AI agents are increasingly capable of discovering new mathematical results, yet their reasoning and generated proofs can be difficult to understand and verify. We introduce an interpretable framework for AI-guided mathematical discovery, focused on improving quantitative bounds on analytic constants. Our key insight is that existing proofs often contain slack in intermediate lemmas, which can be tightened without reconstructing the entire argument. To exploit this structure, we represent proofs as partially-formalized Lean blueprints: human-interpretable dependency graphs of definitions, lemmas, and theorems. An AI agent then iteratively refines the Lean code while preserving the blueprint's dependency structure, allowing local improvements to automatically propagate to a verified final bound. This interpretable structure allows humans to inspect the agent's refinements and understand how local improvements propagate to the final theorem. For Grothendieck's constant (KG), we obtain KG≤2log(1+2)π−10−17, an explicit numerical improvement over Krivine's 1977 bound. For Korenblum's constant (KK), we obtain KK≥0.356565, improving Wang's 2025 bound of 0.3554. Our experiments demonstrate the potential of both partial formalization and iterative agentic refinement for interpretable mathematical discovery.
Figures & tables
Figure 1: AI-guided proof refinement. A human-seeded proof blueprint exposes editable lemmas and numerical certificates. An AI agent proposes local edits, the verifier checks Lean obligations and interval arithmetic certificates, and the result is fed back as the optimization signal.
Table 2
Figure 2: Archimedes’ method for calculating π using inscribed polygons. He used a 96 -gon, by which point the polygon is indistinguishable from the circle.
Figure 3: The proof dependency graph remains invariant while the quantitative estimates improve. Two intermediate inequalities determine the constant κ , which in turn determines the final bound on π .
Figure 4: A proof dependency graph and its partial Lean formalization. The graph exposes dependencies that propagate to the final bound, and the Lean formalization contains background axioms, numerical axioms certified by Arb, definitions, and theorems.
Figure 5: Magnitude of bound improvements over refinement iterations, with red entries denoting the discovery of “interesting” strategies, as judged by domain experts.
Figure 6: The high-level strategies for bounding Grothendieck’s constant KG in Braverman et al. [9] (left) and Korenblum’s constant KK in Wang [47] (right). In each case, the final constant bound is dependent on a number of prior steps with analytical slack, which our framework tightens.
Grothendieck ( Cr↓ )
Korenblum ( c↑ )
Configuration
Ablated
Cr
Reduction
c
Gain
Checkpoint (start)
–
1
–
0.35646
–
Full framework
–
2.5×10−6
400,000×
0.35658
+1.2×10−4
Basic agent
agent, partly
4.0×10−3
250×
0.35648
+2.0×10−5
No agent
agent
7.1×10−2
14×
–
–
No partial formalization
blueprint
–
–
–
–
Table 3: Ablations at a matched budget, one run per configuration, against the full framework (shaded). Arrows give the direction of improvement. Cr enters the KG multiplicatively, so we report its ratio to the checkpoint. A dash in a result column means no accepted candidate improved on the checkpoint.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: The baseline (V0) proof graph for the polygon- π warm-up, read left to right. Grey nodes carry no constant: the two on the left are the trusted surface, and the dashed ones are parts of the graph this candidate has not opened. Amber nodes are editable and each emits one constant; the blue node combines them. With no upper bound on sin , the lower side of the bracket is closed — λ=0 reduces to the trusted Pn<π — so the score is κ alone.
Figure 8: The same proof graph at the final candidate (V5), with both sides of the bracket open. All four editable nodes now carry constants, and the two aggregators combine them into the scored pair. The amber edge is the one piece of shared structure: Bu bounds π3 in κ and π2 in λ , so tightening the upper bound on π lowers κ and raises λ at the same time.
Candidate
κ
λ
Score
Accepted (UTC)
Compile (s)
Type check (s)
V0
1
0
1
19:51
2.5
3.2
V1
2/3
0
2/3
20:03
2.5
3.3
V2
2/3
1/4
5/12
20:29
2.6
3.2
V3
343/768
1/4
151/768
20:52
2.6
3.3
V4
343/768
9/32
127/768
21:14
2.6
3.2
V5
128/375
9/32
721/12000
21:32
2.7
3.3
Appendix
Table 4: The run from Refinement_Log/log.jsonl . The comparison run with Claude Opus 5 as planner and generator reached 9816533/109 in the same number of sprints.
Figure 9: Lean Blueprint for Grothendieck’s constant
Full
Basic agent
No agent
No partial form.
Partial formalization (Lean chain)
yes
as text
as text
no
Numerical library, registered oracles
yes
text; edits allowed
text; edits allowed
no
Source paper
yes
as text
as text
yes
Planner / generator / evaluator agents
yes
no
no
yes
Tools (shell, file edits, Lean LSP, web)
yes
no
no
yes
Feedback from build and evaluator
yes
verbatim, last 3 steps
none
yes
Appendix
Table 5: Inputs and capabilities by configuration.
Case study
Configuration
Wall (s)
Output tokens
Calls
Evaluator runs
Cost (USD)
Grothendieck
Full framework
8,652
544,294
52
9
35.59
Basic agent
11,379
567,988
7
6
54.70
No agent
9,042
781,503
6
5
32.19
No partial formalization
10,093
738,266
80
7
44.06
Korenblum
Full framework
3,285
230,815
14
2
20.65
Basic agent (in budget)
2,721
253,391
2
1
23.44
Appendix
Table 6: Compute per configuration, one run each. Grothendieck rows are the complete runs, gated at the matched window. For the Korenblum basic agent the first row is consumption through the last iteration that began within the token budget, and the second is the whole run; the Korenblum no-partial-formalization run had three times the reference budget. That row sums wall clock, tokens, and cost over its two runs but counts calls only for the second. The first run was cut short by an API credit stall, and retries from that stall dominate its turn log.
