フィールドメダリストが未解決の問題を高校生2名が解明

Internet Shocked: Unsolved Problem by Fields Medalist Breached by Two High School Students with AI

marsbitPublished on 2026-09-14Last updated on 2026-09-14

Abstract

Remarkable breakthrough, two high school students from Oak Park High School, Aayush Bathija and Prince Rohatgi, with guidance from UCLA postdoctoral researcher Daniel Soskin, have solved an open problem related to the theory of Lorentzian polynomials—a field pioneered by Fields Medalist June Huh. Their 75-page paper, "Bounded Ratios for Lorentzian Polynomials," is now available on arXiv. The research investigates the constraints on coefficient ratios within Lorentzian polynomials, extending prior work that characterized bounded ratios for quadratic polynomials to polynomials of arbitrary degree. The core challenge was determining which coefficient ratios have a universal upper bound across all polynomials in the class and identifying the optimal bounds. The students heavily utilized AI assistants, specifically Claude Opus 5 and GPT-5.6 Sol, for computational exploration, proof idea generation, and editing. They emphasize that while AI provided helpful suggestions, some were misleading, and all calculations and arguments were independently verified. This achievement is particularly notable as it follows a recent open letter signed by 25 Fields Medalists, including June Huh, expressing concerns about AI's impact on mathematical rigor. The students' work demonstrates how AI can lower barriers to entry, enabling talented pre-university researchers to engage with advanced mathematical frontiers.

• #AI
• #arXiv
• #Collaboration
• #Fields Medal
• #High School
• #Lorentzian
• #Mathematics
• #Polynomials
• #Proof
• #Research

Can you believe it?

They haven't even started university, yet their research has connected with the work of a "Fields Medalist".

［画像: 高校生がフィールズ賞受賞者の研究分野に貢献したことを示すイメージ］

Just today, stunning news spread through the UCLA mathematics community—

Two high school students, Aayush Bathija and Prince Rohatgi, under the guidance of a postdoctoral researcher, have solved a problem that mathematician June Huh never managed to crack.

Even more eye-catching is that this proof was completed with heavy assistance from AI!

One was Claude Opus 5, the other GPT-5.6 Sol.

［画像: AIツールのロゴやインターフェースを示すイメージ］

Currently, the paper titled "Bounded Ratios for Lorentzian Polynomials" has been published on arXiv, spanning 75 pages.

［画像: arXivの論文ページまたは論文の表紙を示すイメージ］

Paper: https://arxiv.org/pdf/2609.05341

The names of the two high school students and one postdoc are written right there in the author list.

［画像: 論文の著者リストを示すイメージ］

The high school students' mentor, UCLA Postdoctoral Fellow Daniel Soskin

So, what problem exactly did they solve in this research?

To clarify this, we must first start with the work of Fields Medalist June Huh.

What Did They Actually Solve?

June Huh is Korean and won the Fields Medal in 2022.

His journey is quite legendary: he dropped out of high school to become a poet before switching to pure mathematical research, winning the Fields Medal at the age of 39 (the Fields Medal is only awarded to individuals under 40).

［画像: June Huh氏の写真］

In 2020, the theory of "Lorentzian polynomials" that he and his collaborators established was one of his important representative works.

The name might sound unfamiliar, but "polynomials" are something everyone encountered in middle school. The numbers in front of each term are the coefficients.

The special feature of "Lorentzian polynomials" is that there are strict mathematical constraints between these coefficients; they cannot be paired arbitrarily.

This theory connects problems in combinatorics with geometry and inequalities between coefficients.

This research inquires about precisely how strong these constraints are.

For example, if you multiply some coefficients and divide by others, you get a ratio.

As long as the polynomial satisfies the specified conditions, does this ratio have an "impenetrable ceiling"?

［画像: 数式またはグラフで係数比の概念を示すイメージ］

Do you still remember the definition of the classic Hessian matrix from advanced calculus?

Calculating the ratio for a specific set of coefficients is not difficult. The real challenge is: facing an infinite variety of polynomials that meet the conditions, determining which ratios always have an upper bound and which can increase indefinitely.

If there is an upper bound, we must further ask: How low can this upper bound be pushed to obtain the most precise answer?

Previously, June Huh and his collaborators had characterized which ratios have upper bounds for quadratic Lorentzian polynomials and had found the optimal upper bounds for the case of three variables.

But this path was not yet complete.

Moving from quadratic to cubic, quartic, and even arbitrary degrees, do the old rules still hold?

