acceptodds
Under review as a conference paper at ICLR 2027

DASH: Steering Attention Heads for Fine-Grained Concept Control

Abstract

Attention heads are widely used as units of analysis and intervention in transformer models, but treating an entire head as concept-specific can make interventions too coarse. We study concept control at the level of head–direction pairs and introduce DASH, Direction-Aware Steering of Heads. Given a concept dictionary and a selected set of attention heads, DASH identifies concept-related directions within each head and selectively modifies their activation components without retraining the model. Experiments on toxicity suppression, image captioning, image classification, and factual question answering show improved control–preservation trade-offs over whole-head rescaling across a broad range of settings. We further demonstrate simultaneous editing of multiple attributes, including when their selected heads overlap, while largely preserving an untargeted attribute. Together, these results demonstrate the value of within-head directional editing for selective and composable concept control.

Primary area:
generative models
Keywords:
steering, mechanistic interpretability, attention mechanism, representation engineering, generation, control

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9oCCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a method for controlling model outputs by editing only concept-related directions inside selected attention heads instead of scaling whole heads. It builds those directions from token unembedding vectors, uses SOMP to find relevant heads, and then suppresses or amplifies the concept component with either a single mean direction or a multi-direction subspace. The approach is tested on both text and vision-language models across various tasks and shows that the approach preserves unrelated behavior better related approach “Head Pursuit” while providing stronger control.

Critical strengths
  • Although the approach is a natural extension of Head Pursuit, it is a well-motivated refinement that makes the intervention more fine-grained by targeting concept-specific directions within a head instead of modifying the whole head.
  • The ablations show that alternative options for components such as head and direction do not perform worse, supporting that each component of the method contributes meaningfully to its performance.
  • The ability to compose multiple concept interventions is, in my view, the most interesting contribution of this work.
Critical weaknesses
  • The fundamental concern regarding the approach is the limitation of how concepts are defined. The concepts in the dictionary are tied to lexical tokens from the unembedding matrix. It works naturally for toxic content or properties such as color etc. but is not clear how its applicable for general behaviour such as truthfulness, hallucination, refusal, bias [1][2]. Since it is tied to lexical identity, how would it differentiate “river bank” vs “financial bank”? Similar issues arise for negations.
  • Experiments are constructed to control lexical outcomes toxic words, countries etc. The baselines are limited (maybe [1] [2] or other suitable baselines should be added), so it is unclear whether DASH is better than simpler methods. For example, an external classifier could guide generation and suppress the target concept without changing the model or an external embedding dictionary could store vectors for toxic words and penalize generated tokens or continuations based on their similarity to those vectors.
  • There is limited evidence for generalization across models.
  • Some hyperparameters seem to be based on the evaluation dataset.

[1] Inference-Time Intervention: Eliciting Truthful Answers from a Language Model

[2] Discovering Bias in Latent Space: An Unsupervised Debiasing Approach

Questions for authors

The response to the weaknesses is required along with empirical evidence.

1 comment

We thank the reviewer for recognizing the motivation, the ablations and the composition results. We agree that composition is DASH's most distinctive capability and will give it more space. All new results will be in the revision.

Concepts tied to lexical tokens. We agree this is the main scope limit of the dictionary-based variant and will state it in the main text. Dictionary directions work for concepts with a clear vocabulary signature (toxic language, colours, sentiments, quantities, countries, image classes; four models). Truthfulness, hallucination or refusal have none, and for them a contrastive direction is the natural choice. However:

  • The operator does not depend on the direction's source. DASH projects onto any per-head subspace. Appendix E.5 (Table 26) compares six ways of building it: three single directions and three rank-10 subspaces, including a random-vocabulary control. On the same heads we also tested contrastive mean-difference directions (as in ITI): they amplify colours more than the dictionary (3.06× vs 1.73× at CIDEr ≥ 0.95), but on RTP do not lower classifier toxicity as reliably. To save space, we refer the reviewer to our reply to Reviewer zvkc. For truthfulness or refusal, one keeps DASH's head selection, operator and composition and replaces only the dictionary.
  • A fixed direction does not mean a context-free edit. The edit scales the projection of the head activation, which depends on context. We tested context dependence on 31 ambiguous dictionary words, each present in a benign and a harmful prompt ("suffering a heart → attack" vs "launched a surprise → attack"). DASH_mean lowers the harmful sense's log-probability about twice as much as the benign one (−0.59 vs −0.31; difference −0.28 [−0.52, −0.05]) and leaves 30 control words unchanged (−0.02). Negation raises the same question. However, we have not tested negation separately and will list it as an open limitation.

