Show HN: TheoremSmith: Mass-Producing Hard Proof Problems to Train LLMs
Making models better at formal verification will help push the boundaries of math research and significantly reduce the number of bugs in our code. Recently, training models on RL environments (tasks paired with verifiers that reward correct solutions) has driven big gains in model performance, and proofs are unusually easy to verify. Here we introduce TheoremSmith, a pipeline for generating synthetic formal verification environments. It reliably creates multi-hour, frontier-difficulty tasks and can scale throughput effectively.<p>I'd appreciate any feedback, especially on ways models could still cheat the verifier, and on whether synthetic tasks like these are likely to help on real proofs.<p>Code: <a href="https://github.com/abundant-ai/theoremsmith" rel="nofollow">https://github.com/abundant-ai/theoremsmith</a>
竞品与替代品
Show HN: Zero-axiom Lean 4 proof that Yang-Mills on R^4 is a "round square"
On October 26, OpenAI announced a claimed "bare existence" proof in Balaban's RG programme on pure R^4—but notably stop…
Show HN: Fakemetrics.com – "build in public" like a superstar
Hi HN, I am Tigran (github.com/tigranbs) I made Fakemetrics.com as a fun experiment while testing how well Claude Opus…
Show HN: I finetuned 1.5B Qwen to near GPT-4o level bash generation perf
Purely a hobby side project to see how far I can push a really small model, using (mostly) automated training pipelines…
Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights)
My brother and I run a 2-person lab that trains our own generative video models from scratch. The ultimate goal is a se…
Show HN: Vim-like, hyper-efficient, LLM power tool – 20-80k tokens not 350k+
I built a focused TUI tool for efficient parallelized tasks, early independent testers expressed 90%+ API savings and m…