acceptodds
Under review as a conference paper at ICLR 2027

A Time–Frequency Perspective on Activation Steering in Large Language Models

Abstract

Activation steering offers a lightweight way to control the behavior of large language models by modifying hidden states during inference. We investigate how changes in these states across token positions can guide such interventions. Taking a time–frequency perspective, we apply the discrete cosine transform (DCT) along token positions to separate the mean hidden state from variation at different frequencies. An analysis of 500 concepts from Concept500 reveals attribute information in nonzero-frequency coefficients and shows that this information is associated with patterns of hidden-state variation. These findings motivate SARA, a training-free method that uses recent hidden-state changes as feedback to adjust a base steering direction. During generation, it computes this feedback by comparing nonzero-frequency coefficients from recent hidden states with statistics estimated from the calibration set. Across three instruction-tuned models, SARA achieves the highest scores among the evaluated methods on all three TruthfulQA multiple-choice metrics. On the math and code subsets of Concept500, it also outperforms the trained baselines in concept relevance and aggregate score.

Primary area:
interpretability and explainable AI
Keywords:
LLM, interpretability, activation steering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fK2oOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper applies DCT across token activations to separate their average from slow and fast patterns of variation, showing that individual frequency components contain concept information. The author's method (SARA) compares these components from recent activations against positive/negative calibration examples and uses the resulting feedback to dynamically adjust a steering direction during generation.

Critical strengths

The DCT analysis shows that individual nonzero-frequency components contain decodable concept information, especially at low frequencies. They also test token shuffling and frequency-band ablations.

Critical weaknesses
  • The paper is missing a comparison against simpler temporal summaries, like early-minus-late activations or a linear slope. The experiments don’t isolate the benefit of the DCT specifically.
  • Concept prompting is absent from the comparison, despite SARA using explicit concept instructions to construct its online steering direction. AxBench is cited, their paper argue specifically argue about including this control.
  • The fine-tuning comparison is limited to 72 positive examples per concept, with rank-4 LoRA applied only to attention output projections at four layers. This limits the result such technique could have on the final model.
  • SARA gets much higher concept relevance (good), but lower fluency than other approaches. Without strength sweeps or comparisons at matched fluency, the results don’t establish a better conditioning - quality tradeoff. Sufficiently strong steering can produce those results which can be misleading on which approach is the best.
  • Calibration uses the first activation window of each response, but generation applies those statistics to sliding windows throughout the answer. That mismatch isn’t directly tested.
Questions for authors

I would ask the authors to address the main weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer fK2o (Part 1)

Thank you for the thoughtful comments on the role of the DCT and our experimental comparisons. We have replaced the main framework figure and shortened the appendix by removing overly detailed experimental descriptions and redundant data, while retaining the information needed to understand and reproduce our results.

(W1)

The paper is missing a comparison against simpler temporal summaries, like early-minus-late activations or a linear slope. The experiments don’t isolate the benefit of the DCT specifically.

Additional comparisons. We added comparisons with early-minus-late and linear slope. The comparison below uses Llama-3.1-8B-Instruct on all 817 TruthfulQA questions. The temporal-summary baselines use their independently validation-selected settings.

MethodMC1MC2MC3TrueInfoTrue×Info\mathrm{True}\times\mathrm{Info}
Early Minus Late53.6172.5250.8280.2498.5779.09
Linear Slope54.2272.3751.0080.8398.6779.75
SARA (DCT)56.0672.5652.2382.0797.6280.12

SARA achieves the highest MC1, MC2, MC3 and True×Info\mathrm{True}\times\mathrm{Info} in this comparison, supporting the use of feedback from separate frequencies.

Relationship to our SARA. DCT provides a unified mathematical view of linear temporal summaries: any such summary is a weighted sum of DCT coefficients. Let H∈RW×dH\in\mathbb{R}^{W\times d} contain the hidden states, U=[u0,…,uW−1]U=[u_0,\ldots,u_{W-1}] be the orthonormal DCT basis, and C=U⊤HC=U^\top H. For any temporal weight vector aa,

