Automated Theorem Proving Confronts the Bitter Lesson
agents reasoning
| Source: HF Papers | Original article
Automated theorem proving faces a bitter lesson as formal proof work leans increasingly on human‑crafted workflows—planning, decomposition and specialized orchestration—rather than raw model scaling.
A new study shows that a minimalist agentic loop, dubbed **h1**, can push formal theorem‑proving performance to the current state‑of‑the‑art using only Lean 4 feedback and raw test‑time compute. Unlike recent advances in informal mathematics, which have leaned on ever larger language models and sophisticated chain‑of‑thought prompting, the h1 system discards the elaborate, human‑crafted pipelines that have become standard in formal verification. Instead, it relies on a simple execution interface that repeatedly queries a language model, receives Lean code, and feeds the result back for correction. The authors report that this stripped‑down approach matches or exceeds the results of more complex orchestration frameworks that combine planning, decomposition and specialized provers.
The finding matters because it revives the “bitter lesson” of AI research: general methods that scale with computational power eventually outstrip domain‑specific engineering. By showing that a straightforward agentic harness can dominate a field that has traditionally required hand‑tuned workflows, the work suggests that future gains in formal mathematics may come more from scaling compute and improving model‑environment interfaces than from adding layers of bespoke tooling.
Looking ahead, the community will watch whether h1’s approach scales with larger models or transfers to other proof assistants such as Coq or Isabelle. Researchers are also likely to explore tighter integration of the loop with emerging “self‑evolution” techniques we covered earlier, for example the skill‑adaptation mechanisms described on Oct 27, and the self‑improving agents highlighted on Oct 15. If the trend holds, we may see a shift toward leaner, more autonomous AI systems that handle both informal and formal mathematics with comparable ease.
Sources
Back to AIPULSEN