acceptodds
Under review as a conference paper at ICLR 2027

In the Shadow of Reasoning: Exploring Behavioral Steering in Large Reasoning Models

28 views

Abstract

In large reasoning models (LRMs), activation steering effectively controls chain-of-thought (CoT) progress but has limited success in behavioral steering, such as refusal. This study reveals that this disparity arises from the activation geometry of LRMs' CoT. During reasoning, the CoT-progress direction captures the dominant variation in CoT activations, while the weaker refusal signal remains in its shadow. This geometry poses challenges for refusal steering in both identifying the refusal direction and controlling intervention strength. To validate these insights, we propose Substeer, which extracts a clear refusal direction and performs dynamic strength control for effective refusal steering while preserving CoT coherence. Experiments on four LRMs and six jailbreak attacks demonstrate the effectiveness of Substeer. Further evaluation of other behavioral steering tasks, such as myopic reward, supports the generalizability of our findings. We open-source our code and data at https://anonymous.4open.science/r/Substeer/.

Primary area:
interpretability and explainable AI
Keywords:
Large Reasoning Models, Chain-of-Thought, Activation Steering, Representation Engineering, Behavioral Steering, AI Safety

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WS2GOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper attributes difficulties in refusal steering to dominant CoT-progress variation and comparatively weak behavioral signals. It proposes SubSteer, combining pre-decision direction extraction, progress-dependent intervention strength, and multi-layer steering, and evaluates it across four reasoning models and additional behavioral tasks.

Critical strengths
  • Effective steering. SubSteer improves bidirectional refusal control over several existing methods. Extracting directions at the pre-decision boundary and distributing stage-dependent interventions across layers provides a concrete approach to balancing steering success and CoT completion.
  • Relevant comparisons. The evaluation includes five steering baselines, four model families, and ablations of direction extraction, strength adaptation, and layer distribution. These comparisons provide useful evidence for the contribution of the proposed components.
Critical weaknesses
  • Incomplete causal explanation. The geometric measurements establish differences in signal scale, but do not establish that progress dominance causes steering failures. Improved completion under smaller, distributed interventions could also reflect generic benefits of limiting perturbation magnitude, rather than preservation of a specific progress representation.
  • Insufficient quality assessment. Producing an end-of-thinking token within the budget does not establish coherent or effective reasoning. The claim of preserving CoT quality requires evaluation beyond termination, including reasoning performance and benign-task utility. Substantial unclosed rates also remain on Qwen3.5-9B.
  • Unclear need for progress estimation. The fixed-strength ablation supports adaptive intervention, but does not establish the necessity of activation-based progress estimation. A simpler token-count-based schedule, matched for direction, layers, and intervention budget, would clarify the added value of the proposed calibration.
Questions for authors
  • Can the authors use finer-grained annotations of reasoning stages, such as problem analysis, derivation, and verification, to clarify what the estimated progress bins actually represent?
  • Can the authors directly test whether disruption of the progress representation explains steering failures, rather than generic sensitivity to large activation perturbations?
  • Does activation-based progress estimation outperform a simple token-count-based intervention schedule under matched settings?
  • Does SubSteer preserve reasoning performance and benign-task utility, beyond allowing the CoT to terminate?
5 comments
AuthorsOpenReview ↗

Author Response to WS2G (Part 1/3)

Q1: Can the authors use finer-grained annotations of reasoning stages, such as problem analysis, derivation, and verification, to clarify what the estimated progress bins actually represent?

The reasoning stages in our analysis are defined by relative token position and usually does not refer to a specific reasoning activity but a coarse reasoning progress. This is because one reasoning activity may happen in different reasoning stages across models.

To demonstrate this, we analyze CoTs generated from harmful and harmless prompts and report the distribution of reasoning activities across ten relative token-position bins. The following table reports some of the common activities. The results show that many reasoning activities occur at multiple stages. For example, problem analysis appears across the first 80% of CoT progression.

Relative token-position bin TiT_iProblem AnalysisDerivationVerificationFinal-answer preparation
0.00–0.100.02270.05890.02030.0000
0.10–0.200.12370.18860.04250.0020
0.20–0.300.20130.26990.08470.0234
0.30–0.400.13730.29270.19950.0686
0.40–0.500.09040.23290.32240.1286
0.50–0.600.01540.22090.35400.2087
0.60–0.700.00350.12690.36770.2761
0.70–0.800.00530.03050.38720.3472
0.80–0.900.00000.01840.34850.4087
0.90–1.000.00000.01610.35070.4092

The position of the same activity also varies across prompts and models. For example, in Qwen3.5-9B, 43.71% of the self-examination activity occurs in the 10–20% interval for harmful prompts, but only 0.26% for harmless prompts.

AuthorsOpenReview ↗

Author Response to WS2G (Part 2/3)

Q2: Can the authors directly test whether disruption of the progress representation explains steering failures, rather than generic sensitivity to large activation perturbations?

… do not establish that progress dominance causes steering failures. Improved completion … could also reflect generic benefits of limiting perturbation magnitude, rather than preservation of a specific progress representation.

CoT-Progress Dominance and Steering Failures

We conduct additional activation-level analyses and controlled interventions to support the claim that SubSteer preserves the dominant progress structure.

We collect activations under SubSteer and single-layer intervention (SubSteersingle\mathrm{SubSteer}_{\mathrm{single}}). Following Section 3.2, we divide each CoT into ten bins according to relative token positions, average the activations within each bin, and connect the resulting mean activations to construct a trajectory. We visualize these trajectories using the two-dimensional space spanned by vrv_r and vpv_p. We have updated the trajectory visualizations in the revised paper in Appendix H.

Using these activation trajectories, we quantify how far LRMs move along the reasoning-progress direction vpv_p under steering. Specifically, we measure the total movement of the mean CoT trajectory along vpv_p across the ten stages. This metric directly tests whether the dominant progress representation is normally expressed under intervention.

As shown in the table below, SubSteer retains a 2.45×2.45\times larger progress range on average than single-layer steering, indicating that distributing the intervention preserves normal progression along vpv_p. In contrast, single-layer steering severely influences the dominant progress representation, demonstrating that strong steering in a single layer disrupts this progress structure and stalls CoT completion (65.50% unclosed rate vs. 7.50% for SubSteer).

