Pith. sign in

REVIEW 2 major objections 1 minor 1 cited by

MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling

T0 review · 2 major / 1 minor · reviewed 2026-06-27 · grok-4.3

Pith's one-line read MaxProof lets one model exceed human gold-medal scores on IMO 2025 and USAMO 2026 by running tournament selection over populations of candidate proofs at test time.

desk verdict The abstract claims 35/42 on IMO 2025 and 36/42 on USAMO 2026 via population test-time scaling, but the verifier's accuracy on those exact proofs is asserted without shown external checks. read the letter →

arxiv 2606.13473 v1 pith:YX5EWQMS submitted 2026-06-11 cs.LG cs.AIcs.CL

classification cs.LGcs.AIcs.CL
keywords mathematicalprooftest-timescalinggenerativeverifiertournamentselectionIMOUSAMOrepairpopulationsearch
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper presents MaxProof as a test-time framework that merges proof generation, verification, and repair into a single model, then searches a population of proofs and selects the best one via tournament. The central claim is that this population-level scaling, powered by a low false-positive generative verifier, pushes performance past the human gold-medal threshold without further training. A sympathetic reader would care because the approach shows how test-time compute can be applied to formal reasoning tasks that normally require deep search or external solvers. The results are reported on recent competition problems, suggesting the method scales with the number of candidates considered.

What carries the argument

Population-level test-time scaling that treats the model as a combined generator-verifier-refiner-ranker and applies tournament selection over candidate proofs.

What would settle it

Running the same population search on a fresh set of competition problems where the verifier's accuracy is measured independently and the final selected proofs are graded by human experts; if the selected proofs are mostly incorrect despite high reported scores, the claim fails.

Watch

Extended reading notes

Core claim

MaxProof trains the M3 model on three capabilities using a defense-in-depth generative verifier, then at test time deploys the model simultaneously as generator, verifier, refiner, and ranker; it samples populations of candidate proofs and returns a final proof through tournament selection, reaching 35 out of 42 on IMO 2025 and 36 out of 42 on USAMO 2026.

Load-bearing premise

The generative verifier can reliably separate correct proofs from incorrect ones with a low false-positive rate so that tournament selection actually picks the right answer.

Editorial extensions

If this is right

  • A single trained model can reach competition-level performance on proof tasks by allocating more test-time compute to population search rather than larger training runs.
  • Verification and repair capabilities become directly usable at inference time to refine and rank outputs.
  • Tournament selection among many candidates becomes an effective way to leverage the verifier for final answer choice.
  • Performance scales with the size and diversity of the proof population generated at test time.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same population-search pattern could be tested on other verifiable domains such as code synthesis or formal theorem proving in different libraries.
  • If verifier accuracy improves further, the same framework might push scores higher without changing the base model.
  • The approach highlights that verification quality, not just generation quality, is the bottleneck for test-time scaling in reasoning.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 1 minor

Summary. The paper introduces MaxProof, a population-level test-time scaling framework for competition-level mathematical proof. It first trains the M3 model on three capabilities—proof generation, proof verification, and critique-conditioned proof repair—using a defense-in-depth generative verifier engineered for low false-positive rate. These are merged into a single model. At test time, MaxProof uses the model as generator, verifier, refiner, and ranker, searches over populations of candidate proofs, and selects the final proof via tournament selection, reporting 35/42 on IMO 2025 and 36/42 on USAMO 2026.

Significance. If the central claims hold after proper validation, the work would represent a notable advance in scaling AI for formal mathematical reasoning, demonstrating that integrated generative verification plus population-level selection can exceed human gold-medal thresholds on recent contest problems. The RL training of verifier and repair capabilities alongside test-time tournament selection is a coherent technical direction with potential broader impact.

major comments (2)
  1. [Abstract] Abstract and experimental results: The reported scores (35/42 IMO 2025, 36/42 USAMO 2026) are obtained by tournament selection over populations using the same generative verifier trained in the RL stage; however, the manuscript provides no independent human audit, cross-check against official solutions, or false-positive rate measurement on the 71 selected proofs, leaving the low-FP claim unverified and the performance numbers potentially inflated.
  2. [Abstract] Abstract: No experimental setup, baselines, ablation studies, or error analysis are described, so it is impossible to assess whether the method (rather than unstated factors) supports the headline claims of surpassing human performance.
minor comments (1)
  1. [Abstract] The abstract could more explicitly state the total number of problems attempted and any filtering criteria applied before reporting the 35/42 and 36/42 figures.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below.

read point-by-point responses
  1. Referee: [Abstract] Abstract and experimental results: The reported scores (35/42 IMO 2025, 36/42 USAMO 2026) are obtained by tournament selection over populations using the same generative verifier trained in the RL stage; however, the manuscript provides no independent human audit, cross-check against official solutions, or false-positive rate measurement on the 71 selected proofs, leaving the low-FP claim unverified and the performance numbers potentially inflated.

