acceptodds
Under review as a conference paper at ICLR 2027

Where You Steer Shapes the Direction: Operator Conditioned LLM Behavior Control and J-lens Interpretation

Abstract

Activation steering controls LLM behavior by modifying internal representations. However, existing methods primarily focus on identifying behaviorally relevant directions, while overlooking how intervention operators, such as steering during prefill versus decode, shape both the construction and control effectiveness of these directions. This gap motivates us to reformulate activation steering as a local control problem jointly defined by an objective and a specific intervention operator, proposing Operator-Conditioned Gradient Steering (OCGS). OCGS constructs a differentiable behavioral objective from high- and low-scoring target-model responses and naturally aggregates its state gradients over operator-specified token positions. The resulting directions are first-order optimal, explicitly coupling direction extraction with deployment. Experiments reveal that OCGS achieves the highest behavioral-control gains when extraction and intervention operators are matched. Jacobian-lens analysis further shows that distinct steering directions can carry related downstream behavioral content, yet matched OCGS directions yield the strongest layerwise gains in behavior-related readouts. Together, our work shows that the intervention operator should enter the definition of the steering problem itself, with OCGS providing a practical realization of this principle.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLMs, Steering, Jacobian Lens, Behavioral Control, Intervention Operators, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_G6uCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper argues that adding a vector while the model processes a prompt is a different intervention from adding it while the model generates a response, so the direction should be designed for the intended setting. The paper then introduces Operator-Conditioned Gradient Steering (OCGS), which obtains directions for every layer and operator in one backward pass by aggregating the gradient of a likelihood contrast between the target model's own high- and low-scoring responses over the positions the chosen operator writes to. Across three models and six behaviours, matching direction extraction to deployment yields the strongest aggregate behavioural control.

Critical strengths
  1. Clear formulation with direct empirical support

The paper treats the intervention operator as part of steering-direction construction. Table 2 test extraction-deployment compatibility: matched directions achieve the highest reported mean behavioral improvement under every operator across all three models. Operator-matched BiPO shows the same pattern on Llama for two behaviors, suggesting that the principle extends beyond OCGS.

  1. Efficient extraction with careful implementation

One backward pass per minibatch supplies gradients for all candidate layers and operators. The appendix clearly specifies boundary handling, gradient signs, and normalisation. Reported vector-construction times are substantially lower than BiPO's, although they exclude rollout generation and layer selection.

Critical weaknesses
  1. Uncertainty in the main results is not quantified

Table 1-2 lacks reported confidence intervals or run-to-run variability, which matters because some differences are very small. Capability results are also averaged across behaviors, potentially hiding uneven degradation.

  1. Local guarantee is not sufficiently validated at deployment.

Optimality applies to a linearized, teacher-forced likelihood objective, whereas evaluation measures freely generated behavior at a single strength, α=0.08\alpha=0.08. Strength sweeps and approximation-error measurements are needed to establish robustness.

Questions for authors
  1. Can you report confidence intervals and run-to-run variability for Tables 1-2, plus capability sample sizes and per-behavior results?
  2. Does the matching advantage hold across steering strengths, and how well does the first-order prediction track actual gains?
4 comments
AuthorsOpenReview ↗

Response to Reviewer G6uC (1)

Weakness 1 and Question 1 [ Part 1/2 ]


Thank you for this suggestion. In response to your concerns, during the rebuttal, we further show the main results in Tables 1–2 with confidence interval analyses, clarified how the general capability evaluation results are presented, and incorporated the corresponding results into the revised manuscript resubmitted during the rebuttal. Specifically:

(1) Uncertainty in the main experimental results

First, based on the per-prompt scores from the original held-out evaluation prompts, we supplemented the main experimental results in Tables 1–2 with 95% bootstrap confidence intervals. The point estimates reported in the original manuscript remain unchanged and are still calculated using the original evaluation-set procedure. Specifically, within each behavioral dimension, we perform bootstrap resampling with replacement over the 200 evaluation prompts. The same resampling indices are used for different methods and the Base model to preserve the paired structure. We then aggregate the average gains across the six behavioral dimensions with equal weights, following the same procedure as in the main text, to obtain the final 95% confidence intervals.

① Main-text Table 1: Confidence intervals for the main behavioral-control results.

We supplement the average behavioral gains and their 95% CIs for OCGS, PERSONA, and C-K under the same deployment operators.

ModelMethodPrefillDecodeAll
Qwen3.8-27BOCGS3.567 [3.469, 3.661]3.883 [3.750, 4.013]4.304 [4.207, 4.397]
PERSONA2.338 [2.169, 2.496]3.117 [3.012, 3.220]3.488 [3.368, 3.610]
C-K2.904 [2.806, 2.998]3.371 [3.247, 3.488]3.596 [3.488, 3.705]
Llama-3.1-8B-InstructOCGS4.083 [3.979, 4.186]3.767 [3.641, 3.894]4.579 [4.469, 4.690]
PERSONA2.592 [2.452, 2.727]2.921 [2.802, 3.039]3.079 [2.913, 3.243]
C-K3.100 [2.994, 3.205]3.133 [3.007, 3.262]3.271 [3.146, 3.394]
Qwen3.6-35B-A3BOCGS3.200 [3.101, 3.298]3.438 [3.317, 3.557]3.967 [3.865, 4.066]
PERSONA2.917 [2.775, 3.058]2.850 [2.740, 2.963]3.213 [3.092, 3.333]
C-K2.804 [2.704, 2.906]2.825 [2.661, 2.983]3.158 [3.051, 3.266]

Each cell reports Avg. Behavior Δ\DeltaBase [95% CI], with results averaged equally across the six behavioral dimensions.

