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.
arXiv:2607. 20524v1 Announce Type: new Abstract: Mean cross-positional attention degradation is widely reported in transformer interpretability, yet whether it causally limits contextual retrieval remains untested.
By Sagar Dangal, Manoj Shakya
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
arXiv:2606. 07604v1 Announce Type: cross Abstract: Analyzing attention weights has become a standard approach for interpreting the information flow of Large Language Models (LLMs).
By Harry Jake Cunningham, Nicola Muca Cirone
arXiv:2607. 00852v1 Announce Type: cross Abstract: This work studies the hidden-state inversion problem: recovering the original input token sequence of a decoder-only language model from its last-layer hidden states.
By Miko{\l}aj S{\l}owikowski, Maciej Witold Majewski
arXiv:2607. 09204v1 Announce Type: cross Abstract: Pretrained language models often exhibit structured weight spectra, suggesting that training may repeatedly produce similar layerwise and component-wise organization.
By Konstantin Garbers, Nicholas Oh
The paper introduces SAGE, a lightweight method that guides decoder attention in vision-language models away from prompt‑invariant attention sinks (PIS) toward query‑dependent regions of interest (ROIs). By using token‑aligned ROI masks from vision backbones like CLIP, ViT, and DINOv3, SAGE improves visual grounding and reduces hallucinations across various VLM families. Experiments show consistent gains on public vision‑language benchmarks, especially in fine‑grained visual discrimination tasks where localized evidence is essential.
By Jeonghyo Song, YoungJoon Yoo
The paper introduces a three-level evaluation framework—behavioral deployment, LM-head readout, and probe recoverability—to distinguish whether a language model fails a syntactic test by not encoding structure or by failing to use it. Using a trilingual control-dependency benchmark, the authors find that probe recoverability consistently exceeds LM-head readout, which in turn exceeds behavioral deployment across seven models and three languages, with the largest gap observed in Qwen3-0.6B Instruct. Layer-localized activation patching shows that instruction tuning shifts the decoded layer later, suggesting decoding favors surface shortcuts and that behavioral evaluation understates what models encode while probing alone overstates what they deploy.
By Zhenyan Lu, He Wang, Xiaohui Huang
arXiv:2606. 01060v1 Announce Type: cross Abstract: Preference alignment has substantially improved the observable behavior of large language models, yet it remains unclear what alignment changes internally.
By Partha Pratim Saha, Samarth Raina, Mayur Parvatikar, Amit Dhanda, Vinija Jain, Aman Chadha, Amitava Das
The study investigates how layer‑wise intervention responses in language models change over the course of pretraining, using single‑block identity bypass across multiple checkpoints and model‑domain combinations. It finds that while depth ordering of responses persists, their magnitudes shift, with nearby checkpoints showing stronger rank correspondence than distant ones and large changes occurring at positions that recur across samples and transfer across evaluation domains. Controlled experiments reveal that these longitudinal changes cannot be explained by a single downstream sensitivity and depend on perturbation strength and direction, indicating that layer sensitivity is structured but dynamic.
By Shengye Tao, Yinzhu Cheng, Haihua Xie
arXiv:2604. 18738v3 Announce Type: replace Abstract: Diffusion language models (dLLMs) generate text through iterative denoising, filling multiple masked positions at each step.
By Lin Yao
arXiv:2607. 03738v1 Announce Type: cross Abstract: Multimodal large language models (MLLMs) generate responses autoregressively, integrating visual and linguistic information in an evolving context.
By Varun Gupta, Vineet Gandhi, Makarand Tapaswi