The paper investigates what aspects of language model behavior are controlled by activation steering. By introducing Cross‑Encoding Steering Evaluation, the authors show that steering effects often follow the extraction index of answer identifiers rather than the semantic content of the answers, especially at deeper layers. They also find that a low‑rank output‑sensitive component captures most of this effect, and that different datasets (NormBank, MNLI, SC101) exhibit varying preferences for extraction‑index versus semantic‑label following.
Activation steering is often evaluated under the answer encoding used to construct the direction. A reported gain may reflect the intended judgment or compatibility with answer identifiers seen during...
arXiv:2606. 11198v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM outputs, yet the format of injected content -- distinct from its semantic relevance -- can independently distort the model's attention distribution.
arXiv:2608. 02957v1 Announce Type: new Abstract: Steering vectors (SVs) are widely used to influence the expression of concepts (e.
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).
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.
arXiv:2607. 04223v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) reduces but does not eliminate hallucination, and existing detectors return a single answer-level score that does not indicate which sentence is unsupported, or why.
arXiv:2607. 29062v1 Announce Type: new Abstract: Model capabilities have improved in large part due to scaling chain of thought.
The paper demonstrates that a single internal direction in modern language models—called the valence axis (V-axis)—captures how positive or negative a sentence feels. By using only nine emotion category names and 50 short narrative paragraphs per emotion, the authors identify this axis via principal component analysis of frozen encoder embeddings, achieving 93% of supervised performance on SST‑2 and strong correlations with human valence ratings across images, audio, and brain recordings. The method transfers across modalities without target‑modality labels, but works only for continuous attributes and is specific to certain model families.
arXiv:2408.11827v2 Announce Type: replace Abstract: Understanding how language models compose meaning from linguistic input remains a central problem in interpretability research. Mechanistic studies...
arXiv:2607. 01002v1 Announce Type: cross Abstract: In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them.
arXiv:2607. 14111v1 Announce Type: cross Abstract: Can small language models detect and report on perturbations their own internal activations?
arXiv:2608. 10537v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) have helped uncover mechanistic explanations for LLM behaviours such as reasoning, jailbreaking etc.