acceptodds
Under review as a conference paper at ICLR 2027

HeadEdit: Calibrating Language Model Behavior Through the Frozen Unembedding Matrix

Abstract

Alignment does not eliminate behavioral errors in language models. Models may still refuse benign requests, call unnecessary tools, or yield to false user claims. Current methods mitigate such errors as a computation problem, and rarely explore if the desired behavior is already encoded in the model's representation. Motivated by the observation that behavior-relevant information remains linearly decodable from the final hidden state even when the resulting logits produce the undesired behavior, we introduce HeadEdit, a gradient-free method that calibrates model behavior through the unembedding. HeadEdit extracts a low-rank behavioral subspace from paired completions and uses each prompt’s coordinates within it to generate a vocabulary-wide correction, thereby implementing implicitly adaptive steering without manually specified target tokens or parameter updates. HeadEdit improves all nine experimental settings across three tasks and three model families, with negligible inference overhead and no systematic loss of general capabilities. It also reveals a connection to gradient-based alignment. HeadEdit's low-dimensional representation partly predicts how preference tuning changes output logits on unseen prompts. The subspace learned from the model can also be reused after tuning, improving performance without re-extracting or retuning. These results show that HeadEdit provides a practical, lightweight, and interpretable way to calibrate model behavior through the unembedding.

Primary area:
interpretability and explainable AI
Keywords:
steering, unembedding, inference-time intervention, behavioral alignment, large language models

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_mHwvOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes HeadEdit, a gradient-free, inference-time method for correcting residual behavioral errors in instruction-tuned LLMs: over-refusal, tool overuse, and factual sycophancy.

Method.

  • From paired desirable/undesirable completions, it computes differences of final-layer (post-norm) hidden states and keeps the top-rr right singular vectors VV of the centered difference matrix.
  • At inference it adds the logit correction Δzt=−α(WUV)(V⊤ht)\Delta z_t = -\alpha (W_U V)(V^\top h_t). This is equivalent to using a transient effective unembedding WU(I−αVV⊤)W_U(I - \alpha V V^\top). In the main experiments, the edit is applied only at the first generation position.

Validation.

  • Across three ~4B instruction-tuned models (Qwen3-4B, Gemma-3-4B-IT, Llama-3.2-3B-IT) and the three tasks, HeadEdit improves decision accuracy in all nine settings over the unedited model. It is best or second-best among four activation-steering baselines (RepE, CAA, SADI-Head, Spherical Steering).
  • It has near-zero throughput overhead and small changes on MATH-500 and GPQA-Diamond. A Qwen3-14B check is also reported.
  • Additional analyses show that:
    • V⊤hV^\top h is more behavior-predictive than random subspaces of equal rank;
    • a ridge predictor on V⊤hV^\top h explains part of the LoRA-DPO logit update;
    • the base-model subspace can be reused on top of LoRA-DPO.
Critical strengths
  1. Simple, cheap, and carefully evaluated on two-sided behavioral boundaries. The method adds a few lines at the LM head. It needs no training and no layer selection, has two hyperparameters (rr, α\alpha), and its throughput is indistinguishable from the base model (Table 5). Each task is evaluated as a decision boundary, and dev-set selection explicitly protects the opposite side (Eq. 13–14), e.g., keeping unsafe refusals while reducing over-refusal. Retention metrics are reported (Table 6). This makes the reported improvements more credible than one-sided error reductions, and it is the main reason I consider the empirical contribution useful.
  2. Informative diagnostics for the "decodable but not expressed" framing. The probe analyses compare the paired subspace with matched random subspaces and report conditional repair rates with a Fisher test (Fig. 3–4). They show that the subspace used for editing carries nearly all of the full state's linear information about which errors get repaired. This is useful evidence beyond the headline numbers.
Critical weaknesses
  1. The core mechanism is not new, and closely related work is not discussed. The paper claims that interventions "rarely act on the unembedding layer" and that using internal information to correct behavior "remains unexplored" (Sec. 2). LM-Steer (Han et al., ACL 2024) already steers a frozen LM by replacing the output embeddings with (I+ϵW)ev(I + \epsilon W)e_v, with a scalar ϵ\epsilon controlling strength. HeadEdit's Eq. (6), WU(I−αVV⊤)W_U(I - \alpha V V^\top), is algebraically LM-Steer with W=−VV⊤W = -V V^\top. Learned logit-space calibration from the final hidden state (e.g., PALC; Lee & In, ICLR 2026) also uses the same "internal signal, output-stage intervention" design. The genuine contribution is the gradient-free SVD estimator and the evaluation, which is narrower than claimed.