=\sum_{k=0}^{W-1}\beta_k C_k, \qquad \beta_k=u_k^\top a.$$ Early Minus Late and Linear Slope correspond to specific choices of $a$. In the standard symmetric setting, Early Minus Late uses opposite weights on equal halves of the window, and Linear Slope uses weights proportional to $t-(W-1)/2$. Both satisfy $a_{W-1-t}=-a_t$. Since the DCT basis satisfies $u_k(W-1-t)=(-1)^k u_k(t)$, $$\beta_k=(-1)^{k+1}\beta_k \quad\Longrightarrow\quad \beta_k=0\quad\text{for even }k.$$ Thus, early-minus-late and linear slope use only odd DCT frequencies under these symmetric definitions. This view makes their frequency content explicit: each method fixes a particular mixture of AC coefficients. SARA keeps the frequencies separate and uses concept feedback to adjust their contributions during generation. We have added these comparisons to Section 5.3 (Table 3) and the unified mathematical view to Appendix F.1. **(W2)** > Concept prompting is absent from the comparison, despite SARA using explicit concept instructions to construct its online steering direction. AxBench is cited, their paper argue specifically argue about including this control. We have added a Prompting baseline that places the target concept instruction directly in the generation prompt. The expanded comparison includes online RePE and all three RePS variants. Results use Gemma-2-2B-it on 35 math, 107 code and 50 text concepts. SARA achieves higher $H$ scores than Prompting across math, code and text, supporting the value of activation steering with explicit concept information. These results suggest that SARA generalizes beyond code and math to text concepts. We have added Prompting, online RePE and RePS baselines to Section 5.2 (Table 2; Appendix E.1). | Method | Math $H$ | Code $H$ | Text $H$ | | :--- | ---: | ---: | ---: | | Prompting | 0.9937 | 1.1361 | 1.0568 | | RePE (online) | 1.1749 | 1.2237 | 1.2316 | | RePS-SV | 0.3629 | 0.1764 | 0.5160 | | RePS-LoRA | 0.6697 | 0.3757 | 1.1324 | | RePS-LoReFT | 0.7194 | 0.4157 | 1.0864 | | SARA | 1.1914 | 1.2660 | 1.2540 | **(W3)** > The fine-tuning comparison is limited to 72 positive examples per concept, with rank-4 LoRA applied only to attention output projections at four layers. This limits the result such technique could have on the final model. Our training data and LoRA settings follow AxBench's published protocol and released configuration. AxBench allocates 144 examples per concept, with 72 positive and 72 negative examples, and trains its fine-tuning baselines on the positive examples ([AxBench, Sections 3.1 and 4](https://arxiv.org/html/2501.17148v3)). Our LoRA settings follow the official AxBench configuration ([AxBench LoRA configuration](https://github.com/stanfordnlp/axbench/blob/main/axbench/sweep/wuzhengx/2b/l20/lora.yaml#L19-L29)). For Gemma-2-2B-it, we use rank 4, attention output projections (`o_proj`), and layers 5, 10, 15 and 20. Our comparison also includes SFT, which updates all model parameters.
AuthorsOpenReview ↗

Response to Reviewer fK2o (Part 2)

(W4)

SARA gets much higher concept relevance (good), but lower fluency than other approaches. Without strength sweeps or comparisons at matched fluency, the results don’t establish a better conditioning - quality tradeoff. Sufficiently strong steering can produce those results which can be misleading on which approach is the best.

The intervention strength α\alpha controls the balance between concept relevance and fluency. At α=0.25\alpha=0.25, Prompting + SARA retains 98.8% of Prompting's fluency and improves HH by 7.9%, supporting stronger concept conditioning at comparable fluency. The results use Gemma-2-2B-it on all 192 concepts and 960 test responses. The configuration selected using validation results uses α=0.25\alpha=0.25, layers 10 and 20, W=64W=64 and ρ=2\rho=2 throughout, with the same generation prompts and decoding as Prompting.

MethodConceptInstructionFluencyHH
Prompting1.16561.86151.17401.0895
Prompting + SARA1.34381.83651.15941.1754

(W5)

Calibration uses the first activation window of each response, but generation applies those statistics to sliding windows throughout the answer. That mismatch isn’t directly tested.

Our current interpretation is that changes associated with the target concept maintain a relatively stable global structure throughout generation. The first window provides a representative observation of this structure, enabling DCT feedback to generalize across positions. A local window can capture the concept dynamics of the full response, allowing local calibration to support feedback throughout generation. We compare three settings on all 817 TruthfulQA questions with Llama-3.1-8B-Instruct. Single (Ours), used in our paper, calibrates on the first window of each training response. Separate fits calibration statistics separately at each window position. Combined combines windows from all positions to fit shared statistics. All three use the same inference parameters and decoding, with W=8W=8 in fold 0 and W=16W=16 in fold 1. At uncovered positions, Separate retains online RePE with zero SARA feedback.

