🔢 mathematics

Long-Horizon AI Research for Grothendieck Constant: A Case Study in Human-AI Mathematical Collaboration

This paper presents a case study on human-AI collaboration in mathematics, demonstrating how an AI research system successfully generated novel insights to tighten the known bounds of the Grothendieck constant while offering critical reflections on the system's strengths, weaknesses, and the conditions necessary for such breakthroughs.

Original authors: Alan Li, Rahul Saha, Anton Xue, Swarat Chaudhuri, Adam Klivans, Pravesh K Kothari, Raghu Meka

Published 2026-08-12

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Alan Li, Rahul Saha, Anton Xue, Swarat Chaudhuri, Adam Klivans, Pravesh K Kothari, Raghu Meka

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a giant, messy puzzle where the pieces are made of rigid, blocky Lego bricks. You want to arrange them to get the highest possible score. But there's a catch: the rules say you can only move the bricks in specific, jerky ways. This is a "combinatorial" problem—hard, discrete, and frustratingly difficult to solve perfectly. Now, imagine you are allowed to melt those bricks into smooth, flowing clay. Suddenly, the puzzle becomes much easier to solve because you can slide the clay into the perfect shape. This is a "continuous" relaxation. The big question in this corner of mathematics is: how much better can the smooth clay get compared to the rigid blocks? Is the difference tiny, or is it massive? This gap is measured by a famous number called the Grothendieck constant. For over 70 years, mathematicians have been trying to pin down the exact value of this constant. It's like trying to find the exact speed limit of the universe for a specific type of math traffic jam. Knowing this number matters because it tells us the absolute limit of how well we can approximate hard problems using easier, faster methods. If the gap is small, we can trust our fast shortcuts; if it's huge, we need to be very careful.

In this paper, a team of human mathematicians and an artificial intelligence system teamed up to narrow down the range of this mysterious number. Think of the AI not as a robot that solved the puzzle alone, but as a tireless research assistant who can crunch numbers and write proofs at lightning speed, while the humans act as the project managers who decide which dead ends to avoid and which new paths to explore. The team didn't find the exact answer yet, but they did something impressive: they tightened the best-known boundaries for the constant. They proved that the true value is definitely higher than a specific fraction ( $6\pi/11$ , which is about 1.7135) and definitely lower than a slightly improved version of an old record ( $\frac{\pi}{2} \log(1 + \sqrt{2}) - 3.47 \times 10^{-4}$ , which is about 1.7818).

The most exciting part of their journey wasn't just the numbers, but how they got there. The AI system was incredibly good at the "grind"—it could run thousands of experiments, check complex calculations, and even write out long, logical proofs. However, the AI sometimes got stuck in a loop, trying to fix a problem by building a better "blocky" puzzle piece when the real answer lay in a completely different direction. The human researchers had to step in, look at the AI's repeated failures, and say, "Wait, these failures aren't just mistakes; they are actually a clue that proves a rule!" This human insight allowed them to flip the strategy: instead of trying to build a harder puzzle to break the system, they used the AI's failures to prove that no puzzle could ever be that hard. This shift led to a brand-new way of proving the lower limit of the constant, a method that had never been used before.

The paper also serves as a honest report card on how well AI works in real-world math research. The team found that while the AI is a powerhouse for technical execution—like a super-fast calculator that can also write code—it still struggles with "research judgment." It doesn't always know when to stop trying a failed idea or when to change the whole plan. The humans had to constantly steer the ship, curating the AI's notes and making sure it didn't forget important warnings. In the end, this collaboration shows that the future of math isn't about AI replacing humans, but about humans and AI working together, where the human provides the intuition and the AI provides the endless energy to test those ideas.

Technical Summary: Long-Horizon AI Research for the Grothendieck Constant

1. Problem Statement

The paper addresses the Grothendieck constant ( $K_G$ ), a fundamental value in functional analysis and combinatorial optimization that quantifies the integrality gap between a hard discrete optimization problem and its continuous semidefinite programming (SDP) relaxation. Specifically, for a matrix $A$ , the problem seeks the maximum of a bilinear form over sign vectors ( $\pm 1$ ), while the relaxation allows unit vectors in any dimension. The constant $K_G$ is the smallest factor such that the relaxed value never exceeds $K_G$ times the discrete optimum.

Despite over seventy years of study, the precise value of $K_G$ remained unknown. Prior to this work, the best known bounds were:
$1.6769\dots \le K_G \le 1.7822\dots$
The gap was wide enough that even the tenths digit of $K_G$ was undetermined. The authors aimed to tighten these bounds using an AI-assisted research system.

2. Methodology: The AI Research System

The authors engineered an asynchronously steered AI research system designed for long-horizon tasks where goals evolve and rewards are sparse. The system operates as a "harness" coupling two distinct agents:

Reasoning Agent: Acts as the "brain," selecting research directions, developing arguments, specifying experiments, and auditing claims. It utilizes frontier language models (GPT-5.5-Pro and GPT-5.6-Sol).
Coding Agent: Acts as the "executor," maintaining the project repository, implementing experiments, retrieving literature, and driving the reasoning model via a scripted interface. It utilizes coding models (Claude Opus and Claude Fable 5).

