acceptodds
Under review as a conference paper at ICLR 2027

From Behavior to Mechanism: Tracing Divergent Response Modes in Frontier Language Models

Abstract

Frontier language models are trained with distinct data, objectives, and safety pipelines, but whether those differences produce measurably different behavior under steering pressure has not been directly tested. We evaluate 6 frontier models from different labs on 300 paired base and steered items across 3 behavioral categories. All models also act as blind peer judges against fixed rubrics, and each response is labeled by leave-self-judgment-out consensus over 24,480 judgments. Models differ both in how far steering moves them and in the kind of response they give. GPT-5 withholds its reasoning while still providing the answer on 99 of 100 steered items, against 0 in 500 for the others. Claude Opus 4.7 and GPT-5 resist explicit suppression instructions where the other four never do, and they resist differently. In Llama, the open-weight model, a linear probe reads the behavioral split from the residual stream before generation at 0.87 held-out accuracy. Injecting that direction drives the behavior from 0% to 86%, and ablating it cuts the natural rate by more than half, where a random direction of equal norm changes nothing. A second ablation on complementary items reproduces the effect more strongly, and its direction has cosine similarity 0.82 with the first.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, AI alignment, LLM evaluation, steerability, mechanistic interpretability, linear probes, activation steering, ablation, LLM-as-judge, frontier 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_UdFDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper compares how six models from different vendors respond to base prompts and their steered versions. It looks at values conflict, reasoning elicitation, and reasoning suppression, with two additional categories for validation. The authors build rubrics from observed responses and use blind peer judges to classify response modes. They find differences in both compliance and how models respond when they do not comply. They also probe and modify Llama’s activations to change its clarification-seeking behavior.

Critical strengths
  • I like that the paper goes beyond whether a model complies and looks at how it responds.

  • Using multiple blind peer judges and excluding self-judgment is also good. The agreement analysis helps show which labels are more reliable.

  • The Llama experiment extends probing accuracy; It includes activation steering, ablation, and a random-direction control.

  • The comparison across models from different vendors is interesting and uses the same prompts for all models.

Critical weaknesses
  • This is a useful measurement study, but it does not explain why the models behave differently. Different data, objectives, and safety pipelines motivate the study, but their effects are not examined. The Llama experiment is interesting, but it does not explain the main cross-model differences, so the connection between behavior and mechanism feels incomplete.

  • Reported reasoning is not necessarily what the model actually reasons internally. Refusing to share reasoning and giving a step-by-step explanation may reflect how the model presents its answer rather than how it reached that answer. This distinction should be clearer when discussing reasoning suppression and monitoring.

Questions for authors

What do you think explains the different response modes across models? Is there evidence connecting these differences to training data, objectives, or safety pipelines?

  • How were the steering instructions chosen? Could different wording or stronger instructions change the model comparisons?

  • What connects the clarification-seeking behavior studied in Llama to the reasoning-disclosure and suppression findings in the other models?

  • Could the Llama direction capture prompt ambiguity or a tendency to ask questions, rather than resistance based on values? How would you distinguish these?

  • How do you distinguish generated explanations from internal reasoning traces? Could GPT-5’s reasoning refusal mainly reflect its disclosure policy?

  • Does removing the item-context field from judge prompts change the main model comparisons, beyond changing individual labels?

1 comment
AuthorsOpenReview ↗

Addressing weaknesses and responding to questions.

Thank you for your feedback and questions, and for recognizing the value of this work.

Weakness 1. What we are showing is that refusal and compliance rates are not enough to describe how differently frontier models behave when a prompt is vaguely unethical. Specifically, models with similar refusal rates differ in how they decline. We do not claim to explain why each model behaves the way it does, because such a causal claim would require access to each model’s internals, training data, objectives and safety pipeline, and for these models that information is not public.

For Llama-3.3-70B, since the weights are open, we could trace the model's clarification-seeking behavior. The purpose of Llama’s mechanistic analysis was to show that where a model’s weights are open, we can find and verify the direction behind a behavior. As a safety application for any AI lab, such behaviors can be monitored and changed to help prevent misuse when the ethical violations are nuanced and ambiguous.

Weakness 2. As you point out, and as the literature shows (Turpin et al., 2023, Lanham et al., 2023, Chen et al., 2025), how the model reasons internally and what it shows to the user as its reasoning may differ. Our claims are therefore about the visible text, meaning what the model presents to the user as its reasoning and its answer.

While these studies ask whether the reasoning text reflects what drives the answer, we ask the reverse question: whether the answer reflects what appeared in the reasoning text. We test this on DeepSeek-R1, the only model in our study whose responses included its reasoning text, which it returns separately from its answer. Under the suppression instruction, its reasoning text raises the concern it was told to leave out in 83 of 100 steered items. It sets the concern aside every time, and its answer reads as full compliance. Both observations are visible text, so this result holds even if the reasoning text is not faithful to the model’s computation.