    Authors: We agree that the manuscript does not contain an independent human audit, cross-check against official solutions, or explicit false-positive rate measurement performed on the 71 final selected proofs. The low false-positive engineering is described for the verifier training stage and the tournament selection is presented as an additional safeguard, but these do not substitute for post-hoc verification of the contest outputs. We will add a limitations paragraph acknowledging this gap and will include a small-scale human verification study of a random subset of the proofs in the revised manuscript. revision: partial

  2. Referee: [Abstract] Abstract: No experimental setup, baselines, ablation studies, or error analysis are described, so it is impossible to assess whether the method (rather than unstated factors) supports the headline claims of surpassing human performance.

    Authors: The abstract is intentionally concise. The full manuscript describes the RL training of the three capabilities (generation, verification, critique-conditioned repair), the defense-in-depth verifier, the merging into a single model, and the population-level tournament selection procedure at test time. We nevertheless accept that the current version lacks explicit baselines, ablations, and error analysis. We will expand the experimental section in the revision to include these elements. revision: yes

standing simulated objections not resolved
  • Independent human audit and false-positive measurement on the complete set of 71 selected proofs, as this verification step was not performed during the original experiments and cannot be supplied without new work.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity detected

full rationale

The paper describes an empirical pipeline: training M3 on proof generation/verification/repair using a defense-in-depth generative verifier, followed by test-time population search and tournament selection with the same model components to produce final proofs. Reported scores (35/42 IMO 2025, 36/42 USAMO 2026) are presented as direct outcomes of applying this framework to external competition problems. No equations, self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear in the text that would make the performance metrics equivalent to the inputs by construction. The verifier's low false-positive engineering is stated as an assumption enabling the method, not a tautological reduction. The derivation chain is therefore self-contained against the benchmarks.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Since only the abstract is available, no specific free parameters, axioms, or invented entities can be identified from the provided information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling." pith.science (2026). https://pith.science/paper/YX5EWQMS

@misc{pith2026260613473,
  author       = {Pith},
  title        = {Pith review of: MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YX5EWQMS}},
  note         = {Machine review of arXiv:2606.13473}
}
read the original abstract

We present MaxProof, a population-level test-time scaling framework for competition-level mathematical proof in the MiniMax-M3 series. M3 first trains three proof-oriented capabilities -- proof generation, proof verification, and critique-conditioned proof repair -- using a defense-in-depth generative verifier engineered for low false-positive rate. These capabilities are merged into a single released M3 model. At test time, MaxProof treats the model as a generator, verifier, refiner, and ranker, searches over a population of candidate proofs, and returns one final proof through tournament selection. With MaxProof test-time scaling, the M3 model reaches 35/42 on IMO 2025 and 36/42 on USAMO 2026, exceeding the human gold-medal threshold on both.

Figures

Figures reproduced from arXiv: 2606.13473 by the authors.

Figure 1
Figure 1. The MaxProof pipeline. M3 first trains three proof-oriented capabilities—proof generation through verifier-guided proof RL, proof verification through aligned error finding, and critique-conditioned proof repair through refinement augmentation. These capabilities are merged into the M3 release model, which MaxProof scales at test time through population search and tournament selection. © 2026 MiniMax. All rights res… view at source ↗
Figure 2
Figure 2. The training dynamics of proof expert. 2.1. Training Pipeline Overview The Proof Expert is trained around a single central object: a frozen generative verifier that turns a candidate proof into an RL reward. For each competition problem, the rollout policy samples a group of long-form candidate proofs. Each proof is then passed to the verifier, which does not merely check for a final answer; it reads the argument, c… view at source ↗
Figure 3
Figure 3. The verifier pipeline as four defensive layers. The first two layers remove format-driven failure modes; the last two produce a conservative scalar reward. The design is intentionally conservative. A false positive can become a training target that the policy learns to reproduce, while a false negative usually only discards one candidate among many. The pipeline therefore spends its complexity on suppressing high-sc… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The four-layer defense-in-depth verifier. Left: bad-case filtering removes candidates that match well-known failure patterns. Middle: solution normalization reduces the verifier’s sensitivity to surface format. Right: three judges score in parallel; scores are reduced …
Figure 5
Figure 5. Figure 5: shows the resulting training-set distribution after these filters and balancing steps. Algebra 24.7% Combinatorics 30.5% Geometry 27.9% Number Theory 17.0% (a) Domain Distribution L1 (Basic) 10.8% L2 (Intermediate) 50.2% L3 (Advanced) 31.8% L4 (Olympiad) 7.1% (b) Diffi…
Figure 6
Figure 6. Figure 6: Reward-hacking detection dashboard for the M2-cycle Proof RL run. Top left: false positive rate during training. Top right: visible (blue) and thinking (gray) proof length. Bottom left: structural-template rate (step headers, verification section, final-answer block). …
Figure 7
Figure 7. Figure 7: End-to-end MaxProof loop. A population of 𝑁 candidates is initialized, scored, and summarized. Each round selects 𝑀 diverse parents, applies dual PATCH/REWRITE refinement, evaluates offspring, and re-injects them into the archive. The final answer is selected by a pair…
Figure 8
Figure 8. Figure 8: Per-round oracle-best score across the 12 problems. Top: aggregate mean (black) and 25th–75th percentile band. Bottom: per-problem trajectories. Three problems never reach 7/7 within 𝑅 = 10 rounds: IMO P6, USAMO P2, and USAMO P3. 6.3.2. Discussion The MaxProof framewor…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Theoretical Foundations of $\max$@$k$ Reinforcement Learning

    cs.LG 2026-07 conditional novelty 7.0 of 10

    For max@k (best-of-K) finite-horizon MDPs, Markovian policies are suboptimal, a compact (previous-best, current-cumulative) state augmentation restores optimality, exact planning is NP-hard but an FPTAS exists, and th...

