Source-linked AI summary
Robust Reasoning Benchmark
Pavel Golikov, Evgenii Opryshko, Gennady Pekhimenko, Mark C. Jeffrey
TL;DR
Standard mathematical benchmarks may overstate reasoning robustness because performance can depend on textual presentation. The paper introduces deterministic structural perturbations and a sequential-problem protocol, finding fragility in open-weights models and identifying attention dilution within a single query. It argues that contextual resets and appropriate reasoning-task boundaries are important directions for more reliable reasoning architectures.
Problem
Existing mathematical benchmarks and adversarial evaluations provide limited evidence about whether models perform robust reasoning rather than pattern matching under presentation changes.
Method
The paper evaluates eight models with RRB’s 13 deterministic structural transformations and a sequential cognitive overload protocol, supplemented by mechanistic attention analysis.
Results
Open-weights models exhibit structural fragility, and models from 7B to 120B parameters show reasoning degradation as prior chain-of-thought pollutes subsequent problems.
Takeaways & Limitations
Reliable reasoning may require explicit contextual resets within chain-of-thought and architectures that isolate working memory between sequential reasoning steps.
Takeaways & Limitations
The evaluation focuses primarily on competitive mathematics, so generalization to legal, medical, coding, and other high-stakes reasoning tasks remains unexplored.
Abstract
from arXiv · showhide
While Large Language Models (LLMs) achieve high performance on standard mathematical benchmarks, their problem-solving abilities depend on the context and textual formatting. We introduce the Robust Reasoning Benchmark (RRB), a pipeline of 13 deterministic textual perturbations applied to AIME 2024 and AIME 2025. Evaluating 8 state-of-the-art models, we find that frontier models are largely resilient, with the notable exception of Claude, which categorically refuses many transformed prompts. Open-weights reasoning models exhibit a range of failure modes under structural noise (cognitive thrashing, tokenization breakdown, and reasoning collapse), with up to 54% average accuracy drops across perturbations and up to 100% on some. We further study one of these failure modes in isolation: attention dilution caused by the model's own chain-of-thought. By tasking models with solving multiple independent mathematical problems sequentially within a single context window, we identify Intra-Query Attention Dilution. Open-weights models ranging from 7B to 120B parameters exhibit accuracy decay on subsequent problems, suggesting that intermediate reasoning steps progressively pollute standard dense attention mechanisms. We argue that in order to achieve reliable reasoning, future architectures need to integrate explicit contextual resets within models' own chain-of-thought, leading to open research questions regarding the optimal granularity of reasoning tasks.
1 Introduction
The paper introduces deterministic structural perturbations to test whether mathematical reasoning is robust to presentation changes, then isolates degradation caused by prior chain-of-thought within one query.
- Standard benchmark success may reflect brittle pattern matching rather than robust algorithmic reasoning.
- Existing adversarial evaluations often confound reasoning robustness with altered arithmetic, increased difficulty, or irrelevant context.
- RRB applies 13 reversible structural transformations without changing mathematical values, logic, or final answers.
- Open-weights models show cognitive thrashing, tokenization breakdown, and attention dilution, while frontier models are largely resilient and Claude often refuses transformed prompts.
- Sequentially solving independent problems within one query identifies Intra-Query Attention Dilution as accuracy on the final problem degrades after prior reasoning steps.
- The paper argues that reliable reasoning requires explicit contextual resets to isolate working memory within chain-of-thought.
2 Background and Related Work
Prior work shows that static mathematical benchmark accuracy can conceal fragile reasoning, motivating evaluations that test invariance to surface-level and informational perturbations without simply increasing task difficulty.
- GSM8K, MATH, and AIME benchmarks, together with Chain-of-Thought and Tree-of-Thought methods, have driven mathematical reasoning progress.
- Reasoning is framed as multi-step symbol manipulation connecting premises to conclusions, distinct from rapid pattern-based neural inference.
- Robust reasoning should remain invariant to surface-level perturbations while preserving the underlying concepts and operations.
- Prior robustness studies inject irrelevant information or lexical shifts, but such interventions can test more than reasoning alone.
- Numerical and variable changes test memorization but also modify the mathematical constraints of the original problems.
