F
Oct 21, 2026Fix, Not Just Resample
The problem: proof search wastes its best feedback Language models generate Lean proofs by...
Oct 21, 20266 min read0 reactions0 comments
Tag archive
The problem: proof search wastes its best feedback Language models generate Lean proofs by...
Discovery Loop used Claude Fable 5.1 to revise a solver and produced ten circle-packing candidates accepted by Packomania in an eight-hour, $27.72 consumer-PC run.
A new EMNLP demonstration system uses teacher-generated examples to train a compact task-specific adapter from a natural-language specification, reporting 83.6% semantic accuracy on a difficult subset where a fast compiler achieved 22.4% mean LEM.
What Changed The landscape of AI-driven software engineering has long been dominated by...
ReaComp compiles LLM reasoning traces into symbolic solvers that run at zero inference cost—84.7% on PBEBench-Hard with 78% fewer tokens.