Key Architectural Features:

File-Based Memory: Since agent contexts do not persist between sessions, continuity is maintained through files: a problem statement, a summary file (working memory), and append-only transcripts/logs.
Session Structure: Work is organized into 2–5 concurrent sessions per day, each lasting a few hours. Sessions follow a cycle: orientation, focus selection, research loop (computation/literature/reasoning), and a closing audit.
Human Steering: Human operators provide asynchronous directives (approx. 40 over the run) to set priorities, audit claims, and redirect the research agenda.
Verification Protocol: The system employs a calibrated internal verification protocol where claims are tagged as proven, numerically supported, conjectural, or heuristic. Only results independently verified by human authors are stated as theorems.

3. Key Contributions and Results

A. Mathematical Results

The collaboration produced two primary improvements to the bounds of $K_G$ :

New Upper Bound:
- Result: $K_G \le \frac{\pi}{2 \log(1 + \sqrt{2})} - 3.47 \times 10^{-4} \approx 1.7818$ .
- Method: Introduced limiting Krivine schemes, a generalization of classical rounding schemes where the dimension of the partition grows to infinity. The specific construction is a cubic–quintic scheme with cubic boundaries.
- Significance: This is the first improvement obtained by letting the dimension grow, answering a question posed by Braverman et al. [BMMN11] regarding whether higher dimensions aid approximation.
New Lower Bound (The Main Discovery):
- Result: $K_G \ge \frac{6\pi}{11} \approx 1.7135$ .
- Method: This is the first lower bound on $K_G$ not derived from constructing a specific hard instance. Instead, the proof establishes a universal obstruction on the performance of all Krivine schemes.
- Mechanism: The system derived a constraint on the correlation function $H(t) = b_1 t + b_3 t^3 + \dots$ of any scheme: $b_3 \ge 2b_1 - \frac{11}{6}$ . Combined with the optimality theorem of Naor and Regev [NR14], this implies the lower bound.
- Outcome: These results determine the tenths digit of $K_G$ to be 7.

B. Methodological Contributions

The paper documents a methodology for long-horizon AI research, demonstrating how to:

Convert an accumulation of failures into new mathematics.
Maintain a research state across sessions using file-based memory.
Integrate human intuition (research judgment) with AI technical execution.

4. Case Study: From Failure to Breakthrough

The paper reconstructs the discovery of the lower bound to illustrate the division of labor:

Initial Phase (AI Execution): The system initially searched for better upper bounds by exploring variants of limiting Krivine schemes. It repeatedly encountered a tradeoff: suppressing nonlinear terms weakened the leading term.
Human Intervention (Research Judgment): Observing the repeated failures, human operators recognized this as a potential general analytic obstruction rather than a lack of construction. They redirected the system to synthesize these failures and prove a universal obstruction, rather than continuing to search for new constructions.
AI Execution (Technical Proof): Once reframed, the system successfully developed the complete proof, reducing high-dimensional estimates to one-dimensional Gaussian inequalities and producing a computer-assisted certificate.

5. Analysis of Capabilities and Limitations

The paper presents a detailed analysis of the AI system's performance, highlighting a consistent asymmetry:

Strengths (Technical Execution): The system was exceptionally strong at executing well-specified high-level actions. It effectively drew upon vast mathematical knowledge to devise lemmas, construct proofs, and run experiments.
Weaknesses (Research Judgment & State Representation):
- Judgment: The system struggled to decide what to do next at the program level. It failed to recognize when a research direction was exhausted or when a common obstruction could be generalized into a theorem. It often prioritized numerical sharpness over theoretical generality.
- State Representation: The system's "working memory" (the summary file) was fragile. Critical caveats (e.g., that a numerical score was only for exploration) were lost during file merging, leading to the system treating a non-verified numerical result as a record-breaking bound for 25 days.

Origins of the Profile: The authors hypothesize this asymmetry stems from training data. Mathematical literature and training rewards heavily favor technical execution (definitions, proofs, final results) over the process of discovery (failed approaches, strategic pivots, and the evolution of doubt). Consequently, the AI excels at the "product" of mathematics but struggles with the "process."

6. Significance and Claims

The paper claims significance in three areas:

Mathematical Advancement: It provides the first lower bound on $K_G$ that does not rely on constructing a hard instance, fundamentally changing the approach to this problem.
AI Research Paradigm: It serves as a case study for human–AI collaboration in long-horizon research. The authors argue that full autonomy is currently limited by the AI's inability to maintain accurate research states and exercise high-level judgment. The most effective near-term model involves humans steering the agenda and AI executing technical tasks.
Methodological Insight: It highlights the necessity of training AI on records of mathematics as a process (including failures and strategic decisions) rather than just polished final papers to improve research judgment and state management.

The paper concludes that while AI is a powerful tool for technical execution in mathematics, human expertise remains essential for research judgment, agenda setting, and maintaining the integrity of the research state.