Research map / Large language models / Agents, code and math
Not All Invariants Are Equal: Curating Training Data to Accelerate Program Verification with SLMs
Ido Pinto, Yizhak Yisrael Elboher, Haoze Wu, Nina Narodytska, Guy Katz
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 LLM-driven simplification and augmentation (V2) on top of AST normalization (ablation) | Qwen3-4B · InvBench Hard (n=123) | R_correct (%): 33.1 -> 44.4 | Improved |
| add LLM-driven simplification and augmentation (V2) on top of AST normalization (ablation) | Qwen3-8B · InvBench Hard (n=123) | R_correct (%): 30.1 -> 42.8 | Improved |
| add LLM-driven simplification and augmentation (V2) on top of AST normalization (ablation) | Llama-3.1-8B · InvBench Hard (n=123) | R_correct (%): 36.3 -> 45.5 | Improved |
| add LLM-driven simplification and augmentation (V2) on top of AST normalization (ablation) | Mistral-7B · InvBench Hard (n=123) | R_correct (%): 24.9 -> 31.4 | Improved |
| add AST normalization (V1) to raw verifier invariants (ablation) | Qwen3-4B · InvBench Hard (n=123) | R_valid (%): 81.3 -> 97.6 | Improved |
| fine-tune on raw UAutomizer invariants (V0) instead of using the base model (ablation) | Qwen3-4B · InvBench Hard (n=123) | R_valid (%): 99.2 -> 81.3 | Hurt |
| fine-tune on raw UAutomizer invariants (V0) instead of using the base model (ablation) | Mistral-7B · InvBench Hard (n=123) | R_valid (%): 93.8% -> 65.0% | Hurt |
| fine-tune on raw UAutomizer invariants (V0) instead of using the base model (ablation) | Qwen3-0.6B · InvBench Hard (n=123) | R_correct (%): 28.5 -> 18.7 | Hurt |
| fine-tune on raw UAutomizer invariants (V0) instead of using the base model (ablation) | Llama-3.1-8B · InvBench Hard (n=123) | R_correct (%): 14.9 -> 31.2 | Improved |
| fine-tune on Wonda-curated data (V2) instead of the base model | Qwen3-4B · InvBench Hard (n=123) | R_speedup (%): 11.1 -> 24.7 | Improved |
| fine-tune on Wonda-curated data (V2) instead of the base model | Qwen3-0.6B · InvBench Hard (n=123) | R_correct (%): 28.5 -> 27.9 | No clear effect |
| fine-tune on Wonda-curated data (V2) instead of the base model | Qwen3-0.6B · InvBench Hard (n=123) | VBP_E2E (s): 183.0 -> 174.1 | Improved |
Closest papers by meaning
- Towards General Loop Invariant Generation: A Benchmark of Programs with Memory Manipulation (NeurIPS 2024 · similarity 0.90)
- Unsupervised Evaluation of Code LLMs with Round-Trip Correctness (ICML 2024 · similarity 0.88)
- T1: Tool-integrated Verification for Test-time Compute Scaling in Small Language Models (ICLR 2026 · similarity 0.88)
- AlphaVerus: Bootstrapping Formally Verified Code Generation through Self-Improving Translation and Treefinement (ICML 2025 · similarity 0.87)
- rStar-Coder: Scaling Competitive Code Reasoning with a Large-Scale Verified Dataset (NeurIPS 2025 · similarity 0.87)
- Lemur: Integrating Large Language Models in Automated Program Verification (ICLR 2024 · similarity 0.86)
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.