Reference graph

Works this paper leans on

43 extracted references · cited by 1 Pith paper

  1. [1]

    Constructions for𝑘=0,1,3 *𝑘=0.Take the𝑛vertical lines𝑥=1,2, . . . , 𝑛. Every(𝑎, 𝑏) ∈𝑆 𝑛 lies on𝑥=𝑎, so all points are covered. None of these lines is sunny (all are vertical). *𝑘=1.For𝑛≥3take the vertical lines𝑥=1,2, . . . , 𝑛−1and the sunny line through(𝑛,1)and(1,2). Its slope is 1−2 𝑛−1 =− 1 𝑛−1, which is never0,∞or−1; hence it is sunny. The vertical li...

  2. [2]

    Boundary points and a lemma The set𝑆 𝑛 is a right-angled isosceles triangle with vertices(1,1),(1, 𝑛),(𝑛,1). Its boundary consists of three edges: 34 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling 𝐸1 :𝑥=1,1≤𝑦≤𝑛, 𝐸2 :𝑦=1,1≤𝑥≤𝑛, 𝐸3 :𝑥+𝑦=𝑛+1,1≤𝑥, 𝑦. Each edge contains𝑛lattice points, and the three corn...

  3. [3]

    Base case𝑛=3.𝑆 3 ={(1,1),(1,2),(1,3),(2,1),(2,2),(3,1)}

    Inductive step We prove by induction on𝑛that the only possible values of𝑘are0,1,3. Base case𝑛=3.𝑆 3 ={(1,1),(1,2),(1,3),(2,1),(2,2),(3,1)}. The constructions above show that 𝑘=0,1,3are attainable. It remains to prove that𝑘=2is impossible. Assume, for contradiction, that there are three distinct lines covering𝑆3 with exactly two sunny lines. Let 𝐿be the un...

  4. [4]

    Therefore the set of admissible𝑘is {0,1,3}

    Conclusion We have exhibited explicit constructions for𝑘=0,1,3for every𝑛≥3, and we have shown that no other non-negative integer𝑘can be realized. Therefore the set of admissible𝑘is {0,1,3} . 36 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling IMO 2025 Problem 2 (Self-Pick: 7/7) Problem Statement LetΩa...

  5. [5]

    Place𝑀=(0,0), 𝑁=(1,0)and let the radii be𝑟 < 𝑅

    Coordinate set-upThe problem is invariant under scaling, translation and rotation. Place𝑀=(0,0), 𝑁=(1,0)and let the radii be𝑟 < 𝑅. Then Ω:𝑥 2 +𝑦 2 =𝑟 2,Γ:(𝑥−1) 2 +𝑦 2 =𝑅 2. The line𝑀 𝑁(the𝑥-axis) meetsΩat𝐶=(−𝑟,0)andΓat𝐷=(1+𝑅,0); the order𝐶, 𝑀, 𝑁, 𝐷is satisfied because−𝑟 <0<1<1+𝑅. The two intersection points of the circles are 𝐴=(𝑥 0, 𝑦0), 𝐵=(𝑥 0,−𝑦 0), 𝑥 ...

  6. [6]

    Write𝑃=(𝑥 𝑃, 𝑦𝑃)with𝑦 𝑃 unknown

    Circumcenter𝑃of△𝐴𝐶 𝐷 Since𝐶, 𝐷lie on the𝑥-axis, the perpendicular bisector of𝐶 𝐷is the vertical line through the midpoint of 𝐶 𝐷; hence 𝑥𝑃 = −𝑟+ (1+𝑅) 2 = 1+𝑅−𝑟 2 . Write𝑃=(𝑥 𝑃, 𝑦𝑃)with𝑦 𝑃 unknown. The condition𝑃 𝐴=𝑃𝐶(both equal to the circumradius) gives (𝑥 𝑃 −𝑥 0)2 + (𝑦 𝑃 −𝑦 0)2 =(𝑥 𝑃 +𝑟) 2 +𝑦 2 𝑃 . Expanding and using𝑥2 0 +𝑦 2 0 =𝑟 2 yields −2𝑥 𝑃 𝑥0 −2...

