AIMO Interpretability Challenge

arXiv:2607.13899 · cs.AI · Submitted 2026-07-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "AIMO Interpretability Challenge".

Jane: Here's the gist of the paper. The AIMO Interpretability Challenge proposes a competition to distinguish robust from spurious reasoning in frontier mathematical language models by assessing their internal mechanisms against structured variations in problem instances.

Tom: This challenge matters because standard accuracy benchmarks fail to reveal whether models rely on stable reasoning or brittle shortcuts, and the resulting benchmark aims to provide a lasting foundation for mathematical reasoning and interpretability research.

Jane: The gist The proposed competition targets this gap by measuring whether interpretability methods can identify mechanisms that generalize across problem instances and meaningfully distinguish robust models from brittle ones The scientific goal is to measure whether interpretability methods can identify mechanisms that generalize across problem instances and meaningfully distinguish robust models from brittle ones in frontier mathematical language models.

Tom: Competition Description The competition is designed to operationalize the scenario where a research lab may choose among several models with similar benchmark accuracy but different reliability Participants will be asked to create a submission system S that assigns one of two categories to each triple (X, M, ∆).

Title and authors: Tom: So we’re looking at this paper called "AIMO Interpretability Challenge". It’s trying to figure out how to tell the difference between models that are actually reasoning well and just models that are giving us good answers but relying on some kind of trick.

Jane: That makes sense, Tom. Basically, standard accuracy benchmarks don't show us if a model is using stable reasoning or just memorizing specific patterns for the test set. It’s about looking at the internal mechanisms to see if they generalize across different problems.

Lu: Exactly what I find interesting is that this competition moves past just getting a high score on a test and tries to measure whether interpretability methods can actually spot the underlying reasoning structure itself, rather than just observing the output >

Meng: So it’s not just about getting an answer right; it’s about assessing reliability when you have to pick one model out of several that all look equally good on paper >

Lalam: Yeah, I see what they are saying. They want to operationalize that idea where a lab might have three models, two of which are fine for the main task but one is actually brittle >

The paper's summary: Tom: The paper outlines this competition by asking participants to build a system, let's call it S, that takes a problem and assigns it one of two labels: robust or non-robust based on whether the model handles variations correctly >

Jane: So the core task is formal, they are asking if if a model gets one input correct under some perturbations, will it also get another input correct under those same changes >

Lu: They’re using symbolic reasoning annotations from problems like AIMO two and AIME to create these problem shifts, which helps them generate a lot of variations that are structurally related to the original problem > <ref:2607.13899#pg1>

Meng: I see how that creates unconstrained volumes of adversarial problem variants; it means they aren't just testing small tweaks; they’re testing structured variations based on the math itself >

Lalam: The goal is to test whether an interpretability method can identify the mechanism that makes a model robust across these shifts, which is what this whole AIMO Interpretability Challenge is focused on >

Tom: So it’s connecting interpretability methods directly to assessing how robust frontier models are against structured variation in mathematical problems >

The paper's improvements: Jane: What they suggest as an improvement over existing work is shifting the focus from just evaluating methods in simple settings to testing whether those methods can actually determine if a strong performance comes from stable reasoning or brittle shortcuts >

Lu: That’s the novelty here; existing benchmarks tend to look at interpretation in rudimentary or synthetic settings, but this challenge tests if interpretability can actually distinguish robust models from brittle ones on complex reasoning tasks >

Tom: They are using the AIMO two AIMO three JMO problems and two years of AIME submissions as their data foundation, which is a big step up from what most current methods use > <ref:2607.13899#pg1>

Meng: And they’re providing baselines like a simple probing classifier and one based on internal confidence signals during chain-of-thought generation to help participants start measuring anything >

Lalam: The competition results will lead to an open robustness benchmark, which I think is crucial because it provides a foundation for standard benchmarking in mathematical reasoning and interpretability going forward >

