Research map / Large language models / Agents, code and math
SciNav: A Principled Agent Framework for Scientific Coding Tasks
TIANSHU ZHANG, Huan Sun
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 |
|---|---|---|---|
| use SciNav TKCTS instead of Self-Debug | GPT-4o (2024-11-20) · ScienceAgentBench | SR: 15.0 -> 18.6 | Improved |
| use SciNav instead of OpenHands | GPT-4o (2024-05-13) · ScienceAgentBench | SR: 13.1 -> 16.1 | Improved |
| use SciNav instead of Self-Debug | Claude-3.7 · ScienceAgentBench | SR: 22.5 -> 25.5 | Improved |
| use SciNav instead of Self-Debug | Claude-3.7 · ScienceAgentBench | VER: 84.3 -> 72.5 | Hurt |
| use SciNav instead of Self-Debug | GPT-4o (2024-05-13) · ScienceAgentBench | VER: 71.2 -> 66.0 | Hurt |
| use relative judgments instead of random selection (ablation) | SciNav · ScienceAgentBench | SR: 15.2 -> 18.6 | Improved |
| use relative judgments instead of LLM absolute scoring (ablation) | SciNav · ScienceAgentBench | SR: 16.2 -> 18.6 | Improved |
| increase initial solutions from 1 to 5 (ablation) | SciNav · ScienceAgentBench (40 solved tasks) | success rate: 40.5 -> 45.2 | Improved |
| increase initial solutions from 1 to 2 (ablation) | SciNav · ScienceAgentBench (40 solved tasks) | success rate: 40.5 -> 40.5 | No clear effect |
| add self-improvement (ablation) | SciNav · ScienceAgentBench (40 solved tasks) | success rate: 45.2 -> 57.1 | Improved |
| use K=2 instead of larger K in top-K (ablation) | SciNav · DA-Code (30 task preliminary) | average score | Improved |
| use Claude-3.7 or mixture for solution generation (ablation) | SciNav · ScienceAgentBench | SR: 18.6 -> 19.6 | No clear effect |
Closest papers by meaning
- SciCode: A Research Coding Benchmark Curated by Scientists (NeurIPS 2024 · similarity 0.92)
- From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking (ICLR 2026 · similarity 0.91)
- ScienceBoard: Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows (ICLR 2026 · similarity 0.89)
- AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite (ICLR 2026 · similarity 0.89)
- EvoCodeBench: Evaluating Coding Agents in Multi-Turn Iterative Interactions (NeurIPS 2026 · similarity 0.89)
- InnovatorBench: Evaluating Agents’ Ability to Conduct Innovative AI Research (ICLR 2026 · similarity 0.89)
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.