The higher the degree, the more complex the relationships between coefficients become. Conclusions that hold in the quadratic case cannot be directly transferred.

This is the gap where the two high school students entered—

advancing the research on coefficient ratios to higher degrees, clarifying which ratios always have bounds, and further seeking the most precise upper limits.

［画像: 研究の進展を示す図または概念図］

The paper's core formula, the "Main Structural Theorem," extends the relevant conclusions from previous quadratic Lorentzian polynomials to arbitrary degrees. It shows that whether a coefficient ratio has a uniform upper bound can be completely determined using a set of discrete convexity conditions.

Reading this far, you just need to remember one question: How high can these constrained coefficients push the ratio?

Next, we'll use a simple example to explain this "ceiling" clearly.

Lorentzian Polynomials: "Balance Constraints" Between Coefficients

First, look at a polynomial:

［画像: 単純な多項式の例］

Here a, b, c are all positive numbers. Writing the middle term as 2b is to make the subsequent relationship cleaner.

In this simplest case, it is a Lorentzian polynomial precisely requiring:

［画像: ローレンツ多項式の条件を示す数式］

In other words: If the coefficients at both ends are large, the coefficient in the middle cannot be too small. This is an introductory example of the log-concave property embodied by the Lorentzian structure.

For example, if a=4, c=9, then b must be >=6. You cannot arbitrarily change b to 1 and still expect it to remain in this polynomial family.

When there are more variables and higher degrees, these constraints become more complex, requiring various derivative results to collectively satisfy specific conditions. The formal definition involves a matrix describing curvature, which in the strict case has an eigenvalue structure with "one positive, the rest negative." The name "Lorentzian" corresponds to this positive-negative structure. However, simply remembering that "the coefficients are subject to a complete set of interrelated constraints" is enough to understand the main thread of this article.

Transforming the previous inequality slightly:

［画像: 不等式を変形した数式］

This is called a "bounded ratio": no matter how you choose coefficients that meet the conditions, this ratio cannot exceed 1.

But in reverse:

［画像: 逆数の比率を示す数式］

there is no uniform upper bound.

For instance, fix a=c=1, let b continuously increase, still satisfying the Lorentzian condition, but the inverse ratio can become arbitrarily large.

Naturally, one would ask:

Facing more complex polynomials, when multiplying and dividing different coefficients, which combinations have an upper bound, and which do not?

This paper aims to systematically answer this question, with the previous quadratic example belonging to the existing foundational case.

Two High School Students, Teamed Up with AI, Cracked It

The two students who wrote this paper, Aayush Bathija and Prince Rohatgi, both attend Oak Park High School in California.

They are also members of the UCLA Olga Radko Math Circle (ORMC).

Aayush is a sophomore and an AIME qualifier. Prince is currently a junior, also having made AIME, and serves as a main instructor for AMC 10/12 (equivalent to high school freshman) competition classes within the math circle.

In the past, for a high school student to even touch a "Fields Medal-level" field, they would need at least 5-8 years of university and doctoral training.

But this time, under the guidance of UCLA mathematics postdoctoral fellow Daniel Soskin, they directly teamed up with AI to wage a brilliant blitzkrieg.

［画像: UCLA Olga Radko Math Circle (ORMC) のウェブサイトまたは関連画像］

https://circles.math.ucla.edu/circles/index.shtml

The core of their problem-solving was deeply embedding AI and formal computational tools into the central workflow of mathematical discovery.

The overall core idea of the proof can be understood as: To check whether a ratio might run out of control, first study how it changes under extreme conditions.

Transforming Complex Numerical Values into Simple Exponents

Let t approach 0, with the three coefficients taking values:

［画像: 極限状態の係数を示す数式］

Then:

［画像: 指数表現に変換された比率を示す数式］

In reverse, it's t^-4, which tends towards infinity.

The most useful point here is: Multiplication and division become addition and subtraction of exponents. The sign of the exponent directly tells us whether the ratio will shrink or explode.

This idea of extracting the "dominant power" is related to the "tropicalization" method used in the paper. Existing Lorentzian polynomial theory tells us: legitimate power patterns are governed by a discrete convexity rule called M-convexity.

You can think of it as a mathematical rule for "which rates of change can appear simultaneously."

A key step in the paper is to prove: If a ratio can increase indefinitely, a path with this power characteristic can be found to expose the runaway.

Therefore, checking these paths can be upgraded into a complete criterion. The proof uses tools from semi-algebraic geometry like curve selection lemmas.