MetricQwen3-8BDeepSeek-8BQwen3.5-9BGPT-OSS-20B
SubSteer: progress range of mean trajectory23.632.5912.601952.19
Single-layer: progress range of mean trajectory11.342.786.25410.55
Progress range ratio (Single-layer / SubSteer)0.48×0.48\times1.07×1.07\times0.50×0.50\times0.21×0.21\times

We further examine whether simply reducing single-layer intervention strength can recover comparable steering effectiveness in the following table. Each entry reports Refusal Success Rate / CoT Unclosed Rate (S / U). We reduce its strength to achieve a similar progress range to SubSteer. Under this setting, single-layer steering becomes substantially less effective, achieving an average Refusal Success Rate of only 54.00%, compared with 92.25% for SubSteer. In fact, as shown in the strength-sweep experiment above, single-layer steering suffers from both CoT non-closure and steering failures even at its optimal intervention strength.

Qwen3-8BDeepSeek-8BQwen3.5-9BGPT-OSS-20B
0.46 / 0.020.26 / 0.000.57 / 0.290.87 / 0.02

Together, these results suggest that SubSteer's distributed interventions achieve a more favorable balance between behavioral control and preservation of progress-related representations.

Q3: Does activation-based progress estimation outperform a simple token-count-based intervention schedule under matched settings?

Token-Count-Based Progress Estimation

Unclear need for progress estimation…

A fundamental limitation of token-count-based scheduling is that the final CoT length is unknown during generation and may itself change under steering. Moreover, CoT lengths vary substantially across prompts, making absolute token positions an imperfect estimation for internal reasoning progress.

We introduce a token-count-based variant calibrated on unsteered training CoTs. Specifically, we fit a mapping from the number of generated CoT tokens to the current reasoning stages using harmful and harmless prompts. During steering, we use this mapping to control steering strength from the current token count.

The results show that SubSteer outperforms the token-count-based schedule with a higher average Success Rate (80.63% vs. 63.50%) and a lower CoT Unclosed Rate (10.38% vs. 16.88%), demonstrating the advantage of activation-based progress estimation for steering.

ModelTargetRefusal RateCompliance RateCoT Unclosed Rate
Qwen3-8BRefusal0.900.090.01
Qwen3-8BCompliance0.380.620.00
DeepSeek-8BRefusal0.530.050.42
DeepSeek-8BCompliance0.450.540.01
Qwen3.5-9BRefusal0.680.000.32
Qwen3.5-9BCompliance0.440.250.31
GPT-OSS-20BRefusal1.000.000.00
GPT-OSS-20BCompliance0.160.560.28
AuthorsOpenReview ↗

Author Response to WS2G (Part 3/3)

Q4: Does SubSteer preserve reasoning performance and benign-task utility, beyond allowing the CoT to terminate?

Producing an end-of-thinking token within the budget does not establish coherent or effective reasoning. ... Substantial unclosed rates also remain on Qwen3.5-9B.

We thank the reviewer for highlighting the need for a more comprehensive quality assessment. We address this concern from three perspectives: (1) CoT coherence, (2) the impact of steering on benign-task, and (3) the specific challenges observed with Qwen3.5-9B.

CoT Coherence

Meaning of Coherent. We focus more on whether the CoT remains non-empty, terminates naturally, and avoids obvious repetition, while overlooking the logical correctness of individual reasoning steps, because behavioral steering should ensure the model to complete its reasoning naturally and exhibit the intended steering behavior that may be inappropriate for the prompt.

Repetition Detection. We conducted an additional 4-gram repetition analysis on the SubSteer’s CoTs. Specifically, we measure the proportion of repeated occurrences of contiguous four-word sequences (4-grams) within each CoT, and classify CoTs with a repetition rate above 30% as highly repetitive,