  7. [7]

    For a circle with centre𝑂and a point𝐴on it, the second intersection of the line𝐴+𝑡vwith the circle is obtained from|𝐴+𝑡v−𝑂| 2 =𝑅 2

    The points𝐸and𝐹 Letv=𝑃−𝐴and set𝐴𝑃=|v|. For a circle with centre𝑂and a point𝐴on it, the second intersection of the line𝐴+𝑡vwith the circle is obtained from|𝐴+𝑡v−𝑂| 2 =𝑅 2. Since|𝐴−𝑂| 2 =𝑅 2, we get 2𝑡v· (𝐴−𝑂) +𝑡 2|v| 2 =0=⇒𝑡=− 2v· (𝐴−𝑂) |v| 2 . Thus the signed distance from𝐴to that second point along the directionu=v/|v|is 𝑠=𝑡|v|=−2u· (𝐴−𝑂). 37 MaxProof: S...

  8. [8]

    Inthecoordinate system with origin at𝐴, axes alongu(the𝑢-axis) andw(the𝑤-axis), we have 𝐴=(0,0), 𝐸=(𝑒,0), 𝐹=(𝑓 ,0), 𝐵=(𝑏, ℎ), where 𝑏=(𝐵−𝐴) ·u, ℎ=(𝐵−𝐴) ·w

    Orthonormal basis aligned with𝐴𝑃 Letu=v/𝐴𝑃andletwbetheunitvectorobtainedbyrotatingu90 ◦ counter-clockwise. Inthecoordinate system with origin at𝐴, axes alongu(the𝑢-axis) andw(the𝑤-axis), we have 𝐴=(0,0), 𝐸=(𝑒,0), 𝐹=(𝑓 ,0), 𝐵=(𝑏, ℎ), where 𝑏=(𝐵−𝐴) ·u, ℎ=(𝐵−𝐴) ·w. First compute 𝑥𝑃 −𝑥 0 = (1+𝑅−𝑟) − (1+𝑟 2 −𝑅 2) 2 = 𝑌(1+𝑋) 2 = 𝑌𝑇 2 , 𝑦𝑃 −𝑦 0 =− 𝑇 𝑈 𝑋 4𝑦0 , wi...

Show all 43 references
  1. [9]

    The altitude from𝑀to𝑃𝑁has slope− 𝑥𝑃 −1 𝑦𝑃 and equation𝑦=− 𝑥𝑃 −1 𝑦𝑃 𝑥

    Orthocentre𝐻of△𝑃𝑀 𝑁 With𝑀=(0,0),𝑁=(1,0)and𝑃=(𝑥 𝑃, 𝑦𝑃), the altitude from𝑃to𝑀 𝑁is the vertical line𝑥=𝑥 𝑃. The altitude from𝑀to𝑃𝑁has slope− 𝑥𝑃 −1 𝑦𝑃 and equation𝑦=− 𝑥𝑃 −1 𝑦𝑃 𝑥. Intersecting with𝑥=𝑥 𝑃 gives 𝑦= 𝑥𝑃 (1−𝑥 𝑃) 𝑦𝑃 . Hence 𝐻= 𝑥𝑃, 𝑥𝑃 (1−𝑥 𝑃) 𝑦𝑃 . We need the component of𝐻...

  2. [10]

    Hence the circum- centre𝑂has coordinates(𝑆, 𝑘)for some𝑘

    Circumcircle of△𝐵𝐸𝐹 The points𝐸, 𝐹lie on the𝑢-axis, so the perpendicular bisector of𝐸𝐹is the line𝑢=𝑆. Hence the circum- centre𝑂has coordinates(𝑆, 𝑘)for some𝑘. Using|𝑂𝐸|=|𝑂𝐵|we obtain 𝑘= (𝑆−𝑏) 2 +ℎ 2 −𝐷 2 2ℎ . Substituting𝑏=𝑈𝑆,ℎ=−2𝑦 0 𝐷, and𝑆−𝑏=𝑆(1−𝑈)=𝑆𝑌 2 (because1−𝑈=𝑌 2) yiel...

  3. [11]

    Factorising the numerator as a difference of squares, 4𝑟2 − (1+𝑟 2 −𝑅 2)2 = 2𝑟− (1+𝑟 2 −𝑅 2) 2𝑟+ (1+𝑟 2 −𝑅 2)

    Using the coordinates of𝐴, 𝑦2 0 =𝑟 2 −𝑥 2 0 =𝑟 2 − (1+𝑟 2 −𝑅 2)2 4 . Factorising the numerator as a difference of squares, 4𝑟2 − (1+𝑟 2 −𝑅 2)2 = 2𝑟− (1+𝑟 2 −𝑅 2) 2𝑟+ (1+𝑟 2 −𝑅 2) . The first factor equals(𝑅+𝑟−1) (𝑅−𝑟+1)=(𝑋−1) (𝑌+1), the second equals−(𝑌−1)𝑇. Hence 4𝑟2 − (1+𝑟 2...