  2. The gains cannot be attributed to the proposed estimator. All baselines edit intermediate activations, while HeadEdit edits the logits, and only at the first decoding position. Location, timing, and estimator are therefore confounded. The paper does not compare against methods that intervene at the same place, such as LM-Steer or a learned low-rank logit correction trained on the same pairs. Because all three tasks are largely decided by the first token, it is unclear whether the advantage comes from the SVD subspace or simply from editing logits at the decision token.

Questions for authors

All questions are clarifications that could change my assessment.

  1. How does HeadEdit compare with LM-Steer, and with a learned low-rank logit correction trained on the same pairs at matched rank? What does gradient-free SVD estimation buy over learning the same map?
  2. How does a simpler output-stage baseline perform, e.g., CAA/RepE directions applied to the post-norm final hidden state at the first generation position only?
6 comments

We thank Reviewer mHwv for the feedback and questions. Please find our responses below.

Q1.How does HeadEdit compare with LM-Steer, and with a learned low-rank logit correction trained on the same pairs at matched rank? What does gradient-free SVD estimation buy over learning the same map?

Q1.1 Different Hypotheses and Analysis

LM-Steer is conceptually closer to activation steering. Its central hypothesis can be summarized as follows:

Desired behavior can be induced by transforming hidden states within a low-dimensional activation subspace.

HeadEdit starts from a different hypothesis:

The information needed for the correct behavior may already be present in the final hidden state, while some behavioral errors arise because the unembedding interface reads out certain low-dimensional behavioral components incorrectly.

In Section 4, we examine this hypothesis through behavior probing, paired-subspace controls, and related analyses. Our results show that, for the behavioral calibration tasks studied in this paper, modeling these errors as an execution gap between the final representation and the output decision leads to interventions that are both more effective and easier to interpret.

Q1.2. Empirical Comparisons with LM-Steer

We have evaluated the following two versions of LM-Steer and will add a discussion to the revised manuscript.

Official LM-Steer. We implement the original multiplicative-adapter formulation of LM-Steer. With the language model frozen, the method learns two low-rank factors, P1P_1 and P2P_2, and transforms the hidden state as

h′=h+ϵsP2P1⊤h,h' = h + \epsilon s P_2P_1^\top h,

where ss is the steering strength used at inference time. Following the official configuration, we use rank 1,0001{,}000, include an auxiliary dummy steering channel, optimize the conditional likelihood of all completion tokens, and apply the learned transformation at every decoding step. We select the steering strength on the validation set.

Matched LM-Steer. To compare gradient-based learning with HeadEdit’s paired-SVD estimation under controlled conditions, we also construct a matched version of LM-Steer. It uses the same preferred–undesirable completion pairs, task-specific rank, frozen backbone, post-norm final hidden state, first-generation-position intervention schedule, and validation-based strength-selection protocol as HeadEdit. Its correction is still a general low-rank mapping learned through gradient descent:

Δh=sP2P1⊤h.\Delta h = sP_2P_1^\top h.

We train it using a signed conditional-likelihood objective at the first completion token.

We provide the results on Qwen3-4B below. Official LM-Steer performs worse overall than Matched LM-Steer, suggesting that its original high-rank, all-decoding-step intervention may be less suitable for the localized behavioral calibration problems studied here.

The mapping used by Matched LM-Steer is more general than the HeadEdit transformation. Setting P1=VP_1=V, P2=−VP_2=-V, and s=αs=\alpha gives

sP2P1⊤h=−αVV⊤h.sP_2P_1^\top h = -\alpha VV^\top h.

Thus, the learned low-rank mapping used by Matched LM-Steer can represent HeadEdit exactly. Nevertheless, our experiments show that gradient-based learning does not consistently produce an equally effective behavior-related correction, despite having the capacity to represent it.