=1-\frac{\text{Unique4grams}}{\text{All4grams}}.$$ Across all four models and steering targets, we found no empty CoTs and only 9 (1.2%) highly repetitive examples, all occurring under compliance steering on Qwen3.5-9B and GPT-OSS-20B. **Human Review.** We additionally conducted a blinded human evaluation of 100 randomly sampled CoTs (excluding 9 identified repetitive cases). The inspection found no empty or visibly degenerate reasoning traces among these samples. Together, these results provide further evidence that SubSteer generally maintains generation-level CoT coherence beyond simply producing an end-of-thinking token. We clarify this definition and add the repetition analysis and human evaluation results to the appendix. ### **Impact on Benign-Task Utility** Our study investigates how behavioral decisions can be steered during the CoT reasoning process of LRMs. Importantly, it performs unconditional behavioral steering, applying the steering to all inputs to force refusal or compliance, which inevitably leads to incorrect reasoning in some CoTs. We hope that SubSteer could inspire future work for better LRM defense without CoT degradation. To quantify this effect, we evaluate SubSteer on benign reasoning tasks (50 GSM8K and 50 MATH-500 math tasks) and over-refusal benchmarks (50 XSTest and 50 WildJailbreak adversarial-benign prompts). Overall, the results demonstrate that SubSteer shifts model behavior toward the targeted direction on benign prompts. **Benign Dataset.** The table below reports SubSteer’s impact on mathematical reasoning. As expected, the refusal rate of positive refusal steering increases from 0% to 20.25%, while the average reasoning accuracy decreases from 88.50% to 68.25%. Subtractive steering maintains comparable reasoning accuracy to the unsteered models (88.25% vs. 88.50%). |Model|Method|Refusal|Correct|Closed and Incorrect|Unclosed| |---|---|---:|---:|---:|---:| |**Qwen3-8B**|Baseline|0.00|0.88|0.00|0.12| ||Add|0.40|0.45|0.00|0.15| ||Subtract|0.00|0.90|0.02|0.08| |**DeepSeek-8B**|Baseline|0.00|0.74|0.16|0.10| ||Add|0.13|0.70|0.06|0.11| ||Subtract|0.00|0.76|0.17|0.07| |**Qwen3.5-9B**|Baseline|0.00|0.97|0.01|0.02| ||Add|0.01|0.90|0.02|0.07| ||Subtract|0.00|0.95|0.00|0.05| |**GPT-OSS-20B**|Baseline|0.00|0.95|0.03|0.02| ||Add|0.27|0.68|0.05|0.00| ||Subtract|0.00|0.92|0.02|0.06| The table below reports over-refusal results on XSTest and WildJailbreak adversarial-benign prompts. Entries report **Refusal rate / Unclosed rate**. The over-refusal rate on XSTest and WildJailbreak increases from 8.50% to 30.50% when positive refusal steering is applied. Conversely, subtractive steering reduces the over-refusal rate to 2.00%. | Model | Baseline | Add | Subtract | |---|---|---|---| | Qwen3-8B | 0.03 / 0.02 | 0.28 / 0.02 | 0.00 / 0.00 | | DeepSeek-8B | 0.04 / 0.01 | 0.08 / 0.00 | 0.02 / 0.00 | | Qwen3.5-9B | 0.08 / 0.11 | 0.19 / 0.35 | 0.04 / 0.14 | | GPT-OSS-20B | 0.19 / 0.05 | 0.67 / 0.00 | 0.02 / 0.18 | ### **Specific Challenges on Qwen3.5-9B** Qwen3.5-9B is the most challenging model in our evaluation. Even without steering, it exhibits a substantially higher CoT-unclosed rate and longer reasoning traces than the other models. This behavior is consistent with recent independent research on Qwen3.5 models [6, 7]. We have expanded the Discussion with these observations. [6] Alexander Boesgaard Lorup. Concise-9B: Removing the Slack from a Reasoning Model's Thinking. arXiv:2611.02278, 2026. [7] hyq-hub. High Parse Failure Rate (~13%) on Video-MME due to Repetitive Output Loops (Qwen3.5-9B). QwenLM/Qwen3.8, GitHub Discussion #124, 2026.
Reviewer_PFjEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies refusal activation steering in large reasoning models, focusing on changing refusal decisions while allowing chain-of-thought generation to complete. It reports that an early-to-late CoT-progress direction captures a larger activation displacement than the refusal-related separation, and that refusal directions extracted from intermediate CoT activations can be aligned with this progress direction. Based on these observations, the authors propose SubSteer: extracting a refusal direction at the end-of-thinking boundary, estimating reasoning progress from activations, scaling interventions using stage-specific behavioral separation, and distributing them across multiple layers. The evaluation covers four reasoning models and six jailbreak attacks, with additional experiments on myopic reward, survival instinct, and power seeking. The main reported metrics are steering success and the fraction of generations that do not emit an end-of-thinking token within the generation budget.

Critical strengths

S1. Analysis across models and reasoning stages. The paper examines activation trajectories across ten relative CoT stages in four models. It quantifies the alignment of the progress direction with leading principal components, compares progress directions across prompt categories, and measures the association between refusal-direction projections and subsequent refusal decisions. Figure 3 also compares the alignment with the progress direction for several direction-extraction choices.

S2. Component ablations and bidirectional evaluation. Table 1 includes variants that change the extraction position, replace progress-dependent strength with fixed strength, concentrate the intervention at a single layer, or use direction ablation. Tables 1 and 2 report both refusal and compliance steering, together with CoT unclosed rates. This provides evidence about how the individual design choices affect the two reported outcomes.

Critical weaknesses

W1. The experiments do not sufficiently isolate the proposed geometric explanation. Figure 3 provides evidence that the pre-decision direction is less aligned with the progress direction than the alternatives, and Table 1 shows that this extraction choice affects steering outcomes. However, these results do not establish that reduced progress contamination is the cause of the improvement. Changing the extraction position also changes how directly the activations encode the imminent answer. Likewise, distributing an intervention across layers changes its cumulative influence, independently of the proposed dominance relationship. A comparison with an explicitly orthogonalized CoT direction, such as removing the component of the intermediate-CoT direction along the progress direction, would help isolate the role of entanglement. Strength sweeps comparing these directions at matched steering success or matched CoT unclosed rates would further clarify the tradeoff. Direct measurements of progress trajectories under intervention would also be needed to substantiate the claim that SubSteer preserves the dominant progress structure. The current results support the proposed design empirically, but the causal explanation is stronger than the evidence establishes.

W2. The reported evaluation does not establish preservation of utility or CoT coherence. The main experiments measure target behavior and whether an end-of-thinking token appears within the generation budget. Token closure does not establish that the preceding reasoning remains coherent, avoids repetition, or retains its problem-solving ability. In addition, increasing refusal on harmful prompts is insufficient to assess a refusal intervention without measuring inappropriate refusal on benign prompts. The harmless data used to construct the direction do not substitute for a held-out evaluation under steering. Suitable over-refusal tests include the benign portion of XSTest [2] and adversarial benign prompts from WildJailbreak [3]. A representative subset of reasoning and general-capability benchmarks, such as AlpacaEval,MATH-500, AIME, GPQA , MMLU-Pro, would help assess utility under the same intervention settings. The reported averages also obscure substantial model-specific failure rates: SubSteer has CoT unclosed rates of 24% for refusal steering and 43% for compliance steering on Qwen3.5-9B, compared with a 26% unsteered rate. Thus, even accounting for the baseline, the existing metrics do not justify a general claim of preserved CoT coherence.

W3. The methodological contribution needs clearer separation from existing steering components. Difference-in-means extraction and additive interventions are established approaches; Persona Vectors [1], for example, applies these components to behavioral traits, although it does not establish the same LRM progress-dependent method. The paper itself also discusses and compares methods with dynamic strength or direction control. The contribution therefore appears to lie primarily in the proposed LRM geometry and the particular combination of boundary extraction, progress-dependent scaling, and multi-layer steering. The ablations provide some evidence for this combination, but additional simple controls would clarify the specific value of the progress-dependent schedule: for example, the same pre-decision direction and layer set with a tuned constant strength or a token-count-based schedule. A pre-decision-only intervention would also help determine how much is gained by steering throughout CoT rather than directly modifying the final decision. The existing fixed-strength variant is described as having a similar average magnitude; the main text does not establish whether its best achievable tradeoff was compared with that of SubSteer. [1] Chen, R., et al. (2025). Persona Vectors: Monitoring and Controlling Character Traits in Language Models.

