Research map / Large language models / Agents, code and math
Jupiter: Enhancing LLM Data Analysis Capabilities via Notebook and Inference-Time Value-Guided Search
Shuocheng Li, Yihao Liu, Silin Du, Wenxuan Zeng, Zhe Xu, Mengyu Zhou, Yeye He, Haoyu Dong, Shi Han, Dongmei Zhang
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 |
|---|---|---|---|
| finetune on NbQA | Mistral-7B-Instruct-v0.3 · InfiAgent-DABench | accuracy by questions (%): 2.33 -> 59.14 | Improved |
| finetune on NbQA | Llama-3.1-8B-Instruct · InfiAgent-DABench | accuracy by questions (%): 48.25 -> 69.65 | Improved |
| finetune on NbQA | Qwen2.5-7B-Instruct · InfiAgent-DABench | accuracy by questions (%): 43.97 -> 68.09 | Improved |
| finetune on NbQA | Qwen2.5-14B-Instruct · InfiAgent-DABench | accuracy by questions (%): 69.65 -> 77.04 | Improved |
| add value model to Jupiter search (exploration term removed) (ablation) | SFT Qwen2.5-7B-Instruct · InfiAgent-DABench | accuracy by questions (%): 70.04 -> 77.82 | Improved |
| add value model to Jupiter search (exploration term removed) (ablation) | SFT Qwen2.5-14B-Instruct · InfiAgent-DABench | accuracy by questions (%): 79.38 -> 86.38 | Improved |
| remove the PUCT exploration term when using value model (ablation) | SFT Qwen2.5-14B-Instruct · InfiAgent-DABench | accuracy by questions (%): 74.71 -> 86.38 | Improved |
| remove the PUCT exploration term when using value model (ablation) | SFT Qwen2.5-7B-Instruct · InfiAgent-DABench | accuracy by questions (%): 68.87 -> 77.82 | Improved |
| add exploration term without value model (ablation) | SFT Qwen2.5-14B-Instruct · InfiAgent-DABench | accuracy by questions (%): 79.38 -> 75.88 | Hurt |
| add exploration term without value model (ablation) | SFT Qwen2.5-7B-Instruct · InfiAgent-DABench | accuracy by questions (%): 70.04 -> 68.87 | Hurt |
| use Jupiter value-guided search instead of majority voting | SFT Qwen2.5-14B-Instruct · InfiAgent-DABench | accuracy by questions (%): 83.66 -> 86.38 | Improved |
| use value-model-assisted search instead of ReAct | vanilla Qwen2.5-14B-Instruct · DSBench data modeling | task completion rate (%): 66.22% -> 98.65% | Improved |
Closest papers by meaning
- Automated Creation of Reusable and Diverse Toolsets for Enhancing LLM Reasoning (AAAI 2025 · similarity 0.88)
- DiscoveryBench: Towards Data-Driven Discovery with Large Language Models (ICLR 2025 · similarity 0.88)
- DS-Agent: Automated Data Science by Empowering Large Language Models with Case-Based Reasoning (ICML 2024 · similarity 0.88)
- Fleet of Agents: Coordinated Problem Solving with Large Language Models (ICML 2025 · similarity 0.87)
- Token-Supervised Value Models for Enhancing Mathematical Problem-Solving Capabilities of Large Language Models (ICLR 2025 · similarity 0.87)
- TACT: Advancing Complex Aggregative Reasoning with Information Extraction Tools (NeurIPS 2024 · similarity 0.87)
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.
