The Gap Between User Intent and Code Implementation
A new research series on the intent that code leaves implicit: how LLMs infer missing assumptions, what program analysis can check, and when human clarification is needed.
Code does not always express everything its author intends. The same unsigned arithmetic, for example, can be a deliberate wrapping counter or a bug in an allocation size. Checking the implementation alone cannot settle which behavior was intended.
LLMs try to recover this missing intent from names, documentation, and context. This series asks how to make the assumptions they introduce explicit, check them against the program for contradictions, and ask people for the information that remains missing. Finding no contradiction does not establish that an inferred intent is correct.
Talks
LLM Agents as Untrusted Program Components
LMPL 2026 · SPLASH/ISSTA 2026 · October 9, 2026
This talk introduces the direction through array bounds and unsigned wraparound, distinguishing claims that can be checked mechanically from assumptions that require wider analysis or an independent statement of intent.