As shown in the results, OCGS achieves higher average behavioral gains than both PERSONA and C-K across all three models and all three deployment operators, with its corresponding confidence intervals clearly separated from those of the two baselines. Specifically, the average gain advantage of OCGS over PERSONA ranges from 0.283 to 1.500, while its advantage over C-K ranges from 0.396 to 1.308.

② Main-text Table 2: Confidence intervals for Extraction–Deployment Compatibility.

We also conduct the same statistical analysis on the different extraction–deployment configurations in main-text Table 2, supplementing the average behavioral gains and their corresponding 95% CIs for each configuration.

ModelDeploymentPrefill ExtractionDecode ExtractionAll Extraction
Qwen3.8-27BPrefill3.567 [3.469, 3.661]2.3 [2.132, 2.466]2.7 [2.564, 2.834]
Decode3.2 [3.065, 3.330]3.883 [3.750, 4.013]3.0 [2.856, 3.147]
All3.4 [3.261, 3.533]3.8 [3.663, 3.934]4.304 [4.207, 4.397]
Llama-3.1-8B-InstructPrefill4.083 [3.979, 4.186]3.1 [2.945, 3.256]3.5 [3.361, 3.639]
Decode2.7 [2.532, 2.868]3.767 [3.641, 3.894]3.5 [3.273, 3.727]
All4.3 [4.157, 4.447]4.4 [4.213, 4.578]4.579 [4.469, 4.690]
Qwen3.6-35B-A3BPrefill3.200 [3.101, 3.298]1.9 [1.751, 2.056]2.5 [2.349, 2.646]
Decode3.2 [3.031, 3.370]3.438 [3.317, 3.557]2.6 [2.433, 2.764]
All3.5 [3.341, 3.661]3.8 [3.613, 3.990]3.967 [3.865, 4.066]

Each cell reports Mean Δ\DeltaBase [95% CI]. Bold indicates configurations where the extraction operator matches the deployment operator.

The results show that, across all three models and all three fixed deployment operators, OCGS directions matched to the deployment operators achieve the highest average behavioral gains. These results show a consistent trend favoring operator matching across different models and deployment settings, further supporting the central claim of our work.

AuthorsOpenReview ↗

Response to Reviewer G6uC (2)

Weakness 1 and Question 1 [ Part 2/2 ]


(2) Averaging procedure and performance preservation in general capabilities

Regarding capability preservation, our experiments examine both overall performance and changes in model capabilities under different behavioral steering conditions. Since our experiments cover three models, six behavioral dimensions, three deployment operators, and four capability benchmarks (IFEval, GSM8K, CSQA, and MMLU), reporting every combination individually in the main text would result in excessively large tables and make it difficult for readers to grasp the overall findings.

Therefore, in main-text Table 1, we report the average capability changes across behavioral dimensions and capability benchmarks. This metric summarizes a large number of experimental configurations with relatively high information density, allowing readers to directly compare overall capability preservation across different models, methods, and deployment operators. Meanwhile, to preserve the information from individual capability benchmarks, we have separately reported the results for IFEval, GSM8K, CSQA, and MMLU in Table 6 of Appendix C.5.

Furthermore, to directly address your concern that averaging across behaviors might hide localized performance degradation, we supplement the results of capability changes broken down by individual behavioral dimensions. The following table provides an example using the OCGS-All configuration on Qwen3.8-27B.

Target BehaviorIFEval Δ\DeltaGSM8K Δ\DeltaCSQA Δ\DeltaMMLU Δ\DeltaAvg. Capability Δ\Delta
Extraversion−0.13+0.05+0.10+0.03+0.01
Neuroticism−0.35−0.08+0.05−0.05−0.11
Humor−0.08−0.06+0.04−0.02−0.03
Deception−0.55−0.18−0.11−0.12−0.24
Sadness−0.18+0.02+0.08+0.05−0.01
Sycophancy−0.46−0.10+0.04+0.01−0.13

As shown above, the per-behavior results further show that OCGS can achieve effective behavioral control while generally preserving the model's general capabilities.

Thank you again for the comment. We hope that the additional uncertainty analyses, bootstrap confidence intervals, and per-behavior capability results provide a more comprehensive picture of OCGS's behavioral-control advantages and general capability preservation, making the experimental conclusions of our paper clearer, more rigorous, and more reliable.


Weakness 2 and Question 2 [ Part 1/2 ]


Thank you for this comment. Specifically:

(1) First-order optimality and its approximation accuracy under finite interventions

First, we would like to clarify that the first-order optimality established in our paper is a local optimality result obtained by linearizing the operator-conditioned teacher-forced likelihood objective at the zero-intervention point (v=0v=0) (Eqs. 13–14). Specifically, under a fixed L2L_2 perturbation budget, the normalized gradient gives the steering direction that maximizes the gain of the linearized objective.