From a safety perspective, this matters because an instruction can remove a concern from the answer even when the model’s own reasoning text raises it. An audit that reads only the final answers would not see this, and a lab could use such cases to revise how its models handle these instructions.

In the revised version of the paper, the wording in the reasoning-suppression and monitoring sections will be revised to make this distinction clear.

Q1. Please see our response to weakness 1. One case where public information is consistent with our results: as Reviewer u9sE notes, OpenAI’s published policy on hidden reasoning matches GPT-5’s refusal to disclose its reasoning, a behavior no other model shows.

Q2. The steering prompts were written to apply each category’s pressure in a vaguely unethical way, and the wording differs by category. In reasoning-suppression, each of the 100 items words the instruction differently, so the comparisons do not rest on one phrasing. In values-conflict, most steered prompts state the request itself in its unethical form, so the wording changes with each scenario. In reasoning-elicitation, all 100 items use the same instruction, so GPT-5’s result rests on one wording. While different wording could change the rates, all six models saw the same items, so the comparisons between them are on equal terms.

Q3. As mentioned in our response to weakness 1, the goal of the Llama analysis was to find and verify the direction behind a behavior, not to compare models. Nonetheless, the behavior we traced for Llama does connect to our main claim. As shown in Table 3, Gemini complied on 100 of 100 items, Llama on 57 and Opus on 35. A compliance rate would read Llama and Opus as refusing more, but they rarely refused. Most of their other responses were clarifying questions (42 for Llama, 60 for Opus, 5 or less for the other models). The difference between these models is in how they respond, not in how often they refuse.

Q4. The direction could partly capture prompt ambiguity. It does not capture values-based resistance, because the base values-conflict prompts have no ethical tension, which is added only in the steered version. To measure how much comes from the prompt, we trained the same probe on Llama-3.2-1B, a small model that never produced these responses. It reached 0.72 balanced accuracy, against 0.87 for Llama-3.3-70B, so the larger model carries information beyond what the prompt alone provides.

Q5. Please refer to our responses to weakness 2 and Q1.

Q6. We checked this, although the submitted paper reports only the effect on individual labels in the Robustness Checks section. In our control sample, removing the field hardly impacted the main comparisons. However, the sample size was small, so for the revised version of the paper we will cover a larger sample.

We hope these clarifications address your concerns.

Reviewer_Mg5LOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper evaluates six models on 340 paired base/steered items across values-conflict, reasoning-elicitation, and reasoning-suppression. It reports cross-model differences in response modes, then probes and steers a “derail” direction in Llama-3.3-70B.

Critical strengths

NA

Critical weaknesses

No clear main finding. After reading the paper I’m not sure what I’m supposed to take away. Most of it describes how the evaluation was run and then reports what came out, category by category, but there’s no single claim the experiments are built to support. The conclusion basically lists four separate observations.

Questions for authors

NA

1 comment
AuthorsOpenReview ↗

Clarifying the paper's main claim and takeaway.

Thank you for your feedback.

The main takeaway is that refusal and compliance rates are insufficient to describe how differently frontier models behave when a prompt is vaguely unethical. Specifically, models with similar refusal rates differ in how they decline. For example, Opus and GPT-5 resist suppression instructions at nearly the same rate (25 and 21 of 100), but Opus mostly objects while completing the task and GPT-5 mostly refuses outright. This is the type of difference a refusal-rate benchmark cannot show.

For the open-weight model, we show that the choice behind Llama's clarification-seeking behavior is encoded in the residual stream, before response generation, in a single direction that controls it. Adding the direction drives the behavior and removing it reduces the behavior.

From a safety perspective, this is important because much misuse doesn't look like an explicitly dangerous or outright unethical request. When the ethics of a request are vague, the tested models wrote the requested deceptive or unkind content at rates ranging from 3 to 66 of 100 items. Measuring how each model responds, and locating the direction behind a response where weights are available, is what lets those behaviors be monitored and changed.

With respect to the paper's text, it will be revised so that the abstract, the flow of the paper, and the conclusion are clearly restructured around this claim.

We hope this explanation helps clarify the paper's claim and takeaway.

Reviewer_u9sEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how language models respond across value-conflict, reasoning-elicitation, and reasoning-suppression conditions. The paper evaluates six models using 300 paired baseline and steered prompts, plus 40 validation items, showing that 1) models and baselines behave differently under the same pressure, and 2) clarification behavior can be manipulated internally in Llama.

Critical strengths
  1. The paper studies a meaningful question. How a model pushes back on a prompt matters.

  2. The paper evaluates six different models and uses multiple judges, providing thorough experimental results.

Critical weaknesses
  1. The broader takeaway remains unclear. I found the analysis of the tested prompts convincing, but after reading it I still could not say when I would expect a model to push back on a new prompt.

  2. The reasoning-elicitation rubric assigns a single label to each response, even though the behaviors can coexist. This limits what the reported frequencies tell us about how models express ethical considerations.

