acceptodds
Under review as a conference paper at ICLR 2027

WAQS: WEIGHT ABSORPTION FOR QUADRATIC PROBING AND AFFINE STEERING

Abstract

Activation steering provides a lightweight alternative to fine-tuning for controlling large language model behavior. Existing methods, however, often face a trade-off between efficiency and expressiveness: fixed steering vectors incur negligible inference overhead but cannot adapt to the current model input, whereas input-dependent nonlinear interventions provide greater flexibility at additional runtime cost. We introduce **WAQS** (**W**eight **A**bsorption for **Q**uadratic probing and affine **S**teering), a framework that obtains state-dependent steering, without additional inference-time computation. WAQS models a concept with a quadratic probe, whose gradient yields an affine, activation-dependent steering direction. We show that this affine intervention can be folded exactly into adjacent linear transformations through an offline weight update. Moreover, we prove that quadratic functions form the maximal class of smooth concept scores whose gradient-based interventions are affine and therefore exactly absorbable into linear layers. Under a Gaussian representation model, our formulation admits an interpretation as the difference of class-conditional Mahalanobis pulls and naturally generalizes linear discriminant and difference-in-means steering. To scale to high-dimensional representations, we introduce a low-rank parameterization of the quadratic term. Across truthfulness, refusal, and sycophancy steering tasks, quadratic steering matches or improves upon constant linear steering while preserving model utility, with the absorbed implementation introducing no additional inference-time operations. Our results suggest that quadratic probes provide a practical middle ground between static linear steering and computationally expensive dynamic nonlinear interventions.

Primary area:
interpretability and explainable AI
Keywords:
activation steering, weight editing, mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7cWtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WAQS derives activation-dependent steering from the gradient of a quadratic concept probe and absorbs the resulting affine intervention into compatible model projections. The paper characterizes quadratic scores as the smooth functions whose fixed Euclidean-gradient step is affine, provides a Gaussian discriminant interpretation, and introduces low-rank probe estimation. Experiments evaluate truthfulness, refusal, sycophancy, and corrigibility, with ablations of state dependence and measurements of inference throughput.

Critical strengths
  1. A useful connection between probe expressiveness and deployment. A quadratic score induces an affine gradient field, allowing input-dependent interventions to be compiled at compatible affine layers. Proposition 1 correctly characterizes this scalar-gradient construction, and the Gaussian interpretation clarifies when covariance differences introduce state dependence. The low-rank parameterization makes probe estimation and weight folding practical.

  2. Targeted evidence for adaptivity and efficiency. Frozen-QP retains the fitted quadratic probe and replaces its gradient field with its training-set average, providing a meaningful control for state dependence. Table 4 reports higher True*Info and MC1 for dynamic QP. Table 6 reports essentially unchanged prefill and decoding throughput after absorption. Appendix C.1 also specifies held-out validation selection and once-only test evaluation.

Critical weaknesses
  1. The demonstrated compilation scope does not establish the general deployment claim. Equations (7)–(8) require both weight and bias updates; a weight-only edit of a bias-free projection cannot generally realize the nonzero offset. The RMSNorm derivation in Section 3.1(c) likewise retains an additive αwp\alpha w_p after updating the scale. More substantially, Section 5 acknowledges that residual-stream steering generally requires an explicit runtime transformation because it changes the weightless skip path, whereas Appendix C.1 and Table 9 claim absorption across residual-stream locations. Task performance at those locations does not establish equivalence between an intervention hook and a merged checkpoint. The paper needs to identify the compiled operators and reconcile these statements before its central zero-additional-operations claim can be assessed.

  2. The evidence supports useful trade-offs, but not the claimed general superiority without utility loss. Table 2 reports GCG attack success rates of 0 for Spherical Steering and 9 for WAQS, and GSM8K accuracy of 22.59 for DiffMeans versus 22.36 for WAQS. These results contradict the accompanying claims of lowest attack success across all attacks and matching or exceeding both constant baselines on every utility benchmark. Table 1 also reports Llama-2 cross-entropy of 2.41 for WAQS versus 2.13 for Linear Probe. Without unsteered utility results in these tables and uncertainty for the refusal comparisons, the extent of capability preservation and the significance of small gains remain unclear. This limits the evidence for replacing linear steering as the default baseline.

