Research map / Large language models / Agents, code and math
CodeSense: a Real-World Benchmark and Dataset for Code Semantic Reasoning
Monoshi Kumar Roy, Simin Chen, Benjamin Steenhoek, Jinjun Peng, Gail Kaiser, Baishakhi Ray, Wei Le
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 |
|---|---|---|---|
| increase block size from 1 to 3 statements (ablation) | LLMs on output prediction · CodeSense Python/C blocks | accuracy | Hurt |
| predict input from output instead of output from input (ablation) | LLMs on function-level prediction · CodeSense functions | accuracy | Hurt |
| add API definitions to the prompt for API-call statements (ablation) | LLM prompting · CodeSense statement prediction | accuracy | No clear effect |
| add more few-shot examples to the prompt (ablation) | LLM prompting · CodeSense statement and loop prediction | accuracy | Improved |
| use shots relevant to the query (same statement type or same function) instead of random shots (ablation) | LLM prompting · CodeSense statement and loop prediction | accuracy | Improved |
| add simple chain-of-thought prompt (think step by step) (ablation) | LLM prompting · CodeSense statement prediction | accuracy | No clear effect |
| add simple chain-of-thought prompt to loop property prediction (ablation) | LLM prompting · CodeSense post-loop prediction | accuracy | Mixed |
| ask for abstract values with only a definition in a zero-shot prompt (ablation) | LLM prompting · CodeSense approximation tasks | accuracy | No clear effect |
| add 3-shot examples with abstract outputs (ablation) | LLM prompting · CodeSense approximation tasks | accuracy | Improved |
| evaluate C instead of Java or Python for output prediction (ablation) | LLMs · CodeSense input/output prediction | accuracy | Hurt |
| use reasoning models instead of non-reasoning models on statement prediction (ablation) | LLMs · CodeSense statement-level | accuracy | No clear effect |
Closest papers by meaning
- CoRe: Benchmarking LLMs’ Code Reasoning Capabilities through Static Analysis Tasks (NeurIPS 2025 · similarity 0.92)
- CodeMMLU: A Multi-Task Benchmark for Assessing Code Understanding & Reasoning Capabilities of CodeLLMs (ICLR 2025 · similarity 0.91)
- SemCoder: Training Code Language Models with Comprehensive Semantics Reasoning (NeurIPS 2024 · similarity 0.90)
- CodeIO: Condensing Reasoning Patterns via Code Input-Output Prediction (ICML 2025 · similarity 0.89)
- Evaluating Program Semantics Reasoning with Type Inference in System $F$ (NeurIPS 2025 · similarity 0.89)
- DyCodeEval: Dynamic Benchmarking of Reasoning Capabilities in Code Large Language Models Under Data Contamination (ICML 2025 · 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.