And don't forget, all this sophisticated talk was accomplished by two high school students with the help of AI.

Claude and ChatGPT, Participated in "Proof Idea Generation"

In the paper's acknowledgments, the authors lay their cards on the table—

Claude Opus 5 and GPT-5.6 Sol were the core tools for this proof.

［画像: AIツールのロゴまたは使用状況を示すイメージ］

They were used for computation, proof idea generation, and editing assistance. Among these, "proof idea generation" is especially crucial.

This means the models' use has deepened into research exploration: facing an unsolved problem, searching for possible derivation paths that might hold.

However, the authors also clearly wrote the other side: Some suggestions were helpful, others were misleading.

They stated they have independently verified all calculations and take full responsibility for the paper's content.

These few sentences outline the actual boundaries of this collaboration: AI can provide ideas, researchers need to screen, scrutinize, and turn valuable clues into arguments that can withstand scrutiny.

When the model gives a seemingly beautiful answer, the research is not over. Whether conditions are complete, derivations are valid, edge cases are missed—all must be further questioned.

AI has entered the exploration phase, but rigorous verification remains the prerequisite for valid results.

The mentorship of the postdoctoral advisor and the support of the ORMC research program are thus particularly important. The two students were able to access genuine open problems and conduct research under professional guidance. AI was embedded into this collaboration, assisting with computation and idea exploration.

Fields Medalists Collectively Call for a Halt, But It Can't Be Stopped

Dramatically, just one day before this research was announced, the mathematics world was rocked by a major uproar.

Twenty-five Fields Medalists from around the world jointly signed an open letter, issuing the strongest warning: "AI is destroying mathematics itself!"

［画像: 著名な数学者たちの集合写真または関連するイメージ］

June Huh himself was one of the signatories.

［画像: 声明文の一部または関連するイメージ］

Partial list

Who would have thought that two high school students, using AI assistance, produced results in June Huh's own research direction.

The 25 Fields Medalists' concern is that the rigor of pure mathematical research and the purity of scientific inquiry might be crushed by computational power and algorithms.

But the story of Aayush and Prince presents another side—

After AI lowers the research threshold, what's being unleashed is not just "problem-solving speed," but also more people's eligibility to enter the field.

An AIME-qualified high school student, guided by a postdoc, using AI assistance, touched a path a Fields Medalist hadn't finished walking.

Five years ago, this was almost impossible.

The frontiers of mathematics no longer belong only to those who have spent over a decade just to reach the doorstep.

Whether this is good or bad, the mathematical community will debate for a long time. But the door has already been pushed open a crack.

References:

https://circles.math.ucla.edu/circles/index.shtml

https://arxiv.org/pdf/2609.05341

This article is from WeChat Official Account "新智元" (ID: AI_era), authors: Taozi, Mark

Related Questions

QWho are the main high school students mentioned in the article and what is their achievement?［画像: 記事の主要人物を示すイメージ］

AThe main high school students mentioned are Aayush Bathija and Prince Rohatgi. They achieved a proof, with significant AI assistance, that advances the work on bounded ratios of Lorentzian polynomials, a problem left open by Fields Medalist June Huh. Their 75-page paper is posted on arXiv.

QWhat are Lorentzian polynomials, and what specific problem did the research address about them?［画像: ローレンツ多項式の概念図］

ALorentzian polynomials are a special class of polynomials whose coefficients are not independent but are governed by a set of interrelated mathematical constraints, linked to geometric and combinatorial properties. The research addressed the problem of determining which ratios of these coefficients have a universal upper bound and finding the optimal, lowest possible bounds for these ratios, extending the theory from quadratic to higher-degree polynomials.

QWhich AI models were used as core tools in this research, and how were they utilized?［画像: 使用されたAIツールのロゴ］

AThe core AI tools used were Claude Opus 5 and GPT-5.6 Sol. They were utilized for computations, exploring proof ideas, and editorial assistance. However, the authors noted that some AI suggestions were helpful while others were misleading, and they independently verified all calculations, taking full responsibility for the paper's content.

QWhat is the broader context within the mathematical community regarding the use of AI in research, as mentioned in the article?［画像: 数学界のAIに関する議論を示すイメージ］

AThe article mentions that just before this research was publicized, 25 Fields Medalists, including June Huh, signed a joint open letter expressing a strong warning that AI is 'destroying mathematics itself.' They are concerned about the potential erosion of the rigor and purity of mathematical research. This high school students' work presents a contrasting narrative, showing AI can lower barriers to entry and enable new participants in advanced research.