Questions for authors
  1. Which exact weights, biases, and runtime operators are used for the main task results and Table 6? Please provide hook-versus-compiled activation or logit discrepancies for prompt processing and cached decoding, explain the treatment of additive offsets, and identify residual-site results that require hooks. Establishing equivalent merged implementations for the central results, with a consistent architectural scope, would resolve the principal deployment concern.

  2. Can you provide unsteered utility measurements and uncertainty or paired comparisons for the refusal results, and revise the claims to reflect the reported trade-offs? Evidence that the compiled intervention delivers a meaningful steering benefit at comparable capability preservation would support raising the rating; correcting the prose alone would not establish that benefit.

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

This paper presents probing the desired behavior in LLM activations using a quadratic function instead of a linear function. This allows the steering to become context dependent as the gradient becomes a linear function in the current activation. They derive expressions such that the new interventions act as new weights to avoid repeated computations during inference. Results on truthfulness and refusal show that it outperforms linear and angular baselines to some extent, while maintaining utility during refusal.

Critical strengths
  • The paper addresses an important problem of context-aware steering while keeping the inference cost same.
  • The paper is easy to follow and the details are presented in a readable manner.
  • Utility evaluations are also provided along with the refusal metrics for that task.
  • Theoretical analysis is sound and helpful to motivate and ground the method in a simple and easy to understand.
  • Frozen quadratic steering is an interesting ablation design to substantiate the benefits of context-specific steering.
Critical weaknesses
  • Performance variation across different steering locations is not clearly shown for all (Table 9 is incomplete and is not discussed). In fact, Table 9's and Appendix C.1's numbers do not even match the Table 1 results. It is also not clear if such location search also happens for the baselines. Ideally, a comprehensive comparison across locations is required.
  • Given the baselines tend to steer directly on the residual stream activations and WAQS cannot absorb weights in these settings, the efficiency gains may also be limited.
  • Results are only shown on a single model primarily (Llama-7b) with gemma-13b considered only for truthfulness.
  • Datasets are limited to just refusal and truthfulness. They are encouraged to include the full suite of AxBench dataset and the CAA behaviors to show the gains of the proposed probe.
  • Refusal setup is not standard and the numbers are not clear to understand in the community. They are encouraged to explain the metrics clearly or just show the results on standard benchmarks such as HarmBench, AdvBench, and some jailbreaks perhaps. Their reference here of Chhabra and Khalil is also pretty irrelevant as it does not align with the setup here.
  • Probe training is prohibitively expensive given the minimal gains that we get in the results. In particular, their default rank-8 implementation takes 8x more parameters than a linear probe and gives almost within error bounds gains.
  • ODESteer [1] also does gradient of the probe function where the probe function can even be non-linear. This makes WAQS intervention a special case of ODESteer with the major contribution lying in the absorption. This should be discussed properly, and a fair comparison be added.
  • Presentation issues:
    • Tables 8 and 9 are not discussed and it is not clear what they are supposed to show.
    • Appendix C.5 is misplaced as there is no AxBench quantitative results and there is no gemma-2-2b results. Another misplacement is Table 6 which shows results on Llama-3.1-8B when all the other results are on Llama-2-7b.

[1] Zhao, Hongjue, et al. "Odesteer: A unified ode-based steering framework for llm alignment." ICLR 2026

Questions for authors
  • How does the performance vary across locations and alpha for all steering methods?
  • How are the baselines implemented in detail? How are the spherical and angular hyperparameters tuned? as these do not correspond to additive alpha
  • Is the performance generalizable to more models and datasets?
  • How is the validation hyperparameter search conducted and what are the results on this set?
  • How does the method compare against ODESteeer with the same quadratic barrier function?
  • How is the comparison on a standard refusal setup or explain the numbers provided in the paper in more detail?
Reviewer_rRGqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets state-dependent steering for LLMs, with the objective of steering the model away from undesired behaviors without fine-tuning it. Previous steering methods either require additional inference-time computation (input-dependent, non-linear probes) or are linear, and thus do not adjust well to curved decision boundaries. This paper proposes WAQS, a quadratic concept probe whose gradient is affine in the activation, so the resulting input-dependent steering update can be absorbed offline into the LLM’s adjacent linear layers. The authors argue that quadratic functions are the largest class with this property, interpret the probe as QDA under a Gaussian model, and propose a low-rank parameterization. Their method is simple and shows improved performance on truthfulness, refusal and sycophancy steering against static (constant-vector) steering baselines.