HeadEdit directly identifies the shared behavioral contrast through paired SVD and constrains the correction to the symmetric orthogonal projector VV⊤VV^\top. This excludes unconstrained low-rank transformations that may improve token-level likelihood without matching the target behavioral boundary. It also avoids iterative backpropagation, optimizer states, learning-rate selection, and stopping criteria, reducing both estimation cost and optimization choices.

Tool Overuse

MethodConfigurationDecision Acc. ↑Tool Overuse ↓Necessary Keep ↑
Base—57.00%97.80%99.70%
LM-Steer Officialr=1000, s=1r=1000,\ s=144.02%0.51%0.79%
LM-Steer Matchedr=8, s=0.5r=8,\ s=0.560.98%88.07%99.21%
HeadEdit66.10%52.90%80.90%

Over-Refusal

MethodConfigurationAccuracy ↑Safe Answer ↑Unsafe Refusal ↑
Base—60.60%57.30%70.50%
LM-Steer Officialr=1000, s=0r=1000,\ s=069.75%69.83%69.60%
LM-Steer Matchedr=4, s=0.5r=4,\ s=0.569.27%69.11%69.60%
HeadEdit71.60%72.20%70.50%

Factual Sycophancy

MethodConfigurationFactual Acc. ↑Anti-Sycophancy ↑
Base—37.80%37.80%
LM-Steer Officialr=1000, s=0r=1000,\ s=040.24%41.46%
LM-Steer Matchedr=2, s=2r=2,\ s=241.83%42.44%
HeadEdit47.10%47.10%

Q2.How does a simpler output-stage baseline perform, e.g., CAA/RepE directions applied to the post-norm final hidden state at the first generation position only?

We implemented an interface-matched CAA baseline using the same paired completions as HeadEdit. We computed the unit-normalized mean preferred-minus-undesirable direction and added it to the post-normalization final hidden state only at the prompt-final position, which predicts the first generated token. We selected the intervention strength on the development set from {0,0.5,1,2,4,8,12,16,24,32}\{0,0.5,1,2,4,8,12,16,24,32\} and evaluated the selected configuration on the frozen test set.

For tool overuse and factual sycophancy, none of the nonzero intervention strengths improved the development-set selection objective over the unedited model; therefore, matched CAA selected zero intervention. For over-refusal, it selected s=24s=24, but test accuracy decreased by 2.07 percentage points, with both safe-answer and unsafe-refusal rates declining.

For nonzero interventions, we verified nonzero centered logit perturbations and cosine similarity above 0.999 between the actual and expected logit-change directions across all three tasks. This supports that the intervention was active and applied as intended.

These results show that a fixed output-stage mean translation did not improve the evaluated tasks under the matched protocol, supporting HeadEdit’s projection-based calibration design beyond the choice of intervention interface alone.

Thank you for the responses and new controls. The matched comparisons address my two questions on Qwen3-4B:

  • Q2. Matched CAA at the same position selects zero strength on two tasks and hurts the third, and same-rank prompt PCA helps only on tool overuse (62.20 vs. 66.10). Location and timing alone therefore do not explain the gains.
  • Q1. Matched LM-Steer, which as you note can represent HeadEdit exactly, is below HeadEdit on all three tasks.

Taken together, the gains are attributable to the paired-SVD estimator and the projector constraint. Remaining points:

  1. Positioning. The "different hypothesis" framing is an interpretation; the mechanism is LM-Steer's with W = −αVVᵀ, as your own derivation shows. Please revise the Sec. 2 claims that output-stage intervention is "rarely" explored or "unexplored," discuss LM-Steer and the learned logit-space calibration methods raised in my review, and state the contribution as a gradient-free, constrained estimator for this correction that outperforms learning the same map.

  2. Scope of the controls. The matched controls are on Qwen3-4B only. Please report Matched LM-Steer on the other two backbones in the revision, together with its tuning budget (learning rate, steps, selection), so that the learned baseline is not under-tuned.

If the positioning in item 1 is revised accordingly, I would be glad to raise my score.

