Pith. sign in

REVIEW 10 cited by

Improving Activation Steering in Language Models with Mean-Centring

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2312.03813 v1 pith:CG2QVDPK submitted 2023-12-06 cs.CL cs.AIcs.LG

classification cs.CLcs.AIcs.LG
keywords steeringlanguagemean-centringmodelsvectorsactivationactivationsnatural
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Recent work in activation steering has demonstrated the potential to better control the outputs of Large Language Models (LLMs), but it involves finding steering vectors. This is difficult because engineers do not typically know how features are represented in these models. We seek to address this issue by applying the idea of mean-centring to steering vectors. We find that taking the average of activations associated with a target dataset, and then subtracting the mean of all training activations, results in effective steering vectors. We test this method on a variety of models on natural language tasks by steering away from generating toxic text, and steering the completion of a story towards a target genre. We also apply mean-centring to extract function vectors, more effectively triggering the execution of a range of natural language tasks by a significant margin (compared to previous baselines). This suggests that mean-centring can be used to easily improve the effectiveness of activation steering in a wide range of contexts.

Discussion (0). Sign in to comment.

Forward citations

Cited by 10 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Context Is King: How In-Context Specification Shapes the Geometry of Concepts

    cs.LG 2026-07 accept novelty 7.5 of 10

    In capable Gemma and Qwen models, declarative in-context rules set the relational geometry and topology type that the model represents and causally uses, overriding strong pretrained priors.

  2. Toward Fine-Grained Forgetting:Attribute Unlearning for Multimodal Large Language Models

    cs.AI 2026-08 reject novelty 6.0 of 10

    The paper defines attribute-level MLLM unlearning and proposes CLRP, but the method's headline forgetting gains on cloze are partly produced by test-time logit subtraction applied only to the forget and test sets.

  3. Where Steering Signals Come From: Activation Source Selection in Activation Steering

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Activation steering works best when the signal comes from the state where the model is about to produce the target behavior, not from text that already shows it.

  4. Differential syntactic and semantic encoding in LLMs

    cs.CL 2026-01 conditional novelty 6.0 of 10

    Subtracting averaged 'centroid' vectors for syntax or meaning from LLM sentence representations selectively reduces syntactic or semantic similarity, and the two signals peak in different layers.

  5. Multimodal Function Vectors for Visual Relations

    cs.AI 2025-10 conditional novelty 6.0 of 10

    Multimodal function vectors extracted from a handful of attention heads in OpenFlamingo-4B encode spatial relations and can be steered, fine-tuned, and composed to improve zero-shot relational reasoning.

  6. Balancing Stylization and Truth via Disentangled Representation Steering

    cs.CL 2025-08 reject novelty 5.0 of 10

    StyliTruth separates style and truth directions in the activations of selected attention heads, then steers each token along the disentangled subspaces to preserve truthfulness during stylization.

  7. From Emergence to Control: Probing and Modulating Self-Reflection in Language Models

    cs.LG 2025-06 reject novelty 5.0 of 10

    Self-reflection in LLMs can be steered up or down by a single activation-space vector, improving accuracy when amplified and cutting output length when suppressed.

  8. Linear Spatial World Models Emerge in Large Language Models

    cs.AI 2025-06 reject novelty 5.0 of 10

    Spatial relation words in LLaMA and Qwen models form antipodal, roughly orthogonal directions in a low-dimensional subspace, and steering along these directions changes the model's output.

  9. Probing the Robustness of Large Language Models Safety to Latent Perturbations

    cs.LG 2025-06 conditional novelty 4.0 of 10

    Randomized noise injected into hidden layers bypasses safety refusals in 12 open LLMs, and layer-wise adversarial training on the resulting benchmark reduces the attack's success.

  10. Guardians and Offenders: A Survey on Harmful Content Generation and Safety Mitigation of LLM

    cs.CL 2025-08 unverdicted novelty 3.0 of 10

    The submission's abstract promises an LLM safety survey, but the provided body is the opening page of an unrelated arithmetic-dynamics paper, so the artifact is internally inconsistent.

Pith tools