QWhat role did the UCLA Olga Radko Math Circle (ORMC) and the postdoc Daniel Soskin play in this story?［画像: 指導者と学生の協力関係を示すイメージ］

AThe UCLA Olga Radko Math Circle (ORMC) provided the environment where the high school students Aayush Bathija and Prince Rohatgi were members. Postdoctoral researcher Daniel Soskin from UCLA served as their mentor, guiding them in this research project. His supervision and the ORMC's support were crucial in allowing the students to work on an authentic open problem and rigorously validate their AI-assisted findings.

Related Reads

##

TSMC Q3 Signals: Significantly Increasing Investment in the More Mature 3nm Process, More Rational Approach to AI

TSMC reported strong Q3 2026 results, with revenue of NT$1,494.24 billion (about US$46.72 billion), up 50.9% year-over-year, gross margin of 66.4%, and net margin of 51.1%. Growth was driven by faster-than-expected 2nm shipments, which already reached 8% of revenue just two quarters after production ramp-up, as major customers including Apple, Nvidia, AMD, Qualcomm, and MediaTek increased orders by 10–20%. Management raised its full-year 2026 revenue growth guidance to about 44% and plans to raise 2nm pricing by 3–6% starting January 2027. Gross margin declined slightly quarter-over-quarter due to 2nm ramp-up costs but remains at a high level, with Q4 guidance of 65–67%. Capital expenditure remains aggressive: Q3 capex reached US$13.84 billion, up 42.7% year-over-year, and full-year 2026 capex is expected at US$60–64 billion, with 2026–2028 investment significantly higher than the past three years. TSMC is expanding 3nm capacity with new fabs in Tainan, reflecting the long lifecycle and strong demand for mature advanced nodes. HPC products now account for nearly 70% of revenue. Meanwhile, TSMC remains cautious on High NA EUV, planning adoption around 2030, later than Intel and Samsung, prioritizing cost, maturity, and profitability over early adoption. Overall, TSMC is betting heavily on AI but in a pragmatic, manufacturing-driven way, combining capacity expansion, customer diversification, and disciplined technology rollout, making it a relatively comprehensive AI hardware investment choice.

marsbit51m ago

［画像: TSMC Q3 Signals: Significantly Increasing Investment in the More Mature 3nm Process, More Rational Approach to AI］

marsbit51m ago

##

The Financing Flywheel of Bitcoin Treasuries Begins to Stall: Strive's Preferred Stock Yield Approaches 17%, Strategy Trapped in Equity Dilution Dilemma

Bitcoin treasury companies are facing growing financing pressures as traditional capital channels tighten. According to Bloomberg, firms like Strive and Strategy are increasingly turning to high-yield preferred stock to attract fixed-income investors and sustain cash flow for continued crypto purchases. Strive recently completed a $630 million preferred stock offering, about 25% above its original target, offering a 15% nominal annual dividend with a discounted issuance that pushes the effective yield to nearly 17%. Compared with the roughly 4.2% one-year U.S. Treasury rate, the credit spread is close to 1,300 basis points, approaching distressed-debt levels. Meanwhile, Strategy, the largest corporate Bitcoin holder, is also showing strain. Its 12% annual-dividend STRC preferred stock has been suspended since a June secondary-market price decline, and management is now considering a daily dividend mechanism to restore liquidity. With preferred stock sales stalled, the company continues relying on large-scale at-the-market common stock issuance to fund Bitcoin purchases, diluting existing shareholders. The core treasury model relies on low borrowing costs to capture high-volatility asset returns. But as financing costs rise to 15% or even 17%, Bitcoin’s long-term compound gains must consistently outperform these double-digit fixed expenses, or cash flow will be eroded. The shift from zero-interest convertible bonds to distressed-level high-yield preferred stock is pushing corporate balance sheets into a fragile bet highly dependent on a one-way crypto bull market.

marsbit54m ago

［画像: The Financing Flywheel of Bitcoin Treasuries Begins to Stall: Strive's Preferred Stock Yield Approaches 17%, Strategy Trapped in Equity Dilution Dilemma］

marsbit54m ago

##

Harvard's Blockbuster Study: AI Can Predict Cancer Patients' Mortality Risk from a Facial Photo