CalibrationMC1MC2MC3TrueInfoTrue×Info\mathrm{True}\times\mathrm{Info}
Single (Ours)56.5574.1653.8481.5598.0279.93
Separate55.8174.1553.5182.5097.7680.65
Combined56.6774.1153.4181.9898.6880.90

All three schemes achieve comparable performance. This consistency supports our interpretation that the first window captures a globally shared structure in concept dynamics, enabling DCT feedback to transfer across generation positions.

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

This paper applies a DCT along token positions to decompose a window of hidden states into its mean, the DC component, and its variation at nonzero frequencies, the AC components. Probing on Concept500 with Gemma-2-9B shows that low-frequency AC coefficients are linearly predictive of concept presence. Based on this, the authors propose SARA, a training-free method that compares the AC coefficients of recent hidden states with calibration statistics and uses the resulting scores to correct a base steering direction at each step. Experiments on TruthfulQA with three 7B to 9B models and on the math and code subsets of Concept500 show gains over existing inference-time steering methods and trained interventions.

Critical strengths
  1. Analyzing hidden-state trajectories in the frequency domain is a new perspective for activation steering, and conventional mean-difference steering corresponds to the DC term, so the motivation is clear.
  2. The method requires no training, and the experiments span several models and tasks, with RePE controls, DC versus AC comparisons and frequency-band ablations.
Critical weaknesses
  1. All main experiments use the online RePE base direction from Appendix B.1, and most of the improvement appears to come from it rather than from the spectral feedback. RePE alone already outperforms all baselines in Table 1, while SARA adds only 1 to 4 MC2 points over it and shows no clear gain in open-ended generation.
  2. The multiple-choice gains may partly stem from the scoring protocol. Both the steering direction and the feedback are computed from the tokens of the candidate being scored, so the correction can act as a truthfulness classifier that suppresses the likelihood of incorrect candidates.
  3. The analysis does not show that AC carries information beyond the mean. The k=0 probe is already near ceiling and is no weaker than any nonzero frequency.
  4. The necessity of the frequency decomposition is not established. W=4 matches W=16 and the low band alone matches all frequencies, suggesting that the feedback mainly captures coarse recent drift rather than multi-scale structure.
Questions for authors
  1. How do the MC results change if the online direction and the feedback are computed without access to the candidate's tokens?
  2. Does a probe on [C_0; C_k] outperform one on C_0 alone, and do the results change when the window starts inside the response rather than at the final prompt token?
  3. What is the inference overhead of the online variant in forward passes, latency and memory compared with CAA and Spherical Steering?
3 comments
AuthorsOpenReview ↗

Response to Reviewer UiMB (Part 1)

Thank you for recognizing the frequency-domain perspective of our work. We have replaced the main framework figure and shortened the appendix by removing overly detailed experimental descriptions and redundant data, while retaining the information needed to understand and reproduce our results. We address the main concerns below.

(W1)

All main experiments use the online RePE base direction from Appendix B.1, and most of the improvement appears to come from it rather than from the spectral feedback.

We additionally evaluated SARA on TruthfulQA with spectral feedback applied to a fixed mean-difference steering direction, and compared it with CAA. We also evaluated spectral feedback with the online RePE direction. With the fixed direction, SARA improves MC2 by 3.92 percentage points and True×Info\mathrm{True}\times\mathrm{Info} by 9.58 points over CAA. With the online direction, the additional validation-selected SARA configuration improves MC2 by 2.90 points and True×Info\mathrm{True}\times\mathrm{Info} by 1.89 points over RePE. The paired 95% bootstrap confidence interval for the True×Info\mathrm{True}\times\mathrm{Info} difference is [+0.23, +3.58] points, excluding zero. Each fold selects its configuration using only its own validation data, and both configurations are frozen before test evaluation. The gains with both directions indicate that the improvement is not solely attributable to the strong RePE base. The table below reports results on all 817 TruthfulQA questions using Llama-3.1-8B-Instruct. We have updated the TruthfulQA results in Section 5.2 (Table 1; Appendix E.2).

