AI-TechScienceGovernmentBusinessSportsEntertainment
AI-Tech

Frontier models push mathematicians toward the hardest unsolved problems

Human researchers may abandon routine calculations to concentrate on stubborn foundational conjectures while feeding speculative ideas into shared computational systems.

Frontier models push mathematicians toward the hardest unsolved problems
Source: Primary generated image. © 2026
Published16 Nov 2026, 00:13 Last updated16 Nov 2026, 00:13 Sources
Source mode Marks each sentence drawn from a source or a contributor

For centuries, the daily labor of mathematics has centered on quiet persistence in solitary offices, where researchers fill blackboards and notebooks while attempting to bridge gaps between known premises and stubborn conjectures. A working mathematician typically spends months tracking down intermediate lemmas, checking edge cases, and verifying whether existing techniques from adjacent subfields might apply to a difficult question. When an automated system suddenly absorbs those intermediate steps, performing calculations and scanning vast mathematical libraries in seconds, the traditional rhythm of research alters abruptly.

The arrival of frontier reasoning models has disrupted that long-standing division of mental labor. Automated systems have demonstrated an unusual facility with routine deductions, broad literature synthesis, and formal proof verification. This sudden shift raises an urgent question for working scholars: if algorithmic tools can quickly resolve the intermediate technical work that previously occupied graduate students and early-career faculty, human curiosity must migrate toward problems where machines stall.

In an essay posted on November 14, 2026, Michael Freedman, a mathematician at Harvard University's Center of Mathematical Sciences and Applications and Logical Intelligence, examined how the profession should respond to these tools.1 Freedman argued that rather than keeping automated systems at arm's length, researchers should push models to their conceptual limits while refocusing human attention on deeply intractable questions, including the Riemann Hypothesis, the smooth four-dimensional Poincaré conjecture, the Andrews-Curtis conjecture, and the separation of P and NP.1

How does machine reasoning change the daily work of mathematicians?

Machine reasoning alters the research process by shifting human effort away from routine technical verification and toward the framing of new conceptual vantage points. In classical mathematical practice, an investigator formulates a hypothesis, gathers known theorems, and works through mechanical steps to verify whether a chain of logical deductions holds. When automated systems handle those mechanical derivations, the primary bottleneck becomes the choice of problem itself.

Freedman observed that interacting with frontier systems often reveals how much existing mathematics can be brought to bear on a problem, putting researchers into a state of ideal preparation.1 In his account, conversing with artificial intelligence models can be eye-opening regarding the volume of relevant work that already exists across disparate fields.1 He noted that if a model resolves an inquiry on its own without human intervention, the underlying problem was simply not hard enough to warrant human labor.1

This dynamic challenges traditional paths of professional development. Freedman noted that he previously advised graduate students to tackle central problems in their discipline because any genuine progress would inevitably be read and absorbed by the wider community.1 Today, however, he argued that cleaning up loose ends no longer justifies significant human time because automated systems excel at such tasks, leaving human scholars to meditate on larger, unresolved dilemmas.1

Yet pursuing only monumental problems carries severe career risks. Freedman recounted a conversation with computer scientist Adi Shamir, who stated that he stopped working on the P versus NP problem after realizing there was no operational test to distinguish working on that problem from doing nothing at all.1 Without intermediate milestones or publishable derivations, a researcher risks years of silence in an academic culture that demands steady output.

Can artificial intelligence make genuine conceptual leaps?

Artificial intelligence systems excel at deduction and pattern recognition, but researchers disagree on whether these models can generate genuinely original foundational principles. In classical logic, discovery relies on three distinct operations: deduction, which derives necessary results from known rules and cases; induction, which extracts general patterns from observed samples; and abduction, which posits entirely new hypotheses or axioms to explain surprising phenomena.2

Tom Zahavy, in a position paper presented at the International Conference on Machine Learning in 2026, argued that large language models remain structurally incapable of abduction, a cognitive leap he termed the jump.34 Using Albert Einstein's 1907 formulation of the equivalence principle as a case study, Zahavy contended that genuine invention requires intuitive physical grounding and embodied simulation, such as Einstein's thought experiment of an observer falling inside an elevator.34 Zahavy suggested that without multimodal, physically grounded world models, purely textual models merely manipulate linguistic representations without generating novel axioms.34

Other researchers argue that embodiment is not necessary for conceptual leaps. In a preprint posted to the arXiv server in August 2026, Paras Balani and Subhrakanta Panda of the Birla Institute of Technology and Science, Pilani, Hyderabad Campus, analyzed Max Planck's 1900 resolution of blackbody radiation. Classical physics had predicted that a blackbody in thermal equilibrium would emit infinite energy at high frequencies, a contradiction known as the ultraviolet catastrophe.4 Planck resolved this crisis not through sensory experience or embodied simulation, but by introducing the quantum postulate, E equals h multiplied by nu, which had no antecedent in classical physics.