- Chain-of-thought perturbation studies examine derailment of intermediate reasoning, whereas this paper perturbs the initial problem presentation.
- Difficulty-based evaluations increase mathematical or algorithmic complexity, unlike transformations designed to maintain constant task difficulty.
3 The Robust Reasoning Benchmark
The Robust Reasoning Benchmark uses 13 deterministic textual transformations that preserve problem content while testing whether models can decode structurally hostile presentations.
- The benchmark organizes 13 strictly textual, deterministic input perturbations into four broad categories.
- Figure 2 illustrates the transformations on a sample mathematical query, with long mappings and target problems truncated for brevity.
- Semantic and Lexical Substitutions: Semantic and lexical substitutions use logically equivalent aliases, double negations, or explicitly defined wrappers to obscure linguistic structure.
- Interleaved Contexts: Interleaved Contexts weave two distinct problems together at line, word, or symbol granularity while instructing the model to solve only one.
- Reversals: Reversal transformations reverse sentence, word, or symbol order according to delimiters such as periods or spaces.
- Visual and Spatial Encoding: Visual and spatial encodings map one-dimensional problem text onto grids using transformations such as Rail Fence and rectangle-perimeter layouts.
4 Methodology
The methodology preprocesses AIME data, applies reversible transformations under invariance constraints, and evaluates eight models using accuracy and attention-allocation analyses.
- Preprocessing removes information loss from string reversals and avoids measuring low-level parsing or tokenizer artifacts.
- Each transformation must preserve information-theoretic content and be deterministic, algorithmically reversible, and programmatically invertible.
- Models receive a standard step-by-step chain-of-thought prompt, while sequential overload measures accuracy only on the last problem in a single query.
- Problem-solving accuracy is the percentage of strictly correct final answers, with maximum-token outputs counted as failures.
- Attention analysis partitions prompt and generated tokens into instruction, distractor, and target regions, then measures allocation to each region.
- Macro-level distraction averages attention mass across all heads and target tokens at each layer.
- The evaluation covers five open-weights models and three proprietary APIs on AIME 2024 and AIME 2025.
5 Results and Discussion
Across structural perturbations, proprietary models remain comparatively resilient while open-weights models show substantial accuracy degradation. Sequential distractors also reveal intra-query attention dilution, with attention bleeding into prior problems and system instructions.
- Robustness Against Perturbations: 53% and 47% average accuracy drops affect Nemotron-7B and Qwen3-30B-A3B-Thinking-2507, respectively, despite strong baseline accuracy.
- Robustness Against Perturbations: GPT-5.4 and Gemini 3.1 Pro suffer average performance degradations of only 3% and 8%, respectively, under structural noise.
- Robustness Against Perturbations: Open-weights models universally degrade on Syntactic Distortions and Visual Encodings, with Qwen3 and DSR1 reaching 0% accuracy on some character-level tasks.The reported failures are associated with disrupted Byte-Pair Encoding boundaries and subword semantic priors.
- Robustness Against Perturbations: Claude 4.6 Opus suffers a 52% average accuracy drop primarily because it refuses many prompts involving symbol manipulation.The paper attributes the observed drop to refusals rather than reasoning failure and hypothesizes aggressive safety-filter behavior.
- Intra-Query Context Management: All evaluated open-weights models from 7B to 120B parameters experience performance degradation on subsequent problems in a shared context.The paper describes this as intra-query attention dilution within standard dense attention.
- Intra-Query Context Management: Middle and late layers allocate 10% to 16.5% of attention to distractors, while deeper layers send up to 32.9% to the initial System prompt.Figure 5 tracks attention across System, Distractor, and Target regions during target-solution generation.
- Intra-Query Context Management: The findings motivate explicit context management to mitigate attention drift across sequential reasoning steps.The paper connects its intra-query evidence to explicit scaffolding and task-boundary mechanisms in agent frameworks.
6 Limitations
The benchmark evaluates structural reasoning fragility primarily in competitive mathematics, leaving generalization to other high-stakes reasoning domains unresolved.
- The evaluation primarily focuses on competitive mathematics, so generalization to legal, medical, and coding tasks remains to be explored.
