The study investigates how lexical perturbations—such as keyboard noise, character swaps, and filler insertion—affect large language models (LLMs) on reasoning benchmarks. Four open-weight instruction-tuned models and frontier models were evaluated, revealing that character-level perturbations significantly reduce accuracy, especially on multi-step reasoning tasks, while filler insertion has minimal impact. The authors attribute this asymmetry to Attention Diversion, where fragmented subword tokenization draws disproportionate attention in middle and final transformer layers; they demonstrate that both token content and attention allocation are coupled, making it difficult for inference-time repair strategies to fully recover performance.
By Jiaqian Zhu, Yang Zhang, Junhua Ding, Xiaowei Yu
The paper introduces Variance‑Calibrated Modulation (VCM), a training‑free pre‑decoding technique that reshapes language model probability distributions before truncation. VCM uses two dynamic mechanisms: a Contextual Searchlight via PMI to suppress stopwords and highlight context‑relevant tokens, and an Adaptive Self‑Debiasing that applies scale‑invariant penalization based on real‑time logit standard deviation. Experiments on open‑ended generation, factual QA, and mathematical reasoning show that VCM consistently reduces the likelihood trap, improving diversity, coherence, and reasoning accuracy with minimal computational cost.
By Yuanhao Ding, Meimingwei Li, Esteban Garces Arias, Matthias A{\ss}enmacher, Christian Heumann, Chongsheng Zhang
arXiv:2603. 18482v2 Announce Type: replace-cross Abstract: Standard decoding strategies for text generation, including top-$k$, nucleus sampling, and contrastive search, select tokens based on likelihood, restricting outputs to high-probability regions.
By Esteban Garces Arias, Nurzhan Sapargali, Christian Heumann, Matthias A{\ss}enmacher
The study investigates how textual neural models degrade when inputs contain noise such as typos, OCR errors, or dropped words. It finds that model performance decline is largely consistent across architectures under word‑level noise but diverges under character‑level noise, a difference attributed to tokenization rather than architecture. By applying a short contrastive training recipe, diverse encoders converge to a common robustness curve, enabling prediction of a model’s noise resilience and the ability to enhance robustness at specific noise scales through targeted training.
By Yefan Tao, Gerald Friedland, Luyang Kong
The paper investigates how six naturalistic and synthetic input perturbations affect decoder‑only language models at three levels: output behavior, hidden‑state geometry, and attention‑head function. Using GPT‑2 and Qwen2.5 checkpoints, the authors analyze layerwise geometry with centered kernel alignment and intrinsic dimension, and examine attention‑head responses in GPT‑2. They find that perturbation types produce distinct metric profiles that are not fully captured by output measures and vary across checkpoints, highlighting the need for multi‑level evaluation of robustness.
By Dun Li Chan, Emily Liu, Niyathi Allu, Christian Hoang
The paper investigates how six naturalistic and synthetic input perturbations affect decoder‑only language models at three levels: output behavior, hidden‑state geometry, and attention‑head function. Using GPT‑2 and Qwen2.5 checkpoints, the authors analyze layerwise geometry with centered kernel alignment and intrinsic dimension, and examine attention‑head responses in GPT‑2. They find that perturbation types produce distinct metric profiles that are not fully captured by output measures and vary across checkpoints, highlighting the need for multi‑level evaluation of robustness.
The paper re‑examines the GSM‑Symbolic benchmark, arguing that its claim of widespread reasoning failures in 25 LLMs is based on weak statistics. Using bootstrapped Generalised Linear Mixed Models on 20 open‑weight models, only eight show significant performance changes, and a systematic shift toward larger integers in the dataset explains many of these effects. The authors also uncover model‑specific failure modes such as variable binding fragility, arithmetic limits, and dual‑task interference, cautioning against blanket conclusions about LLM reasoning.
By Dominika Agnieszka D{\l}ugosz, Arlindo Oliveira, Natalia D\'iaz-Rodr\'iguez
The study investigates why language models exhibit systematic performance gaps across English dialects, a phenomenon termed the "dialect tax." Using parallel dialect corpora that preserve meaning while altering surface form, the authors confirm that models treat Standard American English and dialectal texts as semantically equivalent, yet find representational disparities that persist through tokenization, pre‑training, post‑training, and inference. Even a character‑level tokenizer does not eliminate input/output asymmetries or accuracy gaps, and dialect pairs produce more divergent gradient updates than unrelated Standard texts, indicating that dialectal content is harder for models to learn.
By Elle
arXiv:2608. 07525v1 Announce Type: cross Abstract: Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications.
By Pengfei Zhou, Jiajun Song, Zhiwei Tang, Yixing Ma, Xiaopeng Peng, Donghui Si, Yuhang Xu, Huiqi Song, Yiyuan Miao, Yichen Qian, Weihua Chen, Wangbo Zhao, Bohan Zhuang, Jiasheng Tang, Yang You
The paper introduces Latent Space Refusal Anchoring (LSR‑Anchoring), a training‑free technique that extracts a refusal direction from English prompts and applies it to the residual stream of instruction‑tuned models at inference time. The primary variant, Mean‑Activation Steering (MAS), works across several architectures (Llama‑3‑8B, Llama‑3.1‑70B, Mistral‑7B‑Instruct, Qwen2.5‑7B), restoring safety for low‑resource African languages with minimal performance loss, while a refined SAE‑Derived Steering (SDS) further reduces KL divergence without degrading legitimate prompt performance. The method shows positive transfer for Yoruba, Igbo, Igala, and Hausa, but fails for Arabic, suggesting a geometric mismatch rather than a data scarcity issue.
By Godwin Abuh Faruna
The paper presents an empirical study of factual errors in human-written text, focusing on corrections in newspaper articles to build a taxonomy of common mistakes such as kanji misconversions and unit errors. It evaluates large language models’ ability to detect these errors, finding that even advanced models like GPT‑5.4 achieve only a 52% word‑level F1 score on synthetic data, underscoring the difficulty of the task. The work highlights the gap in research on factual error detection in human writing compared to LLM hallucinations.
By Kazuma Iwamoto, Kazumasa Omura, Shotaro Ishihara
arXiv:2608. 03803v1 Announce Type: cross Abstract: Multilingual language models are deployed across a hundred or more languages, yet most benchmarks test whether a model can perform a task _in_ a language rather than whether it commands the language itself, conflating fluency with proficiency.
By Tom\'a\v{s} Burkert, Angelika Peljak-{\L}api\'nska, David Zelen\'y