Research map / Large language models / Agents, code and math
ABSINT-AI: Agentic Heap Abstractions for Abstract Interpretation
Michael Wang, Kexin Pei, Armando Solar-Lezama
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 LM-guided per-allocation-site abstraction selection (vs fixed symbolic cascade) (ablation) | ABSINT-AI · 53-benchmark JS suite | # false positives: 820 -> 618.7 (GPT-4.1-mini) | Improved |
| make the LM agentic (interpreter queries, extra loop iterations) rather than single-shot (ablation) | ABSINT-AI · 53-benchmark JS suite, GPT-4.1-mini | mean false positives: 645.7 -> 618.7 | Improved |
| make the LM agentic rather than single-shot (ablation) | ABSINT-AI · 53-benchmark JS suite, GPT-4.1-mini | intersection false positives: 572 -> 555 | Improved |
| add field_value widening strategy (ablation) | ABSINT-AI · 53-benchmark JS suite | mean false positives: 642.5 -> 618.7 (+23.8 when removed) | Improved |
| add recency merging strategy (ablation) | ABSINT-AI · 53-benchmark JS suite | mean false positives: 637.6 -> 618.7 (+18.9 when removed) | Improved |
| add role merging and depth widening strategies (ablation) | ABSINT-AI · 53-benchmark JS suite | mean false positives: role +11.3, depth +9.1 when removed | Improved |
| intersect warnings across multiple runs | ABSINT-AI · 53-benchmark JS suite, GPT-4.1-mini | # false positives: 618.7 -> 555; full intersection 499 | Improved |
| use a newer GPT-5 family model instead of older models | ABSINT-AI · 53-benchmark JS suite | mean false positives: 616.9 (Qwen3-32B) to 640.3 (GPT-5.4-mini) | No clear effect |
| replace TAJS/WALA with ABSINT-AI | static analysis for null/undefined access · 53-benchmark JS suite | # false positives: TAJS 711 -> 618.7; WALA 1327 -> 618.7 | Improved |
| replace TAJS/WALA with ABSINT-AI | run time · 53-benchmark JS suite | run time: about 20 seconds -> about 500 seconds | Hurt |
Closest papers by meaning
- CoRe: Benchmarking LLMs’ Code Reasoning Capabilities through Static Analysis Tasks (NeurIPS 2025 · similarity 0.85)
- Lemur: Integrating Large Language Models in Automated Program Verification (ICLR 2024 · similarity 0.85)
- Evaluating Program Semantics Reasoning with Type Inference in System $F$ (NeurIPS 2025 · similarity 0.85)
- ReGAL: Refactoring Programs to Discover Generalizable Abstractions (ICML 2024 · similarity 0.84)
- Benchmarking Abstract and Reasoning Abilities Through A Theoretical Perspective (ICML 2025 · similarity 0.84)
- The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models (ICML 2025 · similarity 0.83)
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.