To examine the accuracy of this local approximation under finite intervention strengths, we further compare the first-order linear predicted gain (ΔJlinear\Delta J_{\mathrm{linear}}) with the actual gain in the teacher-forced likelihood contrast, recomputed after applying the steering vector (ΔJactual\Delta J_{\mathrm{actual}}). We define the relative Taylor approximation error as:

\frac{ |\Delta J_{\mathrm{actual}}(\alpha) -\Delta J_{\mathrm{linear}}(\alpha)| }{ |\Delta J_{\mathrm{linear}}(\alpha)| }\times 100\%.$$ We evaluate this approximation error under different relative intervention strengths for all three deployment operators (Prefill, Decode, and All). The results are reported in the following table. | Relative Strength | Prefill | Decode | All | |:---|---:|---:|---:| | 4% | 4.2% | 5.1% | 6.0% | | 6% | 7.3% | 8.2% | 9.4% | | 8% | 11.2% | 13.0% | 14.7% | | 10% | 17.5% | 19.4% | 21.6% | | 12% | 25.1% | 28.0% | 30.8% | | 16% | 41.0% | 45.5% | 49.3% | *Each cell reports the average relative Taylor approximation error across the three behavioral dimensions.* The results show that the approximation error between the actual objective changes and the first-order linear predictions gradually increases with intervention strength, consistent with the theoretical expectation that higher-order terms affect the local linear approximation. At the default strength of 8% used in the paper, the approximation errors for the three operators are 11.2%, 13.0%, and 14.7%, respectively, all remaining below 15%. This indicates that the first-order approximation can still reasonably characterize the changes in the actual likelihood objective within this operating range.
AuthorsOpenReview ↗

Response to Reviewer G6uC (3)

Weakness 2 and Question 2 [Part 2/2]


(2) Practical deployment robustness under different intervention strengths

To further verify the effectiveness of the first-order optimal directions in free generation, we conducted an additional Steering Strength Sweep experiment. On Qwen3.8-27B, we selected three behavioral dimensions, Extraversion, Humor, and Sadness, and tested all three deployment operators, Prefill, Decode, and All. We expanded the relative intervention strengths from the original 8% to 4%, 6%, 8%, 10%, 12%, and 16%.

In these experiments, we fixed the original OCGS steering directions, intervention layers, and evaluation prompts, changing only the relative intervention strength (α\alpha). We used the same free-generation settings as in the main paper and employed the Qwen3.8-Max judge to evaluate actual behavioral performance and response quality.

Table R5: Steering Strength Sweep under Different Intervention Operators. Each cell reports Avg. Behavior Δ\DeltaBase [Quality Δ\DeltaBase], representing the average behavioral gain and average response-quality change across the three behavioral dimensions, respectively.

Relative StrengthPrefillDecodeAll
4%2.36 [−0.03]2.20 [+0.02]2.62 [+0.03]
6%3.21 [+0.05]3.14 [−0.07]3.63 [−0.04]
8%3.63 [+0.12]3.71 [−0.05]4.25 [−0.11]
10%3.77 [+0.03]3.84 [+0.01]4.39 [−0.21]
12%3.82 [−0.12]3.88 [−0.18]4.43 [−0.38]
16%3.68 [−0.38]3.74 [−0.46]4.27 [−0.70]

The experimental results show that OCGS achieves effective behavioral control at all tested intervention strengths. In particular, behavioral gains remain stable around 8%. Further increasing the intervention strength can bring additional behavioral gains, but also leads to more pronounced degradation in response quality. For example, at 16%, the average response quality decreases by 0.38, 0.46, and 0.70 under the three deployment operators, respectively. In comparison, 8% achieves relatively strong behavioral control while better preserving response quality.

Combining the first-order approximation error analysis and the free-generation experiments, we consider 8% to be a reasonable default deployment strength: it lies within an operating range where the first-order approximation retains good predictive accuracy, and it achieves effective behavioral control in actual generation while better preserving response quality.

Thank you again for your suggestion. We hope that the additional approximation error analysis, intervention strength sweep experiments, and further clarification of the theoretical scope of our method provide a more comprehensive demonstration of the effectiveness and robustness of OCGS in practical deployment.


We sincerely hope that the additional experiments, analyses, and methodological clarifications provided in this rebuttal address your main concerns. If you still have any questions about our work or require further clarification, we would be more than happy to continue the discussion. We also hope that these additional results and clarifications can further strengthen your confidence in our work and lead to a more favorable assessment of the paper. Thank you again for your careful review, recognition, and helpful feedback!

Thanks for the thorough rebuttal. The new CIs for Table 1 and the strength sweep address most of my concerns, and I'm keeping my positive score. One thing remains open: in Table 2 CIs, the matched direction overlaps with the best mismatched one in several rows (e.g Llama and MoE under Decode/All), and the sweep only covers matched directions. A paired comparison of matched vs mismatched would give stronger evidence for the robustness of operator matching. I'll keep rating 3.

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

This paper proposes Operator-Conditioned Gradient Steering (OCGS), a method for controlling LLM behavior that explicitly considers where and when steering vectors are applied. Unlike previous methods that extract steering directions independently of their deployment, OCGS derives operator-specific directions by aggregating gradients of a behavioral contrast objective over the selected intervention positions. The authors consider three operators: Prefill, Decode, and All. Experiments across three LLMs and six behavioral dimensions show that OCGS outperforms several steering baselines while largely preserving general capabilities and response quality. The authors further use Jacobian-lens analysis to show that different steering directions can capture similar behavioral information but produce different downstream effects, supporting the importance of matching steering-vector extraction with deployment.

Critical strengths

The paper presents an interesting idea by explicitly considering the intervention location during steering-vector extraction, rather than treating it as a separate deployment choice. I particularly appreciate the mathematical derivation, which clearly shows how the gradients at different intervention positions contribute to the final steering vector (though it is not clear the actual effect of each index on the final steering vector/behaviour alignment effect).

Emperically, OCGS consistently outperforms several baselines across three models and six behavioral dimensions, while causing relatively little degradation in general capabilities. The J-lens analysis also provides useful insights into why different steering directions can capture similar behavioral information but differ in their actual steering effectiveness.

