Research map / Large language models / Retrieval and knowledge
MoL: Adaptive Mixture-of-Length Reasoning for Efficient Question Answering with Context
Guocong Li, Jinjian Zhang, Ping Wang, Dongnan Liu, Tian Liang, Qiuyi Qi, Hao Huang, Siyan Guo, Mutian Bao, Wei Zhou, Linjian Mo, Hongxia Xu and 1 more
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 extension reward Rextend (ablation) | MoL · Loong | accuracy: 58.9 -> 62.3 | Improved |
| add compression reward Rcompress (ablation) | MoL · Loong | output tokens: 2862 -> 1374 | Improved |
| use question-relevant key sentence difficulty instead of original difficulty labels (ablation) | MoL · HotpotQA | accuracy: 63.0 -> 67.2 | Improved |
| use key-sentence difficulty instead of whole-passage similarity (ablation) | MoL · HotpotQA | output tokens: 594 -> 316 | Improved |
| add progressive learning strategy lambda(t) (ablation) | MoL · Loong | accuracy: 49.1 -> 62.3 | Improved |
| add progressive learning strategy lambda(t) (ablation) | MoL · HotpotQA | output tokens: 835 -> 316 | Improved |
| raise alpha from 0.1 to 0.9 | MoL · HotpotQA | accuracy and tokens: 64.9 -> 67.3 accuracy, 264 -> 371 tokens | Mixed |
| raise gamma from 0.3 to 0.9 | MoL · HotpotQA | accuracy: 67.2 -> 65.1 | Hurt |
| raise epsilon1 from 0 to 0.8 | MoL · HotpotQA | accuracy: 67.4 -> 65.1 | Hurt |
| set epsilon2 to 0.6 rather than 0 | MoL · HotpotQA | accuracy: 61.7 -> 67.2 | Improved |
| vary initial Ltarget (256 to 1024 sets) | MoL · HotpotQA | accuracy: 66.7 to 67.4 | No clear effect |
| apply MoL instead of GRPO on Qwen3-8B | Qwen3-8B · HotpotQA | accuracy and tokens: 63.7 -> 67.2 accuracy, 747 -> 316 tokens | Improved |
Closest papers by meaning
- Cost-efficient Knowledge-based Question Answering with Large Language Models (NeurIPS 2024 · similarity 0.86)
- QFFT, Question-Free Fine-Tuning for Adaptive Reasoning (NeurIPS 2025 · similarity 0.86)
- LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts (ICLR 2026 · similarity 0.86)
- Chain-of-Action: Faithful and Multimodal Question Answering through Large Language Models (ICLR 2025 · similarity 0.86)
- Fixing the Broken Compass: Diagnosing and Improving Inference-Time Reward Modeling (ICLR 2026 · similarity 0.85)
- Learning to Reason via Mixture-of-Thought for Logical Reasoning (ICLR 2026 · similarity 0.85)
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.