Critical strengths
  1. Simple to implement and generalizable. It only requires a linear layer adjacent to the steered activation, and the result is a standard checkpoint with no hooks or custom inference code.
  2. Experiments show that the method improves downstream performance while keeping the no-inference-cost benefit. The Frozen-QP control (Table 4) uses the same probe with its averaged direction and falls back to the linear-probe level. This supports that the gain comes from input-dependent steering rather than from a better constant direction.
Critical weaknesses
  1. The efficiency-expressiveness claim is not tested against the right baselines or prior work.
  • The previous work does mention non-linear and dynamic steering methods, and the introduction motivates against them, but no experiment compares accuracy or inference cost with them, so the trade-off the paper claims to resolve is never measured. All baselines (linear probe, DiffMeans, Angular, Spherical, CAA) are static or geometric variants.
  • One of their main motivation points is that quadrating is "curved enough to represent a non-flat concept boundary". However, they do not measure this directly and do not compare with other non-linear or higher than quadratic variants. It would be important to compare with a variant of their method with higher polynomial to defend this point.
  • Input-dependent steering absorbed into the weights already exists. Arditi et al. (2024, §4.1, Eq. 5) fold directional ablation into every matrix that writes to the residual stream. Moreover, App. A describes this work but calls it “static, input-independent”, although the removed amount r̂(r̂ᵀx) depends on x.
  • Prop. 1 is a standard fact (an affine gradient means a constant Hessian), and its maximality holds only within single-step gradient steering x + α∇f(x). Other affine, absorbable edits exist (e.g., rotations, LoRA-style edits), so “the entire absorbable class” and “no engineering effort removes [the cost]” overstate it.
  • Absorption is limited to activations next to a linear layer. It does not cover the residual stream (Sec. 5), and norms only with a diagonal W_p. It is unclear how the residual connections are handled when steering the MLP output or the residual stream, where representations are usually read.
  • This matters because the paper’s contribution is input-dependent steering at the cost of static steering. Without the dynamic baselines and correct positioning against prior absorbed edits, neither side of that trade-off is established.
  1. The experimental setup is underspecified and the reported gains do not look significant.
  • The main paper does not say where x is obtained from, or why. It never states whether steering is applied at one layer or several, or which absorption case is used. It does not discuss how several steerings (from different features, or for several behaviors at once) would be combined.
  • How is \alpha selected? What metric is used to tune it and on what set? Is a held-out validation set used for this purpose? If so, how is it constructed?
  • The main paper presents MC3 scores, but does not mention what this is.
  • Many gains are within noise. For example, Table 1 MC1 and MC3 are within noise. Table 2 does not report std. Since they compute 5 seeds, they should provide appropriate std and significance tests for the results.
  • Some claims are not supported. For example, Table 2 mentions "quadratic probe achieves the lowest ASR", which contradicts the table, where spherical steering achieves lower GCG, and it is tied on GPTFuzz.
  • The sycophancy comparison uses the same α for all methods without matching steering norms, so “2.2× / 6.4× stronger” is not meaningful, since activations move by alpha ||v||, not just alpha. The text’s CAA value 0.0032 also disagrees with Table 3's 0.0074.
  • Utility is hard to judge. CE should be reported as perplexity, next to the unsteered base model in order for the values to be compared fairly. On Llama-2-7B, Quadratic’s CE of 2.41 vs. 2.13 for the linear probe is about 11.1 vs. 8.4 perplexity ( around 30% higher), yet it is described as “comparable”, which sounds like an overstatement.
Questions for authors

See weaknesses, especially:

  1. How does WAQS compare with dynamic non-linear steering in both steering quality and inference cost?
  2. How does WAQS relate to the weight-orthogonalized directional ablation of Arditi et al. (2024), which is already input-dependent and absorbed into the weights?
  3. Which location (layer, component and absorption case) is used in each experiment, and how was it chosen?
  4. Is steering applied at a single layer or at several?
  5. When steering the MLP output or the residual stream, would the update have to be absorbed into multiple linear layers?
  6. How is α selected?
  7. What are the perplexities of the unsteered base model and of each steered model? How is the KL in Table 1 computed, given that it stays at 0.005–0.007 while CE differs by 0.28 between methods?
  8. How do the sycophancy results change when steering strength is matched by the norm of the applied update rather than by α?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.