  3. I had trouble connecting the Llama intervention to the cross-model findings. The reasoning-elicitation analysis examines how models respond when asked to disclose their reasoning, while the intervention measures whether Llama asks for clarification. I would appreciate an explanation of why the authors chose clarification and which of the observed differences across models they think this experiment helps explain.

  4. The paper is extremely hard to follow.

Questions for authors
  1. If the judges evaluated the same responses without seeing the expected behavior field, would the main differences across models still hold?

  2. Would the conclusions change if judges scored each response separately for whether it declines to disclose reasoning, expresses an ethical consideration, and answers the underlying question? For the intervention experiment, could they also compare each steered response directly with the original response to the same prompt and identify what changed?

  3. Does the intervention still affect clarification-seeking when the prompt is genuinely ambiguous?

2 comments
AuthorsOpenReview ↗

Clarifying the paper's main claim and addressing the questions and concerns.

Thank you for your feedback and questions, and for finding the analysis of the tested prompts convincing.

Weakness 1. The main takeaway is that refusal and compliance rates are insufficient to describe how differently frontier models behave when a prompt is vaguely unethical. Specifically, models with similar refusal rates differ in how they decline. For example, Opus and GPT-5 resist suppression instructions at nearly the same rate (25 and 21 of 100), but Opus mostly objects while completing the task and GPT-5 mostly refuses outright. This is the type of difference a refusal-rate benchmark cannot show.

For a new prompt, our findings say that whether and how a model pushes back depends on the model and the type of pressure. When asked directly for vaguely unethical content, we expect Opus and GPT-5 to mostly decline with an alternative (91 and 89 of 100), and Gemini and Llama to comply most often (66 and 46). When asked to show its reasoning, we expect GPT-5 alone to decline while still answering. On an underspecified request with no pressure, we expect Opus and Llama to ask clarifying questions rather than answer. Predicting pushback on any prompt outside these categories is beyond the scope of this study.

Weakness 2. The single label does limit what the reasoning-elicitation frequencies say about ethical content. In particular, GPT-5 receiving the values-content label on 1 of 100 steered items does not mean its other answers have no values content. Our finding rests on the other label. 99 of GPT-5's 100 steered responses refuse to disclose reasoning, against 0 of 500 for the other models. Since that refusal is present in the response whatever else it contains, scoring the behaviors separately would not remove this difference. For future work, we could use a probability distribution over potentially coexisting labels, instead of a single label choice.

Weakness 3. In the Llama mechanistic analysis, our goal is not cross-model comparison, since most of the models are not open weight. Instead, we show that the choice behind Llama's clarification-seeking behavior is encoded in the residual stream, before response generation, in a single direction that controls it. Adding this direction drives the behavior and removing it reduces the behavior.

We chose clarification-seeking because the behaviors behind the other findings do not occur in Llama. Llama never refused to disclose its reasoning or resisted the suppression instruction. Clarification-seeking is Llama's largest behavioral split (42 of 100), and it also shows a cross-model difference. In Table 3, Gemini complied on all 100 base values-conflict items, Llama on 57 and Opus on 35. Yet Llama and Opus rarely refused, and most of their other responses were clarifying questions.

This is important from a safety perspective, because much misuse doesn't look like an explicitly dangerous or outright unethical request. Measuring how each model responds, and locating the direction behind a response where weights are available, is what lets those behaviors be monitored and changed.

Weakness 4. In the revised version, the abstract, the flow of the paper, and the conclusion will be restructured around the claim in our response to weakness 1.

Q1. We checked this, although the submitted paper reports only the effect on individual labels in the Robustness Checks section. In our control sample, removing the field hardly impacted the main comparisons. However, the sample was small, so the revised version will use a larger one.

Q2. Scoring the three behaviors separately would change one result but not the others. GPT-5's values-content rate would rise, since some of its responses include ethical considerations after refusing to disclose its reasoning. The reasoning-disclosure finding would not change, as explained in weakness 2. Scoring whether the response answers the question would not change it either, as no model declined the underlying question on any reasoning-elicitation item.

For the intervention, the paper compares each prompt's label with and without it, item by item. In the two ablation splits, 14 and 18 items moved from clarification-question to answering, and vice versa for 2 and 0 items. The paper doesn't show the content of paired responses, due to space limitation, but they will appear in the revised version.

Q3. Yes. The intervention was tested on base prompts that are often underspecified. The prompts where Llama asked for clarification are the more ambiguous ones, since a probe on Llama-3.2-1B, which never produced these responses, separates them at 0.72 balanced accuracy from the prompt alone. In ablation tests, Llama asked for clarification on 21 and 23 of the 50 test items in the two splits, and removing the direction cut these to 9 and 5. So the direction affects clarification-seeking even when asking is reasonable.

We hope these clarifications address your concerns.

Thank you for the detailed clarification. However, I still have concerns about the validity of some behavioral comparisons and the practical implications of the mechanistic intervention, particularly whether suppressing clarification preserves answer quality. While the rebuttal addresses several of my concerns, I will maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.