acceptodds
Under review as a conference paper at ICLR 2027

When One Direction Is Not Enough: GRASP for Conditional Activation Steering

Abstract

Activation steering provides a lightweight way to control LLMs without updating model parameters. However, most existing methods represent a target behavior with a single global direction, implicitly assuming that the same intervention is suitable for every input associated with that behavior. Analysis shows that behavior gradients can be low-dimensional yet multi-directional: different inputs may require distinct, sometimes opposing, steering directions. Averaging these directions into a single global vector may therefore miss useful control signals. To preserve this input-dependent structure, we propose Gradient-Routed Activation Subspace Prototypes (GRASP), a conditional activation steering framework that replaces a single global direction with a small set of representative directions. GRASP first projects calibration gradients into a low-rank subspace and clusters their normalized directions to construct gradient prototypes. It then learns a lightweight router that uses the current input activation to select which prototype to apply at inference time. Across seven behavioral steering tasks and two LLM backbones, GRASP improves seven-task macro accuracy over the unmodified Qwen2.5-7B and Llama-3.1-8B models by 13.37 and 20.23 percentage points, respectively. Ablation studies show that two prototypes outperform a shared direction, while routing controls confirm the need for input-dependent selection.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Activation steering, conditional routing, gradient prototypes, large language models, controllable generation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_y6DLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

GRASP clusters calibration gradients into steering prototypes and trains an activation-based router to choose a prototype for each input. Gradients are needed only during calibration. On seven binary-choice behaviors, the paper reports substantial accuracy gains over two base models and compares against several steering methods.

Critical strengths
  1. The method is computationally simple at inference. Algorithm 1 and Appendix B.1 clearly describe the prototype construction and router training, with separate fitting, tuning, and test examples.

  2. The reported gains span two backbones. Random and shuffled routing, prototype-count ablations, and oracle diagnostics provide useful evidence that selecting a direction for each input matters.

Critical weaknesses
  1. The claimed behavioral geometry has an unresolved answer-label explanation. Equation 1 uses a signed A/B logit difference, so reversing the desired answer letter reverses the gradient sign. Opposing prototypes and balanced clusters could therefore reflect answer letters rather than distinct behavioral directions. The paper does not report cluster association with the desired letter or compare against a direct answer-label probe with a fixed steering axis. Figure 2 already recovers much of the gain with two rank-one prototypes, which are opposite signs of one axis. Since all evaluations use binary-choice scoring, this ambiguity affects the central interpretation of the method.

  2. The comparative evidence is difficult to assess. Several baselines are described as adapted without corresponding implementation details or tuning budgets. Tables B1 and B3 give different unsteered Corrigibility and Hallucination accuracies, 74/60 versus 72/58, without explaining a change in evaluation inputs. Calibration randomness cannot explain that discrepancy on identical inputs. Each task also uses only 50 test examples, and variation across calibration seeds does not quantify uncertainty over test examples. These issues leave both the comparisons and the broader behavioral-steering claim insufficiently supported.

Questions for authors
  1. How strongly do prototype assignments track the desired A/B letter? Does GRASP outperform a direct answer-label probe plus a fixed direction under the same data budget, and does the result persist when answer order is controlled?

  2. What adaptations and tuning budgets were used for each baseline? Which split or preprocessing differences explain Tables B1 and B3, and what uncertainty remains when test examples are resampled?

Reviewer_m7gSOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies conditional activation steering in frozen language models and proposes GRASP. It first computes hidden-state gradients of the target logit difference for calibration samples, constructs a low-rank subspace using uncentered SVD, clusters normalized gradients to obtain multiple prototypes, and trains a linear router to predict which prototype to use from the current activation. The paper reports results on CAA’s seven binary-choice behavioral tasks using Qwen2.5-7B and Llama-3.1-8B, and analyzes the role of input-dependent interventions by varying the number of prototypes, the subspace rank, and routing controls.

