Formalization of Memory, Floating-Point, and Richer Loop Operations

Extend the Trivet formalization and validation workflows to LLVM memory operations, floating-point arithmetic, and richer classes of loop-containing transformations.

Background

The current Trivet formalization targets integer instructions and supports loop-free, single-block transformations together with a restricted class of loop-containing transformations. Consequently, important parts of LLVM semantics and transformation behavior—memory operations, floating-point arithmetic, and more general loop classes—are outside the implemented scope.

The authors explicitly identify extending the formalization to these features as future work. Addressing this problem would increase the range of real-world LLVM transformations for which Trivet can produce machine-checked refinement proofs or counterexample-based refutations.

References

Extending the formalization to memory operations, floating-point arithmetic, and richer loop classes is left as future work.

— LLVM Translation Validation Automated with Large Language Models and Lean  (2609.19583 - Liao et al., 17 Sep 2026) in Section 3.4, “Implementation and Limitations” (label: subsec:implementation)

Native fp16 serving followed by a canonical replay would keep the integer runtime off the serving path; a replay proof still refers to the replayed relation, and acceptance thresholds would need matched benign and adversarial evaluation. We leave that mode as an open design question.

— Seal, Then Sample: Sampled Layerwise Proofs for Verifiable LLM Inference from GPT-2 to 70B  (2609.27367 - Lim et al., 23 Sep 2026) in Section 8.2, “Approximate relations and model authenticity”