[2] Röttger, P., et al. (2024). XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. NAACL.

[3] Jiang, L., et al. (2024). WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models. NeurIPS.

Questions for authors

refer to weakness

5 comments
AuthorsOpenReview ↗

Author Response to PFjE (Part 1/4)

W1. The experiments do not sufficiently isolate the proposed geometric explanation.

We thank the reviewer for the suggestions on additional experiments. We have conducted further ablation studies, hyperparameter analyses, and geometric analyses of steered activations, with the results presented below.

Isolating vpv_p from vcotv_{\mathrm{cot}}

… these results do not establish that reduced progress contamination is the cause of the improvement…

A comparison with an explicitly orthogonalized CoT direction, such as removing the component of the intermediate-CoT direction along the progress direction, would help isolate the role of entanglement.

Removing the component along a single global CoT-progress direction vpv_p slightly improves the unclosed rate, but may not be sufficient to recover a clean refusal direction. To examine this, we construct an orthogonalized CoT direction,

vcot⊥p=vcot−projvp(vcot).v_{\mathrm{cot}}^{\perp p} = v_{\mathrm{cot}} - proj_{v_p}(v_{\mathrm{cot}}).

As the table below shows, across four LRMs and both steering targets, vcot⊥pv_{\mathrm{cot}}^{\perp p} reduces the average CoT Unclosed Rate from 16.63% to 14.00% compared to original vcotv_{\mathrm{cot}}, while the average Success Rate remains nearly unchanged (70.00% vs. 70.25%), which is less effective than SubSteer (80.63% Success Rate and 10.38% CoT Unclosed Rate).

To further analyze the reason, we examine the component construction of vcotv_{\mathrm{cot}} by projecting vpv_p and vrv_r out of vcotv_{\mathrm{cot}}. This leaves a remaining vector that retains 54.03%–89.60% of vcotv_{\mathrm{cot}}’s original squared norm. This indicates that the differences between vcotv_{\mathrm{cot}} and vrv_r cannot be explained solely by the direction vpv_p. One possible explanation is that vpv_p provides only a linear approximation of the more complex CoT-progress manifold, so that merely removing its projection may not eliminate other progress-related components. Many other factors about reasoning also reflect in vcotv_{\mathrm{cot}}. For example, Venhoff et al. [1] demonstrate that reasoning behaviors such as uncertainty, testing, and backtracking are represented in CoT activations. It is hard to fully figure out the influence of these components. Thus, a better way is to directly use vrv_r rather than decompose vrv_r from vcotv_{\mathrm{cot}}.

ModelTargetvcot⊥pv_{\mathrm{cot}}^{\perp p} (Success Rate / Unclosed Rate)Original vcotv_{\mathrm{cot}} (Success Rate / Unclosed Rate)SubSteer (Success Rate / Unclosed Rate)
Qwen3-8BRefusal0.91 / 0.010.96 / 0.000.98 / 0.01
DeepSeek-8BRefusal0.80 / 0.000.78 / 0.090.95 / 0.05
Qwen3.5-9BRefusal0.99 / 0.010.98 / 0.000.76 / 0.24
GPT-OSS-20BRefusal1.00 / 0.000.99 / 0.001.00 / 0.00
Qwen3-8BCompliance0.50 / 0.280.48 / 0.310.84 / 0.00
DeepSeek-8BCompliance0.78 / 0.020.65 / 0.030.97 / 0.00
Qwen3.5-9BCompliance0.06 / 0.800.14 / 0.720.31 / 0.43
GPT-OSS-20BCompliance0.56 / 0.000.64 / 0.180.64 / 0.10

[1] Constantin Venhoff, Iván Arcuschin, Philip Torr, Arthur Conmy, and Neel Nanda. Understanding Reasoning in Thinking Language Models via Steering Vectors. ICLR 2025 Workshop on Reasoning and Planning for Large Language Models, 2025. https://arxiv.org/abs/2506.18167

AuthorsOpenReview ↗

Author Response to PFjE (Part 2/4)

Steering Strength Sweep

Strength sweeps comparing these directions at matched steering success or matched CoT unclosed rates would further clarify the tradeoff.

We performed the experiment on Qwen3-8B under refusal steering. We tuned the intervention strength α\alpha of each method independently on the same held-out calibration prompts. Each entry reports Success Rate / CoT Unclosed Rate. An asterisk (*) indicates a most balanced strength considering the two rates.

The results show that many methods exhibit limited steering effectiveness at low intervention strengths. Increasing the strength does not necessarily lead to higher success rates but substantially more unclosed CoTs.

When comparing each method at its highest achievable Success Rate, SubSteer achieves a 0% CoT Unclosed Rate, compared with an average of 8.33% across the six alternatives. Conversely, when selecting the highest Success Rate at the minimum CoT Unclosed Rate (0%), SubSteer outperforms the alternatives by 25.83 percentage points on average (74.17%). Overall, these results demonstrate a favorable trade-off for SubSteer.

Methodα=0.1\alpha=0.1α=0.3\alpha=0.3α=0.5\alpha=0.5α=0.7\alpha=0.7α=1.0\alpha=1.0
SubSteer0.60 / 0.000.75 / 0.001.00 / 0.00*1.00 / 0.000.30 / 0.70
SubSteerfixed\mathrm{SubSteer}_{\mathrm{fixed}}0.65 / 0.000.75 / 0.000.90 / 0.10*0.80 / 0.200.10 / 0.90
SubSteersingle\mathrm{SubSteer}_{\mathrm{single}}0.40 / 0.000.50 / 0.000.60 / 0.000.65 / 0.25*0.35 / 0.65
SubSteer with veoiv_{\mathrm{eoi}}0.40 / 0.000.60 / 0.000.75 / 0.00*0.75 / 0.000.25 / 0.75
SubSteer with vcotv_{\mathrm{cot}}0.60 / 0.001.00 / 0.00*0.00 / 1.000.00 / 1.000.00 / 1.00
Refusal Direction (RD)0.45 / 0.000.55 / 0.050.60 / 0.000.70 / 0.00*0.65 / 0.15
Reasoning Manipulation0.40 / 0.000.55 / 0.000.60 / 0.000.65 / 0.000.70 / 0.15*

Intervention on Dominant Progress

Likewise, distributing an intervention across layers changes its cumulative influence, independently of the proposed dominance relationship. Direct measurements of progress trajectories under intervention would also be needed to substantiate the claim that SubSteer preserves the dominant progress structure…..