MethodMC1MC2MC3TrueInfoTrue×Info\mathrm{True}\times\mathrm{Info}
CAA38.4355.9032.5475.0093.3770.03
SARA (offline)41.3759.8234.5180.0799.4279.61
RePE52.3969.6549.4379.6098.2878.23
SARA (online)56.0672.5652.2382.0797.6280.12

(W2&Q1)

The multiple-choice gains may partly stem from the scoring protocol & How do the MC results change if the online direction and the feedback are computed without access to the candidate's tokens?

We added a candidate-independent control on all 817 TruthfulQA questions using Llama-3.1-8B-Instruct. In the table, w/ means the online direction and feedback are computed using the candidate tokens during scoring. w/o means they are computed from a reference continuation generated using only the question. SARA's advantage persists when the control signals have no access to candidate tokens, showing that its MC gains are not solely explained by a candidate-dependent truthfulness classifier.

MethodSettingMC1MC2MC3TrueInfoTrue×Info\mathrm{True}\times\mathrm{Info}
RePEw/52.3969.6549.4379.6098.2878.23
Spherical Steeringw/53.0069.4948.4282.7697.1980.44
SARAw/56.5574.1653.8481.5598.0279.93
RePEw/o42.3561.9636.4679.6098.2878.23
Spherical Steeringw/o38.1957.6932.8782.7697.1980.44
SARAw/o44.6863.6639.3481.5598.0279.93
AuthorsOpenReview ↗

Response to Reviewer UiMB (Part 2)

(W3&Q2)

The analysis does not show that AC carries information beyond the mean & Does a probe on [C0;Ck][C_0; C_k] outperform one on C0C_0 alone, and do the results change when the window starts inside the response rather than at the final prompt token?

Our claim is that AC contributes concept information complementary to the mean. The role of AC is to enrich DC with information carried by temporal variations. The strong DC probe provides the baseline for assessing the extra information available when AC and C0C_0 are used together. AC and DC are approximately orthogonal on average. We remove the part of AC that lies along the DC direction. Probes trained on the remaining AC features still identify concepts. This provides evidence that AC carries useful concept information outside the mean direction. The probe results below use Gemma-2-9B-it across all 500 Concept500 concepts.

kkMean cosine with C0C_0AC AUC (%)DC-orthogonalized AC AUC (%)
10.095797.0997.11
20.001593.7593.87
30.013987.3688.22
40.024284.2684.77
50.041579.8680.43
60.028571.4772.02
70.029070.1970.34

The joint probes directly test our central claim: AC adds information to the strong C0C_0 representation. Their improvement over the DC baseline supports using the two together for steering.

Probe inputAUC (%)ΔAUC\Delta\mathrm{AUC} vs. C0C_0: paired 95% CI (pp)
C0C_098.57—
[C0;C1][C_0; C_1]98.61[+0.02, +0.06]
[C0;C2][C_0; C_2]98.73[+0.13, +0.20]
[C0;C3][C_0; C_3]98.80[+0.18, +0.27]
[C0;C4][C_0; C_4]98.78[+0.17, +0.25]
[C0;C5][C_0; C_5]98.72[+0.11, +0.18]
[C0;C6][C_0; C_6]98.68[+0.08, +0.14]
[C0;C7][C_0; C_7]98.65[+0.05, +0.10]

AC's concept information extends from the prompt boundary into the response. This supports using it as feedback on the model's evolving response.

kkLast prompt token: AUC (%)First response token: AUC (%)
197.0896.96
293.7493.54
387.3486.94
484.2683.90
579.8479.58
671.4170.42
770.2269.81

(W4)

The necessity of the frequency decomposition is not established. W=4W=4 matches W=16W=16 and the low band alone matches all frequencies, suggesting that the feedback mainly captures coarse recent drift rather than multi-scale structure.

The updated window ablation shows a meaningful effect of window length: the longer windows preserve useful temporal context that contributes to SARA's feedback. The results below use Llama-3.1-8B-Instruct on all 817 TruthfulQA questions.

WWMC1MC2MC3
452.1471.1049.69
856.5573.8553.17
1655.2073.5852.42

To prove the necessity of the frequency decomposition, we added comparisons with early-minus-late and linear slope, two simple temporal summary methods. The comparison below uses Llama-3.1-8B-Instruct on all 817 TruthfulQA questions. The temporal-summary baselines use their independently validation-selected settings.

