acceptodds
Under review as a conference paper at ICLR 2027

Steering Language Models Without Contrastive Pairs via Unsupervised Latent Region Matching

Abstract

Activation steering methods control language model behavior by intervening on internal activations, but many steering-vector approaches require contrastive prompts or paired good and bad completions. We study a lower-supervision setting where only a behavior dataset and the model's own zero-shot choices are available. We introduce Unsupervised Latent Region Matching (ULRM), which splits examples into successful and unsuccessful subsets and learns a nonlinear map from unsuccessful prompt-side decision activations toward the successful activation region. Our loss objective encourages the intervention to shift activations toward preferred behavior while preserving task-relevant information. We evaluate ULRM on two sycophancy datasets using Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct. We compare two vector baselines. MeanDiff subtracts the mean unsuccessful activation from the mean successful activation in the same split. Contrastive Activation Addition (CAA) averages paired preferred-minus-nonpreferred completion activations. Despite CAA's stronger paired supervision, our best-performing method, MMD-ULRM, leads both baselines in all eight comparisons, averaging 0.604/0.541 accuracy/probability versus MeanDiff's 0.527/0.487 and CAA's 0.503/0.457. These results suggest that behavior-split latent-region matching is a promising route to practical activation steering without paired contrastive completions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Mechanistic Interpretability, Activation Steering, Representation Engineering, Model Alignment

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_r6jLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies activation steering in the setting where neither contrastive prompts nor paired preferred and nonpreferred completions are available. The proposed method, Unsupervised Latent Region Matching (ULRM), first runs the model zero-shot on an A/B behavior dataset and splits the examples by whether the model already gives the preferred answer. Prompt-side decision activations at one layer then form a preferred region and a nonpreferred region. A residual MLP is trained to move nonpreferred activations toward the preferred region while staying close to the identity on preferred activations. The pull toward the preferred region is measured either by a weighted nearest-neighbor distance (kNN-ULRM) or by a multi-scale kernel maximum mean discrepancy (MMD-ULRM). At inference, the learned map replaces the hidden state at the decision position. Experiments on two selections of a model-written sycophancy dataset with Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct compare both variants with the unsteered model, a mean-difference vector built from the same split, and Contrastive Activation Addition. The paper reports that MMD-ULRM achieves the highest average accuracy and correct-token probability, and further studies the identity term, the nonlinearity of the map, the training-set size, unbalanced training pools and computational cost.

Critical strengths

S1. Nonlinear generalization of mean-difference steering

The paper replaces a single global displacement between success and failure activations with a learned map between the two activation distributions. The architecture ablation supports this design. Replacing the MLP with a residual affine map lowers both metrics in every model, pool, and method combination (Table 2, lines 313-316). On Llama with the PAS pool, MMD-ULRM also exceeds the mean-difference baseline built from the same split by a clear margin (Table 1). These results indicate that the difference between preferred and nonpreferred decision states is not fully captured by one translation, which is the main empirical contribution of the submission.

S2. Controlled and transparent protocol

Layers, strengths, and checkpoints are selected on a held-out validation set. The data-scaling study fixes the number of optimizer updates across training-set sizes, which separates data quantity from optimization time (lines 345-351). The paper reports the additional compute of ULRM relative to the vector baselines (Table 4) and lists every selected configuration (App. C-D). This level of detail makes the empirical claims checkable and the method straightforward to reproduce.

Critical weaknesses

W1. Limited methodological novelty

The motivation of steering from observed successes and failures is reasonable, but the methodological novelty is limited. Partitioning examples by baseline success was introduced by Painless Activation Steering, which the paper cites. Learning a map that moves one unpaired activation distribution toward another has been studied in activation-transport steering[1] and in affine distribution-matching steering[2]. Relative to these works, the new element is mainly the use of a residual MLP with a kNN or MMD discrepancy and an identity term on the preferred region. The paper does not discuss these transport-based methods, clarify how region matching differs from them conceptually, or compare against them, even though they address the same unpaired setting. Beyond the choice of discrepancy and mapper, the paper offers limited insight into what the learned map changes in the representation and why this leads to the preferred behavior.

W2. Unclear supervision gap with CAA

The paper presents ULRM as using weaker supervision than CAA, and the title describes the method as unsupervised. However, constructing the split requires knowing, for every example, whether the model's choice is the preferred answer. In a two-choice task, the model's choice together with its correctness identifies the preferred answer, which is the same label CAA uses to build its pairs. The difference between the two methods therefore appears to lie in how the label is used, as an unpaired split versus paired completions, rather than in the amount of supervision. The main result is framed as outperforming CAA despite a supervision gap, so its significance depends on whether such a gap exists in the evaluated setting. Without this gap, the relevant comparison set also extends to methods that learn interventions from the same labeled activations, such as representation finetuning[3] and inference-time intervention[4], none of which is evaluated.

[1]: Controlling Language and Diffusion Models by Transporting Activations

[2]: Representation Surgery: Theory and Practice of Affine Steering

[3]: ReFT: Representation Finetuning for Language Models

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

Questions for authors

Clarifications that could affect the assessment:

In a two-choice task, the behavior split appears to identify the preferred answer for every example. In what sense is the supervision used by ULRM weaker than that used by CAA?

How does ULRM differ from activation-transport steering[1] and affine distribution-matching steering[2], and how does it compare with them empirically on the same data?

The sycophancy data pair a limited set of questions with many different user profiles. Are the training and test sets separated at the question level, and how do the results change when test questions are unseen during training?