Critical strengths
  1. The method is simple and inexpensive to deploy: it uses gradients only during calibration, and inference requires just a linear classifier and one vector addition, making it easy to reproduce.

  2. The report is fairly transparent. The appendix provides results for each seed (Tables B1 and B2).

Critical weaknesses
  1. The two “opposing patterns” may be driven primarily by the answer letters A/B, rather than by a multidirectional structure in the behavior itself. In an A/B forced-choice format, the gradient when the target answer is A is the negative of the gradient when the target answer is B. Thus, even if the behavior is represented by only one direction, as long as answer positions are roughly balanced, the gradients will naturally form two clusters with opposite directions and similar magnitudes.
  2. Several baselines were adapted, but the adaptations and tuning are insufficiently documented, making their relative advantages difficult to interpret accurately. Tables B1–B2 label CAST, K-CAST, Spherical Steering, and K-Steering as adapted, while the main table presents them under the original method names. The current main text and appendix do not adequately describe the training objectives, direction construction, intervention locations, and other details of these adaptations.
  3. The evaluation is too narrow. All results are based on A/B logit comparisons at the final prompt token, with only 50 test examples per behavior (one example equals 2 percentage points). There is no evaluation of open-ended generation, nor any examination of side effects on general capabilities or fluency.
  4. The numbers in the main experiments and ablations do not match, and the relationship between their protocols is unclear. In Table B3(a), Qwen’s Base scores for Corrigibility and Hallucination are 72 and 58, whereas in Table 1 they are 74 and 60. In Table B3(a), GRASP (K=2) scores 84.67 on Corrigibility, but its results for seeds 42/123/456 in Table B1 are 78/84/80, averaging 80.67. Since Base has zero variance across seeds, the test set should be fixed. Where do these differences come from?
Questions for authors

See Weakness above.

Reviewer_T1jrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces GRASP, a conditional activation-steering framework designed to address a fundamental limitation of conventional steering methods that are designed with the assumption that a single steering direction is sufficient to control a LLM behavior. GRASP separates the creation of steering directions from their selection at inference time. At inference time, GRASP uses a trained router to select one of the steering prototypes based on the current input activation.

The experimental evaluation covers seven behavioral steering tasks, namely, AI Coordination, Corrigibility, Hallucination, Myopic Reward, Survival Instinct, Sycophancy, and Refusal and tests two models Qwen2.5-7B and Llama-3.1-8B. Several activation-steering baselines have been compared as a reference point including CAA, BiPO, LinEAS, CAST, K-CAST, Spherical Steering, K-Steering, and Angular Steering.

Critical strengths
  1. The paper is well-written with a clear motivation and conceptual contribution.
  2. The paper addresses an important assumption of many activation-steering methods, that a behavior can be represented only using a single global direction. The observation that calibration gradients exhibit multiple opposing directional modes provides a strong motivation for conditional steering.
  3. The experiments cover seven tasks although with not so many samples per each task.
  4. The improvements over the strongest baselines are consistent.
  5. The authors compare GRASP against several existing steering methods, including CAA, BiPO, LinEAS, CAST, K-CAST, Angular Steering, Spherical Steering, and K-Steering, which is a strength.
  6. The authors evaluate the results using five calibration seeds and report standard deviations.
Critical weaknesses
  1. The evaluation is limited to small, binary-choice test sets, which limits its generalizability and applicability in real-world problems involving open ended generation.
  2. Although the authors evaluate seven behavioral tasks, each behavior uses only 64 calibration examples, 20 tuning examples, and 50 test examples per calibration seed, making the evaluation potentially noisy and not generalizable.
  3. The novelty as compared with existing conditional steering methods remains limited, despite the idea being interesting. Other multi-directional steering approaches already exist and the paper does not draw a clear distinction against them.
  4. There is no experiments conducted about computational efficiency, although the claim has been made.
  5. The intervention analysis is only focused on the Qwen model, while in my experience model families behave very differently which further suggests the results may not generalize.
Questions for authors

Please see the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.