Hallucination-Resilient LLM-Driven Sound and Tunable Static Analysis
summary
Claims a sound, LLM-driven static analyzer called LLMAAM that delegates abstract address allocation to an LLM while guaranteeing soundness by construction and adaptively tuning precision. Implements a Scala prototype using an LChan-for-Java/K framework and evaluates on microbenchmarks with GPT-4o: three runs per benchmark, LLMAAM yields best precision on 3 of 4 cases and close on the fourth. Allocation calls to the LM are currently expensive; authors report promising preliminary results and outline scalability and prompting optimizations.
AI-written summary · about
chapters (7)
- 0:00Static analysis introduction
- 3:25LLM limits for static analysis
- 4:30Meta-analysis with language models
- 6:14Control flow case study
- 9:42Implementation and evaluation
- 11:30Conclusion and future work
- 12:24Q&A
related talks
38mMaking LLMs reliable - A practical framework
53mSemantics-based program manipulation in the age of LLMs
01h 16mHow Will LLMs Change Haskell, Programming and Open Source?
39mLLMs and Hallucinations: Understand, Measure, Control
30mEfficient programming with Scala and LLMs
26mReducing Hallucinations and Evaluating LLMs for Production
49mA Practical Approach to Using LLMs: Peculiarities and Pitfalls
49mPushing LLMs over the Edge: Exploring the Limits of the Possible