Research map / Architectures and efficiency / RNNs and normalization
Beyond Hamming: Query-Aware Decoding of Binary Cosine Sketches
DaeHun Nyang
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 |
|---|---|---|---|
| decode with query-aware probit MLE instead of Hamming agreement | SimHash · ArguAna HARD0.1 (K=100) | Hit@K: 0.000 -> 0.454 | Improved |
| decode with query-aware probit MLE instead of Hamming agreement | SimHash · SciFact HARD0.3 (K=100) | Hit@K: 0.233 -> 0.467 | Improved |
| decode with query-aware probit MLE instead of Hamming agreement | SimHash · FiQA HARD0.3 (K=200) | Hit@K: 0.087 -> 0.357 | Improved |
| decode with query-aware probit MLE instead of Hamming agreement | SimHash · ArguAna HARD0.1 (recall 0.8) | candidates at fixed recall: 766 -> 385 | Improved |
| decode with query-aware probit MLE instead of Hamming agreement | SimHash · FiQA HARD0.2 (recall 0.8) | candidates at fixed recall: 5,831 -> 3,786 | Improved |
| replace search-time scorer in hnswlib with QA-COS (128 bits) | SimHash scorer · Native hnswlib on BEIR | Recall@10: .770 -> .853 | Improved |
| replace search-time scorer in hnswlib with QA-COS (64 bits) | SimHash scorer · Native hnswlib on BEIR | Recall@100: .258 -> .335 | Improved |
| replace search-time scorer in hnswlib with QA-COS (128 bits) | SimHash scorer · Native hnswlib on BEIR | visited nodes: 1364.5 -> 1228.5 | Improved |
| use gated T=1 QA-COS decoder in storage-aware pipeline | SimHash scorer · FiQA matched recall | mean latency (ms): 1.175 -> .997 | Improved |
| use QA-COS instead of SimHash for angle estimation | SimHash · 32 (d,m) settings | MAE: 22.3-28.9% reduction | Improved |
| use Li's practical reduced-statistic estimator instead of SimHash | SimHash · 32 (d,m) settings | MAE: 11.8-18.3% reduction | Improved |
| use full query-aware likelihood (QA-COS) instead of Li's reduced statistic | Li practical estimator · 32 (d,m) settings | MAE reduction: 11.8-18.3% -> 22.3-28.9% | Improved |
Closest papers by meaning
- Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment (ICLR 2025 · similarity 0.83)
- Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match (ICLR 2026 · similarity 0.83)
- TriAttention: Efficient Long Reasoning with Trigonometric KV Compression (ICML 2026 · similarity 0.83)
- Sampling-Efficient Test-Time Scaling: Self-Estimating the Best-of-N Sampling in Early Decoding (NeurIPS 2025 · similarity 0.83)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.82)
- Geometry-Aware Decoding with Wasserstein-Regularized Truncation and Mass Penalties for Large Language Models (ICML 2026 · similarity 0.82)
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.
