The study investigates how task context reshapes the roles of shared attention heads in large language models. By applying identical graph facts and measuring interventions on attention heads and downstream activations, the authors show that task changes redistribute influence across five open‑weight models, causing the same heads to respond to different facts and affect later components differently. Linear response maps before and after activation edits reveal significant reductions in prediction error, indicating that task context alters both the facts processed and the downstream impact of shared components.
By Sovesh Mohapatra, Dani S Bassett
Large language models (LLMs) exhibit a human‑like working‑memory pattern on the N‑back task, with performance dropping as memory load increases and errors mainly on lure items. Two hypotheses—precision limits of softmax attention and interference from competing items—were tested across 15 pretrained LLMs and minimal transformers. Results show that interference, not precision, constrains LLMs’ working‑memory performance, as suppressing non‑target item identities improves sensitivity and larger models benefit more from reducing interference.
By Dongyu Gong
The paper introduces Declarative Attention (DA), a protocol that lets language models explicitly declare which parts of their context to focus on during generation. By partitioning decoding into full-context, region-specific, and recent-output-only modes, the inference engine can skip large portions of the KV cache, dramatically reducing attended tokens. Experiments on 15 long-context tasks with off-the-shelf models show significant savings (52.0% and 31.1% reductions) with only modest accuracy drops that diminish as model size increases.
By Namgyu Ho, Huzama Ahmad, Woosung Koh, Se-Young Yun, Tal Schuster, Cicero Nogueira dos Santos
The paper investigates how different prompt adaptation techniques—discrete prompt optimization, prompt tuning, and prefix tuning—affect the internal processing of dense and mixture-of-experts language models. It finds that the adaptation location determines how task information reaches the output: text prompts mainly influence response construction, input-level prompts modify content representations, and layer-wise prefixes exert direct control via attention. In MoE models, these differences also lead to distinct expert‑routing patterns, while generic representation‑level metrics fail to reliably distinguish effective adaptations from placebo effects.
By Mikhail Aleksandrov, Nail Bashirov, Maksim Ivanov, Andrew Tsukanov, Timofey Karyagin, Alexander Miasnikov, Gleb Molodtsov, Valentin Samokhin, Aleksandr Beznosikov
arXiv:2604. 09670v2 Announce Type: replace-cross Abstract: Intelligent systems must maintain and manipulate task-relevant information online to adapt to dynamic environments and changing goals.
By Hua-Dong Xiong, Li Ji-An, Jiaqi Huang, Robert C. Wilson, Kwonjoon Lee, Xue-Xin Wei
Instruction tuning is often thought to give language models a universal ability to follow instructions, but this study shows otherwise. By probing nine tasks across three models, the authors find that general probes reveal selective, not uniform, deficits, cross‑task transfer is weak and skill‑similar, and causal ablation uncovers sparse, asymmetric dependencies. The results suggest instruction following is a coordinated use of diverse linguistic skills rather than a single shared mechanism.
By Elisabetta Rocchetti, Alfio Ferrara
The study investigates how post‑training of large autoregressive language models (ARMs) into masked diffusion models (MDMs) affects their internal computation. Across two 7B ARM‑MDM families and four diagnostic tasks, the authors find that MDMs retain much of the ARM’s high‑attribution pathways on prefix‑dominant tasks, but reorganize computation toward earlier layers on globally constrained tasks. Component‑level probes reveal that ARMs depend on sharply specialized components, whereas MDMs show weaker specialization and more diffuse output‑space alignment.
By Injin Kong, Hyoungjoon Lee, Yohan Jo
arXiv:2601. 04098v2 Announce Type: replace-cross Abstract: Transformer language models systematically prefer tokens at specific input positions regardless of semantic relevance---a phenomenon known as positional bias.
By Maryam Rahimi, Mahdi Nouri, Yadollah Yaghoobzadeh
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:2610.06897v1 Announce Type: new
Abstract: Localizing latent structures in the activation space of language models (LMs) is central to understanding and controlling their behavior. Yet, localize...
By Or Shafran, Mor Geva
arXiv:2606. 08682v1 Announce Type: cross Abstract: Activation steering has emerged as a popular inference-time technique for modulating the behavior of large language models (LLMs).
By Qi Cao, Jian Lou, Meiting Liu, Wenjie Feng, Dan Li, See-Kiong Ng, Anh Tuan Luu
arXiv:2507. 18043v2 Announce Type: replace-cross Abstract: Inference-time steering methods offer a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying internal activations at test time without updating model weights.
By Duy Nguyen, Archiki Prasad, Elias Stengel-Eskin, Mohit Bansal