Reviewer_WBWAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HeadEdit, a training-free method for calibrating LLM behavior. The method identifies subspaces using PCA on hidden-state differences between desirable and undesirable responses. During inference, it subtracts the projection of the final hidden state onto the identified subspace before generating the next token, without modifying model parameters.

Critical strengths
  1. The paper focuses on an important and intriguing research question.
  2. The paper conducts experiments on different LLM tasks, such as tool use and sycophancy.
Critical weaknesses
  1. Unclear Motivation for the Projection-Based Steering. The observation that LLM hidden states encode preference-related information that can be analyzed using PCA is not new, as explored in prior work [1]. More importantly, unlike conventional steering methods that explicitly move hidden states toward a preferred direction, HeadEdit subtracts their projections onto a behavioral subspace. However, it is unclear why this operation can encourage desirable behavior. For example, if the desirable and undesirable responses were swapped, the resulting PCA subspace would remain unchanged, leading to exactly the same intervention. Why should removing these components necessarily steer the model toward the preferred behavior rather than potentially suppressing it? In other words, why not directly steer the hidden states toward the preferred direction instead of subtracting their projections onto the behavioral subspace?

[1] On Prompt-Driven Safeguarding for Large Language Models. ICML 2024.

  1. Ambiguity of using unembedding editing.I find the term "Unembedding Editing" somewhat confusing. As I understand it, HeadEdit simply performs low-rank projection-based steering on the final hidden state before passing it through the original LM head for decoding. Since the unembedding matrix itself remains unchanged, I am not convinced that this constitutes a new editing mechanism. The authors should clarify how this differs from existing activation steering methods.

  2. Limited Generalization to Advanced LLMs. The experiments mainly focus on relatively small models, including Qwen3-4B, Gemma-3-4B, and Llama-3.2-3B. I encourage the authors to evaluate HeadEdit on larger, more advanced models, such as Qwen3.5, to further demonstrate its effectiveness and generalization across stronger LLMs.

Questions for authors

Please refer to the weakness section.

5 comments

We thank Reviewer WBWA for the feedback and questions. Please find our responses below.

Q1.1. Unclear Motivation for the Projection-Based Steering.

Thank you for the comment. Our motivation is not simply that hidden states contain behavioral information that can be analyzed using PCA. We propose and examine a hypothesis about a possible mismatch between existing representations and the final readout. As discussed in the Introduction, glitch tokens illustrate that some vocabulary tokens are insufficiently learned and cannot be reliably read out. This motivates the opposite question: could tool overuse, over-refusal, and factual sycophancy arise from the overexpression of behavior-related vocabulary coordinates in the final readout? If so, calibrating their contributions could correct behavior without learning a new preference mapping.

Q1.2. More importantly, unlike conventional steering methods that explicitly move hidden states toward a preferred direction, HeadEdit subtracts their projections onto a behavioral subspace. However, it is unclear why this operation can encourage desirable behavior. For example, if the desirable and undesirable responses were swapped, the resulting PCA subspace would remain unchanged, leading to exactly the same intervention. Why should removing these components necessarily steer the model toward the preferred behavior rather than potentially suppressing it?

As discussed in Sec. 3.3, paired-completion differences identify a behavioral subspace rather than a fixed “undesirable direction.” HeadEdit applies