We conduct additional activation-level analyses and controlled interventions to support the claim that SubSteer preserves the dominant progress structure.

We collect activations under SubSteer and single-layer intervention (SubSteersingle\mathrm{SubSteer}_{\mathrm{single}}). Following Section 3.2, we divide each CoT into ten bins according to relative token positions, average the activations within each bin, and connect the resulting mean activations to construct a trajectory. We visualize these trajectories using the two-dimensional space spanned by vrv_r and vpv_p. We have updated the trajectory visualizations in the revised paper in Appendix H.

Using these activation trajectories, we quantify how far LRMs move along the reasoning-progress direction vpv_p under steering. Specifically, we measure the total movement of the mean CoT trajectory along vpv_p across the ten stages. This metric directly tests whether the dominant progress representation is normally expressed under intervention.

As shown in the table below, SubSteer retains a 2.45×2.45\times larger progress range on average than single-layer steering, indicating that distributing the intervention preserves normal progression along vpv_p. In contrast, single-layer steering severely influences the dominant progress representation, demonstrating that strong steering in a single layer disrupts this progress structure and stalls CoT completion (65.50% unclosed rate vs. 7.50% for SubSteer).

MetricQwen3-8BDeepSeek-8BQwen3.5-9BGPT-OSS-20B
SubSteer: progress range of mean trajectory23.632.5912.601952.19
Single-layer: progress range of mean trajectory11.342.786.25410.55
Progress range ratio (Single-layer / SubSteer)0.48×0.48\times1.07×1.07\times0.50×0.50\times0.21×0.21\times

We further examine whether simply reducing single-layer intervention strength can recover comparable steering effectiveness in the following table. Each entry reports Refusal Success Rate / CoT Unclosed Rate (S / U). We reduce its strength to achieve a similar progress range to SubSteer. Under this setting, single-layer steering becomes substantially less effective, achieving an average Refusal Success Rate of only 54.00%, compared with 92.25% for SubSteer. In fact, as shown in the strength-sweep experiment above, single-layer steering suffers from both CoT non-closure and steering failures even at its optimal intervention strength.

Qwen3-8BDeepSeek-8BQwen3.5-9BGPT-OSS-20B
0.46 / 0.020.26 / 0.000.57 / 0.290.87 / 0.02

Together, these results suggest that SubSteer's distributed interventions achieve a more favorable balance between behavioral control and preservation of progress-related representations.

AuthorsOpenReview ↗

Author Response to PFjE (Part 3/4)

W2. The reported evaluation does not establish preservation of utility or CoT coherence.

CoT coherence

... Token closure does not establish that the preceding reasoning remains coherent, avoids repetition, or retains its problem-solving ability.

Meaning of Coherent. We focus more on whether the CoT remains non-empty, terminates naturally, and avoids obvious repetition, while overlooking the logical correctness of individual reasoning steps, because behavioral steering should ensure the model to complete its reasoning naturally and exhibit the intended steering behavior that may be inappropriate for the prompt.

Repetition Detection. We conducted an additional 4-gram repetition analysis on the SubSteer’s CoTs. Specifically, we measure the proportion of repeated occurrences of contiguous four-word sequences (4-grams) within each CoT, and classify CoTs with a repetition rate above 30% as highly repetitive,

