Fine-Tuning GPT Models on Riemann Hypothesis Bounds: AI Mathematical Breakthrough or Hallucination?

A Casual Prompt Sparks an AI Mathematics Experiment

Recently, a Reddit post ignited a heated discussion at the intersection of AI and mathematics. A user employed an extremely casual prompt—"do ur thang"—to get a ChatGPT variant (referred to in the post as "GPT 5.6 Sol") to attempt improving a numerical bound that Anthropic had previously established on a problem related to the Riemann Hypothesis.

According to the post, the AI model produced a candidate "unconditional improvement": raising the lower bound on the proportion of Riemann zeta function zeros that simultaneously lie on the critical line and are simple zeros, from Anthropic's 67.250070% to approximately 67.25214%—a seemingly negligible +0.002% improvement.

The Riemann Hypothesis and the Research Lineage of Zero Distribution

To understand the significance of this result, some mathematical background is necessary. The Riemann Hypothesis is one of the most famous unsolved problems in mathematics, proposed by German mathematician Bernhard Riemann in 1859. It asserts that all non-trivial zeros of the Riemann zeta function ζ(s) lie on the line in the complex plane where the real part equals 1/2—the so-called "critical line." While the conjecture remains unproven, mathematicians have long been trying to prove "at least what proportion of zeros actually lie on the critical line." In 1942, Selberg first proved that a positive proportion of zeros lie on the critical line; subsequently, Levinson (1974) proved more than 1/3, and Conrey (1989) proved more than 40%. Recent work has continually pushed this proportion higher, with Anthropic's 67.25% result representing the latest advance along this research path. In this context, even an improvement in the third decimal place signifies a substantive breakthrough in extremely refined analytical techniques.

As a notable detail, as the conversation progressed through the night, this number kept being "refreshed": from 67.25214% to 67.252625%, then to 67.26229%, and finally reaching 67.2787614454% in the early morning hours—an improvement of approximately +0.023% over Anthropic's original constant.

Sol's Technical Core: Identifying Discarded Gram Matrix Structure

Information Loss in Anthropic's Original Method

According to the technical description in the post, Sol's key insight was that Anthropic's original argument discards structural information.

Specifically, when Anthropic handled the simple-zero contribution term (denoted P), it primarily relied on rank and trace bounds for estimation. However, P is actually a Gram matrix determined by the geometric positions of the zeros. This means Anthropic's method, during its approximation process, discarded the additional information encoded in the matrix's spectral structure.

To appreciate the depth of this insight, it helps to review the mathematical significance of Gram matrices. A Gram matrix is a fundamental concept in linear algebra: given a set of vectors v₁, v₂, ..., vₙ, the (i,j) entry of their Gram matrix G is defined as the inner product ⟨vᵢ, vⱼ⟩. Gram matrices are naturally positive semi-definite, and their eigenvalues (spectrum) encode geometric relationship information about the vector set—including angles between vectors, degrees of linear dependence, and more. In analytic number theory, when certain kernel functions are used to detect zeros of the zeta function, the associated quadratic forms naturally take on a Gram structure. If one estimates using only the matrix's rank and trace, it's equivalent to knowing only "how many non-zero eigenvalues there are" and "what the eigenvalues sum to," while losing information about the specific distribution of eigenvalues—precisely the optimization space Sol claims to have found.

GPT Model's Improvement Path

Sol's approach is to retain the resulting non-negative Gram spectral deficiency (denoted Δ(P)), then leverage Anthropic's existing Montgomery–Taylor kernel to prove:

• When three sufficiently close simple zeros form a "triple," they necessarily contribute a positive spectral penalty term;
• Through a packing argument, once the simple-zero density exceeds 1/2, a linear number of such triples is forced to exist;
• Substituting these additional deficiencies into Anthropic's optimized inequality yields a stronger lower bound.

Regarding the Montgomery-Taylor kernel: this technique originates from Hugh Montgomery's pioneering 1973 work on the pair correlation of zeta function zeros. Montgomery proposed the famous "pair correlation conjecture," describing the statistical distribution of zero spacings (interestingly, this matches the eigenvalue distribution of GUE ensembles in random matrix theory, hinting at a deep connection between number theory and quantum physics). In practical work estimating the proportion of zeros on the critical line, researchers must choose an appropriate "detection kernel"—a function satisfying specific conditions for use in mollifier or moment methods. Different kernel choices lead to different numerical results, which is why improvements in this field often manifest as advances in decimal places.