=h_\perp+(1-\alpha)h_\parallel.$$ This operation does not always remove the projection: - $\alpha=0$: no intervention; - $0<\alpha<1$: attenuate the subspace component; - $\alpha=1$: remove it; - $1<\alpha<2$: reverse it with reduced magnitude; - $\alpha=2$: reflect it with unchanged magnitude; - $\alpha>2$: reverse and amplify it. HeadEdit therefore calibrates the contribution already present in the current state, with the correction depending on $V^\top h$. The subspace determines where to intervene, while the normative development objective selects the gain, including whether to reverse the component. The reviewer points out that swapping desirable and undesirable completions leaves the projector unchanged. This is mathematically correct and consistent with our design: swapping labels does not change the behavioral subspace to be calibrated. However, please note that the development-set labels also change in this case, so the calibration objective changes as well. **Q2.I find the term "Unembedding Editing" somewhat confusing. As I understand it, HeadEdit simply performs low-rank projection-based steering on the final hidden state before passing it through the original LM head for decoding. Since the unembedding matrix itself remains unchanged, I am not convinced that this constitutes a new editing mechanism. The authors should clarify how this differs from existing activation steering methods.** We use “unembedding editing” to emphasize our central hypothesis: behavioral errors such as tool overuse, over-refusal, and factual sycophancy may arise from the overexpression of behavior-related components in the original unembedding readout, and calibrating their contributions may correct these errors. In contrast, the core hypothesis of activation steering is that preferred directions exist in the model’s hidden-state space, and moving hidden states along these directions can encourage desirable behavior. In Section 4, we conduct experiments around these two hypotheses. Table 1 and Figure 6 show that the compared activation steering methods do not achieve consistent improvements across these tasks, often struggle to preserve correct decisions on both sides of the behavioral boundary, and degrade performance on unrelated tasks. Section 4.3 further examines behavioral information that is decodable but not appropriately expressed in the output, providing additional support for our hypothesis. Section 4.5 and Appendix D.2 also demonstrate two advantages of HeadEdit over activation steering baselines: 1. Implicit adaptivity: Although $V$ and $\alpha$ are fixed, the input-dependent coordinates $V^\top h$ determine the direction and magnitude of each component’s correction, without an additional learned gate or controller. 2. Practicality and low overhead: HeadEdit operates at the fixed final output interface without intermediate-layer selection. It caches $W_UV$ and computes the low-rank correction only at the first generation position, requiring neither auxiliary models nor online gradient optimization. Its throughput remains close to that of the unedited model. We wish to clarify that our contribution is to propose and empirically investigate a behavioral readout-calibration hypothesis and develop a lightweight, effective calibration method based on it, rather than merely algebraically reformulating existing activation steering.

Q3.Limited Generalization to Advanced LLMs.

Here we provide the results on Qwen3.5-9B.

Qwen3.5-9B

Tool overuse

MethodConfigurationAccuracy ↑Tool overuse rate ↓Necessary keep ↑
Base—57.9195.4598.13
HeadEditr=4, α=2r=4,\ \alpha=267.0748.1978.57

Over-refusal

MethodConfigurationAccuracy ↑Safe answer ↑Unsafe refusal ↑
Base—37.2111.3588.35
HeadEditr=2, α=8r=2,\ \alpha=842.3720.4085.80

Factual sycophancy

MethodConfigurationFactual accuracy ↑Anti-sycophancy ↑
Base—39.5540.11
HeadEditr=8, α=4r=8,\ \alpha=460.4560.45

All metrics are percentages.

Dear Reviewer WBWA,

As the author–reviewer discussion period is drawing to a close, we wanted to check whether our response has addressed your concerns, and whether there is anything further we can clarify or provide. We welcome any comments and suggestions you may have at any time, and would be glad to discuss with you how to further improve our paper.

Thank you again for your time and thoughtful review.

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

HeadEdit learns a low-rank subspace from pairs of desirable and undesirable responses. It adjusts the part of the hidden state in this subspace before mapping it to token scores, without retraining the model. The experiments apply the edit when generating the first token. Results on three behavioral tasks and three model families improve over the original models, with little change in inference speed. The paper also studies what the subspace captures, effects on other abilities, and reuse after DPO fine-tuning.

Critical strengths
  1. The method is simple and adds little inference cost. Accuracy improves in all nine model–task combinations, and HeadEdit is best among the tested gradient-free methods in seven (Table 1). With caching, inference speed stays close to the original model (Table 5).
  2. The paper includes useful controls and analysis. The random-subspace comparison matches the size of changes to token scores, making it a stronger control than an arbitrary random edit. The probes test what information is available before generation. The paper also reports both gains and failures when reusing the edit after DPO.
Critical weaknesses
  1. The edit sometimes harms behavior that was already correct. Retention of necessary tool calls drops by 18.8 percentage points on Qwen and 11.7 on Llama (Table 6). Tool-call accuracy also does not show whether the model completes the task. Qwen's factual-sycophancy edit lowers MATH-500 accuracy by 5.2 points, and adding HeadEdit to Qwen DPO raises unsafe compliance from 25.0% to 38.6% (Figure 6; Table 2). These trade-offs need more attention when claiming that the edit preserves other behavior.
  2. The comparison with related projection methods could be closer. Equations 4–7 are equivalent to editing the hidden state just before the output layer. Related projection methods already exist[1]. Appendix C.4 tests CAA at the final block, but both the operation and its position relative to normalization differ. Testing simpler edits at exactly the same position would help explain where HeadEdit's gains come from.