Baselines, including output-level control. We added comparisons with ITI, CAA, Arditi et al.'s directional ablation and Style Modulation Heads. To save space, we refer the reviewer to our reply to Reviewer APiS. Residual-stream methods are stronger than DASH on RTP toxicity, but DASH suppresses caption colours far more than any of them (0.03 vs 0.33–0.65).

SteerFair also steers without labelled data, but derives its direction from demonstrations of a bias rule (e.g. preferring the first answer option) by permuting unlabelled inputs. Order bias has no vocabulary signature, so the two are complementary: a SteerFair direction could be used with DASH's head selection and operator.

For output-level control, we added a dictionary logit penalty, up to a full ban, with toxicity scored on a held-out word list and perplexity on the modified distribution. It does not transfer beyond its list: lexical toxicity stays ≥ 0.96 at every strength (ban: 1.10), judged toxicity is unchanged (+0.01), and the ban makes perplexity unbounded because dictionary subwords occur in ordinary text. DASH, built from the same list, reaches 0.56 and lowers judged toxicity by 0.20. Output filtering remains strong for blocking listed words, but DASH covers settings that it cannot:

  • Grounded amplification in VLMs. A logit bonus raises colour mentions (2.91×) with nothing tying them to the image. DASH's amplification stays grounded: a VLM judge finds accuracy changes of only −0.03 (DASH_mean) and −0.07 (DASH_multi), vs −0.10 for Head Pursuit (to save space, see our reply to Reviewer APiS). We will judge the logit-bonus captions in the revision.
  • Opposing edits inside the model, e.g. amplifying one colour family while suppressing two others that share most of its heads, where DASH moves every family as requested and Head Pursuit cannot (to save space, see our reply to Reviewer APiS).
  • DASH uses no external classifier at every decoding step.

Generalization across models. We ported the pipeline to Llama-3.1-8B-Instruct and Qwen3.5-9B. To save space, we refer the reviewer to our reply to Reviewer zvkc.

Hyperparameters selected on evaluation data. We re-selected each method's strength on 1,000 held-out RTP prompts. This recovers the paper's default strengths and matches the evaluation-sweep choice in 11 of 15 cells (the neighbouring strength otherwise).To save space, we refer the reviewer to our reply to Reviewer APiS. The RTP strengths and rank were reused unchanged on all other tasks.

Reviewer_zvkcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DASH, a framework for steering large language models at the level of attention heads. Given a target property, the method builds on prior work by combining an existing head identification strategy with a novel sub-head direction identification technique. The method is tested over different tasks spanning language generation and image understanding, and proves effective in steering across multiple properties at the same time.

Critical strengths
  • The paper is well written and pleasant to read, and necessary details are provided. The experimental reporting is thorough and negative results on TriviaQA are reported candidly.
  • The DASH method is simple and training-free, and it shows promising steering results across different tasks and modalities.
Critical weaknesses
  • Novelty is limited. The works largely builds on Head Pursuit (HP), as rightfully acknowledged by the authors. The novel methodological contribution of the present work over HP appears slim, since the head scoring and experimental evaluation pipelines exactly match those of the existing work, and the new component, rescaling the projection onto a concept subspace within each head, reduces at β=1\beta=1 to directional ablation from [a], applied per head.
  • Architectural scope is limited. DASH is evaluated on only two 7B models from early 2024, Mistral-7B and LlavaNext 7B (the paper does not specify which variant, the two also share a backbone if it is the Mistral-based one). Evaluating more recent families such as Gemma 4 or Qwen3.5 would test generality more convincingly. Results at a larger scale would also help.
  • Steering baselines are missing. All evaluations compare DASH only against HP and randomized controls. However, several head-level steering methods have been proposed, and comparing against them would clarify the impact of various design choices of DASH. ITI [b] shifts activations along data-derived directions within selected heads, offering an alternative operator and direction construction, while Style Modulation Heads [c] identifies heads with a different selection criterion. Comparisons with standard steering methods that are not head-specific would also help contextualize the results. Even evaluating simple difference-of-means steering would be informative. While DASH does not require contrastive data, the two sets of examples such a baseline needs (with and without the target concept, unpaired) are cheap to obtain in the paper’s settings, e.g., captions with and without color terms, or RTP continuations flagged by the concept dictionary.
  • Preservation metrics are weak. Coherency collapse under steering often goes undetected by perplexity (e.g., see [c]). To claim improved control-preservation trade-offs I suggest adding at least one coherency metric (e.g. LLM judge based coherency) and one external benchmark (e.g. MMLU).

