acceptodds
Withdrawn from ICLR 2027

A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Steering Vectors in Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) possess impressive generative capabilities but remain opaque and can exhibit unsafe or undesired behaviors. Existing control methods rely on supervised fine-tuning or curated prompt-response datasets, which limits their scalability. We propose SteerCLR, an unsupervised method that simultaneously discovers a bank of diverse and disentangled steering vectors directly from unlabeled prompts. By optimizing a novel contrastive objective over internal model activations, SteerCLR learns vectors that correspond to distinct behavioral shifts. Injecting these vectors into a frozen LLM enables fine-grained, low-latency control over generation, including suppressing toxicity, modulating sentiment, and uncovering subtle stylistic dimensions, without relying on labeled data, classifiers, or attribute-specific supervision. We demonstrate that optimizing for activation magnitude and activation diversity yields a rich set of interpretable directions. Experiments on instruction-tuned Llama-2-13B-chat model show SteerCLR discovers diverse interpretable steering vectors in a single training run, significantly advancing the scalability of mechanistic interpretability and enabling practical interventions for safety, alignment, and model auditing.

Primary area:
interpretability and explainable AI
Keywords:
Steering vectors, interpretability, large language models

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_KtgfOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SteerCLR, an unsupervised method for jointly learning a bank of steering vectors on a frozen LLM. The method combines downstream activation-shift maximization with a contrastive objective that encourages consistent effects across prompts for the same vector and distinct effects across different vectors, without requiring behavioral labels during training.

Critical strengths

Unlike conventional methods such as CAA, which require predefined behavioral concepts and corresponding examples, SteerCLR jointly optimizes a bank of vectors using unlabeled prompts. The contrastive objective operates on the downstream activation changes induced by steering, which is a reasonable design choice, since geometrically distinct vectors do not necessarily produce distinct downstream effects.

Critical weaknesses
  1. The main evaluation measures steering effects along seven predefined behavioral axes. Although SteerCLR achieves higher functional effective rank and lower behavioral collision rates, these metrics only describe how the learned vectors respond along those seven axes. They do not establish how many distinct, interpretable concepts the method actually discovers. The qualitative examples also provide limited support for the central discovery claim. Most demonstrated behaviors, such as switching languages, adopting a pirate persona, or changing writing styles, are familiar concepts that can already be controlled using methods like Activation Addition and CAA.
  2. SteerCLR maximizes the magnitude of downstream activation shifts and encourages separation between their directions. However, large activation changes do not necessarily imply effective steering. They may also introduce undesirable changes in generation quality or unrelated behaviors. Likewise, distinct activation shifts do not necessarily correspond to distinct semantic effects. The main evaluation focuses on the distribution of steering effects across seven behavioral axes, measured using multiple-choice answer probabilities. It does not directly establish whether the learned vectors provide stronger, more consistent, or more reliable behavioral control on natural generation tasks.
Questions for authors
  1. This paper uses different bank sizes, model architectures, and training prompt distributions for the results. It does not explain these particular choices. Do the authors have results showing how the number of vectors affects behavioral coverage or redundancy? In particular, does increasing the bank size discover additional concepts, subdivide existing concepts, or primarily introduce redundant vectors?
  2. The qualitative analysis relies on Qwen2.5-72B-Instruct to assign labels based on a small number of paired generations. Have the authors examined whether different labeling models or different prompt subsets produce consistent labels for the same vector?
  3. For concepts identified by the labeling model, such as language switching or refusal, could the authors compare SteerCLR vectors against vectors obtained through CAA on the same base model? This would help clarify what unsupervised discovery provides beyond established methods.
  4. The abstract states that discovering steering vectors with distinct, interpretable effects remains difficult. However, existing approaches can construct effective steering vectors when the target concepts and corresponding examples are specified. The difficulty addressed here appears to be discovering diverse behavioral directions without predefined concepts, rather than constructing steering vectors in general. I suggest making this distinction more explicit.
Reviewer_SntNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes SteerCLR, an unsupervised method for jointly learning a bank of steering vectors through contrastive learning. The method injects steering vectors into an intermediate layer of a LLM and measures the effects through activation changes at a later layer. It combines magnitude objective and contrastive objective to promotes large downstream activation changes. Experiments show SteerCLR improves the functional diversity of learned steering vectors and achieving higher functional effective rank and lower behavioral collision rates.