  4. [12]

    The distance from the centre𝑂=(𝑆, 𝑘)to this line is|𝑘−𝐻 𝑤 |

    Tangency condition The line through𝐻parallel to𝐴𝑃is, in the(𝑢, 𝑤)system, the horizontal line𝑤=𝐻 𝑤. The distance from the centre𝑂=(𝑆, 𝑘)to this line is|𝑘−𝐻 𝑤 |. The radius of the circle(𝐵𝐸𝐹)is|𝑂𝐸|= √ 𝐷2 +𝑘 2. Tangency is equivalent to |𝑘−𝐻 𝑤 |= √︁ 𝐷2 +𝑘 2 ⇐ ⇒𝐻 2 𝑤 −2𝑘𝐻 𝑤 =𝐷 2.(...

  5. [13]

    - Taking𝑎=𝑏=1gives𝑓(1) |1−𝑓(1) 𝑓(1)

    Basic properties Let𝑓:ℕ→ℕbe a bonza function. - Taking𝑎=𝑏=1gives𝑓(1) |1−𝑓(1) 𝑓(1) . Since𝑓(1) |𝑓(1) 𝑓(1) , we have𝑓(1) |1; hence𝑓(1)=1. - Setting𝑏=𝑎yields𝑓(𝑎) |𝑎 𝑎 −𝑓(𝑎) 𝑓(𝑎) . Because𝑓(𝑎) |𝑓(𝑎) 𝑓(𝑎) , it follows that 𝑓(𝑎) |𝑎 𝑎 for every𝑎∈ℕ.(1)

  6. [14]

    As𝑝 𝑝 is a prime power, 𝑓(𝑝)=𝑝 𝑘 with0≤𝑘≤𝑝

    Behaviour on primes For a prime𝑝, (1) implies𝑓(𝑝) |𝑝 𝑝. As𝑝 𝑝 is a prime power, 𝑓(𝑝)=𝑝 𝑘 with0≤𝑘≤𝑝. Define 𝑃={𝑝prime|𝑓(𝑝)>1}

  7. [15]

    Substituting𝑎=𝑝into the bonza condition gives 𝑝𝑘 |𝑏 𝑝 −𝑓(𝑏) 𝑝𝑘 for all𝑏∈ℕ

    Congruence modulo𝑝∈𝑃 Fix𝑝∈𝑃and write𝑓(𝑝)=𝑝 𝑘 with𝑘≥1. Substituting𝑎=𝑝into the bonza condition gives 𝑝𝑘 |𝑏 𝑝 −𝑓(𝑏) 𝑝𝑘 for all𝑏∈ℕ. Work modulo𝑝. By Fermat’s little theorem,𝑥𝑝 ≡𝑥(mod𝑝)for every integer𝑥. An easy induction shows 𝑥 𝑝𝑘 ≡𝑥(mod𝑝) (base𝑘=1is Fermat; if𝑥 𝑝𝑘 ≡𝑥(mod𝑝),the...

  8. [16]

    Therefore𝑓is the identity function

    Structure of𝑃 -If𝑃is infinite, then for each𝑏the number𝑓(𝑏) −𝑏is divisible by infinitely many distinct primes (the primes in𝑃), which forces𝑓(𝑏)=𝑏. Therefore𝑓is the identity function. -Now assume that𝑃is finite and non-empty. We first prove that𝑃cannot contain any odd prime. S...

  9. [17]

    Fix an arbitrary𝑛∈ℕ

    Case analysis Case A:𝑃=∅.Then𝑓(𝑝)=1for every prime𝑝. Fix an arbitrary𝑛∈ℕ. For any prime𝑏we have 𝑓(𝑛) |𝑏 𝑛 −1, because the bonza condition with𝑎=𝑛,𝑏=𝑏gives𝑓(𝑛) |𝑏 𝑛 −𝑓(𝑏) 𝑓(𝑛) =𝑏 𝑛 −1. Hence𝑓(𝑛)divides𝑏 𝑛 −1 foreveryprime𝑏. The greatest common divisor over all primes𝑏of𝑏 𝑛 −1is...

  10. [18]

    - If𝑛is even, write𝑛=2 𝑗𝑚with𝑗≥1,𝑚odd; then𝑓(𝑛) ≤2 𝑗+2 ≤4𝑛

    Upper bound𝑓(𝑛) ≤4𝑛 Collecting the three cases: - If𝑛is odd,𝑓(𝑛)=1≤4𝑛. - If𝑛is even, write𝑛=2 𝑗𝑚with𝑗≥1,𝑚odd; then𝑓(𝑛) ≤2 𝑗+2 ≤4𝑛. - The identity function also satisfies𝑓(𝑛)=𝑛≤4𝑛. Hencefor every bonza function𝑓and every𝑛∈ℕ, 𝑓(𝑛) ≤4𝑛.(4)

  11. [19]

    (Equivalently, for even𝑛≠2, write𝑛=2 𝑗𝑚with𝑗≥1,𝑚odd, and set𝑓(𝑛)=2 𝑗+2.) We verify that this𝑓is bonza

    Construction achieving the ratio4 Define𝑓by 𝑓(𝑛)=    1,if𝑛is odd, 4,if𝑛=2, 2 𝑗+2,if𝑛=2 𝑗𝑚with𝑗≥1, 𝑚odd, and𝑛≠2. (Equivalently, for even𝑛≠2, write𝑛=2 𝑗𝑚with𝑗≥1,𝑚odd, and set𝑓(𝑛)=2 𝑗+2.) We verify that this𝑓is bonza. Case 1:𝑎odd.Then𝑓(𝑎)=1; the condition holds trivial...