References:

[a] Arditi et al., Refusal in Language Models Is Mediated by a Single Direction, NeurIPS 2024

[b] Li et al., Inference-Time Intervention: Eliciting Truthful Answers from a Language Model, NeurIPS 2023

[c] Izawa et al., Steering at the Source: Style Modulation Heads for Robust Persona Control, ICML 2026

Questions for authors
  • In Section 3.2, the authors define the pullback vector so that the scalar product between the unembedding direction u(c)u^{(c)} and the head contribution Wl,hOzl,hW^O_{l,h} z_{l,h} can be computed directly in head space. However, in standard LLMs there is an RMSNorm between the final residual hidden state and the unembedding. Appendix C includes the RMSNorm weight when scoring heads with SOMP, but the editing directions appear to be built from the raw unembedding vectors. Why this choice, and does incorporating the weight into the directions change the results?
  • How does DASH deal with concept vocabulary entries that get tokenized into more than one token?
  • Is there a rule of thumb to decide which variant should be used in a given setting between DASH multi and DASH mean?
2 comments

We thank the reviewer for the expert review: the RMSNorm and tokenization questions were especially useful. All new results will be added to the revision.

Novelty and directional ablation. We agree that at β = 1, DASH is per-head directional ablation (Arditi et al.), and the revision will clarfy that. DASH differs in three ways: directions come from a dictionary rather than contrastive data, β is continuous and can be negative (amplification), and several concepts can be edited additively in shared heads. Ablation alone fails here: at β = 1, RTP lexical toxicity is 0.91 (DASH_mean) and 0.86 (DASH_multi), against 0.56 and 0.49 at the operating points. To separate direction from operator, we compared four combinations on the same 32 heads. The table shows captions at CIDEr ≥ 0.95. "Contrastive" is a difference of means between outputs with and without vocabulary words.

Direction × operatorRTP lexical / classifier ↓ (Δppl ≤ 5%)Colour supp. ↓Colour amp. ↑
Dictionary × scaling (DASH_mean)0.56 / 0.750.031.73×
Dictionary × shift0.52 / 0.750.082.89×
Contrastive × scaling (per-head Arditi)0.81 / 0.810.403.06×
Contrastive × shift (ITI operator)0.52 / 1.090.224.18×

The direction matters more than the operator: on RTP only dictionary directions lower both metrics, while contrastive directions amplify colours more. A dictionary does not win everywhere, but its advantage is no need of labelled examples. Appendix E.1 already shows that heads, directions and their assignment all matter. Thus, DASH's contribution is not the projection operator but label-free, per-head concept directions that can be scaled continuously in either sign and composed.

Architectural scope. Indeed, LLaVA-NeXT-7B (llava-v1.6-mistral-7b-hf) uses a Mistral-7B backbone. We applied DASH to Llama-3.1-8B-Instruct (GQA, byte-level BPE) and Qwen3.5-9B (hybrid linear-attention/attention):

Model, methodTET lexical / classifier ↓ (Δppl ≤ 10%)Δppl at default (RTP, TET)
Llama, Head Pursuit—+79%, +84%
Llama, DASH_mean0.26 / 0.68+2.6%, +3.1%
Llama, DASH_multi1.02 / 1.02+170%, +417%
Qwen, Head Pursuit0.84 / 0.94+47%, +65%
Qwen, DASH_mean0.17 / 1.14+81%, +16%
Qwen, DASH_multi0.39 / 0.67+34%, +18%

DASH_mean transfers to Llama at its default strength, and Head Pursuit's whole-head scaling is far costlier on both models. DASH_multi's rank-10 subspace transfers worse. On Qwen, lexical control works through linear-attention heads too, but judged toxicity barely changes because Qwen already refuses 78% of TET prompts. In Qwen captioning (CIDEr ≥ 0.95), DASH_mean cuts colour mentions to 0.20× and DASH_multi raises them to 2.45×. Head Pursuit has no suppressing setting here.

Baselines. We added ITI, SMH, CAA, Arditi et al.'s residual ablation and a logit penalty. To save space, we refer the reviewer to our reply to Reviewer APiS for the tables. Residual-stream methods are stronger on RTP toxicity. DASH suppresses caption colours far more (0.03 vs 0.33–0.65). SMH at its automatic layer lowers judged toxicity more than DASH_mean (−0.26 vs −0.20) at an MMLU cost (0.574 vs 0.584), but at the near-tied runner-up layer it raises it (+0.11).