Critical weaknesses

There are three main weaknesses: (1) limited exploration of intervention operators, (2) limitations of the J-lens analysis, and (3) unclear quality of the constructed high–low response pairs.

The paper only considers Prefill, Decode, and All, although the proposed formulation seems applicable to more fine-grained interventions, such as selected prompt tokens or the first few decoding steps. I wonder why the authors only consider these three operators, and whether more selective interventions could achieve better control with fewer perturbations. It is also not clear how the gradient at each index contribute to the final steering vector/behaviour alignment effect, e.g. are there any index that is more important than the others.

I'm also not fully convinced by the J-lens analysis in Figure 1, which relies on a fixed vocabulary-based readout to measure behavioral changes. Behaviors like Extraversion depend on sentence-level context and may not be well captured by individual words. It would be more convincing to show whether these readout changes correlate with actual behavioral scores from an LLM judge.

Another concern is the quality of the high–low response pairs. The method selects the highest- and lowest-scoring responses from eight candidates, but the paper does not report the distribution of their scores or score differences. What happens when all responses have very similar scores, or when no valid pair exists? I think this deserves more discussion, as the quality of these pairs directly affects the extracted gradients.

Questions for authors

I am curious whether the proposed method would generalize to other behaviors, such as jailbreak-related behavior or refusal. The current experiments focus on six behavioral dimensions, so it would be interesting to see whether the benefit of matching extraction and deployment still holds in these settings.

I also noticed a few differences between the paper and the released code that make reproduction difficult. The scoring rubrics in judge_trait.py and the template in stance_judge.py differ from Table 8 and Appendix C.7. The validation-set size and construction details are unclear, and I could not identify the exact validation membership or layer-search implementation. Also, the paper specifies 8% steering strength, while the script defaults to 16%. Could the authors clarify which scoring templates, validation prompts, and settings were actually used for the reported experiments, and the validation probing result?

9 comments
AuthorsOpenReview ↗

Response to Reviewer LpCG (1)

Weakness 1


Thank you for your suggestions. We use three general intervention operators (Prefill, Decode, and All) in the main text to demonstrate OCGS across different stage-level intervention settings, not because the method is limited in positional granularity. In fact, Eq. 11 naturally extends OCGS to arbitrary subsets of token-time positions by aggregating state gradients over the specified positions. Moreover, a single backward pass provides gradients at all positions, which can then be aggregated using different positional masks without additional backward passes.

Following your suggestion, we conducted supplementary experiments on Qwen3.8-27B to examine both sparse intervention effectiveness and position-wise gradient contributions. These experiments are included in Appendix C.12 of the revised manuscript. Specifically:

(1) Position selection substantially affects sparse intervention effectiveness.

We first compare sparse Prefill interventions at the beginning and end of the prompt. Under the same write budget, we independently extract and deploy steering directions using the first or last 10 effective prompt-token positions.

BehaviorFirst 10 PositionsLast 10 PositionsGain of Last Window
Extraversion+0.320+1.270+0.950
Humor+0.215+2.420+2.205
Sadness+0.040+1.430+1.390

Table 1: Behavioral gains of OCGS using the first versus the last 10 prompt positions under the same write budget (Δ\DeltaBase).

The last window consistently outperforms the first across all three behaviors. This advantage also persists with window sizes of 20 and 30 positions. Notably, the last 30 positions require only 24%–42% of the writes used by full Prefill while retaining 63%–82% of its behavioral gains. These results show that sparse intervention effectiveness depends not only on the number of writes but also on position selection, allowing substantial behavioral gains to be retained with fewer writes.

(2) Position-wise gradient analysis reveals non-uniform contributions to the final steering direction.

We project the aggregated gradients at each position onto the corresponding full-stage OCGS direction and calculate their signed projection contribution ratios. For Prefill and Decode, we use their respective full-stage directions as references.

BehaviorFirst 10 Prefill PositionsLast 10 Prefill PositionsFirst 10 Decode Steps
Extraversion0.86%43.11%17.44%
Humor0.60%29.84%57.01%
Sadness0.24%47.05%39.81%

Table 2: Projection contribution ratios of different position windows to the full OCGS direction.

Prefill contributions are strongly concentrated toward the end of the prompt: the last 10 positions account for 29.84%–47.05% of the total projection, compared with only 0.24%–0.86% for the first 10 positions. The first 10 Decode steps also make substantial gradient contributions.

We further visualize the position-wise gradient contributions (anonymous link: https://anonymous.4open.science/r/iclr2027_rebuttal_24276/position_profiles.pdf). The three rows correspond to Extraversion, Humor, and Sadness, respectively. The columns show Prefill positions aligned to the prompt beginning, Prefill positions aligned to the prompt end, and Decode positions aligned by generation step. Blue curves represent signed projection contribution ratios, while brown curves show the proportion of extraction trajectories covering each position. These curves show that the important gradient contributions in Prefill are not uniformly distributed across the entire prompt, but are instead more concentrated in prompt positions closer to the beginning of response generation.

Notably, these findings are consistent with Experiment (1): the last prompt window contributes more strongly to the full direction and produces better behavioral control when used for sparse steering under the same write budget.

Finally, Leave-one-out analysis further demonstrates the unequal influence of individual positions on the final direction. For Prefill, the most influential position across all three behaviors is the second-to-last token. Removing its gradient changes the direction by 20.77°. For Decode, the most influential position is the first generation step, producing angular changes of 12.78°.

