acceptodds
Withdrawn from ICLR 2027

Conditional Intervention Selection for Efficient Activation Steering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Activation steering adapts frozen models by intervening on internal representations, yet conventional relevance scores do not reveal whether neurons remain useful when intervened on jointly. We formulate sparse steering-site localization as budgeted conditional selection and introduce SCOUT, which constructs a relevance-ranked working set, accepts neurons according to their measured utility conditioned on the current selection, and calibrates the resulting displacement into static neuron-wise scaling for single-forward deployment. Across the evaluated model scales, SCOUT achieves positive mean accuracy gains of up to 3.44 percentage points while selecting only 11–16 neurons on average. This corresponds to an approximately 84–99.6% reduction in intervention sites relative to baselines using 100 to 4096 sites. Selection ablations further show that conditional utility consistently improves over relevance-only and independent-utility alternatives. These results establish conditional intervention selection as an accurate and highly sparse approach to activation steering.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Representation Repair, Activation Steering, Frozen Language Models, Stochastic Probing, Hidden-State Intervention, Task Adaptation, Representation Geometry

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XzGWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the problem of selecting sparse intervention locations in Activation Steering and points out that traditional neuron relevance scores cannot accurately reflect the actual utility of jointly intervening on multiple neurons. The authors propose SCOUT, which first uses perturbation sensitivity to construct a candidate neuron set, then accepts neurons one by one according to their conditional marginal benefit under the currently selected set, and finally approximates the dynamic intervention as static scaling coefficients to enable a single forward inference. Experiments cover Qwen3-0.6B to 8B and show that the method can achieve certain classification accuracy improvements with fewer intervention locations.

Critical strengths

S1. The research problem is clear and has certain research value. The paper distinguishes Representation Relevance from Intervention Utility and points out that neurons with higher relevance do not necessarily bring positive benefits in joint interventions. This provides a reasonable motivation for location selection in sparse Activation Steering.

S2. The ablation experiments on the conditional selection strategy are convincing. Table 4 compares different selection strategies under fixed candidate pools, intervention operators, budgets, and calibration methods. Conditional Utility outperforms Contrastive, Sensitivity, and Random Selection in all nine experiments across three datasets, which fairly well validates the value of conditional utility selection.

Critical weaknesses

W1. The methodological novelty is limited, and the core selection strategy is a relatively direct greedy search. The main idea of SCOUT is to measure, under a fixed candidate order, the marginal benefit of adding each neuron to the current set one by one, and to accept candidates with positive benefit. This is essentially an application of the Greedy Forward Selection idea to Activation Steering. Proposition 1 only proves that accepting positive gains makes the search objective monotonically increase, which is a direct algebraic conclusion and does not provide stronger optimization guarantees. Can the authors further clarify the substantive technical innovation of SCOUT relative to standard Greedy Forward Selection, and compare it with other budget-constrained greedy search or forward selection methods?

W2. The main performance improvements are limited and do not consistently surpass existing methods. Table 2 shows that SCOUT achieves +3.44 pp on Qwen3-1.7B, but this is lower than CAA’s +4.36 pp; on Qwen3-4B and 8B, the average accuracy improvements are only +0.01 pp and +0.26 pp, respectively. This indicates that its actual gains on larger models are very limited. Although SCOUT uses significantly fewer intervention locations, the paper mainly compares it with baselines using 100–4096 locations, making it difficult to determine whether its advantage holds under the same sparse budget. Can the authors conduct a fair comparison with CAA, SADI, and AUSteer under the same number of intervention locations, and explain why SCOUT’s gains almost disappear on 4B and 8B models?

Questions for authors

See weaknesses.

Reviewer_GYY4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a sparse steering method SCOUT. SCOUT steers a small number of MLP neurons (chosen by greedy selection on forward passes on a training set) to improve the performance of Qwen3 0.6B-8B on a multiple-choice dataset. The method improves performance similar to non-sparse steering method but achieves much higher sparsity.

Critical strengths

Cheap inference: The neuron scaling, once computed, is cheap to add to the model forward pass.

High sparsity: Only a few neurons are changed. This would in principle allow inspection and interpretation of these neurons.

The paper highlights the difference between relevance ranking and steering utility.

Critical weaknesses

Major: It is not clear how meaningful the performance improvements are: SCOUT optimizes neurons to improve performance on a specific distribution. It is well-known that fine-tuning LLMs on a narrow distribution can improve their performance there. This is particularly relevant as many of the LLMs used are very small. The paper should compare to common task-adaptation baselines such as few-shot prompting and cheap LoRA fine-tuning on the training set.

Major: SCOUT does not clearly outperform simple steering methods such as CAA. Table 2 shows that CAA frequently beats SCOUT.

Questions for authors

What benefits does the high sparsity of SCOUT, compared to dense methods like CAA, bring? In what use-cases would one prefer SCOUT over CAA or other dense steering methods?

The paper refers to a “workbook”, “git commit is recorded as unknown”, “author-provided full-14 ledger”. Please clarify what these refer to.

Reviewer_gDw9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work approaches the task of increasing accuracy of pretrained models on downstream classification tasks by steering neurons. Existing methods select intervention targets by relevance score but the authors claim this metric is not aligned with the actual objective: accuracy maximization on the task. They propose SCOUT, a method that, after selecting a subset of neurons by relevance, decides one by one which neurons to add to the final intervention set based on the marginal gain on the task. Each neuron is added to the set if it contributes, given the already selected set of neurons. The method operates in a budget-constrained setup. The authors show that it is possible to get similar accuracy to the existing methods by only steering a handful of SCOUT selected neurons.

Critical strengths

The paper finding about relevance score not being the correct metric to choose intervention sites, along with the proposal of a neuron-contribution based method, are important contributions that could guide the community towards better intervention design.

Critical weaknesses

While the findings are relevant, I have several concerns about the proposed method and its usefulness.

The authors point out that relevance score, in this case sensibility, is not aligned with the actual objective of maximizing accuracy, nonetheless this score is utilized not only to select the starting subset but also for choosing the order of neurons inspected by SCOUT. It is not clear how results would change by selecting a different starting point or different ordering criteria. Also, SCOUT performs a greedy selection potentially discarding neurons that could offer accuracy gains if added to a different set. While this point is addressed by the authors in the Limitations section, it would be interesting to understand how performance change by examining again rejected neurons.

It is not clear to me if sparsity would give a strong advantage over dense interventions performed by previous methods. The only advantage that comes to my mind is an interpretability argument about which minimal subset of neurons causally steers a model. This angle, however, is not explored and in general a compelling argument for the practical need for sparsity is not given.

Classification seems a weak domain of application for this method, why should someone undergo this pipeline instead of training a small LoRA or a linear probe?

Collateral effects of the method on the capabilities of the pretrained models are not investigated.

For these reasons I am leaning toward rejecting the paper. I am open to changing my score if the authors produce convincing arguments for the weaknesses I highlighted and answers to the questions.

Questions for authors

Could the authors show how the algorithm would perform under different ordering criteria or how the starting point affects performance? How does re-evaluating the contribution of previous discarded neurons affect the final selected set?

Is there a clear practical advantage that sparse intervention has over existing methods?

How does the method accuracy gain on classification tasks compare with using LoRA? is there some reason why this method should be preferred?

How does the intervention impact on the original capabilities (e.g. perplexity) of the model?

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

Related papers

Discussion (0)

Sign in to comment.