Preservation. We additionally computed 5-shot MMLU for each method. Results show it is unchanged for DASH, Head Pursuit, CAA and Arditi (0.583–0.585 vs 0.584); only ITI loses 2.5 points. LLM-judged coherence falls by 0.04 for DASH_mean vs 0.12 for Head Pursuit (to save space, see our reply to Reviewer APiS). DASH_mean's TriviaQA F1 drops 27%, mostly from longer answers (median 15 vs 4 words): its answers contain the gold answer almost as often as unsteered (0.591 vs 0.602), the highest of any steered method.

Q1 (RMSNorm γ). This was an unintended inconsistency. Folding γ into the directions (same heads and dictionary) gives nearly identical directions (cosine 0.97) and equal or better results at the same perplexity (RTP, DASH_mean: 0.47 / 0.72 vs 0.56 / 0.79). We will add γ-folding into the revision and report both.

Q2 (multi-token entries). Each entry contributes all its subword tokens as atoms, deduplicated (see Appendix B–C: e.g. 197 countries → 374 atoms). Removing 18 generic words (e.g. "a", "you", "her") from the TET dictionary changes half the selected heads but not the results (0.56 / 0.63 vs 0.59 / 0.66). For further results on dictionary perturbations, to save space, we refer the reviewer to our reply to Reviewer APiS.

Q3 (mean vs multi). DASH_mean is generally better for tight preservation budgets: it wins RTP at every perplexity budget and MNIST, SVHN, GTSRB at CIDEr ≥ 0.99. DASH_multi for strong edits, especially amplification: 4.2× colour and 2.0× quantity at CIDEr ≥ 0.90, vs 1.9× and 1.5×. DASH_multi uses rank 10 by default, and larger ranks cost perplexity faster than they help (Appendix E.2).

Thank you for your response and for the additional experiments. I believe they strengthen the experimental soundness of the paper, and I strongly encourage the authors to include them in the next revision.

Overall, my final assessment of this paper is genuinely mixed. On the one hand, the paper has clear merits, particularly in terms of clarity and presentation, and with the new results I now consider its coverage of steering baselines sufficient. On the other hand, I believe its novelty is limited: it proposes a somewhat incremental improvement to an existing method, and this improvement is itself a direct adaptation of techniques already known in the literature. To a lesser extent, model coverage also remains limited (a single model and its instruction-tuned counterpart in the manuscript, plus limited experiments on two additional models in the rebuttal). My assessment is truly borderline, but these concerns, especially the one regarding novelty, lead me to maintain my current score, which leans negative.

Reviewer_APiSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DASH, a training-free framework for editing concept-specific directions within attention heads. It extends Head Pursuit by replacing whole-head rescaling with directional projection.

Critical strengths
  1. Directional editing is more selective than whole-head rescaling in several reported settings.
  2. The experiments include intervention-strength sweeps and randomized head/direction controls.
Critical weaknesses
  1. The methodological novelty is limited. DASH retains the head-selection procedure from Head Pursuit and mainly replaces whole-head scaling with projection-based directional editing.
  2. Projection-based activation steering is already well established. The paper does not sufficiently distinguish its contribution from existing directional intervention methods such as ITI and representation steering.
  3. The main comparisons focus heavily on Head Pursuit. Direct comparisons with other activation-steering methods under matched conditions are missing.
  4. The fixed intervention strengths are not matched by target effect or preservation cost. Using different manually selected strengths for DASH and Head Pursuit can make the comparison difficult to interpret, especially for K = 8 and K = 16.
  5. Lexical toxicity is an incomplete measure of harmful content. Reducing toxic-word occurrences does not necessarily reduce semantic toxicity, and alternative expressions may remain unaffected.
  6. The image-captioning evaluation relies on vocabulary counts and CIDEr. Neither metric directly establishes whether the intervention preserves the factual correctness of generated descriptions.
  7. The method depends on manually specified concept dictionaries and calibration samples. Its effectiveness for ambiguous, context-dependent, or poorly represented concepts remains unclear.
Questions for authors
  1. How does DASH compare with ITI or other activation-steering methods under matched preservation constraints?
  2. Can the authors evaluate semantic toxicity and caption fidelity using human judgments or independent evaluators?
  3. How stable are the results when concept dictionaries and calibration distributions change?
  4. What happens when three or more concepts with highly overlapping subspaces are edited simultaneously?
  5. Can intervention strengths be selected on a separate validation split?
2 comments

We thank the reviewer for the constructive review. We added steering baselines, matched-budget comparisons, independent semantic judges and robustness experiments. All will be included in the revision. We divide our response into two parts for readability.

