Hardness Optimized Batch Sampling for InfoNCE Training
Himanshu Dutta, Lokesh Nagalapatti, Yashoteja Prabhu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HOBIT batching instead of Random batching | dual encoder (roberta-base) · MSMARCO Passage | MRR@10: 24.4 -> 27.4 | Improved |
| use HOBIT batching instead of Random batching | dual encoder (roberta-base) · NQ | Recall@10: 45.1 -> 52.4 | Improved |
| use HOBIT batching instead of NGAME | dual encoder (e5-large-unsupervised) · MSMARCO Passage | MRR@10: 28.6 -> 30.1 | Improved |
| use HOBIT batching instead of NGAME | dual encoder (e5-large-unsupervised) · NQ | Recall@100: 90.5 -> 89.9 | Hurt |
| use fixed (static) teacher-ordered batches (B3) instead of adaptive batching | dual encoder (roberta-base) · MSMARCO Passage | MRR@10: 24.4 -> 23.4 | Hurt |
| add HOBIT batching on top of ANCE hard negatives | ANCE (roberta-base) · MSMARCO Passage | MRR@10: 27.33 -> 27.29 | No clear effect |
| add HOBIT batching on top of TriSampler | TriSampler (roberta-base) · MSMARCO Passage | MRR@10: 26.26 -> 26.65 | Improved |
| add HOBIT batching on top of ANCE | ANCE (roberta-base) · NQ | Recall@10: 59.27 -> 56.43 | Hurt |
| add HOBIT batching on top of ANCE | ANCE (roberta-base) · NQ | Recall@100: 84.63 -> 85.41 | Improved |
| add HOBIT batching on top of TriSampler | TriSampler (roberta-base) · NQ | Recall@10: 53.80 -> 57.13 | Improved |
| add HOBIT batching on top of ANCE | ANCE (roberta-base) · TriviaQA | Recall@100: 86.34 -> 88.34 | Improved |
| use cached embeddings (HOBIT-C) instead of per-epoch recomputed embeddings (ablation) | HOBIT (roberta-base) · NQ | Recall@10: 52.4 -> 51.7; 29.4% per-epoch speedup | No clear effect |
Closest papers by meaning
- InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning (ICLR 2024 · similarity 0.85)
- Exploring Learning Complexity for Efficient Downstream Dataset Pruning (ICLR 2025 · similarity 0.84)
- Critical Batch Size Revisited: A Simple Empirical Approach to Large-Batch Language Model Training (NeurIPS 2025 · similarity 0.84)
- Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty (ICML 2025 · similarity 0.84)
- Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation is Wasteful (NeurIPS 2025 · similarity 0.84)
- Batch Normalization Is Blind to the First and Second Derivatives of the Loss (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