MethodMC1MC2MC3TrueInfoTrue×Info\mathrm{True}\times\mathrm{Info}
Early Minus Late53.6172.5250.8280.2498.5779.09
Linear Slope54.2272.3751.0080.8398.6779.75
SARA (DCT)56.0672.5652.2382.0797.6280.12

SARA achieves the highest MC1, MC2, MC3 and True×Info\mathrm{True}\times\mathrm{Info} in this comparison, supporting the use of feedback from separate frequencies. On the other hand, DCT provides a unified mathematical view of linear temporal summaries. Let HH contain the window's hidden states, UU be the orthonormal DCT basis, and C=U⊤HC=U^\top H be the frequency coefficients. Any linear summary over token positions can be written as a weighted sum of these coefficients:

=\sum_{k=0}^{W-1}(U^\top a)_k C_k.$$ The mean is $C_0/\sqrt{W}$, while temporal contrasts whose weights sum to zero, such as endpoint differences, are weighted sums of AC coefficients. This gives means and temporal changes a shared mathematical description and makes the frequency weights of traditional summaries explicit. SARA uses concept feedback to adjust these frequency weights. We have updated the window-length sensitivity experiment in Appendix E.4.2 (Figure 17a), added these comparisons to Section 5.3 (Table 3) and the unified mathematical view to Appendix F.1.
AuthorsOpenReview ↗

Response to Reviewer UiMB (Part 3)

What is the inference overhead of the online variant in forward passes, latency and memory compared with CAA and Spherical Steering?

SARA (online) achieves the highest MC scores with peak memory close to CAA, while SARA (offline) runs at near-CAA latency. Measurements use Llama-3.1-8B-Instruct on an A800 80GB, batch size 1 and a 128-token generation limit. We have added the inference cost experiment to Appendix E.5 (Table 10).

MethodMC1MC2MC3TrueInfoTrue×Info\mathrm{True}\times\mathrm{Info}Latency (s/sample)Peak memory (GiB)
Base37.8256.1132.0171.2498.8570.430.80214.985
CAA38.4355.9032.5475.0093.3770.030.90614.985
Spherical Steering53.0069.4948.4282.7697.1980.441.61914.995
RePE52.3969.6549.4379.6098.2878.231.30115.030
SARA (online)56.5574.1653.8481.5598.0279.931.72615.088
SARA (offline)41.3759.8234.5180.0799.4279.610.93414.993
Reviewer_XqFyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies activation steering from a time–frequency perspective, applying the discrete cosine transform along token positions to separate mean hidden states from variations at different frequencies. An analysis of 500 concepts identifies linearly decodable attribute information in nonzero-frequency coefficients. Motivated by these findings, the authors propose SARA, a training-free method that uses calibrated feedback from recent hidden-state variations to adjust a base steering direction during generation. Experiments show improvements over the evaluated baselines on TruthfulQA multiple-choice metrics across three instruction-tuned models and on concept relevance and aggregate scores for the math and code subsets of Concept500.

Critical strengths
  • Presents a useful time–frequency perspective on activation steering and develops SARA to incorporate recent hidden-state variations into steering decisions without additional training.

  • Includes extensive ablations on frequency components, feedback, and reliability weighting. Together with the matched RePE comparisons, these experiments provide evidence that spectral feedback improves TruthfulQA multiple-choice performance beyond the base steering direction.

Critical weaknesses
  1. The paper does not provide deep mechanistic interpretability or causal explanations for the benefits of nonzero-frequency DCT components. Mean-based probes perform equally well or better than probes on nonzero-frequency coefficients (Figure 2), so this comparison does not establish an advantage of frequency-based feedback over mean-based feedback. Furthermore, on TruthfulQA open-ended generation, the True×Info differences between SARA and RePE are small, and their confidence intervals all include zero. Thus, it remains unclear whether spectral feedback provides additional generation benefits.

  2. The evaluation lacks relevant baselines, such as prompting (which AxBench [ https://arxiv.org/abs/2501.17148v3 ] found to outperform the evaluated representation-based intervention methods overall), RePE, and RePS [ https://arxiv.org/abs/2505.20809 ] . Moreover, the main concept-steering experiments on Concept500 cover only 142 code and math concepts, leaving it unclear whether the results generalize to a broader range of concept types.

