Research map / Architectures and efficiency / Attention and transformers
Less Is More: Fast and Accurate Reasoning with Cross-Head Unified Sparse Attention
Lijie Yang, Zhihao Zhang, Arti Jain, Shijie Cao, Baihong Yuan, Yiwei Chen, Zhihao Jia, Ravi Netravali
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 |
|---|---|---|---|
| unify top-k token selection across heads and keep 25% recency window (LessIsMore) vs TidalDecode | LessIsMore · AIME-24 (Qwen3-8B, 2K budget) | accuracy (%): 53.3 -> 73.8 | Improved |
| use LessIsMore instead of Quest | Qwen3-8B · AIME-24 (budget 2K) | accuracy (%): 18.2 -> 73.8 | Improved |
| sparse attention baselines lengthen generation | Quest at 2K budget · AIME-24 Qwen3-8B | generation length (K tokens): 14.8 -> 30.0 | Hurt |
| LessIsMore keeps generation length near full attention | LessIsMore · AIME-24 Qwen3-8B | generation length (K tokens): 14.8 -> 15.8 | No clear effect |
| cross-head unified aggregation vs head-to-head and randomized top-k, selection only at Layer 2 (ablation) | LessIsMore · attention recall, Qwen3-8B | attention recall | Improved |
| add a recency window (25%) to cross-head selection (ablation) | LessIsMore · AIME-24, 4K budget, Qwen3-8B | attention recall | Improved |
| use only recent tokens (r=100%) instead of 25% recency window (ablation) | LessIsMore · AIME-24, 4K budget, Qwen3-8B | attention recall | Hurt |
| remove recency window (0% recent) with cross-head selection (ablation) | LessIsMore · AIME-24, 4K budget, Qwen3-8B | correctness | Hurt |
| unified selection kernel vs TidalDecode | LessIsMore · kernel latency 2K/16K | latency: 32.1us -> 20.1us | Improved |
| end-to-end decoding with LessIsMore at longer context | LessIsMore-2K · TBT speedup | speedup: 1.11 -> 1.51 | Improved |
| increase token budget 2K to 4K in serving stack | LessIsMore · TBT speedup at 64K | speedup: 1.51 -> 1.48 | Hurt |
| LessIsMore vs TidalDecode on LongBench with 1024 budget | LessIsMore · LongBench avg | average score: 40.89 -> 43.82 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- TidalDecode — TidalDecode: Fast and Accurate LLM Decoding with Position Persistent Sparse Attention (ICLR 2025) · better on 6 of 6 shared results
- Quest — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Efficient Reasoning via Constrained Optimization in Latent Space (NeurIPS 2026 · similarity 0.90)
- Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models (ICML 2026 · similarity 0.89)
- Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (NeurIPS 2025 · similarity 0.89)
- BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference (ICML 2026 · similarity 0.89)
- FROST: Filtering Reasoning Outliers with Attention for Efficient Reasoning (ICLR 2026 · similarity 0.88)
- Multipole Attention for Efficient Long Context Reasoning (NeurIPS 2025 · similarity 0.88)
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.
