Research map / Large language models / Agents, code and math
TASE: Token Awareness and Structured Evaluation for Multilingual Language Models
Chenzhuo Zhao, Xinda Wang, Yue Huang, Junting Lu, Ziqian Liu
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 |
|---|---|---|---|
| fine-tune with GRPO on synthetic TASE data (fine-grained reward) | Qwen2.5-14B-Instruct · TASE average | accuracy (%): 11.72 -> 20.40 | Improved |
| use fine-grained reward instead of coarse binary reward in GRPO (ablation) | qwen2.5-14b-grpo · TASE average | accuracy (%): 16.08 -> 20.40 | Improved |
| GRPO with coarse binary reward (ablation) | Qwen2.5-14B-Instruct · TASE average | accuracy (%): 11.72 -> 16.08 | Improved |
| GRPO fine-tuning | Qwen2.5-14B-Instruct · TASE LENOP | accuracy (%): 4.87 -> 27.22 | Improved |
| add chain-of-thought prompting | o1-mini · TASE average | accuracy (%): 24.56 -> 37.27 | Improved |
| add chain-of-thought prompting | gpt-3.5-turbo · TASE average | accuracy (%): 12.80 -> 18.45 | Improved |
| add chain-of-thought prompting | doubao-pro · TASE average | accuracy (%): 19.87 -> 22.94 | Improved |
| add chain-of-thought prompting | doubao-lite · TASE average | accuracy (%): 16.31 -> 16.44 | No clear effect |
| add chain-of-thought prompting | qwen2.5-14B · TASE average | accuracy (%): 11.72 -> 13.69 | Improved |
| evaluate in Chinese or Korean instead of English | o3 · TASE language split | accuracy (%): 86.71 -> 69.12 (Chinese), 67.83 (Korean) | Hurt |
| evaluate in Chinese or Korean instead of English | gpt-4.1 · TASE language split | accuracy (%): 39.89 -> 27.87 (Chinese), 25.75 (Korean) | Hurt |
| increase model size within Yi-1.5, Qwen2.5 and Qwen3 families | Yi-1.5, Qwen2.5, Qwen3 · TASE average | accuracy (%) | Mixed |
Closest papers by meaning
- Soft Token Alignment for Cross-Lingual Reasoning (NeurIPS 2026 · similarity 0.87)
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.86)
- MME-SCI: A Comprehensive and Challenging Science Benchmark for Multimodal Large Language Models (AAAI 2026 · similarity 0.86)
- Data Mixture Inference Attack: BPE Tokenizers Reveal Training Data Compositions (NeurIPS 2024 · similarity 0.86)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.86)
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.86)
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.
