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 DRY, a sampling-time logit adjustment that penalizes token generation only when it would extend the current suffix into an exact repetition of an earlier span, thereby preventing verbatim loops in large language model outputs. Experiments across models ranging from 1.5B to 120B parameters and various prompt families show that DRY cuts suffix-extension rates by 47% and improves lexical diversity, while preserving benchmark performance. The method has been adopted by popular open-source LLM inference frameworks, indicating its practical relevance.
By Philipp Emanuel Weidmann, Allen Roush, Judah Goldfeder, Sanjay Basu, Ravid Shwartz-Ziv
The paper introduces ADU, a fine‑grained training framework that unlearns sensitive information from large language models by decoupling contextual attention pathways instead of erasing tokens. ADU exploits the distinction between local and global attention heads to identify and suppress attention paths that retrieve persistent sensitive anchors, while preserving local‑attention structure and overall language modeling performance. Evaluation on the TOFU and WMDP benchmarks shows ADU achieves superior forget quality (0.93 on TOFU) and retains 92.9% of model utility compared to 81.9% for existing baselines, with fewer side effects in benign contexts.
By Xunlei Chen, Qirui Ye, Yuang Li, Yi Gong, Zhaokun Wang, Wenyi Li, Shiyao Guo, Jinyu Guo
arXiv:2606. 08432v1 Announce Type: new Abstract: On-policy distillation (OPD) has become a central post-training tool for large language models (LLMs), providing dense per-token teacher supervision along the student's own rollouts.
By Li Jiang, Haoran Xu, Yichuan Ding, Amy Zhang
arXiv:2607. 22694v1 Announce Type: new Abstract: Attention collapse in autoregressive language models -- manifested as repetitive token loops where the model becomes trapped in self-reinforcing attractors -- is a persistent pathology that existing decoding-time heuristics fail to address at its root cause.
By Wenjie Fan, Bin Ma, Dong Li
arXiv:2607. 01792v1 Announce Type: cross Abstract: While decoder-only LLMs excel at a vast array of natural language tasks, it suffers from an asymmetric information flow induced by causal attention: later tokens are richer in contextual grounding than earlier ones.
By Andikawati P Widjaja, Yongjun Kim, Hyounghun Kim, Jaeho Lee
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
The paper introduces DLM-One, a score‑distillation framework that enables one‑step sequence generation with continuous diffusion language models (DLMs). By aligning a student model’s outputs with a pretrained teacher DLM’s score function in the forward‑diffused noisy space, DLM-One removes the need for iterative refinement. Experiments across various DLM architectures show up to ~2000× speedup in sampling steps and ~500× in wall‑clock time while retaining competitive performance, and the authors propose an adversarially‑regularized two‑stage training scheme to mitigate student degeneration.
By Tianqi Chen, Shujian Zhang, Mingyuan Zhou
arXiv:2606. 15521v1 Announce Type: cross Abstract: Tokenization introduces representational redundancy: under a fixed token vocabulary, every byte string admits many valid token encodings, or segmentations, that decode to the same surface string.
By Kanishk Jain, Matthew Day, Tankut Can
While decoder-only LLMs excel at a vast array of natural language tasks, it suffers from an asymmetric information flow induced by causal attention: later tokens are richer in contextual grounding than earlier ones. A simple and effective remedy is prompt repetition -- just appending a second copy of prompt before generation can redistribute grounding across positions and improve reasoning performance.
arXiv:2605. 26436v2 Announce Type: replace-cross Abstract: Discrete masked diffusion language models such as LLaDA generate text through iterative denoising, where mask tokens are progressively replaced with predicted tokens.
By Lin Yao
arXiv:2606. 09456v1 Announce Type: new Abstract: On-Policy Distillation (OPD) has become a core technique in the post-training of Large Language Models (LLMs) for transferring knowledge from domain experts to student models.
By Yifan Niu, Han Xiao, Dongyi Liu, Zelong Wang, Dihong Gong, Yasheng Wang, Jia Li