  12. [20]

    Therefore the smallest such real constant is 4

    Conclusion From (4) we have𝑓(𝑛) ≤4𝑛for every bonza function𝑓and every positive integer𝑛, and the construction above shows that4is the smallest possible constant. Therefore the smallest such real constant is 4 . 45 MaxProof: Scaling Mathematical Proof with Generative-Verifier R...

  13. [21]

    Thus 𝑒𝑛+1 =𝑒 𝑛 −2, 𝑓 𝑛+1 =𝑓 𝑛 −1, 𝑟 𝑛+1 =13𝑟 𝑛.(1)

    The case𝑒𝑛 ≥2.From (*) we have 𝑎𝑛+1 =𝑓(𝑎 𝑛)= 13 12 𝑎𝑛 =13·2 𝑒𝑛 −2 ·3 𝑓𝑛 −1 ·𝑟 𝑛. Thus 𝑒𝑛+1 =𝑒 𝑛 −2, 𝑓 𝑛+1 =𝑓 𝑛 −1, 𝑟 𝑛+1 =13𝑟 𝑛.(1)

  14. [22]

    47 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling The factor2disappears, and the remaining product is odd

    The case𝑒𝑛 =1,5|𝑟 𝑛.Then 𝑎𝑛+1 = 31 30 𝑎𝑛 =31·3 𝑓𝑛 −1 · 𝑟𝑛 5 . 47 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling The factor2disappears, and the remaining product is odd. Hence𝑎𝑛+1 is odd, so6∤𝑎 𝑛+1. By Lemma 1, the sequen...

  15. [23]

    Reaching a fixed point Suppose we start with a term𝑎 1 for which the sequence is infinite

    The case𝑒𝑛 =1,5∤𝑟 𝑛.Here𝑓(𝑎 𝑛)=𝑎 𝑛, i.e., the sequence has reached a fixed point. Reaching a fixed point Suppose we start with a term𝑎 1 for which the sequence is infinite. Then we must never enter the forbidden branch𝑒=1,5|𝑟. Consequently, after some number of steps we must a...

  16. [24]

    +𝜆.(2) Since𝜆 > √ 2 2 we have2𝜆− √ 2>0; thus the right-hand side of (2) tends to+∞as𝑘→ ∞. Consequently there exists an integer𝐾such that for all𝑘≥𝐾, 50 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling 𝐴2 2𝑘+1 >(2𝑘+1) +1=2𝑘...

  17. [25]

    Thus 𝑎≤𝜆 < √ 2 2 < √ 2, so𝑎 2 <2and Bazza’s move𝑏𝑘+1 = √ 2−𝑎 2 is well defined

    +𝜆.(4) Because𝜆 < √ 2 2 we have2𝜆− √ 2<0; hence the right-hand side of (4) is decreasing in𝑘and its maximum over𝑘≥0is𝜆. Thus 𝑎≤𝜆 < √ 2 2 < √ 2, so𝑎 2 <2and Bazza’s move𝑏𝑘+1 = √ 2−𝑎 2 is well defined. Then 𝑄2(𝑘+1) =𝑄 2𝑘 +𝑎 2 +𝑏 2 𝑘+1 =2𝑘+𝑎 2 + (2−𝑎 2)=2(𝑘+1). The increase in th...

  18. [26]

    Now consider Alice’s next turn, number2(𝑘+1) +1=2𝑘+3

    Consequently 𝑆2(𝑘+1) =𝑆 2𝑘 +Δ𝑆≥𝑘 √ 2+ √ 2=(𝑘+1) √ 2, completing the induction. Now consider Alice’s next turn, number2(𝑘+1) +1=2𝑘+3. At that moment the sum is𝑆2(𝑘+1) . Using the lower bound𝑆2(𝑘+1) ≥ (𝑘+1) √ 2, it suffices to show 51 MaxProof: Scaling Mathematical Proof with Ge...

  19. [27]

    Consequently Alice is always forced to choose𝑎≤ √ 2 2

    Thus (5) holds for all𝑘. Consequently Alice is always forced to choose𝑎≤ √ 2 2 . Therefore 𝑄2𝑘+1 =𝑄 2𝑘 +𝑎 2 ≤2𝑘+ 1 2 <2𝑘+2=(2𝑘+1) +1, so she never makes𝑄2𝑘+1 >2𝑘+2, which would be necessary to win. Bazza always has a legal move (because𝑎 2 ≤ 1 2 <2) and never loses. Hence Alic...

  20. [28]

    Conclusion Alice has a winning strategy if and only if𝜆 > √ 2 2 , Bazza has a winning strategy if and only if𝜆 < √ 2 2 , Neither player has a winning strategy if𝜆= √ 2 2 . 53 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scalin...

