aiqbench analysis
This study empirically demonstrates that character-level lexical corruption degrades LLM reasoning by fragmenting subword tokens, which disproportionately distracts attention mechanisms in deeper transformer layers. The research reveals a critical coupling between corrupted token content and misallocated attention, explaining why standard inference-time repair strategies like spell-checking or self-correction often fail to restore performance. These findings provide essential insights for improving model robustness against realistic input noise and inform the development of more resilient reasoning architectures.
Lexical Perturbations Disrupt LLM Reasoning: An Empirical Study of Attention Diversion
arXiv:2608.22140v1 Announce Type: cross Abstract: Large Language Models (LLMs) achieve strong reasoning performance, but their robustness to realistic lexical corruption remains poorly understood. We evaluate four open-weight instruction-tuned models and frontier models across four reasoning benchmarks under keyboard noise, character swaps, and filler insertion. Character-level perturbations substantially degrade accuracy, especially on multi-step reasoning tasks, while filler insertion has little effect. We trace this asymmetry to Attention Diversion: lexical corruption fragments subword tokenization, and the resulting fragments attract disproportionate attention mass, concentrated in middle and final transformer layers. Length-matched controls confirm that fragmentation, not prompt length, drives the loss. A factorial intervention then shows why the damage is hard to undo: fragmentation corrupts token content and attention allocation together, and the two are coupled. Restoring clean attention while the content remains corrupted is actively harmful, restoring content alone is insufficient, and only restoring both recovers a substantial share of the gap. This coupling explains why inference-time strategies, including chain-of-thought prompting, spell-checking, self-repair, and stronger repair models, fail to consistently recover performance: each addresses one channel at a time. Code and data are available at https://github.com/Jiaqian-Janelle/Attention-Diversion
Abstract:Large Language Models (LLMs) achieve strong reasoning performance, but their robustness to realistic lexical corruption remains poorly understood. We evaluate four open-weight instruction-tuned models and frontier models across four reasoning benchmarks under keyboard noise, character swaps, and filler insertion. Character-level perturbations substantially degrade accuracy, especially on multi-step reasoning tasks, while filler insertion has little effect. We trace this asymmetry to Attention Diversion: lexical corruption fragments subword tokenization, and the resulting fragments attract disproportionate attention mass, concentrated in middle and final transformer layers. Length-matched controls confirm that fragmentation, not prompt length, drives the loss. A factorial intervention then shows why the damage is hard to undo: fragmentation corrupts token content and attention allocation together, and the two are coupled. Restoring clean attention while the content remains corrupted is actively harmful, restoring content alone is insufficient, and only restoring both recovers a substantial share of the gap. This coupling explains why inference-time strategies, including chain-of-thought prompting, spell-checking, self-repair, and stronger repair models, fail to consistently recover performance: each addresses one channel at a time. Code and data are available at this https URL
Accepted to EMNLP 2026 (Main Conference). 9 pages main text, 12 figures, 20 tables
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2608.22140 [cs.CL]
(or arXiv:2608.22140v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2608.22140 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 296 words
Related coverage
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · Oct 8, 2026
PIRAMID: Progress and Plans
LessWrong · Jul 27, 2026
Reflections on unlearning and inoculation
LessWrong · Sep 20, 2026
Boosting LLM Exploration via Weak-Model Guidance in RLVR
arXiv · Aug 28, 2026