In summary, these experiments demonstrate that OCGS not only supports sparse interventions at arbitrary positional granularities, but also reveals the different contributions of individual token positions to the final steering direction. In particular, positions near the end of the prompt and the beginning of response generation exhibit more concentrated gradient contributions in our experiments. The matched-budget intervention experiments further confirm that sparse interventions at these positions can achieve better behavioral control.

AuthorsOpenReview ↗

Response to Reviewer LpCG (2)

Weakness 2


Thank you for your comments on our J-lens analysis. We provide responses from two perspectives: First, we further clarify the definition of the J-lens readout in Section 5.2. Second, following your comment, we conduct an experiment to directly examine the correspondence between J-lens readouts and the actual behavioral performance of the model. The experimental results show that J-lens readouts across different layers exhibit highly consistent trends with the actual behavioral scores. We have also added these results to Appendix D.4 of the revised version. Specifically:

(1) On the validity and scope of applicability of the J-lens readout

First, we would like to clarify that Section 5.2 does not determine whether a model exhibits extraversion based on individual words. Instead, we use two fixed sets of lexical anchors, each containing 10 high-pole or low-pole words. At each response-prediction position, we map the hidden state into the vocabulary space using the context-averaged Jacobian and unembedding matrix, compute the average readout scores for both sets, and take their difference (Eq. 16). Combined with the definition of the J-lens, this index measures whether the internal state, after propagation through the subsequent network, is more aligned with the lexical content associated with one behavioral pole, rather than detecting the occurrence of specific words.

Regarding your concern that behaviors such as extraversion depend on sentence-level context, we would like to emphasize that the J-lens reads out contextualized hidden states formed during the model's actual generation process, rather than isolated lexical representations. During autoregressive generation, each response-prediction hidden state is influenced by the original prompt and previously generated response tokens. Its downstream lexical readout therefore reflects behavior-related information within the corresponding generation context. In addition, we average the readouts across all response positions (Eq. 17), integrating information from different generation contexts into a behavior-related representation of the entire generation trajectory. Thus, we do not use a fixed vocabulary to replace sentence-level behavioral understanding. Instead, we use it as an interpretable downstream readout space to analyze the association between context-dependent internal states and the target behavior, and to examine the effects of different steering methods on the model's internal behavior-related representations.

(2) Further validating the relationship between J-lens readouts and actual behavioral scores

Furthermore, we conducted additional experiments to validate the relationship between J-lens readouts and the actual behavioral performance scores of the model. We reused the 8 naturally generated responses from Qwen3.8-27B for each of 100 prompts, resulting in a total of 800 responses, and used Qwen3.8-Max to assign an extraversion score to each complete response. We then divided the responses into three score ranges: 1–3, 4–7, and 8–10. For each response, we used exactly the same Jacobian lens, vocabulary, and readout definition as in Section 5.2, computed the average J-lens readout across its response positions through teacher forcing, and compared the average readouts across the different score ranges.

The experimental results are shown in the figure at the following anonymous link: (https://anonymous.4open.science/r/iclr2027_rebuttal_24276/J_lens_readout.png). The left panel shows the average J-lens readout at each layer from layers 45–64, while the right panel further presents the average across these 20 layers. As can be easily observed, as the actual extraversion scores increase, the J-lens readouts exhibit a clear and consistent increasing trend across all layers. Specifically, the average readouts across layers 45–64 for the three score ranges are 0.75, 1.76, and 3.47, respectively.

This result shows that J-lens readouts not only reveal behavior-related characteristics of steering directions in the downstream vocabulary space, but their changes on naturally generated responses also exhibit consistent trends with actual behavioral scores. This provides additional empirical support for using J-lens as a tool for analyzing behavior-related representations.

We also believe that introducing the J-lens into the interpretability analysis of LLM behavior control is a promising research direction. It enables us to go beyond the geometric differences between steering directions in the original residual space and further understand the behavior-related information carried by these directions after subsequent network transformations. Our analyses, together with the additional rebuttal experiments, provide evidence of the consistency between J-lens readouts and actual behavioral performance. We hope these explorations offer new analytical perspectives and valuable insights for this research area.

AuthorsOpenReview ↗

Response to Reviewer LpCG (3)

Weakness 3


Thanks for the question. To address this concern, we further analyzed the score distributions of candidate responses and the behavioral differences between the selected response pairs, and clarified how to handle cases where valid behavioral contrasts are unavailable. We have included these additional analyses and clarifications in Appendix C.13 of the revised version. Specifically:

(1) Analysis of candidate response score distributions and the quality of high/low response pairs

First, we analyzed the score distributions of all candidate responses across the six behavioral dimensions. As reported in the main text, for each prompt, the target model samples 8 candidate responses, which are scored on a 1–10 scale by Gemma-4-12B-IT according to the corresponding behavioral rubric. The following table reports the proportion of responses falling into each score range for each behavioral dimension.

BehaviorScore 1–4Score 5–7Score 8–10
Extraversion24.5%13.6%61.9%
Neuroticism60.7%13.1%26.2%
Humor80.2%3.8%16.0%
Sadness71.6%9.9%18.5%
Sycophancy68.2%0.2%31.6%
Deception79.8%0.9%19.3%

As shown above, candidate responses exhibit different score distributions across behavioral dimensions, but overall, their scores are not concentrated within a narrow range. This indicates that candidate responses generated by the target model itself can exhibit different levels of the target behavior, providing a basis for constructing behavioral contrasts.

