Research map / Label-efficient and robust learning / Recognition and long-tail
Behavioral Embeddings of Programs: A Quasi-Dynamic Approach for Optimization Prediction
Haolin Pan, Dong Jinyuan, Hongbin Zhang, Hongyu Lin, Mingjie Xing, Yanjun Wu
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 scale-invariant logarithmic ratio quantification (vs absolute differences) (ablation) | Behavioral-PQ · -Oz Benefit Prediction (test) | MAE (%): 10.96 -> 8.19 | Improved |
| add scale-invariant logarithmic ratio quantification (vs absolute differences) (ablation) | Behavioral-PQ · Best Pass Prediction (test) | Top-5 accuracy (%): 94.33 -> 89.55 | Hurt |
| add Transformer encoder (vs pooling sub-embeddings) (ablation) | Behavioral-PQ · -Oz Benefit Prediction (test) | MAE (%): 10.08 -> 8.19 | Improved |
| add Transformer encoder (vs pooling sub-embeddings) (ablation) | Behavioral-PQ · Best Pass Prediction (test) | Top-5 accuracy (%): 87.46 -> 89.55 | Improved |
| use Product Quantization instead of K-Means on full vectors (ablation) | Behavioral-PQ · -Oz Benefit Prediction (test) | MAE (%): 8.24 -> 8.19 | Improved |
| use Product Quantization instead of K-Means on full vectors (ablation) | Behavioral-PQ · Best Pass Prediction (test) | Top-5 accuracy (%): 93.43 -> 89.55 | Hurt |
| use Product Quantization instead of K-Means on full vectors (ablation) | Behavioral-PQ · Best Pass Prediction (validation) | Top-5 accuracy (%): 98.48 -> 99.08 | Improved |
| use behavioral embedding instead of static embeddings | embedding · Best Pass Prediction (test) | Top-1 accuracy (%): 39.27 -> 64.48 | Improved |
| use behavioral embedding instead of static embeddings | embedding · -Oz Benefit Prediction (test) | MAE (%): 16.23 -> 8.19 | Improved |
| use behavioral embedding instead of InstCount for K-NN | embedding · Best Pass K-NN | Top-1 accuracy (%): 75.82 -> 79.70 | Improved |
| use specialized Behavioral-PQ instead of zero-shot LLMs | predictor · Best Pass Prediction (test) | Top-5 accuracy (%): 59.10 -> 89.55 | Improved |
| use zero-shot LLM instead of Autophase features | predictor · Best Pass Prediction (test) | Top-1 accuracy (%): 38.51 -> 35.52 | Hurt |
Closest papers by meaning
- CODE REPRESENTATION LEARNING AT SCALE (ICLR 2024 · similarity 0.86)
- Can Large Language Models Understand Intermediate Representations in Compilers? (ICML 2025 · similarity 0.85)
- Emergent Representations of Program Semantics in Language Models Trained on Programs (ICML 2024 · similarity 0.85)
- Towards A Generalist Code Embedding Model Based On Massive Data Synthesis (NeurIPS 2025 · similarity 0.84)
- REASONING COMPILER: LLM-Guided Optimizations for Efficient Model Serving (NeurIPS 2025 · similarity 0.84)
- MM-OptBench: A Solver-Grounded Benchmark for Multimodal Optimization Modeling (NeurIPS 2026 · similarity 0.84)
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.