Conclusion: Tom: So the final summary is that this challenge aims to measure whether interpretability methods can identify those mechanisms that generalize across problem instances and meaningfully distinguish robust models from brittle ones >

Jane: It sets up a way for researchers to audit high-stakes reasoning systems because, in a real lab setting, you might have several models with similar accuracy but different reliability levels >

Lu: The implication is that this work will support future research focused on developing more robust frontier models by giving us a concrete way to test their generalization capabilities under pressure >

Meng: From my side, the main point is that we need these types of benchmarks because models shipped in production need to be reliable and not just lucky on a specific set of inputs >

Lalam: So, by running the AIMO Interpretability Challenge, they hope to create a lasting foundation for understanding what makes mathematical reasoning truly robust in AI systems >

Michal štefšanik, Philipp Mondorf, Andreas Waldis, Qianying Liu, Chuan Yang, ♈ Liu, ▷ Frieder, Κ Kucharżˇ

National Institute of Informatics, Japan · Munich Center for Machine Learning / MaiNLP LMU · University of Tübingen · Fuzhou University · University College London

cs.AI

Submitted: 2026-07-15

Updated: 2026-10-20

Comments: Accepted Competition at NeurIPS 2026; v2 with a completed list of contributors

Code: https://github.com/aimo-interp/baselines

Project page: https://aimo-interp.github.io

License: http://creativecommons.org/licenses/by/4.0/

The gist: The AIMO Interpretability Challenge proposes a competition to distinguish robust from spurious reasoning in frontier mathematical language models by assessing their internal mechanisms against

Key concepts

Robust vs. Brittle Reasoning: This refers to the core question: does a model solve mathematical problems reliably across different but related problem instances, or does it rely on specific, fragile patterns that fail when the problem structure is slightly changed? Robust models use stable reasoning mechanisms.
Distribution Shifts ($\Delta$): These are structured variations applied to original math problems using their symbolic representations. They create a set of perturbations that test how well a model's reasoning holds up when the input problem is systematically altered in predictable ways.
Interpretability Methods: These are techniques used to examine the internal workings or mechanisms of a language model. The challenge aims to see if these methods can uncover underlying, generalizable reasoning patterns that explain why a model performs well or fails under various conditions.

Terminology

Interpretability — 4× in this paper · explained in 15 episodes
Benchmark — 2× in this paper · explained in 2 episodes
Robustness — 2× in this paper · explained in 6 episodes
Accuracy — 1× in this paper · explained in 1 episode
Adversarial Robustness — 1× in this paper · explained in 4 episodes
Benchmarking — 1× in this paper · explained in 4 episodes
Generalization — 1× in this paper · explained in 13 episodes

Summary

The AIMO Interpretability Challenge proposes a competition to distinguish robust from spurious reasoning in frontier mathematical language models by assessing their internal mechanisms against structured variations in problem instances. This challenge matters because standard accuracy benchmarks fail to reveal whether models rely on stable reasoning or brittle shortcuts, and the resulting benchmark aims to provide a lasting foundation for mathematical reasoning and interpretability research.

The gist

The proposed competition targets this gap by measuring whether interpretability methods can identify mechanisms that generalize across problem instances and meaningfully distinguish robust models from brittle ones The scientific goal is to measure whether interpretability methods can identify mechanisms that generalize across problem instances and meaningfully distinguish robust models from brittle ones in frontier mathematical language models.

Competition Description

The competition is designed to operationalize the scenario where a research lab may choose among several models with similar benchmark accuracy but different reliability Participants will be asked to create a submission system S that assigns one of two categories to each triple (X, M, ∆).

The task is formally defined as determining whether M(∆X) will also be correct for all perturbations under ∆ if M(P) is correct.

The distribution shifts ∆ are instantiated by applying one or more perturbation types on the original problem using its symbolic representation.

