Research map / Large language models / Agents, code and math
InteractScience: Programmatic and Visually-Grounded Evaluation of Interactive Scientific Demonstration Code Generation
Qiaosheng Chen, Yang Liu, Lei Li, Kai Chen, Qipeng Guo, Gong Cheng, Fei Yuan
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 reference snapshots as input (ablation) | GPT-5 · InteractScience PFT | Overall pass rate (%): 39.47 -> 42.53 | Improved |
| add reference snapshots as input (ablation) | GPT-5 · InteractScience VQT | CLIP: 71.95 -> 73.51 | Improved |
| add reference snapshots as input (ablation) | Qwen2.5-VL-72B-Instruct · InteractScience | PFT and VQT | Hurt |
| add reference snapshot to VLM judge input (ablation) | VLM-Judge (Igen+L) · Human alignment | Spearman correlation: 0.8224 -> 0.8827 | Improved |
| add checklist to VLM judge input (ablation) | VLM-Judge (Igen+Iref) · Human alignment | Spearman correlation: 0.3837 -> 0.8827 | Improved |
| add checklist to VLM judge with generated image only (ablation) | VLM-Judge (Igen) · Human alignment | Spearman correlation: 0.7360 -> 0.8224 | Improved |
| add reference snapshot to generated image only judge (ablation) | VLM-Judge (Igen) · Human alignment | Spearman correlation: 0.7360 -> 0.3837 | Hurt |
| give judge rendered screenshot instead of code only (ablation) | VLM-Judge · Human alignment | Spearman correlation: 0.1408 -> 0.7360 | Improved |
| swap judge from Gemini-2.5-Pro to GPT-5 / Kimi-K2.5 (ablation) | VQT ranking · InteractScience | Spearman rho between rankings: 0.937, 0.919, 0.956 | No clear effect |
| scale Qwen3 from 8B to 235B | Qwen3 · InteractScience PFT | Overall pass rate (%): 20.00 -> 33.33 | Improved |
| use closed-source instead of open-source models | LLMs · InteractScience PFT | Overall pass rate (%): 41.47 -> 33.87 | Improved |
| PFT ability vs visual fidelity | per-problem scores · InteractScience | Spearman PFT vs CLIP / VLM-Judge: 0.023 -> 0.312 range | No clear effect |
Closest papers by meaning
- SciCode: A Research Coding Benchmark Curated by Scientists (NeurIPS 2024 · similarity 0.87)
- Unveiling Fine-Grained Visual Traces: Evaluating MultiModal Interleaved Reasoning Chains in Multimodal STEM Tasks (NeurIPS 2026 · similarity 0.85)
- PhysGym: Benchmarking LLMs in Interactive Physics Discovery with Controlled Priors (NeurIPS 2025 · similarity 0.84)
- ScienceBoard: Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows (ICLR 2026 · similarity 0.84)
- From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking (ICLR 2026 · similarity 0.84)
- SciReason: A Controllable Benchmark for Scientific Reasoning in LLMs (NeurIPS 2026 · similarity 0.84)
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.