To further verify whether the response pairs selected by the original evaluator exhibit stable behavioral differences, we also used an additional model, Qwen3.8-Max, to independently score the selected high/low responses. The results show that the independent evaluator achieves a pair ordering accuracy of 94.0%, with an average score gap of 2.1 points between high- and low-scoring responses. In addition, repeated scoring of the same responses yields a score standard deviation of only 0.3, indicating that the constructed behavioral contrasts not only maintain high consistency across different evaluators but also exhibit good scoring stability.

(2) Adaptive Resampling and Alternative Contrastive Data Sources

For the small number of prompts where the initial 8 sampled responses exhibit a small score gap, we can continue sampling additional candidate responses from the same prompt, for example, increasing the number of candidates to 16 or 32, and then selecting the highest- and lowest-scoring responses from the expanded candidate pool. For common behavioral dimensions, increasing the number of candidates can improve the chance of obtaining response pairs with sufficient behavioral differences.

If a valid behavioral contrast still cannot be obtained after reaching a predefined sampling limit, the corresponding prompt can be skipped to avoid forcing samples without reliable behavioral supervision signals into gradient extraction.

Further, if the target model itself has difficulty exhibiting sufficient variation in the target behavior during natural generation(the most extreme case), our method is not limited to target-model rollouts. We can also use externally provided positive/negative responses to construct behavioral contrasts and extract steering directions through the same teacher-forced likelihood objective and operator-conditioned gradient extraction procedure.

In fact, we have specifically validated this in Appendix C.8 (Table 10). This experiment replaces the contrastive responses generated by the target model itself with externally provided positive/negative responses from PERSONA, while keeping the prompts, target model, optimization objective, deployment operator, intervention strength, and evaluation protocol consistent. The results show that even without using target-model rollouts, OCGS can still effectively extract steering directions and achieve behavioral control. However, across all evaluated settings involving three target models, two behavioral dimensions, and three deployment operators, using target-model rollouts consistently achieves higher behavioral gains.

This indicates that externally provided contrastive responses are a feasible alternative source of contrast data, while responses generated by the target model itself can provide behavioral contrasts that are closer to its actual generation trajectories, thereby facilitating the extraction of more effective control directions.

In summary, these design choices allow OCGS to flexibly handle behavioral contrast samples of different qualities and from different sources, while retaining the core mechanism of operator-conditioned gradient extraction.

Reviewer_1XE4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors apply ideas from control theory to design a gradient-based optimization framework for jointly learning an activation steering vector and at which token positions to apply it at test-time. They coin their method Operator-Conditioned Gradient Steering. In prior work, token position is generally selected rather arbitrarily, often with a mismatch between train- and test-time, or searched over via an expensive grid search. The authors find that, 1) the method outperforms other baselines for all 3 tested token positions, and 2) intuitively, steering works best when the "extraction" (i.e., training-time) and test-time locations are the same.

Critical strengths
  1. The activation steering literature is currently fairly sloppy and unprincipled in how both train-time extraction and test-time intervention positions are selected. This is an important research question.
  2. I appreciate adopting methodologies from control theory, as the analogies between LLMs and dynamical systems are currently fairly under-explored in the literature. This direction seems promising.
  3. I also appreciate the careful notation used throughout the paper.
  4. The experiments are comprehensive.
Critical weaknesses
  1. Lack of most relevant baselines (gradient-based optimization ones, listed below-- ReFT, RePS, NTP, etc.). It is unclear if the gains in Table 1 are due to the gradient-based optimization technique or the formulation that also factors in inference-time application token position (the paper's claim is that it is the latter). Only by comparing to a gradient-based method that does not include operator conditioning as a baseline can the main paper's claim be supported.
  2. The paper has too short of a related works section in the main paper, but thankfully there is a substantially expanded version in the appendix (which should be referenced in main). However, there are still a number of missing works which have inspected the role of intervention location on steering vector success OR used gradient-based behavioral optimization methods for steering vector construction. The authors could enumerate the choices of where to intervene made in prior work (which differs a lot-- either in motivating a specific prompt or response position, or in treating this as a hyperparameter to optimize; also see https://arxiv.org/abs/2406.17563). As for gradient-based SV optimization, see ReFT (https://arxiv.org/abs/2404.03592), RePS and NTP (https://arxiv.org/abs/2505.20809). For using models' own rollout generations as opposed to teacher forcing, see https://arxiv.org/abs/2602.16080, https://arxiv.org/abs/2507.21509. I understand all of these works may be subtly different from the proposed method, but I would appreciate additional clarification from the authors on this in the rebuttal.
  3. Presentation clarity recommendations: Section 3 (Method) presents both background and the proposed formulation, but it's slightly unclear which components are background/setup from related work vs. your contributions. While I appreciate the full mathematical formalization, I believe the novel components that differ from traditional gradient-based steering vector optimization begin around ~Eqn. 10, which is quite a ways into the section. Perhaps the authors could clarify this for me during the rebuttal. In general, Section 3 (and the paper overall) is very dense. A figure providing an overview of your methods, or perhaps a bulleted summary of your contributions at the end of the intro, would also help. I also found the results table very hard to read.
Questions for authors
  1. The intervention is applied at a very course level of granularity (all prompt tokens, all response tokens, or all tokens (both)). I believe one strength of this direct optimization framework would be the ability to very carefully select even sparser intervention locations-- is there a reason why this is not computationally feasible?
  2. See above question about Section 3 and the formalization contribution.
5 comments
AuthorsOpenReview ↗

Response to Reviewer 1XE4 (1)

Weakness 1


Thank you for your comment. Regarding whether the gains reported in Table 1 arise from gradient-based optimization itself or from accounting for inference-time intervention positions when extracting steering directions, we would first like to clarify that the original manuscript already includes three groups of experiments and analyses specifically examining this question. Furthermore, following your suggestion, we have conducted additional comparative experiments with other gradient-based optimization methods during the rebuttal. Specifically:

(1) Cross-operator deployment experiments with OCGS directly demonstrate the importance of operator matching.