Frontier models push mathematicians toward the hardest unsolved problems
Source: Genghizkhan91 (CC BY-SA 3.0)

Balani and Panda argued that genuine abduction requires an architecture where epistemic errors impose physical and computational costs.4 In their empirical tests, transformer models maintained nearly constant output entropy across tasks of increasing causal difficulty, even while their accuracy fell from 100 percent to 17 percent across the evaluated test cases. Because fixed-weight inference lacks a thermodynamic penalty for generating false world models, Balani and Panda contended that current models generate the appearance of abduction without paying the physical cost required to force a radical theoretical revision.

A contrasting mathematical perspective emerged in a preprint posted to arXiv in September 2026 by Dai Shi, Xiaoyu Li, and José Miguel Hernández-Lobato.2 The authors evaluated whether four frontier language models could execute an abductive override when tested on nine certified category-theoretic extension problems.2 In these tests, the canonical completion determined by Kan extensions was deliberately ruled out by task constraints.2 Without constraints, the models produced canonical completions in 98 percent of samples; under strict constraints, the models achieved a Kan-default rate of zero across all 248 constrained trials, discovering the unique alternative completion that required positing unobserved structures.

Freedman noted that his own perspective on model capabilities evolved over 2026.1 In the spring of that year, he viewed models as forming the closure of human mathematical knowledge; by summer, he regarded them as spanning its convex hull; and following an October 6 release from OpenAI, he concluded that their capabilities extended beyond that boundary.1 Freedman cited Scott Aaronson's concept of statements that are true, but for no good reason, suggesting that many number-theoretic conjectures may simply be true without possessing neat, compact axiomatic proofs.1

How would an open research workspace track mathematical credit?

An open research workspace would track intellectual credit by recording informal suggestions and using models to monitor how widely those ideas propagate through the community. Under the system Freedman envisioned, researchers would interact with frontier models in a shared digital environment where queries and speculative prompts are recorded in real time.1 These contributions would simultaneously train the underlying models and populate a searchable library of human research concepts.

In this framework, mathematicians would not need to write exhaustive formal papers to claim credit for preliminary breakthroughs.1 Instead, an investigator could share an intuitive formulation or a conjectural stepping stone directly with an automated agent.1 If that intuition proved fruitful, the system would transmit the concept to other scholars who are actively working on related questions, appending an explicit citation of the originator.1

Freedman suggested that this mechanism could establish a parallel publication process.1 Automated systems could log the frequency and impact of a scholar's contributions worldwide, generating a verifiable metric to support tenure reviews, hiring decisions, and grant proposals.1 While admitting that researchers might risk being scooped by openly sharing preliminary ideas, Freedman maintained that accelerating communal progress would compensate for individual priority disputes.1

This collective approach would also enable what Freedman described as bulk math, the simultaneous statistical analysis of vast families of mathematical structures.1 Similar computational methods have already illuminated phase transitions in k-satisfiability problems, where truth values shift abruptly across parameter thresholds.1 Freedman anticipated extending such macroscopic investigations to Diophantine equations, knot invariants, and three-manifold topology.1

What are the boundaries of automated mathematical discovery?

These theoretical frameworks and benchmark evaluations describe specific algorithmic behaviors under laboratory conditions, not complete models of human creativity. The category-theoretic tests devised by Shi, Li, and Hernández-Lobato provided fixed constraints and predefined codomains to the models, evaluating how algorithms behave within human-designed formal systems rather than testing their ability to invent entirely new mathematical languages.2 Furthermore, the thermodynamic arguments raised by Balani and Panda rely on theoretical analogies between Gibbs distributions and model activations, which remain open to empirical debate.4

Preprint papers describing these capabilities represent early exploratory findings that have been posted to public servers without formal peer review.42 Mathematical creativity itself remains elusive, and existing benchmarks cannot determine whether an artificial intelligence model is generating a fundamentally novel perspective or merely executing a sophisticated recombination of its training data. As Freedman observed, citing Albert Einstein's famous remark to a reporter, a human being is fortunate to encounter one or two truly original ideas in an entire lifetime.1

For the broader scientific community, these shifts suggest that professional training must emphasize conceptual formulation and interdisciplinary synthesis over routine derivation. As automated systems shoulder the burden of intermediate calculation, human researchers will increasingly serve as curators of deep questions, guiding algorithms toward fruitful conjectures. The open question is whether future model architectures will remain sophisticated interpolation engines or eventually develop the capacity to formulate entirely new mathematical universes.

The record

34 sentences trace to 4 sources.

  1. 1 M mikefreedman Document · lesswrong.com · 14 Nov 2026 Mathematicians: Let’s Surf This Wave See the source
  2. 2 A arxiv.org Paper · arxiv.org When the Canonical Completion Is Wrong:Formalizing and Measuring the Jump in LLMs Preprint · may not have been peer reviewed See the source
  3. 3 T tomzahavy.com Third-party reporting · tomzahavy.com LLMs can't jump — Tom Zahavy See the source
  4. 4 A arxiv.org Paper · arxiv.org LLMs Don’t Pay for the Jump Preprint · may not have been peer reviewed See the source