acceptodds
Under review as a conference paper at ICLR 2027

Steer Responsibly: Towards A Grounded Library of Interpretable Steering Vectors

Abstract

Can concepts grounded in human feedback become reliable activation-steering controls? We build Steer Responsibly to test this path: discover signed preference features, independently confirm their descriptions, extract model-specific vectors, and qualify their behavioral effects against quality and control criteria. Twelve of 32 features pass automated semantic confirmation, yielding 48 candidate tensors across Qwen3-8B and Gemma-3-4B-IT, each tested in both intervention directions. None meets the behavioral qualification criteria. Of 96 signed candidates, 86 select the unsteered fallback. Among ten held-out interventions, five miss the target margin in their means, two more fail random-control specificity in their means, and three meet all mean margins but fail uncertainty bounds. Concrete case studies expose limits on attribution: 36 signed candidates have no baseline score headroom; a random direction better expresses a supportive cue than an empathy vector; and near-identical answers receive opposite labels. A separate public-vector audit also finds order-sensitive judgments. We diagnose the gap between semantic grounding and qualified control. The findings make measurement validity part of the library itself to support future research.

Primary area:
datasets and benchmarks
Keywords:
LLM Steering

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_CRmNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a pipeline for discovering concepts in human preference data, confirming their descriptions, and turning them into model-specific steering vectors. The authors evaluate 48 candidate vectors across two language models and audit 84 existing vectors, recording their meanings, provenance, and evaluation results in a library. None of the newly constructed vectors meets the paper’s behavioral qualification criteria; the authors use this result to discuss limitations in both steering interventions and their evaluation.

Critical strengths
  1. Transparent reporting of unsuccessful results: The paper retains failed candidates and distinguishes semantic confirmation from behavioral qualification. Its analysis of score saturation, truncation, and judge instability provides useful documentation of potential evaluation pitfalls.
Critical weaknesses
  1. Unvalidated contrast signal: Human preference pairs can differ in correctness, length, tone, and other qualities at once. Assigning a concept label to a feature does not establish that the underlying pairs isolate that concept or provide a strong contrast for extracting its direction. The paper does not sufficiently validate this assumption, so the negative results cannot be attributed to the overall construction approach.

  2. Limited experimental scope: The study uses a narrow range of data, extraction settings, and evaluation prompts. Without broader experiments or stronger ablations, it remains unclear whether the findings reflect limitations of the construction pipeline or the particular settings tested.

Questions for authors
  1. Extraction validity: Do the 32 extraction pairs isolate the intended concept, or do they also differ systematically in length, content, and other attributes? Can the authors compare against controlled contrastive pairs and examine how results change with more extraction data?
  2. Robustness and statistical power: Do the findings persist across a broader layer and strength search, different prompt panels, and repeated extraction samples? More independent test coverage would also help distinguish small behavioral effects from judge variability.
  3. Presentation: The paper should more clearly distinguish its different evaluation criteria and cohorts, and consistently separate failure to meet the qualification standard from evidence that a steering direction is ineffective. Also the in-text citations appear to have a LaTeX formatting issue. Please check the citation commands and correct the formatting throughout the paper.
  4. Data construction detail: It would be better if the authors could provide representative preference-pair examples for the extracted concepts, and explain how they determined that these pairs isolate the intended concept rather than other response differences.
Reviewer_Z9DvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper attempts to build a library of interpretable steering vectors grounded in human preference data. It is pretty hard to understand the motive of the work beyond this framing.

Critical strengths
  • Important goal: Steering LLMs towards desired behaviors is an important problem.

  • Negative results: Majority of the results seem to be negative and the paper mentions that.

  • AI usage statement: The AI usage statement declares that AI was substantially used for every aspect of the work, it is good that this has been declared.

Critical weaknesses
  • Novelty: Even with multiple readings, it is hard to understand what the paper wants to propose and what the evaluation setup is, it is impossible to judge the novelty of the work in the current state.
  • What is being proposed?: There is no coherent formulation for what is being proposed.
  • Weak evaluation setup: Since the formulation is not coherent, it follows that the evaluation setup makes little sense.
  • Results: Although it is hard to understand what the results actually convey, on top of that the paper itself mentions that most of these are negative.
Questions for authors
  1. Can you clarify what is being proposed? A well-defined formulation of the methodology is required.
  2. Can you justify the positioning of the work with respect to the existing literature?
  3. Can you properly define the experimental setup and provide easy to understand tables?
  4. Can you re-write the various sections such that it becomes easy to understand? If AI is being used, please review the statements such that they coherently follow for us humans to be able to understand.

Note - I believe that the work needs a major overhaul in multiple aspects which might be beyond the scope of a rebuttal, though I would love to be proven wrong.

Reviewer_ip4hOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper appears to provide a negative result about the reliability of steering vectors, but it's very unclear.

Critical strengths

The questions of whether steering vectors are reliable and interpretable are significant.

Critical weaknesses

I found the paper extremely poorly written.

For instance, the title and abstract reference “Steer Responsibly” as a contribution, but this phrase appears nowhere else in the paper. Similarly, the title uses the words “interpretable” and “grounded”, but I wasn't able to determine what they are supposed to mean in this context.

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.