Regarding the packing argument: this is a classical technique from combinatorics and metric geometry. Its core idea is that if "too many" objects are placed in a finite space, then some objects must be close together, thereby forming specific local structures—think of it as a continuous version of the pigeonhole principle. In Sol's argument, if the density of simple zeros exceeds 1/2 (meaning simple zeros on the critical line account for more than half of all zeros), then in any sufficiently long interval, the average spacing between zeros is compressed, necessarily producing a linear number of "tight triples." Each such triple corresponds to a positive spectral deficiency in the Gram matrix, thereby providing additional positive-term contributions to the inequality.

This logical chain is structurally self-consistent: it introduces no new assumptions but rather "squeezes out" overlooked information from the existing framework. This is also why it's called an "unconditional improvement"—it doesn't depend on any unproven conjectures.

Critical Caveat: A Candidate Result, Not a Proven Theorem

It must be emphasized that Sol itself does not treat this result as a proven theorem.

The post explicitly states that while the finite-dimensional linear algebra portions, the packing argument, and the computer-assisted kernel bounds appear "compelling," the analytical steps transitioning from finite windows/truncations to ideal Gram blocks still require publication-standard rigorous auditing and independent verification.

Across multiple updates, the author consistently uses the phrasing "computer-assisted candidate theorem" rather than "established results." This careful wording deserves commendation—it distinguishes between "AI-generated seemingly plausible derivations" and "peer-verified mathematical truths."

How to Assess AI's Real Capabilities in Mathematical Research

Positive Signals for LLM Mathematical Ability

If this candidate result ultimately passes rigorous verification, it represents an important trend: large language models are moving from reciting known mathematics toward discovering structural optimizations overlooked by humans.

Sol's core contribution isn't inventing entirely new tools, but identifying the information loss point in Anthropic's method where "rank and trace bounds substitute for the complete Gram structure." This ability to "see through simplified steps in a proof" is precisely one of the most valuable forms of intuition in mathematical research.

Hallucination Risks in AI Mathematical Reasoning

However, AI's greatest risk in mathematics is precisely plausible-looking hallucination. A derivation might appear "correct" at each local step while containing fatal logical leaps in the overall argument—especially at critical junctures transitioning from finite computation to infinite analysis.

The phenomenon of "numbers being continually refreshed" throughout the post also deserves reflection: climbing from 67.252% all the way to 67.278%, this continuous fine-tuning could reflect genuine iterative optimization, or it might suggest the model is "pleasingly" producing increasingly attractive numbers. Without formal verification, these numbers can only be considered candidates awaiting examination.

Formal Verification: The Bridge from "Seemingly Plausible" to "Certainly Correct"

The formal verification mentioned here refers to using proof assistant systems like Lean and Coq to transform mathematical proofs into computer-checkable programs. Lean, developed by Microsoft Research, has been used to formalize large amounts of undergraduate and research-level mathematics; Coq, developed by France's INRIA, is based on constructive type theory. The core value of formal verification lies in eliminating oversights that can occur in human verification. Recent milestones include the formalization of Perfectoid spaces in Lean, the community's completion of verification for Peter Scholze's liquid tensor experiment, and the complete proof of the Four Color Theorem in Coq. For AI-generated mathematical results, formal verification provides a clear path from "seemingly plausible" to "certainly correct"—which is why an increasing number of AI mathematics projects (such as Google DeepMind's AlphaProof) incorporate formalization as a core verification component.

Conclusion: Verification Is the Last Line of Defense in AI-Assisted Mathematics

This post provides an excellent observation window into the true capability boundaries of current AI models on frontier mathematical problems:

It can understand complex analytic number theory arguments, identify structural weaknesses in existing proofs, and even construct seemingly plausible improvement paths. But it equally cannot replace the rigorous peer verification of the mathematical community—especially on subtle points involving infinity, limits, and truncation errors.

Regardless of whether the +0.023% improvement ultimately holds up, it reminds us: AI-assisted mathematical research has entered a new phase. The real value lies not in the specific numbers AI produces, but in whether it can provide human mathematicians with new directions worth deeply verifying. And verification will forever remain mathematics' last line of defense.