AutoAssumption: Searching over Assumptions Discovers ML-Systems Algorithms
Summary
- Researchers introduce AutoAssumption, an automated discovery harness that searches over explicitly stated algorithm assumptions written as proofs.
- By formalizing 89 ML-systems algorithms across KV-cache eviction, weight pruning, and speculative decoding into proof graphs checked in Lean, the framework systematically relaxes assumptions.
- The harness automatically re-derives, re-proves, and generates code for modified algorithms, producing new state-of-the-art methods in KV-cache eviction and weight pruning.
Why it matters
Shifting automated algorithm design from program-level search to explicit assumption-level searching improves reliability and enables the discovery of state-of-the-art ML systems techniques.