Novelty and baselines. We agree that projection-based editing itself is not new. The most distinctive contribution of DASH is compositional control: it applies separate, concept-specific edits within shared heads, including opposing edits that whole-head scaling cannot express. DASH selects attention heads for a concept and edits only a per-head concept direction derived from a dictionary. Head Pursuit localizes dictionary concepts to heads but rescales whole heads. ITI edits head directions learned from labelled examples. SteerFair targets a different goal (biases from superficial input variations, using directions from bias-rule demonstrations). DASH adds dictionary-derived concept subspaces that act as separate controls inside shared heads, so concepts can be suppressed and amplified simultaneously without labelled contrastive data. We will state this contribution explicitly.

We also added four baselines, fitted on the same calibration outputs, to the comparison. Unlike DASH, they also use labels for whether each calibration output contains concept vocabulary. Each cell is the strongest effect within a preservation budget (|Δppl| ≤ 5% on RTP, normalized CIDEr ≥ 0.95 in captioning):

MethodRTP lexical ratio ↓Colour suppression ↓Colour amplification / Δquantity
DASH_mean0.560.031.73× / +2%
DASH_multi0.620.012.38× / +4%
ITI0.710.651.18× / −2%
SMH (auto layer)0.700.473.25× / −37%
CAA0.380.332.49× / −20%
Arditi et al. (residual ablation)0.211.77—

DASH suppresses colours far more than any baseline and amplifies them with little change to the untargeted quantity attribute. SMH and CAA amplify further but shift quantity by 20–37%. The residual-stream methods (CAA, Arditi) are stronger on RTP toxicity. We also compared dictionary and contrastive directions on the same heads: to save space, we refer the reviewer to our reply to Reviewer zvkc. We also note that the appendix already compares several intervention scopes (selected heads, all heads in their layers, all heads).

Smaller head budgets (K = 8, 16). We added comparisons for K = 8 and K = 16. With full strength sweeps and matched preservation budgets on all dataset–model pairs, a DASH variant shows better results than Head Pursuit in 63/83 task–metric–budget combinations at K = 8 and 66/83 at K = 16.

Semantic toxicity. We agree lexical rates are incomplete. The paper also reports RoBERTa toxicity, where DASH shows better results than HP on TET at every budget, and Head Pursuit shows better results on RTP at budgets ≥ 3%. We also now added an LLM judge (gpt-4o-mini) that scores each RTP continuation's toxicity (0–10) and coherence (1–5) at each method's operating point:

MethodΔ judged toxicity ↓Δ coherence
DASH_mean−0.20 [−0.29, −0.11]−0.04
DASH_multi−0.09 [−0.19, 0.00]−0.19
Head Pursuit−0.10 [−0.19, 0.00]−0.12
ITI−0.37−0.06
SMH (auto layer)−0.26 [−0.34, −0.17]−0.01
CAA−0.36 [−0.45, −0.27]−0.03
Arditi et al.−0.58 [−0.67, −0.48]−0.02

Paired differences from the unsteered model (toxicity 0.94, coherence 4.97), 95% bootstrap intervals.

Lexical rates overstate semantic suppression for every method. DASH_mean lowers judged toxicity twice as much as Head Pursuit at a third of its coherence cost; the residual-stream methods lower it more. On TET, DASH_mean and Head Pursuit tie (−0.33 vs −0.34) and DASH_mean keeps more coherence (−0.10 vs −0.18).

Caption fidelity. Beyond CIDEr (Appendix D.3) and visual examples (Figs 9–10, Appendix D.3.3), we added a VLM judge (gpt-4o-mini) that sees the image and one caption and rates factual accuracy (1–5), on all 1,000 evaluation images. The table below shows paired Δ accuracy (unsteered baseline 4.38, 95% intervals ≈ ±0.05)

EditDASH_meanDASH_multiHead Pursuit
Suppress colours / sentiments / quantity−0.04 / −0.05 / −0.02−0.06 / −0.02 / −0.13−0.23 / −0.06 / −0.21
Amplify colours / sentiments / quantity−0.03 / −0.01 / −0.36−0.07 / −0.12 / −0.32−0.10 / −0.09 / −0.17
Suppress all three−0.10−0.34−0.97

DASH_mean stays within 0.05 of the unsteered captions for every suppression and for colour and sentiment amplification, while Head Pursuit loses up to 0.23 (0.97 when suppressing all three). CAA and ITI also keep colour captions accurate (−0.05, −0.01) but suppress colours far less (0.33, 0.65 vs 0.03). Amplifying quantity is the exception: DASH sometimes adds counts the image does not support.

