Skip to main content
Search

# Activation Steering

23 related articles

Subscribe to updates
  1. arXiv

    Few-step diffusion students lose directions, not variance: ReDiverse restores diversity at one intermediate block

    Distilling a diffusion model into a four-step student makes different random seeds produce nearly identical images and videos for the same prompt. By comparing teacher and student activations block by block, ReDiverse finds that the student keeps its total activation variance but concentrates it into far fewer directions, at one intermediate block during the first denoising step. ReDiverse-FT then uses LoRA at that block to regress the student's first-step activations onto the teacher's, while ReDiverse-TF perturbs activations along teacher principal directions; across four teacher-student pairs on SDXL, SD3, Wan 2.1 and MiniMax-H3, both restore diversity while largely preserving generation quality.

  2. bioRxiv

    Protein language model steering drops the global mean: PULSE pairs similar backgrounds, with the largest predicted-score gain on toxicity

    Protein language models can be steered toward a target property without changing their weights, but existing activation steering builds the direction from the difference between the global means of property-positive and property-negative proteins, ignoring that the useful direction shifts with sequence background. PULSE first matches positive and negative proteins by representation similarity, then aggregates similarity-weighted, unit-normalized differences between matched pairs into a steering vector. On solubility, immunogenicity and toxicity optimization, PULSE beats global centroid steering in both final predicted property scores and optimization success rates, with the largest score gain on toxicity and the largest success gain on immunogenicity.

  3. arXiv

    CANS gives every feed-forward neuron its own scaling coefficient, steering safety, toxicity and hallucination while holding general capability

    Behavior control in large language models often costs general capability; CANS instead assigns each feed-forward network neuron a scaling coefficient that accounts both for its contribution to the target behavior and for its disruption of general outputs. Target behavior is proxied by the log-probability of target tokens in model responses, capability preservation by output-distribution KL divergence on general-purpose text, and local approximations of the two yield a closed-form solution with no iterative training and no added inference-time computation. On Qwen3-4B and Llama models, CANS matches or beats SFT, CAA and neuron-level baselines on safety, toxicity and hallucination control while better preserving MMLU, BBH and IFEval performance.

  4. arXiv

    Where to inject activation steering: TopoSteer ranks layers by the connectivity length of displacement clouds, skipping held-out trials

    Activation steering adds a vector to a large language model's internal representations at inference time, but where to add it has mostly been settled by trial: existing selectors score how separable the two activation classes are or how well individual displacements align with their mean, without ever steering the model to check. TopoSteer instead scores each layer by the total length of the Euclidean minimum spanning tree of its per-example displacement cloud, measuring whether the mean-difference vector can stand in for those displacements. Across five instruction-tuned models and six binary persona-steering tasks it beats LayerNavigator in most model-budget comparisons, but the margin depends on model and layer budget and is smaller and more mixed on Qwen-7B and Qwen-32B.

  5. arXiv

    Sparse-autoencoder steering gets an information-theoretic design: DESSIRT beats a contrastive baseline on most question modes in Llama and Gemma

    Sparse-autoencoder (SAE) steering has relied on contrastive activation statistics to pick features, although feature interpretability and steering utility are not necessarily aligned. DESSIRT represents each dataset as a vector of SAE feature firing probabilities, measures deviation from a target distribution with Jensen–Shannon divergence, and constructs steering vectors analytically from that divergence. In a seven-mode question-generation task, DESSIRT-2 beat the SDCV baseline on all but one mode in Llama 3.1-8B-Instruct and on most modes in Gemma-2-9B-IT.

  6. arXiv

    Diffusion language models stop reusing one steering vector: feeding the previous prediction back lifts TruthfulQA multiple-choice average from 30.89 to 45.42

    TraceSteer feeds a diffusion language model's previous denoising prediction back into activation steering, adapting both the correction direction and its strength step by step, and improves factual reliability on TruthfulQA for LLaDA-8B-Instruct and Dream-7B-Instruct.

  7. Ai2 Blog

    Fairness beyond the final pick: one persuasive argument shifts models' willingness to abandon the right candidate by up to 42 points

    Ai2 released FairStress, a benchmark that measures demographic fairness in large language models beyond which candidate a model finally picks, by testing whether it abandons that choice under persuasion. Its 300 two-candidate scenarios span 29 domains and expand to more than 13 million test cases when identity presentation and matched adversarial prefixes are combined; across 15 models, the average gap in abandoning correct answers was 6.9 percentage points with implied identity and 9.3 points with explicit labels.

  8. arXiv

    No weight updates at inference: steering neurons per query lifts seven models from 3B to 32B

    Instead of a fixed intervention chosen in advance, Q-NeuS reads each query's own internal activations to decide which feed-forward neurons to amplify, improving every one of seven models from 3B to 32B parameters across three benchmarks.

  9. arXiv

    Before the story is written, event types are decodable: four open-weight models look ahead at least four events, but only the next one beats text-only prediction

    Across four 24–32B open-weight models, hidden states at event ends decode the state-change type of at least four upcoming events, yet only the next event significantly beats text-only predictors; in OLMo, sustained steering along that direction shifts subsequent event composition in both directions.

  10. Topic Models Drop the Word List: Directions in LLM Activations Track Short and Abstract Texts Better

    Topic models can describe topics with interpretable directions learned by sparse autoencoders instead of word lists, and an LLM-based topic judge prefers that representation on eight datasets, with the largest margins on short and abstract texts.