Suggestions for future work:

The objective preserves per-example information only on the preferred region. An analysis of what the map changes in the representation may help clarify whether it moves activations toward the correct answer or toward regions associated with easier questions.

[1]: Controlling Language and Diffusion Models by Transporting Activations

[2]: Representation Surgery: Theory and Practice of Affine Steering

I would be happy to raise the score if the author could provide more evidence to answer my question.

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

This paper proposes ULRM, a method for steering LLMs through activation interventions without requiring contrastive pairs. The main idea behind the paper is to exploit the model's own behavior on a dataset to identify two groups of hidden representations: those corresponding to successful (preferred) behavior and those corresponding to unsuccessful (non-preferred) behavior. Instead of learning a single steering direction, ULRM learns a nonlinear transformation that maps unsuccessful activations toward the region associated with successful behavior.

The propose method involves 3 stages: 1) evaluate the unmodified model and partition the dataset according to whether the model produces the preferred answer. 2) extract hidden states at and train a residual multilayer perceptron to transform unsuccessful states toward successful ones. 3) at inference time, the learned transformation replaces the model's hidden state at a selected layer.

The approach has been tested on two datasets (effectively one domain), and two model families.

Critical strengths
  1. The paper addresses an important limitation of activation-steering methods, namely, their dependence on explicitly constructed contrastive examples.
  2. The experiments show improvements over the evaluated baselines.
  3. The method is presented clearly, and its objective is sufficiently explained.
Critical weaknesses

My main reservations regarding this paper which are rather critical are:

  1. The evaluation is very limited, less than a typical ICLR paper. Thus, its unclear if the results are generalizable. The paper only evaluates sycophancy. Although the paper uses two datasets, both are derived from the same broad evaluation setting and involve closely related decision tasks.
  2. The evaluation is restricted to single-token multiple-choice predictions.
  3. The first and second weakness also means that some claims made in the paper such as "introducing a general framework for activation steering" are not backed by evidence.
  4. The comparison with existing steering methods is not sufficiently comprehensive. Further baslines could be considered.
  5. The paper also lacks statistical significance tests.
Questions for authors

Please see the weaknesses.

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

The paper introduces Unsupervised Latent Region Matching (ULRM), which divides examples into successful and unsuccessful subsets and learns a nonlinear mapping that shifts prompt-side decision activations from unsuccessful examples toward the activation region associated with successful examples.

The work runs the model on behavior dataset and then categorize two groups that judges whether the model exhibits the prefered behaviors. Next, they collect the layer activation for further analysis.

Critical strengths
  1. Simplicity and effectiveness.The method leverages lightweight activation signals and achieves improvements over the two baselines considered in the paper. Its core idea is intuitive: divide examples according to whether the model exhibits the preferred behavior and use the resulting groups to guide activation steering. However, the notion of preferred behavior remains somewhat underspecified. It would be helpful to clarify how preferred behaviors are defined and how this definition can be generalized to different behavioral objectives.
Critical weaknesses

Weaknesses

The following are my concerns not AI:

  1. Limited experimental validation and baseline comparisons. I do not find the experimental evaluation sufficiently extensive. The paper compares ULRM with only two baselines, and the number of comparative experiments is limited. My main concern is that the evaluation relies on two relatively simple baseline settings rather than a broader set of established methods from prior work. The authors emphasize that ULRM eliminates the need for contrastive pairs through unsupervised latent region matching, but the current experiments do not sufficiently isolate or validate the benefits of these design choices. I recommend including more competitive and representative baselines, together with targeted ablation studies and additional experiments that investigate the contribution of latent region matching itself. These additions would provide stronger evidence for the paper's central claims.

  2. Insufficient analysis of Figure 2. I find Figure 2 difficult to interpret, as it is unclear how the visualization supports the paper's main contributions and claims. The authors state that Figure 2 visualizes a learned mapping in a fixed PCA projection, but provide little further discussion or analysis of what the visualization reveals. I encourage the authors to explain the key observations from this figure and clarify how they provide insights into the effectiveness of ULRM, particularly regarding how unsuccessful activations are mapped toward the successful activation region.

  3. Insufficient discussion of training-size results. The paper presents training-size results, but the relationship between training-set size and the proposed ULRM method is not sufficiently explained. It remains unclear what these results reveal about the method's effectiveness, data efficiency, or dependence on training data. The authors should provide a more thorough analysis of these results and explain how the observed trends support their claims. In particular, comparisons with the baselines across different training-set sizes could help clarify whether ULRM offers consistent advantages under limited-data conditions and how its performance scales with additional training data.

Questions for authors
  1. As mentioned in the weaknesses, the authors should provide a clearer justification for their baseline selection. Why do they choose only these two baselines? Given that the proposed method aims to steer model behavior without contrastive pairs, why are these particular methods or settings appropriate for evaluating its effectiveness? A more comprehensive explanation and additional comparisons with relevant prior methods would strengthen the evaluation.

  2. What insights does Figure 2 provide in relation to the paper's main claims? It is currently unclear how the visualization supports the proposed method or demonstrates the effectiveness of ULRM. The authors should provide a more detailed discussion of the observed patterns and explain how they relate to the underlying mechanism and claimed benefits of the proposed approach.

  3. How are preferred behaviors defined in the proposed method? Do these behaviors vary across datasets or experimental settings, and, if so, how does ULRM accommodate such variations? The authors should clarify how successful and unsuccessful examples are identified and discuss how the definition of preferred behavior affects the applicability and generalizability of the method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.