Research map / Large language models / Instruction and fine-tuning
DARE-bench: Evaluating Modeling and Instruction Fidelity of LLMs in Data Science
Fan Shu, Yite Wang, Ruofan Wu, Boyi Liu, Zhewei Yao, Yuxiong He, Feng Yan
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 MM data to IF data in SFT-DV training (ablation) | Qwen3-32B SFT-DV · DARE-bench class-MM | score: 22.38 -> 43.82 | Improved |
| add IF data to MM data in SFT-DV training (ablation) | Qwen3-32B SFT-DV · DARE-bench class-IF | score: 11.71 -> 38.58 | Improved |
| train on MM data only instead of baseline (ablation) | Qwen3-32B SFT-DV · DARE-bench class-IF | score: 17.11 -> 11.71 | Hurt |
| train on IF data only instead of baseline (ablation) | Qwen3-32B SFT-DV · DARE-bench class-MM | score: 30.71 -> 22.38 | Hurt |
| SFT with all-valid rejection sampling | Qwen3-32B · DARE-bench Total | total score: 23.25 -> 42.91 | Improved |
| SFT with Duo-Valid rejection sampling | Qwen3-32B · DSBench | accuracy: 32.38 -> 42.41 | Improved |
| use inclusive sampling (DV) instead of strict filtering (FV) (ablation) | Qwen3-32B SFT · DSBench | accuracy: 37.82 -> 42.41 | Improved |
| use Fastest-Valid instead of All-Valid sampling (ablation) | Qwen3-32B SFT · DARE-bench Total | total score: 42.91 -> 42.42 | Mixed |
| GRPO reinforcement learning | Qwen3-4B · DARE-bench Total | total score: 4.39 -> 37.40 | Improved |
| GRPO reinforcement learning | Qwen3-4B · DSBench | accuracy: 18.23 -> 40.00 | Improved |
| increase interactive turns from 3 to 5 (time 300 s to 200 s) (ablation) | o4-mini · DARE-bench class-IF | score: 37.16 -> 67.56 | Improved |
| increase interactive turns from 5 to 15 (time 200 s to 100 s) (ablation) | o4-mini · DARE-bench class-IF | score: 67.56 -> 76.80 | Improved |
Closest papers by meaning
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.89)
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.88)
- Automating Dataset Updates Towards Reliable and Timely Evaluation of Large Language Models (NeurIPS 2024 · similarity 0.88)
- Prescriptive Scaling Reveals the Evolution of Language Model Capabilities (ICML 2026 · similarity 0.88)
- ResearchCodeBench: Benchmarking LLMs on Implementing Novel Machine Learning Research Code (NeurIPS 2025 · similarity 0.88)
- Active Evaluation Acquisition for Efficient LLM Benchmarking (ICML 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.