  21. [29]

    Let𝑁, the midpoint of the minor arc𝐵𝐶, be(−1,0); the tangent at𝑁is then the line𝑥=−1

    Coordinate setupPlace the circumcircle𝜔as the unit circle. Let𝑁, the midpoint of the minor arc𝐵𝐶, be(−1,0); the tangent at𝑁is then the line𝑥=−1. Because𝑁bisects the minor arc𝐵𝐶, the chord𝐵𝐶is vertical. Write 𝐵=(−cos𝜃,−sin𝜃), 𝐶=(−cos𝜃,sin𝜃),0< 𝜃 < 𝜋 2 . The vertex𝐴lies on the m...

  22. [30]

    * The equilateral triangleΔ𝐵 has its centre at the origin, so the midpoint of the two vertices different from 𝐵is− 𝐵

    Equations of the linesFor a point𝑃on the unit circle the tangent at𝑃has equationX·𝑃=1. * The equilateral triangleΔ𝐵 has its centre at the origin, so the midpoint of the two vertices different from 𝐵is− 𝐵

  23. [31]

    ℓ𝐵 :𝑥cos𝜃+𝑦sin𝜃= 1 2

    The lineℓ𝐵 through those two vertices is parallel to the tangent at𝐵; henceℓ𝐵 :X·𝐵=− 1 2, i.e. ℓ𝐵 :𝑥cos𝜃+𝑦sin𝜃= 1 2 . * Similarly,ℓ𝐶 :𝑥cos𝜃−𝑦sin𝜃= 1 2. * The tangent at𝑁isℓ 𝑁 :𝑥=−1

  24. [32]

    Line𝐴𝐶meetsℓ 𝐵 at𝑌.Write𝑌=𝐴+𝑢(𝐶−𝐴)

    Points𝑌and𝑍Set 𝐶1 =cos(𝛼−𝜃)=cos𝛼cos𝜃+sin𝛼sin𝜃, 𝐶2 =cos(𝛼+𝜃)=cos𝛼cos𝜃−sin𝛼sin𝜃, 𝑄=cos 2𝜃, 𝑀=𝑄+ 1 2 . Line𝐴𝐶meetsℓ 𝐵 at𝑌.Write𝑌=𝐴+𝑢(𝐶−𝐴). Substituting into the equation ofℓ 𝐵 gives 𝐶1 +𝑢(𝑄−𝐶 1)= 1 2 =⇒𝑢= 𝐶1 − 1 2 𝑄+𝐶 1 ,1−𝑢= 𝑀 𝑄+𝐶 1 . Hence 𝑌=(1−𝑢)𝐴+𝑢𝐶. Line𝐴𝐵meetsℓ 𝐶 at𝑍.Analog...

  25. [33]

    Because𝐴lies on the unit circle,𝑎 2 +𝑏 2 =1, so 𝐹=−1−𝐷𝑎−𝐸𝑏,where𝑎=cos𝛼, 𝑏=sin𝛼

    Circumcircle of△𝐴𝑌 𝑍We look for the circle 𝑥2 +𝑦 2 +𝐷𝑥+𝐸 𝑦+𝐹=0 through𝐴, 𝑌 , 𝑍. Because𝐴lies on the unit circle,𝑎 2 +𝑏 2 =1, so 𝐹=−1−𝐷𝑎−𝐸𝑏,where𝑎=cos𝛼, 𝑏=sin𝛼. Substituting𝑌and𝑍into the circle equation and using𝐹yields the linear system    𝐷(𝑌𝑥 −𝑎) +𝐸(𝑌 𝑦 −𝑏)=1− |𝑌| 2, ...

  26. [34]

    Substituting𝐷, 𝐸 and simplifying yields 𝑂= − 𝑀(1−2 cos𝛼cos𝜃) cos𝛼+cos 3𝜃 , 2𝑀sin𝛼cos𝜃 cos𝛼+cos 3𝜃

    Centre and radius of the circumcircleΓof𝐴𝑌 𝑍The centre is𝑂=(−𝐷/2,−𝐸/2). Substituting𝐷, 𝐸 and simplifying yields 𝑂= − 𝑀(1−2 cos𝛼cos𝜃) cos𝛼+cos 3𝜃 , 2𝑀sin𝛼cos𝜃 cos𝛼+cos 3𝜃 . Let 𝐿=cos𝛼+cos 3𝜃, 𝑁=2 cos 2 𝜃−2 cos𝛼cos𝜃+ 1 2 . Using𝑀=2 cos 2 𝜃− 1 2, a direct computation (expanding t...