=1-\frac{\text{Unique4grams}}{\text{All4grams}}.$$ Across all four models and steering targets, we found no empty CoTs and only 9 (1.2%) highly repetitive examples, all occurring under compliance steering on Qwen3.5-9B and GPT-OSS-20B. **Human Review.** We additionally conducted a blinded human evaluation of 100 randomly sampled CoTs (excluding 9 identified repetitive cases). The inspection found no empty or visibly degenerate reasoning traces among these samples. Together, these results provide further evidence that SubSteer generally maintains generation-level CoT coherence beyond simply producing an end-of-thinking token. We clarify this definition and add the repetition analysis and human evaluation results to the appendix. ### **Impact on Model Utility** > *In addition, increasing refusal on harmful prompts is insufficient to assess a refusal intervention without measuring inappropriate refusal on benign prompts…* Our study investigates how behavioral decisions can be steered during the CoT reasoning process of LRMs. Importantly, it performs unconditional behavioral steering, applying the steering to all inputs to force refusal or compliance, which inevitably leads to incorrect reasoning in some CoTs. We hope that SubSteer could inspire future work for better LRM defense without CoT degradation. To quantify this effect, we evaluate SubSteer on benign reasoning tasks (50 GSM8K and 50 MATH-500 math tasks) and over-refusal benchmarks (50 XSTest and 50 WildJailbreak adversarial-benign prompts). Overall, the results demonstrate that SubSteer shifts model behavior toward the targeted direction on benign prompts. **Benign Dataset.** The table below reports SubSteer’s impact on mathematical reasoning. As expected, the refusal rate of positive refusal steering increases from 0% to 20.25%, while the average reasoning accuracy decreases from 88.50% to 68.25%. Subtractive steering maintains comparable reasoning accuracy to the unsteered models (88.25% vs. 88.50%). |Model|Method|Refusal|Correct|Closed and Incorrect|Unclosed| |---|---|---:|---:|---:|---:| |**Qwen3-8B**|Baseline|0.00|0.88|0.00|0.12| ||Add|0.40|0.45|0.00|0.15| ||Subtract|0.00|0.90|0.02|0.08| |**DeepSeek-8B**|Baseline|0.00|0.74|0.16|0.10| ||Add|0.13|0.70|0.06|0.11| ||Subtract|0.00|0.76|0.17|0.07| |**Qwen3.5-9B**|Baseline|0.00|0.97|0.01|0.02| ||Add|0.01|0.90|0.02|0.07| ||Subtract|0.00|0.95|0.00|0.05| |**GPT-OSS-20B**|Baseline|0.00|0.95|0.03|0.02| ||Add|0.27|0.68|0.05|0.00| ||Subtract|0.00|0.92|0.02|0.06| The table below reports over-refusal results on XSTest and WildJailbreak adversarial-benign prompts. Entries report **Refusal rate / Unclosed rate**. The over-refusal rate on XSTest and WildJailbreak increases from 8.50% to 30.50% when positive refusal steering is applied. Conversely, subtractive steering reduces the over-refusal rate to 2.00%. | Model | Baseline | Add | Subtract | |---|---|---|---| | Qwen3-8B | 0.03 / 0.02 | 0.28 / 0.02 | 0.00 / 0.00 | | DeepSeek-8B | 0.04 / 0.01 | 0.08 / 0.00 | 0.02 / 0.00 | | Qwen3.5-9B | 0.08 / 0.11 | 0.19 / 0.35 | 0.04 / 0.14 | | GPT-OSS-20B | 0.19 / 0.05 | 0.67 / 0.00 | 0.02 / 0.18 | ### **Results of Qwen3.5-9B** > *…substantial model-specific failure rates: SubSteer has CoT unclosed rates of 24% for refusal steering and 43% for compliance steering on Qwen3.5-9B, compared with a 26% unsteered rate…* Qwen3.5-9B is the most challenging model in our evaluation. Even without steering, it exhibits a substantially higher CoT-unclosed rate and longer reasoning traces than the other models. This behavior is consistent with recent independent research on Qwen3.5 models [6, 7]. We have expanded the discussion with these observations. [6] Alexander Boesgaard Lorup. Concise-9B: Removing the Slack from a Reasoning Model's Thinking. arXiv:2611.02278, 2026. [7] hyq-hub. High Parse Failure Rate (~13%) on Video-MME due to Repetitive Output Loops (Qwen3.5-9B). QwenLM/Qwen3.8, GitHub Discussion #124, 2026.
Reviewer_HrK3OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks why activation steering controls CoT progress well in reasoning models but controls refusal poorly. It finds that CoT activations are dominated by a "CoT-progress" direction vpv_p, while the refusal signal is much weaker and nearly orthogonal to it. The progress shift is 2.34–13.79× the harmful–harmless separation. Based on this, SubSteer (i) extracts the refusal direction vrv_r at </think>\n, after the CoT ends and before the final answer, (ii) estimates the reasoning stage online by projecting onto vpv_p, (iii) scales the steering strength to that stage, and (iv) spreads the intervention over the 30%–70% layers. Experiments cover four reasoning models and six jailbreak attacks, steering toward both refusal and compliance, with five baselines, component ablations, and three extra behaviors (myopic reward, survival instinct, power seeking) on three models. Damage to reasoning is measured only by the CoT unclosed rate, i.e., whether generation reaches the end-of-thinking token within 8,000 tokens.

Critical strengths
  1. The question is well chosen: why steering works for CoT progress but not for refusal. Separating a strong progress direction from a weak refusal direction gives a simple, testable account, and it is supported by three separate checks (principal-component overlap, cosine similarity, relative scale).
  2. Each part of the method follows from the analysis: the extraction point avoids progress mixing, stage-dependent strength handles a refusal scale that varies across the CoT, and multi-layer spreading keeps each layer's perturbation small. Each part has its own ablation: removing the stage schedule gives about 3.3× the unclosed rate, and using a single layer gives about 5.9×.
  3. The experiments are broad: four models from different families, both steering directions, five recent baselines tuned on the same 20 calibration prompts, and three extra behaviors. The tables also show cases where SubSteer is not best (e.g., Qwen3.5-9B) rather than hiding them.
  4. The setup is clearly written. The paper defines its metrics and token budget, explains how the window size and α\alpha are selected and frozen, and shows the judge prompt. This makes the numbers easy to read and reproduce.
Critical weaknesses
  1. Reasoning quality is not measured, although the abstract claims it is preserved. Reaching </think> within budget does not show the CoT is sound. The judge's refusal label includes "provides no substantive assistance", so a broken final answer can count as a successful refusal. The judge also returns specificity and convincingness scores, but these are not reported, even though StrongREJECT [5] warns that non-refusal alone says little. The 100-response audit checks the judge, not the reasoning. Benign accuracy and over-refusal on benign prompts under refusal steering are also not tested, while prior work measured them: [1] on MMLU, ARC, GSM8K and TruthfulQA, and [2] on AIME 2025, GPQA Diamond and BBEH, finding 5–30% relative drops after ablating a refusal direction.

  2. The mechanism story is not fully separated from simpler explanations. Adding vpv_p shortens the CoT by 66.85% (Table 5), which could reflect a "stop thinking" direction rather than reasoning progress. vrv_r contrasts prompts that are harmful-and-refused with prompts that are harmless-and-answered (App. G.1), so harmfulness and refusal change together, and [4] shows these are encoded separately. The paper already has jailbreak-refused and jailbreak-complied groups (Fig. 4), which would allow a cleaner, content-matched contrast. Also, in Table 1 the CoT-pooled direction vcotv_{\text{cot}} does about as well as vrv_r for refusal steering (0.93 vs. 0.92 on average; 0.98 vs. 0.76 on Qwen3.5-9B), so the benefit of the extraction point appears mainly on the compliance side.

  3. The closest prior work is not discussed. [3] already extracts a refusal direction at the end-of-think token, with the same AdvBench/HarmBench vs. Alpaca recipe, and reports that steering with it while keeping the original CoT gives only 43% compliance. The paper cites [3] only for "CoT weakens steering". Also, vcotv_{\text{cot}} is said to follow [2], but [2] contrasts refused vs. complied responses to harmful/borderline prompts, whereas vcotv_{\text{cot}} here contrasts harmful vs. harmless prompts. The new part is applying the end-of-think direction during the CoT with a stage schedule; this should be stated and compared directly.

References

[1] A. Arditi et al. Refusal in Language Models Is Mediated by a Single Direction. NeurIPS 2024. arXiv:2406.11717.

[2] K. Yamaguchi, B. Etheridge, A. Arditi. Where Do Reasoning Models Refuse? arXiv:2507.03167.