[1] Arditi et al., 2024 https://arxiv.org/abs/2406.11717

Questions for authors
  1. How much can HeadEdit reduce unnecessary tool calls while preserving necessary ones? Does it improve task completion after the tools are actually run?
  2. Does HeadEdit beat a mean-direction baseline and a projection baseline of the same rank, when all edits are applied after normalization at the first-token position with the same tuning budget?
  3. After DPO, do the gains in six of nine settings remain if the edit is tuned to avoid increasing unsafe compliance?
3 comments

We thank Reviewer GHdY for the thoughtful review. We address your questions below.

Q1.How much can HeadEdit reduce unnecessary tool calls while preserving necessary ones? Does it improve task completion after the tools are actually run?

Table 1 and Table 6 report the reduction in unnecessary calls and the corresponding necessary tool-call rates across all three models. For example, on Qwen3-4B, tool overuse decreases from 97.8% to 52.9%, while necessary tool use decreases from 99.7% to 80.9%; overall decision accuracy improves from 57.0% to 66.1%. We select configurations by development-set decision accuracy, so these results reflect a trade-off, not a guarantee of no loss in necessary tool use.

The intervention effect and the opposite-side loss should be considered together. For example, SADI-Head greatly reduces tool overuse on Llama-3.2-3B, but also sharply reduces necessary tool use, leading to lower overall accuracy. When effective intervention harms the opposite side, accuracy-based selection may favor weak or even zero intervention, as illustrated by CAA and RePE on Qwen3-4B in Table 1. The general-capability results in Figure 6 also show that HeadEdit preserves the original model’s performance more consistently than the other steering baselines, although it is not lossless.

For Tool Overuse task, we define the cases using the same model under two conditions:

  1. Tool overuse: The model answers correctly without tools but calls a tool when tools are available.
  2. Necessary tool use: The model answers incorrectly without tools and calls a tool when tools are available.

On the tool-overuse side, we count a correction as successful only when HeadEdit both avoids the tool call and answers correctly. On the necessary-tool side, we measure whether the model still calls a tool.

This evaluation tests whether HeadEdit can remove unnecessary calls while preserving correct answers and necessary tool use. Final task success also depends on tool selection, arguments, execution, and the subsequent response. We therefore did not measure whether task completion rates improved in necessary-tool-use scenarios.

Q2.Does HeadEdit beat a mean-direction baseline and a projection baseline of the same rank, when all edits are applied after normalization at the first-token position with the same tuning budget?

We implemented an interface-matched CAA baseline using the same paired completions as HeadEdit on Qwen3-4B model. We computed the unit-normalized mean preferred-minus-undesirable direction and added it to the post-normalization final hidden state only at the prompt-final position, which predicts the first generated token. We selected the intervention strength on the development set from {0,0.5,1,2,4,8,12,16,24,32}\{0,0.5,1,2,4,8,12,16,24,32\} and evaluated the selected configuration on the frozen test set.

For nonzero interventions, we verified nonzero centered logit perturbations and cosine similarity above 0.999 between the actual and expected logit-change directions across all three tasks. This supports that the intervention was active and applied as intended.

For tool overuse and factual sycophancy, none of the nonzero intervention strengths improved the development-set selection objective over the unedited model; therefore, matched CAA selected zero intervention. For over-refusal, it selected s=24s=24, but test accuracy decreased by 2.07 percentage points, with both safe-answer and unsafe-refusal rates declining.

