Reductive Analysis with Compiler-Guided Large Language Models for Input-Centric Code(…)
summary
Uses a compiler-guided LLM pipeline that applies reductive analysis over the call graph, one function at a time, with prompt templates and a map-reduce token strategy to find seminal behaviors and map them to inputs; it auto-generates an XGBoost extractor snippet to extract features at runtime. Claims a 44× speedup with remote LLMs (9.6 hours → 13 minutes) and 450× with local LLMs (→77 seconds), 92.6% accuracy for adaptive OpenMP decisions, and 20–30% serverless speedup while cutting resource use 50–60%.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:27Input-adaptive optimization
- 1:54Input characterization challenges
- 3:22Limitations of prior approaches
- 4:47LLMs and compilers for input features
- 6:36RCL code analyzer overview
- 7:39Call graph analysis
- 9:51Handling recurrence
- 11:57Extraction models for input features
- 13:01Experimental results
- 15:11Conclusion and future work
- 16:00Q&A
related talks
17mRLM: Recursive Language Models for Large Codebases
19mAutomated Program Refinement: Guide and Verify Code Large Language Model with(…)
40mSurviving the Agentic Hype with Small Language Models
15mInterleaving Large Language Models for Compiler Testing
33mDesigning LLM Applications for Massive Datasets
44mAccelerating Software Development: The LLM (R)evolution
36mProduction-Ready AI Agents: From LLMs to Small Language Models
38mThe Art of Compression: Crafting Insightful Summaries with LLMs