Research map / Large language models / Agents, code and math
MALICE: Memory-aware Loop Invariants Generation on Symbolic Execution Traces
Tong Chen, Siyu Liu, Hongyi Zhong, Liao Zhang, Lixiang Wang, Xiwei Wu, Junchi Yan, Qinxiang Cao
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 the full Prediction and Refinement pipeline over a vanilla prompt (ablation) | MALICE DeepSeek-V3.2 · LIG-MM+ (pass@8) | overall pass rate: 15.77% -> 82.69% | Improved |
| add the Prediction stage (ablation) | MALICE DeepSeek-V3.2 · LIG-MM+ (pass@8) | overall pass rate: 66.15% -> 82.31% | Improved |
| add the Refinement stage (ablation) | MALICE DeepSeek-V3.2 · LIG-MM+ (pass@8) | overall pass rate: 34.62% -> 82.31% | Improved |
| add refinement versus 150 samples without it (ablation) | MALICE DeepSeek-V3.2 · LIG-MM+ (pass@8) | overall pass rate: 47.31% -> 82.69% | Improved |
| add symbolic traces and detailed QCP feedback over binary feedback (ablation) | MALICE refinement · LIG-MM+ (pass@8) | overall pass rate: 36.15% -> 82.31% | Improved |
| add few-shot examples (ablation) | MALICE DeepSeek-V3.2 · LIG-MM+ (pass@8) | overall pass rate: 41.15% -> 82.31% | Improved |
| replace Qwen3-Max with DeepSeek-V3.2 backbone (ablation) | MALICE · LIG-MM+ (pass@8) | overall pass rate: 75.00% -> 82.69% | Improved |
| use reasoning model DeepSeek-V3.2-Thinking instead of DeepSeek-V3.2 (ablation) | MALICE · LIG-MM+ (pass@8) | overall pass rate: 82.69% -> 82.69% | No clear effect |
| replace ICL refinement with SFT-trained Llama-3.1-8B refiner (ablation) | MALICE · LIG-MM+ (pass@8) | overall pass rate: 82.69% -> 86.15% | Improved |
| use all-8B pipeline (Qwen3-8B prediction, SFT refinement) instead of DeepSeek-V3.2 (ablation) | MALICE · LIG-MM+ (pass@8) | overall pass rate: 82.69% -> 81.54% | No clear effect |
| limit to 8 total LLM calls (3 prediction + 5 refinement, pass@1) (ablation) | MALICE · LIG-MM+ (pass@8) | overall pass rate: 82.69% -> 59.62% | Hurt |
| use MALICE instead of LIG-SE | LIG-SE · LIG-MM+ (pass@8) | overall pass rate: 41.92% -> 82.69% | Improved |
Closest papers by meaning
- Towards General Loop Invariant Generation: A Benchmark of Programs with Memory Manipulation (NeurIPS 2024 · similarity 0.91)
- CLEVER: A Curated Benchmark for Formally Verified Code Generation (NeurIPS 2025 · similarity 0.84)
- Chain of Execution Supervision Promotes General Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.84)
- Lemur: Integrating Large Language Models in Automated Program Verification (ICLR 2024 · similarity 0.83)
- StepCodeReasoner: Aligning Code Reasoning with Stepwise Execution Traces via Reinforcement Learning (ICML 2026 · similarity 0.83)
- Evaluating Program Semantics Reasoning with Type Inference in System $F$ (NeurIPS 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.