In Section 4.3, Table 2 presents an Extraction–Deployment Compatibility experiment, where we keep the behavioral objective, contrast data, and gradient extraction method fixed, while varying the extraction operator. For each deployment operator (Prefill, Decode, and All), we evaluate OCGS directions extracted under all three operators.

The results show that the operator-matched direction consistently achieves the highest average behavioral gain across all deployment settings. This demonstrates that the effectiveness of a steering direction depends not only on the gradient extraction method, but also on whether the extraction operator matches the intended deployment operator, supporting the central design principle of OCGS.

(2) Additional BiPO experiments further demonstrate the importance of operator matching.

To examine whether Extraction–Deployment Matching generalizes beyond OCGS, we extend BiPO, an iterative steering-vector optimization method based on bidirectional preference optimization, to the three intervention operators (Prefill, Decode, and All) in Appendix C.10.

Table 13 shows that BiPO exhibits the same Extraction–Deployment Compatibility pattern: the operator-matched direction achieves the highest behavioral gain for each deployment operator. Furthermore, Table 12 shows that OCGS achieves higher behavioral gains than BiPO across all evaluated behaviors. Together, these results demonstrate that operator matching is beneficial beyond OCGS and that OCGS achieves stronger behavioral control than an alternative iterative optimization method.

(3) J-lens analysis provides complementary evidence at the representation level.

Section 5.2 of the main paper uses J-lens analysis to investigate the downstream propagation effects of different steering directions within the model. The results show that, under each of the three fixed deployment settings (Prefill, Decode, and All), the OCGS direction matched to the deployment operator produces the largest vocabulary-level Extraversion-related J-lens readout shift. This provides complementary evidence at the representation level for the behavioral-control results observed in Table 2 and supports our main claim.

(4) Additional comparisons with gradient-based optimization baselines in rebuttal.

Regarding your suggestion to include comparisons with other gradient-based optimization methods, we have conducted additional experiments during the rebuttal on Qwen3.8-27B, comparing OCGS with NTP-SV and RePS-SV across three behavioral dimensions: Extraversion, Humor, and Sadness.

For the deployment setting, we follow the original full-sequence intervention schemes of the two baseline methods, applying their steering vectors to all intervention positions. For OCGS, we use the All-conditioned steering direction to ensure consistency between extraction and deployment. All methods use the same setting, following a unified intervention-strength and layer-selection protocol. The following table reports behavioral score improvements over the unsteered model (ΔBase).

BehaviorNTP-SVRePS-SVOCGS (All-matched)
Extraversion+2.775+3.050+3.525
Humor+2.350+3.125+4.675
Sadness+3.100+3.575+4.550
Average+2.742+3.250+4.250

Table 1: Behavioral control performance of different gradient-based steering-vector methods under the All deployment operator on Qwen3.8-27B. All values report behavioral score improvements over the unsteered model (ΔBase), with higher values indicating stronger target-behavior control.

As shown in the table, OCGS achieves higher behavioral gains than both NTP-SV and RePS-SV across all three behavioral dimensions. These additional experiments further support the central claim of our work. We have included the corresponding baseline comparison in Appendix C.11 of the revised manuscript resubmitted during the rebuttal, and have briefly introduced and cited these results in the main text when discussing Table 2. Thank you again for your valuable suggestion!

AuthorsOpenReview ↗

Response to Reviewer 1XE4 (2)

Weakness 2


Thank you for these valuable suggestions. In the revised manuscript submitted during the rebuttal, we have expanded the Related Work section: we added a paragraph in Section 2 of the main text discussing how previous studies select intervention positions, and added more corresponding related work discussion in Appendix A. The newly added paragraph in Section 2 is shown below:

Intervention Positions. Steering methods intervene at different token positions. CAA adds a shared direction from the instruction boundary onward (Rimsky et al., 2024). ReFT treats the number of edited positions at each end of the prompt as a hyperparameter (Wu et al., 2024). Dynamic Activation Composition (DAC) compares first-step and sustained steering and adapts intervention strength during decoding (Scalena et al., 2024). However, positions for direction extraction and inference-time intervention are often chosen using fixed rules or empirical tuning. Prior work offers limited guidance on how to choose extraction positions for a given set of intervention positions.

Regarding the differences between OCGS and the works you mentioned, we would like to provide the following clarifications:

(1) Selection of intervention positions

First, regarding intervention positions, CAA applies the steering direction during inference to generation positions after the instruction boundary; ReFT treats the number of prompt tokens at the beginning and end that are subject to intervention as hyperparameters; and Dynamic Activation Composition (DAC) compares strategies such as intervention at the first generation step, sustained intervention, and dynamically adjusting intervention strength. These works highlight the importance of intervention positions and timing. In contrast, we ask: given a fixed inference-time intervention operator, can extracting a steering direction specifically for its designated token positions improve behavioral control?

To address this, Section 3 formulates activation steering as a local control problem jointly defined by a behavioral objective and an intervention operator. As shown in Table 2, OCGS consistently achieves the highest average behavioral gain when its extraction operator matches the deployment operator across all three settings. Thus, unlike prior approaches that primarily specify where to intervene, OCGS directly incorporates deployment positions into steering direction extraction.