A Harvard-affiliated study at Mass General Brigham's cancer institute, published in The Lancet Digital Health, introduced FAHR-Face, an AI model that predicts cancer patients' survival from a single facial photo. Trained on over 40 million unlabeled faces (ages 0-116) from the WebFace42M dataset, it was fine-tuned into two tools: FAHR-FaceAge, estimating biological age (mean absolute error 5.1 years), and FAHR-FaceSurvival, trained on 34,363 cancer patient photos. In a test set of 3,820 patients, 52.5% of the highest-risk group (122 patients) died within 90 days of their photo, versus 2.1% in the lowest-risk group. Five-year survival was 78.1% for lowest-risk patients but only 17.7% for highest-risk. Looking older than one's actual age raised mortality risk by 21%, while looking younger lowered it by 17%. Attention mapping showed the survival model focused on the eye sockets and nasal bridge, echoing the ancient Hippocratic facies signs of impending death. The model proved robust against makeup, expression, lighting, and cosmetic surgery, except rhinoplasty, which slightly skewed age estimates. Smoking, alcohol, and drug abuse correlated with older biological age. Researchers stress the tool is meant to aid, not replace, clinical judgment, while ethicists warn of risks like insurance or employment discrimination.

marsbit55m ago

［画像: Harvard's Blockbuster Study: AI Can Predict Cancer Patients' Mortality Risk from a Facial Photo］

marsbit55m ago

##

Morgan Stanley Research Interpretation: HDD Cycle Is Not Over, but the Market Is Pricing It as Peaking

Morgan Stanley’s October 20, 2026 report argues the HDD cycle is not over, contrary to a market pricing of about 8x P/E that implies a FY28 peak. Demand is accelerating: nearline HDD spot prices are 4–7 cents/GB versus blended nearline contract prices of about 1.5 cents/GB, and industry supply shortages are seen at ~400EB in CY27/CY28 and 300EB in CY26. Customers are expanding from core CSPs to AI labs, which want to buy 100% of EB. Morgan Stanley reiterates Overweight on Seagate (STX) and Western Digital (WDC), with STX back as IT hardware Top Pick. Target prices are $1,200 for STX and $643 for WDC. The 8x P/E appears to assume HDD earnings peak in FY28 and an 80% SSD price decline from C2Q26, which Morgan Stanley views as too punitive. Target multiples were cut to 15x FY28 EPS for STX and 14x for WDC on market structure and AI financing uncertainty. Non-nearline HDD prices rose 15–30% q/q. Seagate is more aggressive than WDC in capturing spot pricing; Toshiba’s CY27 capacity is sold out. December-quarter EPS is expected above consensus: STX at $9.54, 12% above consensus, possibly 5% higher with more aggressive spot assumptions; WDC at $5.25, 10% above consensus, potentially $5.67. WDC’s tactical setup is better due to lower buy-side expectations. STX remains 10% above the “no-brainer” valuation level, WDC only 3%. The catalyst path is less clear, but demand checks support upside and raised guidance. Morgan Stanley prices the cycle as continuing, while the market prices it as ended.

marsbit56m ago

［画像: Morgan Stanley Research Interpretation: HDD Cycle Is Not Over, but the Market Is Pricing It as Peaking］

marsbit56m ago

Trading

ルール

• 技術用語は適切にカタカナ表記または原語のまま残し、それ以外の英単語は日本語に訳してください
• 数値と単位は原文の意味を正確に保って訳してください（例: trillion は「兆」）
• 本文は Markdown で与えられます。translatedBody も Markdown で書き、原文の構造（見出し・段落・箇条書き・番号付きリスト・引用・表・強調・取り消し線）をそのまま保ってください。段落の間には空行を入れ、全文を1つの塊にまとめないでください
• コードブロックとインラインコードは翻訳せずそのまま残してください
• リンク テキスト はテキストだけを翻訳し、URL は一字も変えずに残してください
• ［画像: …］［動画: …］［音声: …］［図: …］は原文に画像・動画・音声・図があった位置を示す印です。印の形と位置はそのまま残し、「…」の説明文だけを翻訳してください
• 原文にない見出しや装飾を追加しないでください
• タイトル先頭の「Ask HN:」「Show HN:」「Tell HN:」「Launch HN:」は翻訳せず、そのまま titleJa の先頭に残してください
• titleJa と summary は Markdown を使わないプレーンテキストで書いてください（太字、箇条書きの記号、# 見出し、テキスト のリンクなどの記法は使わない）
• 本文が取得できない場合は、タイトルから推測した簡潔な説明のみを記述し、「本文が取得できません」等の取得状況には一切言及しないでください
• 記事の順番を維持してください