  27. [35]

    By symmetry its incenter lies on the𝑥-axis

    Incircle ofRThe triangleRhas sidesℓ 𝐵,ℓ 𝐶,ℓ 𝑁. By symmetry its incenter lies on the𝑥-axis. Solving the equal-distance condition from a point(𝑥,0)toℓ 𝑁 (𝑥=−1) and toℓ 𝐵 (𝑥cos𝜃+𝑦sin𝜃= 1

  28. [36]

    gives 𝑥𝐼 =− 1 2(1+cos𝜃) , 𝑟=𝑥 𝐼 +1= 1+2 cos𝜃 2(1+cos𝜃)

  29. [37]

    Set𝑆=1+cos𝜃

    Tangency verificationWe compute𝑂𝐼2. Set𝑆=1+cos𝜃. Then 𝑂𝑥 −𝑥 𝐼 =− 𝑀(1−2𝑎𝑡) 𝐿 + 1 2𝑆 , 𝑂𝑦 = 2𝑀𝑏𝑡 𝐿 . Hence 𝑂𝐼 2 = − 𝑀(1−2𝑎𝑡) 𝐿 + 1 2𝑆 2 + 2𝑀𝑏𝑡 𝐿 2 = 𝑀2(1−2𝑎𝑡) 2 +4𝑏 2𝑡2 𝐿2 − 𝑀(1−2𝑎𝑡) 𝑆𝐿 + 1 4𝑆2 . Using𝑎 2 +𝑏 2 =1, (1−2𝑎𝑡) 2 +4𝑏 2𝑡2 =1−4𝑎𝑡+4𝑡 2. Thus 𝑂𝐼 2 = 𝑀2 (1−4𝑎𝑡+4𝑡 2) 𝐿2 − 𝑀...

  30. [38]

    62 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling Verification of (ii): From (i),𝑀2𝑃=2𝑁 𝑀 2, so 𝑀2𝑃−𝑁 2 =𝑁(2𝑀 2 −𝑁)

    Since4𝑡 2 =2(2𝑡 2)=2(𝑀+ 1 2 ), we have 𝑃=1−4𝑎𝑡+2𝑀+1=2(𝑀−2𝑎𝑡+1)=2𝑁 . 62 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling Verification of (ii): From (i),𝑀2𝑃=2𝑁 𝑀 2, so 𝑀2𝑃−𝑁 2 =𝑁(2𝑀 2 −𝑁). A direct computation gives2𝑀2 −𝑁=2𝑡...

  31. [39]

    Let𝑐𝑖 be the carry into the𝑖-th digit (𝑐0 =0), and let𝑛𝑖 be the𝑖-th digit of𝑛(𝑛0 units)

    Addition with carries Consider the addition𝑎+𝑏=𝑛digit by digit, from the units digit upwards. Let𝑐𝑖 be the carry into the𝑖-th digit (𝑐0 =0), and let𝑛𝑖 be the𝑖-th digit of𝑛(𝑛0 units). We must choose𝑎𝑖, 𝑏𝑖 ∈𝐷and a carry𝑐 𝑖+1 ∈ {0,1} such that 𝑎𝑖 +𝑏 𝑖 +𝑐 𝑖 =𝑛 𝑖 +10𝑐 𝑖+1. The left...

  32. [40]

    *𝑐 𝑖 =0:then𝑥=𝑛 𝑖 +10𝑐 𝑖+1

    Possible transitions of the carry We examine the two possible values of𝑐𝑖. *𝑐 𝑖 =0:then𝑥=𝑛 𝑖 +10𝑐 𝑖+1. –𝑐𝑖+1 =0possible iff𝑛 𝑖 ≠1. –𝑐𝑖+1 =1possible iff𝑛 𝑖 ≤8. *𝑐 𝑖 =1:then𝑥=𝑛 𝑖 +10𝑐 𝑖+1 −1. –𝑐𝑖+1 =0possible iff𝑛 𝑖 ∈ {1,3,4,5,6,7,8,9}(i.e.𝑛 𝑖 ≥1and𝑛 𝑖 ≠2). – 𝑐𝑖+1 =1is always po...

  33. [41]

    The transitions are: *From𝐴:1↦→𝐵,9↦→𝐴,all other digits↦→𝐶

    From the NFA to a DFA The NFA has only two states, so the subset construction yields a DFA with states𝐴={0}, 𝐵={1}, 𝐶= {0,1}. The transitions are: *From𝐴:1↦→𝐵,9↦→𝐴,all other digits↦→𝐶. 64 MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Tes...

  34. [42]

    The accepting states are those containing0, i.e.𝐴and𝐶

    When is a number solitary? A number𝑛(written with𝐿digits) is accepted by the DFA iff it can be expressed as a sum of two1-free numbers. The accepting states are those containing0, i.e.𝐴and𝐶. Therefore𝑛isnotsolitary (i.e. it belongs to𝐴+𝐴) iff the DFA computation ends in𝐴or𝐶; i...

  35. [43]

    Hencethenumberofsolitarynumberswiththe1atposition𝑡is2 𝐿−𝑡−1

    Counting the solitary numbers For a fixed𝑡we have: * the𝑡trailing digits are forced to be9(one way); * the digit at position𝑡is1(one way); * the remaining 𝐿−𝑡−1higher digits can be chosen independently as0or2(2 𝐿−𝑡−1 possibilities). Hencethenumberofsolitarynumberswiththe1atpos...

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.