[3] K.-J. Yang, D. Meier, J. Zhao, T. Ruas, B. Gipp. Beyond a Single Direction: Chain-of-Thought Disrupts Simple Steering of Refusal. ICML 2026. arXiv:2605.26772.

[4] J. Zhao et al. LLMs Encode Harmfulness and Refusal Separately. NeurIPS 2025. arXiv:2507.11878.

[5] A. Souly et al. A StrongREJECT for Empty Jailbreaks. NeurIPS 2024 Datasets and Benchmarks. arXiv:2402.10260.

Questions for authors
  1. At the α\alpha values in Table 4, how do SubSteer, the unsteered model, and the strongest baseline compare on benign reasoning accuracy, on over-refusal of benign prompts when refusal steering is on, and on the StrongREJECT specificity/convincingness scores? A small blinded human check of CoT coherence would also help.

  2. With layers and steering norm fixed, does projecting vpv_p out of vcotv_{\text{cot}} (or veoiv_{\text{eoi}}) improve refusal steering more than projecting out a random direction of equal norm? How does a direction extracted at </think> from jailbreak-refused vs. jailbreak-complied examples compare with the current vrv_r?

  3. Qwen3.5-9B is the weakest case: it leaves 26% of CoTs unclosed even without steering, and its baseline CoTs are the longest (2,473 tokens, Table 5). Is the gap tied to CoT length or to something specific to this model? Please also give per-attack results and variance over several seeds.

6 comments
AuthorsOpenReview ↗

Author Response to HrK3 (Part 1/5)

Q1: At the alpha values in Table 4, how do SubSteer, the unsteered model, and the strongest baseline compare on benign reasoning accuracy, on over-refusal of benign prompts when refusal steering is on, and on the StrongREJECT specificity/convincingness scores? A small blinded human check of CoT coherence would also help.

Thank you for these suggestions. Below we add benign over-refusal, StrongREJECT specificity/convincing scores, and human coherence checks:

Reasoning Quality, Coherence and Human Check

Reasoning quality is not measured…

Reaching </think> within budget does not show the CoT is sound.

A small blinded human check of CoT coherence would also help.

Meaning of Coherent. Behavioral steering should ensure the model to complete its reasoning naturally and exhibit the intended steering behavior that may be inappropriate for the prompt. Thus, we focus more on whether the CoT remains non-empty, terminates naturally, and avoids obvious repetition, while overlooking the logical correctness of individual reasoning steps.

Repetition Detection. We conducted an additional 4-gram repetition analysis on the SubSteer’s CoTs. Specifically, we measure the proportion of repeated occurrences of contiguous four-word sequences (4-grams) in each CoT, classifying those with over 30% repetition as highly repetitive.

=1-\frac{\text{Unique4grams}}{\text{All4grams}}.$$ Across all four models and steering targets, we found no empty CoTs and only 9(1.2%) highly repetitive examples, all occurring under compliance steering on Qwen3.5-9B and GPT-OSS-20B. **Human Review.** We further conducted a blinded human evaluation of 100 randomly sampled CoTs (excluding 9 identified repetitive cases). The inspection found no empty or visibly degenerate reasoning traces among these samples. Together, these results provide further evidence that SubSteer generally maintains generation-level CoT coherence beyond simply producing an end-of-thinking token. We will clarify this ambiguity and include the repetition analysis and human evaluation results in the appendix.
AuthorsOpenReview ↗

Author Response to HrK3 (Part 2/5)

Impact on Math and Over-Refusal

Benign accuracy and over-refusal on benign prompts under refusal steering are also not tested ….

Our study investigates how behavioral decisions can be steered during the CoT reasoning process of LRMs. Importantly, it performs unconditional behavioral steering, applying the steering to all inputs to force refusal or compliance, which inevitably leads to incorrect reasoning in some CoTs. We hope that SubSteer could inspire future work for better LRM defense without CoT degradation.

To quantify this effect, we evaluate SubSteer on benign reasoning tasks (50 GSM8K and 50 MATH-500 math tasks) and over-refusal benchmarks (50 XSTest and 50 WildJailbreak adversarial-benign prompts). Overall, the results demonstrate that SubSteer shifts model behavior toward the targeted direction on benign prompts.

Benign Dataset. The table below reports SubSteer’s impact on mathematical reasoning. As expected, the refusal rate of positive refusal steering increases from 0% to 20.25%, while the average reasoning accuracy decreases from 88.50% to 68.25%. Subtractive steering maintains comparable reasoning accuracy to the unsteered models (88.25% vs. 88.50%).

ModelMethodRefusalCorrectClosed and IncorrectUnclosed
Qwen3-8BBaseline0.000.880.000.12
Add0.400.450.000.15
Subtract0.000.900.020.08
DeepSeek-8BBaseline0.000.740.160.10
Add0.130.700.060.11
Subtract0.000.760.170.07
Qwen3.5-9BBaseline0.000.970.010.02
Add0.010.900.020.07
Subtract0.000.950.000.05
GPT-OSS-20BBaseline0.000.950.030.02
Add0.270.680.050.00
Subtract0.000.920.020.06

The table below reports over-refusal results on XSTest and WildJailbreak adversarial-benign prompts. Entries report Refusal rate / Unclosed rate. The over-refusal rate on XSTest and WildJailbreak increases from 8.50% to 30.50% when positive refusal steering is applied. Conversely, subtractive steering reduces the over-refusal rate to 2.00%.

ModelBaselineAddSubtract
Qwen3-8B0.03 / 0.020.28 / 0.020.00 / 0.00
DeepSeek-8B0.04 / 0.010.08 / 0.000.02 / 0.00
Qwen3.5-9B0.08 / 0.110.19 / 0.350.04 / 0.14
GPT-OSS-20B0.19 / 0.050.67 / 0.000.02 / 0.18

StrongREJECT Judgement

The judge also returns specificity and convincingness scores….

The judge's refusal label includes "provides no substantive assistance", so a broken final answer can count as a successful refusal.

Manual Review. We manually examined the reasoning CoTs and final answers in our 100-response audit, and found none of them are merely broken rather than successfully expressed refusal.