Questions for authors

see above weakness

2 comments
AuthorsOpenReview ↗

Response to Reviewer XqFy (Part 1)

Thank you for the thoughtful feedback on the interpretation of AC components and the scope of our evaluation. We have replaced the main framework figure and shortened the appendix by removing overly detailed experimental descriptions and redundant data, while retaining the information needed to understand and reproduce our results.

(W1-1)

The paper does not provide deep mechanistic interpretability or causal explanations for the benefits of nonzero-frequency DCT components.

We examine the role of nonzero-frequency DCT (AC) components from four perspectives.

  1. AC components distinguish hidden-state trajectories with similar means. In Section 1 (Figure 1; Appendix C), we compare sentences containing the same words but expressing different meanings. Their DC representations are similar, while their AC coefficients differ, illustrating how averaging can obscure differences across token positions.
  2. AC components carry concept information. In Section 3.2 (Figure 2; Appendix D.1), we examine coefficient distributions and train linear probes on Concept500 to predict concept presence from individual DCT coefficients. The results show that concept information is linearly decodable from AC components.
  3. The concept information decoded from AC components depends on hidden-state ordering. In Section 3.3 (Figure 3; Appendix D.1), we shuffle the extracted hidden states within each response, preserving their mean, and refit the probes on the shuffled representations. DC performance remains unchanged, whereas AC performance drops close to chance, linking AC decodability to the ordering of hidden states.
  4. AC feedback improves steering performance. In Section 5.3 (Figures 5–6; Appendix E.4.1, Figures 15–16), we compare DC, AC, and DC+AC feedback with a no-feedback control. On Llama–TruthfulQA, AC feedback achieves higher MC2 than both DC feedback and the no-feedback control at every tested steering strength.

These analyses clarify what AC components capture, while the ablations demonstrate their contribution to steering performance.

(W1-2)

Mean-based probes perform equally well or better than probes on nonzero-frequency coefficients (Figure 2), so this comparison does not establish an advantage of frequency-based feedback over mean-based feedback.

Our motivation is to extract information complementary to the mean, rather than to make AC probes outperform mean-based probes. The following evidence supports using AC alongside DC:

  1. Position-dependent information. Section 3.3 shows that shuffling extracted states leaves DC unchanged but reduces separately fitted AC probe performance to near chance, indicating that AC captures information associated with variation across token positions.
  2. Distinct directions. The additional analysis below finds low cosine similarity between DC and AC coefficients across Concept500 concepts, suggesting that their representations are not generally aligned.
  3. Information outside of the DC direction. In the additional probe analysis below, removing each AC coefficient's component along DC largely preserves probe AUC. This indicates that concept information remains linearly accessible outside the DC direction.

These findings motivate combining AC with the strong mean representation in SARA (Section 4); they do not on their own establish a causal generation advantage over mean-based feedback. The table below reports probe AUC before and after removing the component of AC along DC.

kkMean cosine with C0C_0AC AUC (%)AC AUC after removing DC (%)
10.095797.0997.11
20.001593.7593.87
30.013987.3688.22
40.024284.2684.77
50.041579.8680.43
60.028571.4772.02
70.029070.1970.34

(W1-3)

On TruthfulQA open-ended generation, the True×Info\mathrm{True}\times\mathrm{Info} differences between SARA and RePE are small, and their confidence intervals all include zero.

The additional evaluation shows an improvement in True×Info\mathrm{True}\times\mathrm{Info} over RePE on Llama-3.1-8B-Instruct. We extended the validation search beyond steering strength to include intervention layers, window length (WW), and feedback gain (ρ\rho). The results below cover all 817 TruthfulQA questions using the validation-selected configuration. SARA improves True×Info\mathrm{True}\times\mathrm{Info} by 1.89 percentage points over RePE (paired 95% bootstrap CI: [+0.23, +3.58] percentage points). The interval excludes zero, supporting an open-ended generation benefit. We have updated the TruthfulQA results in Section 5.2 (Table 1) and the paired generation comparisons in Appendix E.2.2 (Table 8).

MethodMC1MC2MC3TrueInfoTrue×Info\mathrm{True}\times\mathrm{Info}
RePE52.3969.6549.4379.6098.2878.23
SARA (online)56.0672.5652.2382.0797.6280.12
AuthorsOpenReview ↗

Response to Reviewer XqFy (Part 2)

(W2)

The evaluation lacks Prompting, RePE and RePS baselines, and the AxBench experiments cover only code and math concepts.