Grothendieck (Lemma 5.4)
Korenblum
Numbered sprints
34
102
Sprints graded by the evaluator
29
86
accepted
27
85
rejected
2
1
Sprints ending without a candidate
4
16
Distinct accepted candidates
27 (V1, V3 to V28)
69 (V0 to V68)
Appendix
Table 7: Accounting of the two main runs from their retained records. For Grothendieck the wall clock covers sprints 3 to 31, run back to back on one day; sprints 33 and 34 were run on 2026-05-03 after the truncation audit of Appendix F . The Korenblum evaluator was also run by the generator while it worked, so its 213 logged runs exceed its 86 verdicts. The Korenblum Lean time is the sum of the logged compile, type-check, and axiom-scan stages (68, 49, and 56 min); oracle recomputation time was not logged. Twenty-eight Grothendieck candidate files exist but only 27 were accepted: V2 was rejected in sprint 4 and never resubmitted, whereas V12 was rejected in sprint 16 and accepted in sprint 17 (Table 9 ).
Run
Iterations
Accepted
Rejected
Best certified Cr
Orders below start
Hours
1
5
4
1
1.60⋅1010
4.7
1.7
2
6
5
1
1.54⋅107
7.7
1.9
3
8
6
2
7.16⋅106
8.1
3.3
4
8
6
2
1.71⋅109
5.7
3.9
5
7
7
0
2.76⋅104
10.5
3.8
Appendix
Table 8: Five independent restarts of the Lemma 5.4 loop from the seed Cr≈8.20⋅1014 . “Orders below start” is log10 of the ratio between the starting and the best certified Cr .
Sprint
Outcome
Candidate
Cr
Contract (abbreviated)
1
no verdict
–
–
harness shakedown, no candidate
2
no verdict
–
–
contract refiled as sprint 3
3
PASS
V1
9.56⋅1012
Switch to the tight Rouché-certified rectangle (contract 2)
4
FAIL
V2
–
score expression not evaluable
5
PASS
V3
1.45⋅109
Define Cr_val as the rational upper bound 1451520000 (…
6
PASS
V4
1.23⋅109
Tighten the rational R upper bound from 4/3 to 6/5 (…
Appendix
Table 9: The Lemma 5.4 run, one row per numbered sprint. Contract titles are truncated; the full contracts and verdicts are in the repository under ProofRefinement/Lemma54/ .
Refinement
Origin
Exact inverse-function identities
sprints 14 to 17 (V10 to V12), inside the loop
Joint suprema instead of products
sprints 19 to 21 (V14 to V16), inside the loop
Merged-numerator interference
sprints 29 and 30 (V24, V25), inside the loop
Truncation-free evaluation, Gaussian moments, affine denominator bound, Taylor pads with corner subdivision, contour symmetry
the audit and patch of 2026-05-03, carried out outside the sprint loop with a human in the loop (Section 4.4 ); then re-baselined by sprints 33 and 34
Analyticity-domain lift
sprint 13, blocked; filed as a proposal, not approved
Korenblum per-mode kernel and window partition
sprints 84 to 88 (V64 to V66), inside the loop
Appendix
Table 10: Origin of each refinement. “Inside the loop” means the planner proposed it and the generator implemented it with no human edit; the sprint contract and verdict are the record.
AI agents are increasingly used in mathematics research, but it is often unclear how to use them effectively. Towards this, we present an extensive case study of how AI was used to improve bounds on the Grothendieck constant KG, which captures the hardness between combinatorial problems and their continuous relaxations. Specifically, while the precise value of KG is not known, we recently tightened the best known bounds to
116π≤KG≤2log(1+2)π−10−4.
Crucially, these improvements were achieved using an AI research system that could arrive at insights deemed novel by domain experts. We give a detailed discussion of our experience using AI for mathematics research, particularly touching upon its strengths and weaknesses, as well as our experience with creating ideal conditions for AI to arrive at breakthrough insights.
Building on recent progress in machine learning for formal theorem proving in Lean, we introduce LeanDojo-v2: an end-to-end framework that unifies repository data extraction, LLM fine-tuning, proof search, AI autocompletion, and LLM-assisted theorem proving in a single software library. In addition, we provide Lean4Code, a lightweight IDE that integrates the various features of LeanDojo-v2 through a simple, mathematician-friendly UI. These tools will enable mathematicians and researchers to easily fine-tune language models on domain-specific Lean codebases and experiment with LLM-assisted theorem proving without a complex setup.
Ryan Hsiang, William Adkisson, Robert Joseph George +1
National Taiwan University · Washington University in St. Louis · California Institute of Technology
Large language models (LLMs) increasingly excel at mathematical reasoning, but their unreliability limits their utility in mathematics research. A mitigation is using LLMs to generate formal proofs in languages like Lean. We perform the first large-scale evaluation of this method's ability to solve open problems. Our most capable agent autonomously resolved 9 of 353 open Erdős problems at the per-problem cost of a few hundred dollars, proved 44/492 OEIS conjectures, and is being deployed in combinatorics, optimization, graph theory, algebraic geometry, and quantum optics research. A basic agent alternating LLM-based generation with Lean-based verification replicated the Erdős successes but proved costlier on the hardest problems. These findings demonstrate the power of AI-aided formal proof search and shed light on the agent designs that enable it.
George Tsoukalas, Anton Kovsharov, Sergey Shirobokov +18