Critical strengths
  1. This is a novel approach to learning diverse steering vectors. It encourages diverse downstream activation effects rather than simply maximizing activation changes. It is an interesting method for unsupervised steering vector discovery.
  2. The unsupervised nature of the proposed method is a promising strength, as it may enable the discovery of useful steering vectors without requiring predefined behavioral labels. This also suggests the potential to scale up the discovery.
Critical weaknesses
  1. The proposed method optimizes entirely based on downstream activation changes. However, large activation changes do not necessarily correspond to meaningful behavioral changes and may instead disrupt the model's original capabilities. This concern is further amplified by the lack of systematic evaluation of the model's behavior and performance after applying the learned steering vectors. The authors only provide a few examples, which are insufficient to demonstrate that the learned vectors induce meaningful behavioral changes.
  2. The motivation behind the method is not clear. Activation steering is typically used to induce specific, desired behaviors in language models. However, this paper focuses on discovering a diverse set of steering vectors without knowing what behaviors these vectors actually induce. It is unclear why maximizing the diversity of steering effects is a meaningful objective, especially when the learned vectors may not correspond to useful or interpretable behaviors. The authors should better justify the practical value of discovering diverse steering vectors without explicitly considering their behavioral effects.
Questions for authors
  1. How do the learned steering vectors affect the model's behavior? For example, can the discovered vectors reliably steer the model to respond in French, improve its safety, or induce other meaningful behavioral changes? Could the authors provide a more systematic evaluation using existing benchmarks? There are many benchmarks for evaluating different aspects of model behavior, so such an evaluation should be feasible. The few qualitative examples currently provided are not sufficient to demonstrate that the discovered steering vectors are genuinely useful or that their effects are consistent across different prompts.
  2. Among the steering vectors learned in a single training, what proportion actually induces meaningful and interpretable behavioral changes? Since the method aims to discover a diverse set of steering vectors, it would be useful to understand how many of them correspond to identifiable and useful behaviors, rather than merely producing distinct activation changes.
Reviewer_vMaoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces SteerCLR, which jointly learns a bank of fixed-norm steering vectors on a frozen language model without behavioral labels in the training objective. It combines downstream activation-shift maximization with a Circle Loss objective: effects of the same vector across prompts form positive pairs, and effects of different vectors form negative pairs.

The main experiment compares SteerCLR with a matched magnitude-only ablation on Llama-2-13B-chat across six paired training seeds. On seven behavioral panels, SteerCLR increases functional effective rank from 4.659 to 5.487 and reduces dominant-axis concentration from 0.850 to 0.426. Separately trained banks provide qualitative examples on three models.

Critical strengths
  1. The research question is well motivated.

Discovering interventions with distinct effects is a useful extension of finding interventions that merely cause large activation changes. Contrasting downstream effects provides a natural way to encourage both consistency across prompts and separation across vectors.

  1. The paired ablation provides credible evidence for the narrow empirical claim.

Functional rank improves and dominant-axis concentration decreases in all six pairs. The paper appropriately treats training seeds as independent replications.

  1. Several evaluation choices improve interpretability of the results.

The mass-aware score addresses a limitation of conditioning solely on the two answer options. The paper also clearly separates the controlled experiment from descriptive examples and discloses the use of behavior-specific prompt pools for the latter.

Critical weaknesses
  1. Diverse activation changes do not establish interpretable behavior.

The objective encourages large, consistent, and distinct activation shifts, but these could also produce consistent off-topic or incorrect answers. Some examples show this problem: the division-by-zero response adopts a style but loses the explanation. Please test whether descriptions of vector effects predict behavior on held-out prompts, and report how often those effects remain coherent and relevant to the task.

  1. It is unclear how to select a useful vector for a given task.

The paper mentions labeled selection questions and separately reports LLM-based vector labeling, but does not fully specify how to choose a vector, its sign, and its strength. Higher bank-wide diversity does not show that this selection succeeds. Could you define the selection rule, state any human or LLM assistance it requires, and evaluate the selected vectors on held-out questions? This concerns selecting a vector in advance, not an assumed need to call another LLM during every inference.

  1. The comparisons do not establish the method's added value or practical cost.

For the named behavioral targets, the paper also lacks an effectiveness and efficiency comparison with a label-based method such as CAA. Such a comparison should cover the full selection-and-steering procedure, response quality, labeled-data requirements, and training, labeling, selection, and inference costs. This would clarify the trade-off for targeted control; it need not establish superiority over supervised methods.

  1. The discovery process still uses behavior-informed prompts and semantic screening.

Training uses 1,932 questions drawn from seven behavioral data sources, and the qualitative banks use behavior-specific prompts. These choices are disclosed, but their contribution to the discovered behaviors is unclear. A small comparison with general prompts, or with one behavioral source omitted, would help assess how much the results depend on prompt curation. The prompts used for configuration screening should also be identified.

  1. The controlled results are limited to Llama-2-13B-chat.

The other models provide selected examples rather than matched quantitative comparisons. This matters because a major gain is reduced concentration on refusal, which may depend on Llama-2's alignment characteristics. Repeating the core ablation on one more recent strong model would test whether the finding extends beyond this setting.

Questions for authors
  1. Do descriptions assigned to vectors predict their effects on held-out prompts, while preserving coherent, task-relevant answers?

  2. What exact rule selects a vector, its sign, and its strength for a target behavior? What labels, human inspection, or LLM analysis does it require?

  3. How does SteerCLR compare with a geometric-diversity control and, for named targets, CAA? What are the complete selection-and-steering costs and held-out effects?

  4. How sensitive are the findings to the curated prompt pool, and which prompts informed LLM-based configuration screening?

  5. Does the core quantitative improvement persist on a more recent strong open-sourced model?

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

Related papers

Discussion (0)

Sign in to comment.