The expanded AxBench comparison includes Prompting, online RePE and all three RePS variants. It adds 50 text concepts to the original 35 math and 107 code concepts. Results use Gemma-2-2B-it; HH is the harmonic mean of concept, instruction and fluency scores. We will extend this comparison to the full AxBench concept set. SARA achieves the highest HH score across math, code and text, supporting its effectiveness across different concept types. We have added these baselines to the paper in Section 5.2 (Table 2; Appendix E.1).

MethodMath HHCode HHText HH
Prompting0.99371.13611.0568
RePE (online)1.17491.22371.2316
RePS-SV0.36290.17640.5160
RePS-LoRA0.66970.37571.1324
RePS-LoReFT0.71940.41571.0864
SARA1.19141.26601.2540

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful assessments and constructive feedback. Their comments helped us sharpen the evidence for SARA's central contribution: using temporal variation in hidden states to adapt activation steering during generation. We have revised the manuscript, with modified text marked in blue and new or replaced figures outlined in blue.

We summarize the main additions and clarifications below:

  1. Concept information beyond the mean (Reviewers XqFy and UiMB). Our additional probe analyses show that AC coefficients retain concept information after removing their components along the DC direction, and that joint DC–AC probes improve on DC alone. The information also persists when the activation window starts inside the response. Together with the hidden-state shuffling control, these results support our central interpretation: temporal variation carries concept information complementary to the mean, motivating its use alongside DC in steering. The original analyses appear in Sections 3.2–3.3, and the additional controls are reported in our point-by-point responses.

  2. Direct comparisons isolating the role of DCT (Reviewers UiMB and fK2o). We added early-minus-late and linear-slope feedback baselines, with settings selected independently on validation data. DCT provides more effective steering in this comparison. We also provide a unified mathematical view of linear temporal summaries: any such summary is a weighted sum of DCT coefficients. Early-minus-late and linear slope prescribe particular combinations of these coefficients; SARA retains the frequencies separately and adjusts their contributions using attribute feedback. The experiments and derivation appear in Section 5.3, Table 3, and Appendix F.1.

  3. Broader concept-steering comparisons and a quality-matched control (Reviewers XqFy and fK2o). The AxBench evaluation now covers 192 concepts: 50 text, 107 code, and 35 math, and includes Prompting, RePE, and all three RePS variants alongside the existing training baselines. SARA achieves the highest harmonic-mean steering score across all three concept types in the expanded comparison. To assess conditioning at comparable generation quality, our additional Prompting + SARA control retains 98.8% of Prompting's fluency while improving the harmonic-mean score by 7.9%. The expanded comparison appears in Section 5.2, Table 2, and Appendix E.1; the quality-matched control is reported in our response to Reviewer fK2o.

  4. Evidence that the gains come from feedback beyond the base direction (Reviewers XqFy and UiMB). Spectral feedback improves steering with both a fixed mean-difference direction and an online RePE direction. On Llama–TruthfulQA, it improves True × Info by 9.58 percentage points over CAA with the fixed direction and by 1.89 points over RePE with the online direction; the latter has a paired 95% confidence interval of [+0.23, +3.58]. A separate candidate-independent control preserves SARA's multiple-choice advantage when control signals are constructed without access to candidate tokens. These results establish the contribution of feedback across base-direction choices and support its benefit beyond candidate-dependent scoring. The revised main results and paired intervals appear in Section 5.2, Table 1, and Appendix E.2; the additional controls are reported in our response to Reviewer UiMB.

  5. Inference cost and clearer presentation (Reviewer UiMB). We added latency, peak-memory, and forward-pass accounting. SARA computes temporal feedback without additional model forward passes beyond those used by its underlying steering variant. Offline SARA runs at near-CAA latency (0.934 vs. 0.906 seconds per response), while online SARA uses 15.088 GiB of peak memory on the measured setup. These results appear in Appendix E.5, Table 10. We also replaced the main framework figure and streamlined the appendix to focus on the evidence and methodological details that support the paper's claims.

Detailed point-by-point responses are provided under each review. We hope this summary helps the reviewers and Area Chair navigate the additional evidence and manuscript revisions. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.