- 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- right singular vectors of the centered difference matrix.
- At inference it adds the logit correction . This is equivalent to using a transient effective unembedding . 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:
- is more behavior-predictive than random subspaces of equal rank;
- a ridge predictor on explains part of the LoRA-DPO logit update;
- the base-model subspace can be reused on top of LoRA-DPO.
Critical strengths
- 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 (, ), 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.
- 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
-
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 , with a scalar controlling strength. HeadEdit's Eq. (6), , is algebraically LM-Steer with . 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.
-
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.
- 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?
- 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
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 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 , 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:
-
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.
-
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.
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:
HeadEdit starts from a different hypothesis:
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, P1 and P2, and transforms the hidden state as
h′=h+ϵsP2P1⊤h,
where s is the steering strength used at inference time. Following the official configuration, we use rank 1,000, 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.
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=V, P2=−V, and s=α gives
sP2P1⊤h=−αVV⊤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⊤. 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
Over-Refusal
Factual Sycophancy