7 Conclusion
The paper argues that mathematical reasoning in current LLMs is overfit to standard textual presentation and that dense attention fails to isolate working memory across sequential reasoning steps. It therefore advocates native contextual resets and task-boundary compartmentalization for more reliable multi-step deduction.
- Current LLM reasoning is often overfit to standard textual presentation rather than abstract logical manipulation.
- Intra-query attention dilution appears in open-weights models from 7B to 120B parameters as intermediate reasoning steps pollute the context window.
- Standard dense attention struggles with robust algorithmic reasoning because it fails to isolate working memory between sequential reasoning steps.
- Reliable multi-step deduction requires native contextual resets and task-boundary compartmentalization to flush intermediate states and preserve logical coherence.
A.1 A: Methodology Details
The methodology applies deterministic textual transformations that rearrange, reverse, remap, or spatially encode user queries while preserving the underlying task. It also documents development artifacts and measures output-token efficiency by task.
- Reversal transformations: Word, sentence, and symbol reversals redefine the query ordering at progressively different textual units.Words are sequences separated by spaces, while sentences are sequences separated by periods.
- Transformation families: 13 transformations include reversals, interleaved contexts, symbol-level rearrangements, semantic remappings, and visual grid encodings.The transformations operate at word, sentence, symbol, line, semantic, and spatial levels.
- Interleaved contexts: Interleaved-context tasks combine two problem statements and instruct the model to solve only Problem A.Interleaving is evaluated at line, word, and symbol granularity, with repeated padding when statements differ in length.
- Semantic remappings: The evaluation also includes Opposites and Wrappers, which remap terms using definitions supplied inside a ’defyn’ block.The remappings are embedded within the user query rather than changing the mathematical problem itself.
- Visual encodings: Visual encodings include Rail Fence, Rectangle Perimeter, Snake Vertical, and Snake Horizontal layouts.These transformations provide the encoded query as a grid or perimeter representation with explicit structural markers where specified.
- Efficiency measurement: Figure 6 reports average output token length by task, with each bar labeled in thousands of tokens.The figure is used to compare reasoning efficiency across transformations.
A.2 Prompts
The prompt protocol first directs the model to decode the supplied transformation, then solve the recovered mathematical problem and format the final answer in boxed notation.
- Baseline prompt: The baseline system prompt requests step-by-step mathematical reasoning and a final answer within \\boxed{}.This establishes the common response format for the evaluation.
- Decoding protocol: The solution protocol instructs the model to read the transformation rule and reverse it on the transformed input.Decoding precedes mathematical problem solving.
- Answer protocol: After recovering the original problem statement, the model is instructed to solve it and place the final answer within \\boxed{}.The protocol separates transformation reversal from mathematical reasoning.
A.3 Cognitive Thrashing
Cognitive thrashing describes unproductive reasoning loops triggered by structural noise, while the paper frames context partitioning as a trade-off between interference and computational overhead. The authors leave scaling and generalization as open questions.
- Cognitive thrashing: Nemotron-7B generates 27K tokens on Rectangle Perimeter while achieving 0% accuracy.This exemplifies a long, unproductive output rather than a clean failure.
- Cognitive thrashing: Qwen generates 13K tokens on Snake Vertical with 1% accuracy.The result illustrates that structural noise can produce extensive reasoning with almost no successful solutions.
- Failure mechanism: The paper attributes these loops to brittle stopping criteria and logical-progression heuristics under structural noise.The proposed failure pattern is hallucination loops rather than structured problem-solving.
- Open questions: The optimal sub-task granularity and scaling of atomic reasoning capacity remain unknown.The authors propose quantifying safe cognitive load and designing architectures that determine sub-task boundaries dynamically.
- Granularity trade-off: Insufficient decomposition permits attention dilution and cognitive thrashing, whereas excessive decomposition incurs overhead from repeated state summaries and context flushing.The paper presents this as a trade-off in choosing contextual reset granularity.
- Scope boundary: The paper cautions that generalization beyond mathematical reasoning and the maximum safe cognitive load of one transformer forward pass require further study.This defines the current scope boundary of the reported findings.