Research map / Large language models / Agents, code and math
Gistify: Codebase-Level Understanding via Runtime Execution
Hyunji Lee, Minseon Kim, Chinmay Singh, Matheus Pereira, Atharv Sonwane, Isadora White, Elias Stengel-Eskin, Mohit Bansal, Zhengyan Shi, Alessandro Sordoni, Marc-Alexandre Côté, Xingdi Yuan and 1 more
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 |
|---|---|---|---|
| prompt the model to trace execution (ablation) | SWE-agent Claude-4 · Gistify (50 pylint tests, SWE-agent Claude-4) | execution fidelity: 42.0 -> 48.0 | Improved |
| prompt with explicit reading strategy (ablation) | SWE-agent Claude-4 · Gistify (50 pylint tests, SWE-agent Claude-4) | execution fidelity: 42.0 -> 50.0 | Improved |
| add RepoGraph tool (ablation) | SWE-agent Claude-4 · Gistify (50 pylint tests, SWE-agent Claude-4) | execution fidelity: 42.0 -> 52.0 | Improved |
| add gold execution tracing tool (ablation) | SWE-agent Claude-4 · Gistify (50 pylint tests, SWE-agent Claude-4) | execution fidelity: 42.0 -> 56.0 | Improved |
| give full bash execution access (ablation) | SWE-agent Claude-4 · Gistify (50 pylint tests, SWE-agent Claude-4) | execution fidelity: 42.0 -> 52.0 | Improved |
| restrict execution to edit-and-execute instead of full bash (ablation) | SWE-agent Claude-4 · Gistify (50 pylint tests, SWE-agent Claude-4) | execution fidelity: 52.0 -> 56.0 | Improved |
| provide gold test function in prompt (ablation) | SWE-agent Claude-4 · Gistify (pylint) | execution fidelity: 42.0 -> 60.0 | Improved |
| enable execution tools in main setting (ablation) | 3 frameworks x 4 models · Gistify | execution fidelity: e.g. mini-SWE-agent GPT-5-mini 17.1 -> 24.0; Copilot GPT-5-mini 58.0 -> 55.3 | Mixed |
| use agentic multi-step model instead of static LLM with gold files (ablation) | Claude-4 · Gistify subset | execution fidelity | Improved |
| select hardest tests by trace length and files (Gistify-hard) | Claude-4 · Gistify-hard | execution fidelity: 43 -> 21 | Hurt |
| static LLM given gold files vs agentic | static coding LLM · Gistify | line existence | Improved |
| move from bash-only mini-SWE-agent to Copilot toolset | GPT-5-mini · Gistify | execution fidelity: 17.1 -> 58.0 | Improved |
Closest papers by meaning
- GitTaskBench: A Benchmark for Code Agents Solving Real-World Tasks Through Code Repository Leveraging (AAAI 2026 · similarity 0.87)
- RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades (NeurIPS 2026 · similarity 0.85)
- SemCoder: Training Code Language Models with Comprehensive Semantics Reasoning (NeurIPS 2024 · similarity 0.85)
- R2E: Turning any Github Repository into a Programming Agent Environment (ICML 2024 · similarity 0.85)
- BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions (ICLR 2025 · similarity 0.85)
- NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents (ICML 2026 · similarity 0.85)
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.