(2) Gradient-based steering-vector optimization methods

Second, regarding gradient-based methods, ReFT iteratively learns low-rank transformations using a teacher-forced next-token prediction loss, while RePS compares NTP and preference-based objectives under different intervention parameterizations, including rank-1 steering vectors (NTP-SV and RePS-SV).

The main difference between OCGS and these methods lies in the optimization objective and its corresponding control variable. OCGS directly incorporates the deployment operator into the behavioral objective and gradient computation. Specifically, it uses the length-normalized log-likelihood difference between high- and low-behavior responses generated by the target model to optimize a shared steering vector over the operator-specified write positions. Under a fixed L2L_2 perturbation budget, its normalized gradient yields the optimal steering direction for the specified operator under the first-order linearized objective.

(3) Direction extraction using responses generated by the target model itself

We also note that both Persona Vectors and Generative Causal Mediation (GCM), which you mentioned, utilize responses generated by the model itself. Persona Vectors uses system prompts that respectively encourage and suppress the target trait to obtain contrastive responses, and then computes the mean activation difference at response positions. GCM uses contrastive prompts and their generated responses to construct a sequence-level likelihood difference, in order to identify attention heads that mediate the target behavior.

OCGS also utilizes responses generated by the target model itself. However, these responses are used not only to construct high- and low-behavior contrasts, but, more importantly, to provide the target model's own generation trajectories for gradient computation. Specifically, we sample multiple responses from the same behavior-neutral prompt, select high- and low-scoring responses, and compute state gradients at each position through teacher forcing along the target model's own generation trajectories. This ensures that gradient contributions directly correspond to the internal states actually encountered during the model's own generation process, thereby more accurately characterizing the contributions of different positions to target-behavior control, rather than relying on externally constructed responses and the potentially out-of-distribution states they may induce relative to the model's own generation distribution.

AuthorsOpenReview ↗

Response to Reviewer 1XE4 (3)

Question 1.


Thank you for your question! We choose the three relatively coarse-grained intervention operators, Prefill, Decode, and All, in the main text primarily to introduce and validate OCGS under several representative and general intervention settings, rather than because the method itself has any limitations in terms of positional granularity. In fact, OCGS can naturally be extended to arbitrary specified sets of sparse intervention positions: according to Eq. 11, we only need to sum the state gradients over the specified position set to extract a steering direction matched to that intervention set, and deploy it at the same positions during inference. This is straightforward both mathematically and computationally. Furthermore, we can obtain the gradients at all positions through only a single backward pass and then aggregate them using different positional masks, without requiring a new backward pass for each sparse position set.

Furthermore, during the rebuttal, we conducted two additional sparse intervention experiments. For each setting, we extract the steering direction using gradients from the selected positions and deploy it at the same positions during inference. We consider the first 10, 20, and 30 positions for both Decode (starting from the assistant-boundary state used to predict the first response token) and Prefill (prompt-side token positions). The experimental results are shown in the following table. All values report behavioral score improvements over the unsteered model (Δ\DeltaBase).

BehaviorDecode-10Decode-20Decode-30Full DecodePrefill-10Prefill-20Prefill-30Full Prefill
Extraversion1.3001.5751.8252.7500.3200.4000.5752.375
Humor1.0251.8002.3254.4250.2150.3250.6254.425
Sadness0.5750.8500.8253.9500.0400.2000.2754.100
Average over three behaviors0.9671.4081.6583.7080.1920.3080.4923.633

Table 2: Behavioral control performance under different sparse intervention positions. Decode-kk denotes the first kk decoding positions starting from the assistant boundary, while Prefill-kk denotes the first kk prompt-side positions. For each setting, the steering direction is extracted and deployed using the corresponding position set.

The results show that Decode and Prefill exhibit different behavioral-control characteristics under sparse intervention. Decode achieves noticeable behavioral gains even when intervening at only the first 10 generation positions, and the average gain further increases as the intervention range expands. This indicates that interventions at early generation steps can influence subsequent autoregressive computation and thereby affect the generated behavior. In contrast, intervening at only a small number of positions at the beginning of the prompt during Prefill has relatively limited effects, whereas full Prefill intervention produces substantially larger gains.

Further positional selection experiments reveal that the effectiveness of Prefill is highly sensitive to the specific intervention positions. Under the same write budget, intervening at the last 10 prompt positions achieves substantially higher behavioral gains than intervening at the first 10 positions across all three behavioral dimensions, with improvements of 0.950, 2.205, and 1.390, respectively. The specific results are shown below:

BehaviorFirst 10 PositionsLast 10 PositionsGain from Last Window
Extraversion+0.320+1.270+0.950
Humor+0.215+2.420+2.205
Sadness+0.040+1.430+1.390

Table 3: Behavioral control performance when intervening at the first versus the last 10 prompt positions under the same write budget (Δ\DeltaBase).

Furthermore, when the windows are expanded to 20 and 30 positions, the last-position windows consistently achieve higher behavioral gains in matched-budget comparisons across different window sizes and intervention strengths. In particular, the last 30 positions require only 24%–42% of the writes used by full Prefill, while retaining 63%–82% of its behavioral gains. This demonstrates that the effectiveness of sparse Prefill intervention depends not only on the number of write positions, but also critically on selecting more appropriate prompt positions.

These additional experiments demonstrate that OCGS is not only applicable to the three general intervention operators introduced in the main text, but can also flexibly support finer-grained sparse intervention positions and construct corresponding operator-conditioned steering directions for arbitrary given intervention position sets. We have organized and presented all these additional results in Appendix C.12 of the revised manuscript resubmitted during the rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.