Novelty and Data

The novelty of this competition lies in connecting interpretability with an objective that standard benchmarks do not directly target, which is assessing whether frontier models that succeed on most complex reasoning tasks do so robustly Existing interpretability benchmarks focus on evaluating methods in rudimentary and/or synthetic settings, whereas this challenge tests whether interpretability can determine whether strong performance reflects stable, generalizable reasoning mechanisms or brittle strategies.

The competition uses three main data components

Original olympiad-level mathematical problems, drawn from AIMO 2, AIMO 3, JMO (Japanese Mathematical Olympiad), and two last years of AIME.
Symbolic reasoning annotations created for the original problems, similar to the functional variation introduced by GSM-Symbolic.

Tasks and Metrics

The central task is to decide if a model responds to a given problem robustly with respect to a pre-defined set of distribution shifts.

The primary metric will be accuracy on the held-out test set, which is the fraction of examples for which a submission correctly classifies the model as robust or non-robust.

In addition to global accuracy, final reports will provide evaluations dissecting performance across several dimensions: (1) for public and private sets; (2) for each perturbation type; (3) for each model; and (4) for the previously published and unpublished problems.

Baselines and Feasibility

The competition provides at least three baselines, including a simple probing classifier, a classifier based on model-internal confidence signals during chain-of-thought generation, and a data attribution method assessing robustness as the impact and scale of memorization.

These baselines have shown that both the probing classifier and uncertainty-based classifier achieve up to 58.37 (±7.5) and 69.23% (±10.13) of accuracy, respectively.

The challenge is deemed tractable because the competition provides a clear extrinsic signal of success through labelled datasets that support experimentation with a wide range of possible methods and strategies.

Code of Ethics and Resources

The competition is aligned with the NeurIPS Code of Ethics, aiming to improve reliability assessment by helping practitioners identify brittle systems before they are used in consequential applications.

The organizers will provide resources such as large-scale compute, unrestricted access to best-performing AI systems from AIMO in standalone containers, and a standardized evaluation infrastructure with eight large-scale GPUs.

Participants can request up to 10,000 H200 GPUhours of compute if justified in their proposal.

The competition schedule spans from July 15 to December 12, with a final submission deadline on November 15.

Improvements for AI systems

Develop a method to identify which models solve problems robustly by using interpretability methods that can identify mechanisms that generalize across problem instances. This allows systems to distinguish robust models from brittle ones by examining their internal representations rather than just final answer accuracy.
Create an open robustness benchmark and baseline systems based on the competition results. This provides a lasting foundation for standard benchmarking in mathematical reasoning and interpretability so future frontier models can be reliably assessed for generalization.
Implement a system that uses symbolic reasoning annotations to generate unconstrained volumes of adversarial problem variants. This enables the system to test whether a model's success reflects stable, generalizable reasoning mechanisms or brittle strategies that fail under structured variation.

Abstract

We propose the AIMO Interpretability Challenge, a competition on distinguishing robust from spurious reasoning in frontier mathematical language models based on the models' internal mechanisms. The challenge is motivated by a central limitation of standard reasoning benchmarks: strong final-answer accuracy does not reveal whether a model relies on stable reasoning mechanisms or exploits brittle reasoning shortcuts. Building on AI Mathematical Olympiad (AIMO) problems and submissions, together with resources from the Fields Model Initiative, the competition will provide (1) newly-published olympiad-level math reasoning problems and their symbolic representations, allowing generation of novel functional variants, (2) access to frontier reasoning models, and (3) assessments of models' adversarial robustness on these problems. Participants will use these resources, along with our computing infrastructure support, to develop methods for identifying which models solve problems robustly. Our competition will also create a new, open robustness benchmark and baseline systems, aiming to provide a lasting foundation for standard benchmarking in mathematical reasoning and interpretability. Scientifically, the competition connects interpretability and generalization research around a central question in AI research: can we determine if, and to what extent, the decision-making of frontier AI models is generalizable and thus, reliable?