TaskBaseline Acc.Matched CAA Acc.Selected Strength
Tool overuse57.09%57.09%γ=0\gamma=0
Over-refusal60.55%58.48%γ=24\gamma=24
Factual sycophancy37.80%37.80%γ=0\gamma=0

On Qwen3-4B, we also constructed a same-rank prompt-PCA projection baseline. We extracted post-normalization final hidden states from the unedited model on the training prompts, before generating the first token. We mean-centered these states and used PCA to obtain an orthonormal subspace QQ with the same rank as HeadEdit. At inference time, we applied the same projection edit as HeadEdit at the first generation position: h′=h−αQQ⊤h.h'=h-\alpha QQ^\top h. After scanning configurations on the development set, both over-refusal and factual sycophancy selected α=0\alpha=0, corresponding to no intervention. Only tool overuse showed an effect, with a consistent tendency to suppress tool calls. We report the results below.

MethodTool Overuse ↓Necessary Tool Retention ↑Decision Acc. ↑
Unedited97.80%99.70%57.00%
Prompt PCA42.64%65.97%62.20%
HeadEdit52.90%80.90%66.10%

Q3.After DPO, do the gains in six of nine settings remain if the edit is tuned to avoid increasing unsafe compliance?

Yes. We repeated DPO+HeadEdit model selection while imposing unsafe compliance as a hard constraint. For each model–task setting, we considered only configurations whose unsafe-compliance rate on an independent development set did not exceed that of DPO alone, and selected the configuration with the best target-task performance among them. Under this constraint, all six previously observed task gains over DPO remained. We believe that this safety-constrained selection protocol further improves the practical applicability of HeadEdit by preventing task gains from being achieved through increased unsafe compliance. We will include this analysis in the revised manuscript.

Dear Reviewer GHdY,

As the author–reviewer discussion period is drawing to a close, we wanted to check whether our response has addressed your concerns, and whether there is anything further we can clarify or provide. We welcome any comments and suggestions you may have at any time, and would be glad to discuss with you how to further improve our paper.

Thank you again for your time and thoughtful review.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for highlighting the value of our controlled analyses, including matched random-subspace comparisons and pre-generation probes, which provide evidence for the “decodable but not expressed” framing (GHdY, mHwv). The reviewers also recognized our evaluation across different models and behavioral tasks, including checks for performance degradation on general capability benchmarks (GHdY, WBWA). We also appreciate their recognition of HeadEdit’s simplicity, low inference cost, and near-base throughput (GHdY, mHwv).

We next address the reviewers’ two main concerns.

1. Methodological novelty and contribution (mHwv, WBWA). We clarified the difference in the core hypotheses of HeadEdit and activation steering. Activation steering assumes that a desirable direction exists in hidden-state space and that moving hidden states along this direction can encourage desirable behavior. HeadEdit investigates whether errors such as tool overuse, over-refusal, and factual sycophancy arise from the overexpression of behavior-related components in the original unembedding readout. Section 4 examines this hypothesis through behavioral probing, controlled subspace interventions, and vocabulary-level analysis, and demonstrates the method’s effectiveness across multiple models and tasks. “Unembedding editing” emphasizes this readout-calibration perspective: we extract a behavioral subspace and calibrate its contribution through the frozen unembedding, without identifying a fixed desirable direction.

2. Fair comparison (GHdY, mHwv). We added matched CAA, same-rank prompt-PCA projection, and learned low-rank controls, all applied after final normalization at the first generation position. Matched CAA provided no improvement. Prompt PCA selected zero intervention strength on two tasks. On the remaining task, the development-selected edit reduced test accuracy relative to the unedited model. These controls support the importance of the paired behavioral subspace beyond the intervention location, timing, and projection operation.

We also compared Official and Matched LM-Steer. The matched version achieved better overall performance than the official version, showing that steering settings designed to induce desirable behavior do not directly guarantee two-sided behavioral calibration. The reviewer noted that Matched LM-Steer can be equivalent to HeadEdit in a special case. Nevertheless, despite having the capacity to represent this correction, gradient-based learning did not consistently produce an equally effective behavioral correction in our experiments. HeadEdit directly identifies shared behavioral contrasts through paired SVD and constrains the correction to gain calibration based on the symmetric orthogonal projector VV⊤VV^\top. This excludes general low-rank transformations that may improve token-level likelihood without matching the target behavioral boundary. It also avoids iterative backpropagation, optimizer states, learning-rate selection, and stopping criteria, reducing estimation cost and optimization choices.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.