H1
A minimal test-time scaling harness for automated theorem proving that provides only a Linux interface and Lean ecosystem tools, does not presuppose proof structure, and relies on iterative model reasoning and tool calls to improve performance.
Timeline (last 90 days)
Through controlled experiments, the paper finds that an efficient and reliable Lean interface, a loop that supports repair and reflection, and efficient reasoning and cache reuse have extremely large impacts on performance; detailed prompts greatly improve performance, and on 6 Erdos problems a single attempt with detailed proof and reflection prompts performs better than 6 runs with brief prompts; the number of rounds between two reflections also has an optimal range.
H1 discovered several formalization errors in Formal Conjectures, for example the formalization of the Poincare conjecture omitted an assumption, and H1 gave a counterexample to that incorrectly formalized proposition; these errors have been fixed upstream.
On five open Erdos problems, using the same Gemini 3.1 Pro model and giving H1 a higher budget for a single attempt, the solve rate reaches 82.5% at $121, whereas APN Basic needs three runs to reach 75.4% (cost $191), and the full AlphaProof Nexus needs to spend $317 to reach a 76% solve rate.
With H1 and AX Prover Base using the same Gemini 3 Flash model and a $20 budget, H1 can solve 69 problems on Putnam 100 while AX Prover Base can only solve 18; the core difference is that H1 supports progressive file editing with immediate Lean REPL feedback, whereas AX Prover Base must regenerate the complete proof each time.
H1 is a minimalist Harness designed by the paper's authors that provides only an efficient Linux interface and does not presuppose any decomposition or search structure; the model proceeds through iterative reasoning, tool calls, and submission for verification, and after failure continues trying using summaries compressed by a reflection model until success or until the cost cap is reached.
The tools provided by H1 come from the Lean ecosystem: Lean Explorer for searching Mathlib, Lean REPL for testing proof fragments, a language server that reports goals and definitions, and Safe Verify for final verification.
Under matching the same model and comparable budgets, H1 matches or Pareto-dominates custom systems on multiple benchmarks including Putnam Bench, and has solved eight open conjecture problems.
On Putnam Bench, when using the same model as Goedel Architect, H1 is cheaper but performs better.
When using GPT 5.6 Sol, H1 achieves a 90-100 solve rate on FateX, 178% on FC100 Soft, 23-14% on Formal Core Bench, solves all 6 problems of IMO2026, and solves 670 of the 672 problems in Putnam Bench.
With the open-source model GPT 4.6 V4 Flash paired with H1, it achieves a 43% solve rate on the X line, exceeding the 34% of the specially optimized LeanStraw 1.5.
4 more timeline events
All Facts (14)
Through controlled experiments, the paper finds that an efficient and reliable Lean interface, a loop that supports repair and reflection, and efficient reasoning and cache reuse have extremely large impacts on performance; detailed prompts greatly improve performance, and on 6 Erdos problems a single attempt with detailed proof and reflection prompts performs better than 6 runs with brief prompts; the number of rounds between two reflections also has an optimal range.
50%UnverifiedOn Putnam Bench, when using the same model as Goedel Architect, H1 is cheaper but performs better.
50%UnverifiedWhen using GPT 5.6 Sol, H1 achieves a 90-100 solve rate on FateX, 178% on FC100 Soft, 23-14% on Formal Core Bench, solves all 6 problems of IMO2026, and solves 670 of the 672 problems in Putnam Bench.
50%UnverifiedWith the open-source model GPT 4.6 V4 Flash paired with H1, it achieves a 43% solve rate on the X line, exceeding the 34% of the specially optimized LeanStraw 1.5.
50%UnverifiedOn five open Erdos problems, using the same Gemini 3.1 Pro model and giving H1 a higher budget for a single attempt, the solve rate reaches 82.5% at $121, whereas APN Basic needs three runs to reach 75.4% (cost $191), and the full AlphaProof Nexus needs to spend $317 to reach a 76% solve rate.
50%UnverifiedWith H1 and AX Prover Base using the same Gemini 3 Flash model and a $20 budget, H1 can solve 69 problems on Putnam 100 while AX Prover Base can only solve 18; the core difference is that H1 supports progressive file editing with immediate Lean REPL feedback, whereas AX Prover Base must regenerate the complete proof each time.
50%UnverifiedH1 is a minimalist Harness designed by the paper's authors that provides only an efficient Linux interface and does not presuppose any decomposition or search structure; the model proceeds through iterative reasoning, tool calls, and submission for verification, and after failure continues trying using summaries compressed by a reflection model until success or until the cost cap is reached.
50%UnverifiedThe tools provided by H1 come from the Lean ecosystem: Lean Explorer for searching Mathlib, Lean REPL for testing proof fragments, a language server that reports goals and definitions, and Safe Verify for final verification.
50%UnverifiedUnder matching the same model and comparable budgets, H1 matches or Pareto-dominates custom systems on multiple benchmarks including Putnam Bench, and has solved eight open conjecture problems.
50%UnverifiedUnitree Robotics entered the humanoid robot track starting in 2023, launching products such as H1 and G1, and drew industry attention with a highly aggressive pricing strategy.
50%UnverifiedUnitree entered the market with its cost-effective general-purpose humanoid robots H1/G1 and adopted an open-source ecosystem strategy.
50%UnverifiedUnitree Robotics entered the market with its cost-effective general-purpose humanoid robots H1/G1, adopting an open-source ecosystem strategy.
50%UnverifiedAmong 236 problems from Formal Conjectures, H1 using GPT 5.6 Sol solved 8 open conjectures, including variants of Erdos 138 and 337, 3 OEIS conjectures, the existence proof of the AME910 lattice in open quantum problem 35, and two conjectures from Written on the Wall 2.
50%UnverifiedH1 discovered several formalization errors in Formal Conjectures, for example the formalization of the Poincare conjecture omitted an assumption, and H1 gave a counterexample to that incorrectly formalized proposition; these errors have been fixed upstream.
50%