Robust Reasoning Benchmark
| Authors | Pavel Golikov et al. |
| Year | 2026 |
| HF Upvotes | 1 |
| arXiv | 2604.08571 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
While Large Language Models (LLMs) achieve high performance on standard mathematical benchmarks, their underlying reasoning processes remain highly overfit to standard textual formatting. We propose a perturbation pipeline consisting of 14 techniques to evaluate robustness of LLM reasoning. We apply this pipeline to AIME 2024 dataset and evalute 8 state-of-the-art models on the resulting benchmark. While frontier models exhibit resilience, open weights reasoning models suffer catastrophic collapses (up to 55% average accuracy drops across perturbations and up to 100% on some), exposing structural fragility. To further disentangle mechanical parsing failures from downstream reasoning failures, we strictly isolate the models' working memory capacity by forcing models to solve multiple unperturbed mathematical problems sequentially within a single context window. Our results indicate that open weight models ranging from 7B to 120B parameters and Claude Opus 4.6 exhibit accuracy decay on subsequent problems. This degradation demonstrates that intermediate reasoning steps permanently pollute standard dense attention mechanisms. We argue that to achieve reliable reasoning, future reasoning architectures must integrate explicit contextual resets within a model's own Chain-of-Thought, leading to fundamental open questions regarding the optimal granularity of atomic reasoning tasks.
Engineering Breakdown
Plain English
This paper introduces a systematic robustness evaluation framework for LLMs by creating 14 different perturbation techniques applied to the AIME 2024 mathematical reasoning dataset. The authors tested 8 state-of-the-art models and found a stark divide: frontier models (like GPT-4) maintain reasonable performance under perturbations, while open-weight reasoning models experience catastrophic failures with accuracy drops up to 55% on average and 100% on specific perturbations. The core finding is that many LLMs have learned to exploit superficial patterns in standard problem formatting rather than developing robust reasoning, making them fragile when the presentation of problems changes even slightly.
Core Technical Contribution
The primary technical contribution is a principled perturbation pipeline with 14 distinct techniques designed to stress-test the reasoning robustness of language models without changing the underlying mathematical problem. Rather than just applying random noise, the authors systematically vary problem formatting, notation, working memory constraints, and presentation order to isolate whether failures stem from parsing brittleness versus genuine reasoning limitations. This provides a structured methodology that goes beyond existing robustness benchmarks by targeting the specific vulnerability pattern of mathematical reasoning—overreliance on textual surface features. The secondary contribution is the empirical discovery of a large capability gap between closed frontier models and open-source reasoning models in robustness, suggesting that current fine-tuning and scaling approaches for open models don't transfer robustness properties effectively.
How It Works
The pipeline starts with the original AIME 2024 problem set and applies 14 orthogonal perturbations such as reformatting mathematical notation, reordering problem steps, injecting irrelevant information, and constraining working memory through forced intermediate output. For each perturbation, models are evaluated in isolation to measure the accuracy delta from the clean baseline. A critical component is the working memory isolation step mentioned in the abstract—models are forced to generate intermediate reasoning steps and solve subproblems sequentially, which prevents them from leveraging shortcut heuristics tied to problem length or formatting cues. The pipeline produces a robustness score aggregating performance across all 14 perturbations, allowing comparison of models along both average robustness and failure mode specificity. This design enables attribution of failures to either shallow parsing issues (caught early) or deep reasoning deficits (persisting across perturbations).
Production Impact
For teams deploying LLMs in real-world mathematical, scientific, or logical reasoning tasks, this work signals a critical risk: models that score well on standard benchmarks may fail silently when input formatting deviates from training distribution, which is inevitable in production where users, systems, and data sources introduce natural variation. Adopting this framework means implementing a multi-perturbation evaluation stage before deploying any reasoning-critical model, adding evaluation latency but catching fragile models before they reach users. For teams building internal tools, this suggests that fine-tuning or prompt engineering should include adversarial formatting variations in the training/validation set to improve robustness—a measurable but non-trivial engineering investment. The frontier model resilience suggests that larger models and better training procedures (possibly including robustness-focused objectives) do improve this property, so model selection becomes more critical: upgrading from open-weight to closed frontier models may be justified purely for robustness, not just raw capability.
Limitations and When Not to Use This
The paper is limited to mathematical reasoning on AIME problems, so findings may not generalize to other reasoning domains like code generation, logical deduction, or natural language inference where different shortcut patterns may dominate. The 14 perturbations are heuristically designed rather than derived from a formal theory of robustness, meaning some critical failure modes may be missed and the perturbation set may not be minimal or complete. The abstract is truncated and doesn't fully explain how working memory constraints are enforced or whether that technique itself introduces confounding factors (e.g., does forcing intermediate steps help some models and hurt others?). Additionally, the paper doesn't propose a method to improve robustness—it only measures and diagnoses it—so practitioners learn where models fail but lack concrete techniques to fix those failures beyond retraining, which limits immediate actionability.
Research Context
This work builds on a growing line of robustness research in NLP and ML that examines out-of-distribution generalization, building on earlier work like adversarial robustness in vision and distribution shift studies in NLP. It contributes directly to the benchmark evaluation ecosystem by extending the mathematical reasoning evaluation paradigm from single-answer accuracy to robustness-aware scoring, similar to how other papers added stress tests to standard benchmarks. The findings align with recent empirical observations that LLMs exhibit brittle behavior under distribution shift and support the hypothesis that scaling alone doesn't solve robustness—a key open question in AI safety and reliability. This work opens a research direction toward robustness-by-design training methods and could inform future architectural choices for reasoning-focused models, particularly for open-weight systems that currently lag significantly.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::