Research map / Architectures and efficiency / Quantization
CrossQ: Task-Aligned Cross-Token Conditional Quantization for Late Interaction Retrieval
Rohit Kumar Salla, Manoj Saravanan, Ramya Manasa Amancherla
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 |
|---|---|---|---|
| add document context h(D) to code selection (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.371 -> 0.386 | Improved |
| add listwise KL distillation (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.377 -> 0.386 | Improved |
| add ranking losses and context over reconstruction-only training (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.358 -> 0.386 | Improved |
| add context and listwise together (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.364 -> 0.386 | Improved |
| add context pathway versus parameter-matched no-context (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.372 -> 0.386 | Improved |
| use correct document context instead of shuffled context (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.370 -> 0.386 | Improved |
| use untied assignment keys (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.383 -> 0.386 | Improved |
| replace mean pooling with DeepSets context encoder (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.378 -> 0.386 | Improved |
| replace DeepSets with 1-layer Transformer context (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.386 -> 0.387 | No clear effect |
| initialize codebooks from OPQ instead of random Gaussian (ablation) | CrossQ · MS MARCO dev (4 B/token) | MRR@10: 0.386 -> 0.388 | No clear effect |
| add light fine-tuning of last two blocks | CrossQ · MS MARCO dev (2 B/token) | MRR@10: 0.369 -> 0.374 | Improved |
| apply fine-tuning to OPQ versus CrossQ | OPQ · MS MARCO dev (4 B/token) | MRR@10: +0.002 vs +0.006 | Mixed |
Closest papers by meaning
- BTR: Binary Token Representations for Efficient Retrieval Augmented Language Models (ICLR 2024 · similarity 0.86)
- Rep2Text: Decoding Full Text from a Single LLM Token Representation (NeurIPS 2026 · similarity 0.85)
- Deep Evidential Hashing for Trustworthy Cross-Modal Retrieval (AAAI 2025 · similarity 0.84)
- RSQ: Learning from Important Tokens Leads to Better Quantized LLMs (NeurIPS 2026 · similarity 0.84)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.84)
- One Token Can Help! Learning Scalable and Pluggable Virtual Tokens for Retrieval-Augmented Large Language Models (AAAI 2025 · 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.