AuthorsOpenReview ↗

Continuing the rebuttal

Dictionary and calibration robustness.* We kindly note that Appendix E.4.2 shows that directions from random half-dictionaries align closely with the full ones (cosine 0.909) and that 50 calibration prompts suffice (lexical rate 0.049 vs 0.047 with 1,000; Appendix E.3). We also conducted additional analysis, using random subsets containing 25–75% of dictionary entries (3 seeds each). This gives normalized lexical toxicity 0.39–0.60 (full dictionary: 0.56). We also checked calibrating on TET while evaluating on RTP, or vice versa: this changes scores by at most 0.09.

Three or more overlapping concepts. The paper's three caption concepts share few heads (2–6 pairwise, none jointly); suppressing all three with DASH_mean leaves colours at 0.03 and quantity at 0.78 at CIDEr 0.94, against colours at 1.06 and CIDEr 0.53 for Head Pursuit. For a hard case, we did additional experiment splitting the colour dictionary into three subsets: warm, cool and achromatic-related colours. Their 32-head sets share 16–22 heads pairwise and 15 jointly, though their directions on shared heads differ (median |cos| 0.11–0.32). We amplified one family and suppressed the other two. The table shows warm / cool / achromatic mentions relative to the unsteered model @ normalized CIDEr.

AmplifiedHead Pursuit (two tie rules)DASH_meanDASH_multi
Warm2.00 / 2.39 / 2.20 @ 0.97 or 1.20 / 0.02 / 0.41 @ 0.781.71 / 0.02 / 0.00 @ 0.952.51 / 0.02 / 0.01 @ 0.95
Cool1.86 / 2.49 / 2.20 @ 0.94 or 0.02 / 0.02 / 0.51 @ 0.750.02 / 9.59 / 0.00 @ 0.920.07 / 5.00 / 0.00 @ 0.89
Achromatic0.67 / 1.02 / 1.28 @ 0.88 or 0.07 / 0.00 / 0.79 @ 0.780.00 / 0.00 / 2.56 @ 0.940.00 / 0.00 / 2.51 @ 0.89

Head Pursuit must pick one scale per shared head, so under either tie rule some family moves the wrong way or not at all. DASH edits each family's direction separately and moves every family as requested (CIDEr 0.89–0.95). Thus, DASH can make opposing edits to concepts that share most of their heads, which whole-head rescaling cannot.

Strengths selected on a validation split. We drew 1,000 new RTP prompts, disjoint from calibration and evaluation, with perplexity on WikiText-2 validation. For each method and perplexity budget we took the strength with the lowest held-out lexical toxicity and scored it on the evaluation prompts. The table shows normalized lexical toxicity / Δppl on the evaluation prompts at chosen strength.

BudgetHead PursuitDASH_meanDASH_multi
≤ 1%α = 0: 0.89 / +0.5%β = 7: 0.58 / +0.2%β = 2: 0.70 / +1.1%
≤ 3%α = −0.5: 0.85 / +2.8%β = 10: 0.56 / +1.7%β = 2: 0.70 / +1.1%
≤ 5%α = −0.5: 0.85 / +2.8%β = 10: 0.56 / +1.7%β = 3: 0.62 / +4.6%
≤ 10%α = −1: 0.75 / +7.8%β = 10: 0.56 / +1.7%β = 5: 0.49 / +11.5%

Across five budgets (1–15%), the held-out choice equals the evaluation-sweep choice in 11 of 15 cells and is the neighbouring strength otherwise. It recovers the default strengths used in the paper (DASH_mean β = 10 at every budget ≥ 3%; DASH_multi β = 5 and Head Pursuit α = −1 at 10%), and DASH_mean suppresses more than Head Pursuit at every budget. Other tasks reuse the RTP strengths without tuning.

Ethics. We will add an ethics statement. DASH is dual-use: the mechanism that suppresses toxic vocabulary can also amplify it. It requires white-box access and adds no capability beyond fine-tuning or prompting, though it is cheaper. We release only aggregate metrics, no amplified-toxicity generations.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.