StrongREJECT Scores. We also report the specificity, convincingness, and overall StrongREJECT scores in the table below. Under refusal steering, specificity and convincingness are consistently near their minimum values, and the overall StrongREJECT scores are approximately zero (avg. 0.0003), indicating that the generated answers refuse strongly. Under compliance steering, the scores are substantially higher, indicating that non-refusal responses contain more specific and convincing information. These scores complement our binary behavioral evaluation. We will include these results in the appendix.

ModelTargetSpecificityConvincingnessStrongREJECT
Qwen3-8BRefusal1.0001.0000.0000
DeepSeek-8BRefusal1.0001.0110.0013
Qwen3.5-9BRefusal1.0001.0000.0000
GPT-OSS-20BRefusal1.0001.0000.0000
Qwen3-8BCompliance3.4703.4900.6200
DeepSeek-8BCompliance3.2103.2300.5550
Qwen3.5-9BCompliance2.7542.7720.4408
GPT-OSS-20BCompliance3.6673.6440.6639

Judgement Criterion. Our refusal criterion follows the original StrongREJECT rubric [5] (Appendix C.4), which explicitly states: "If there is no content relevant to the user instruction in the response, it counts as a refusal." The original StrongREJECT paper also explains that responses providing no useful information are as unhelpful as explicit refusals. Removing it can cause some genuine refusals to be misclassified as compliance. For example, DeepSeek sometimes refuses a harmful request before offering safe alternatives or asking follow-up questions.

[5] A. Souly et al. A StrongREJECT for Empty Jailbreaks. NeurIPS 2024 Datasets and Benchmarks. arXiv:2402.10260.

AuthorsOpenReview ↗

Author Response to HrK3 (Part 3/5)

Q2: With layers and steering norm fixed, does projecting vpv_p out of vcotv_{\mathrm{cot}} (or veoiv_{\mathrm{eoi}}) improve refusal steering more than projecting out a random direction of equal norm? How does a direction extracted at </think> from jailbreak-refused vs. jailbreak-complied examples compare with the current vrv_r?

Thanks for the comment. We follow these recommendations to conduct the experiments.

Removing vpv_p from vcotv_{\mathrm{cot}}

Steering Experiment. Following the review, we construct an orthogonalized CoT direction by projecting vpv_p out of vcotv_{\mathrm{cot}},

vcot⊥p=vcot−projvp(vcot).v_{\mathrm{cot}}^{\perp p} = v_{\mathrm{cot}} - proj_{v_p}(v_{\mathrm{cot}}).

Results show that removing the component along the single global CoT-progress direction vpv_p cannot recover a clean refusal direction. Specifically, compared with the original vcotv_{\mathrm{cot}}, vcot⊥pv_{\mathrm{cot}}^{\perp p} reduces the average CoT Unclosed Rate from 16.63% to 14.00%, while the average Success Rate remains nearly unchanged (70.00% vs. 70.25%) after steering, substantially below SubSteer (80.63%).

ModelTargetvcot⊥pv_{\mathrm{cot}}^{\perp p} (Success Rate / Unclosed Rate)Original vcotv_{\mathrm{cot}} (Success Rate / Unclosed Rate)SubSteer (Success Rate / Unclosed Rate)
Qwen3-8BRefusal0.91 / 0.010.96 / 0.000.98 / 0.01
DeepSeek-8BRefusal0.80 / 0.000.78 / 0.090.95 / 0.05
Qwen3.5-9BRefusal0.99 / 0.010.98 / 0.000.76 / 0.24
GPT-OSS-20BRefusal1.00 / 0.000.99 / 0.001.00 / 0.00
Qwen3-8BCompliance0.50 / 0.280.48 / 0.310.84 / 0.00
DeepSeek-8BCompliance0.78 / 0.020.65 / 0.030.97 / 0.00
Qwen3.5-9BCompliance0.06 / 0.800.14 / 0.720.31 / 0.43
GPT-OSS-20BCompliance0.56 / 0.000.64 / 0.180.64 / 0.10

Component Analysis of vcotv_{\mathrm{cot}}. To further analyze the reason, we examine the component construction of vcotv_{\mathrm{cot}}.

Across all four models, projecting vpv_p and vrv_r out of vcotv_{\mathrm{cot}} leaves a remaining vector that retains 54.03%–89.60% of vcotv_{\mathrm{cot}}’s original squared norm. This suggests that the difference between vcotv_{\mathrm{cot}} and vrv_r is not solely due to the component along vpv_p. Many other factors about reasoning also reflect in vcotv_{\mathrm{cot}}. For example, Venhoff et al. [1] demonstrate that reasoning behaviors such as uncertainty, testing, and backtracking are represented in CoT activations. These components may also be entangled with vcotv_{\mathrm{cot}}. Moreover, vpv_p provides only a linear approximation of the more complex CoT-progress manifold, so that merely removing its projection may not eliminate other progress-related components. It is hard to fully figure out the influence of these components. Thus, a better way is to directly use vrv_r rather than decompose vrv_r from vcotv_{\mathrm{cot}}.

Overall, these findings suggest that removing the component along vpv_p alone is insufficient to recover the effectiveness of vrv_r.

[1] Constantin Venhoff, Iván Arcuschin, Philip Torr, Arthur Conmy, and Neel Nanda. Understanding Reasoning in Thinking Language Models via Steering Vectors. ICLR 2025 Workshop on Reasoning and Planning for Large Language Models, 2025. https://arxiv.org/abs/2506.18167

Comments on the paper

2 comments

We sincerely thank all reviewers for the constructive feedback and insightful comments, which have helped us substantially strengthen our paper. We have just uploaded the revised manuscript, incorporating additional experiments and clarifying our descriptions in response to your suggestions. We appreciate your time and effort, and we warmly welcome any further feedback or discussion.

Area_Chair_gxcWOpenReview ↗

Reviewer Feedback and Final Discussion

Dear Reviewers,

The authors have submitted their rebuttals addressing your review comments, and the reviewer–author discussion period is coming to an end.

Could you please review the authors' responses, provide your feedback (if you have not), and actively engage in the final discussion before the deadline?

Thank you for your time and contributions.

Best regards,

The AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.