Research map / Large language models / Agents, code and math
ECCO: Evidence-Driven Causal Reasoning for Compiler Optimization
Haolin Pan, Lianghong Huang, Dong Jinyuan, 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 chain-of-thought rationale training (vs w/o CoT) (ablation) | ECCO-1.5B (w/o GA) · Average of 7 suites, greedy N=1 | cycle reduction vs -O3 (%): 1.27 -> 4.38 | Improved |
| add chain-of-thought rationale training (vs w/o CoT) (ablation) | ECCO-1.5B (w/o GA) · Average of 7 suites, Best-of-32 | cycle reduction vs -O3 (%): 16.25 -> 19.02 | Improved |
| add execution evidence to the CoT data (vs w/o Evidence) (ablation) | ECCO-1.5B (w/o GA) · Average of 7 suites, greedy N=1 | cycle reduction vs -O3 (%): 3.52 -> 4.38 | Improved |
| add execution evidence to the CoT data (vs w/o Evidence) (ablation) | ECCO-1.5B (w/o GA) · Average of 7 suites, Best-of-32 | cycle reduction vs -O3 (%): 17.55 -> 19.02 | Improved |
| scale model from 1.5B to 7B (ablation) | ECCO (w/o GA) · Average of 7 suites, Best-of-32 | cycle reduction vs -O3 (%): 19.02 -> 19.43 | No clear effect |
| scale model from 3B to 7B (ablation) | ECCO (w/o GA) · Average of 7 suites, greedy N=1 | cycle reduction vs -O3 (%): 7.03 -> 5.04 | Hurt |
| add GA tactician guided by LLM intent to the LLM policy (ablation) | ECCO-7B · Average of 7 suites, Best-of-32 | cycle reduction vs -O3 (%): 19.43 -> 24.44 | Improved |
| replace GA with iterative LLM refiner (ablation) | ECCO · Average of 7 suites | cycle reduction vs -O3 (%): 24.44 -> 19.36 | Hurt |
| use ECCO (LLM strategist + GA) instead of CFAST | best traditional baseline · Average of 7 suites | cycle reduction vs -O3 (%): 23.70 -> 24.44 | Improved |
| use direct LLM prompting instead of TPE | Direct LLM prompting Best-of-32 · Average of 7 suites | cycle reduction vs -O3 (%): 19.88 -> 16.06 | Hurt |
Closest papers by meaning
- REASONING COMPILER: LLM-Guided Optimizations for Efficient Model Serving (NeurIPS 2025 · similarity 0.87)
- Chain of Execution Supervision Promotes General Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.85)
- From I/O to Code with Discovery Agent (NeurIPS 2026 · similarity 0.85)
- Compiler-R1: Towards Agentic Compiler Auto-tuning with Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Eco Search: A No-delay Best-First Search Algorithm for Program Synthesis (AAAI 2025 · similarity 0.84)
- EquivaMap: Leveraging LLMs for Automatic Equivalence Checking of Optimization Formulations (ICML 2025 · 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.
