- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper introduces soft model editing, where an edit specifies a distribution over multiple plausible answers rather than a single target response. The proposed method, SoftEdit, learns a row-sparse update to the final vocabulary projection and stores it with a retrieval key, applying the update to sufficiently related queries. Its objective combines first-token distribution alignment and full-answer continuation, graph-Laplacian regularization of the first-token probability Jacobians using a pairwise co-plausibility graph, and synthetic text-only locality samples to preserve distributional shifts. The evaluation covers three language models and three vision-language models on curated ProtoQA and VQAv2 data, including soft-label adaptations of existing editors, sequential editing, ranked feedback, and component ablations. The paper reports improved alignment, soft generality, and soft locality compared with existing editors, while maintaining competitive model editing performance and high computational efficiency.
Critical strengths
S1. A meaningful editing objective beyond single-answer correctness.
The paper addresses an important limitation of conventional single-answer model editing. Using multi-answer human feedback and its target distribution can distinguish successful correction from excessive concentration on a single response. The additional distinction between generalizing calibration to distributionally similar queries and preserving real distributional shifts makes the problem formulation more informative than simply replacing hard labels with soft labels.
S2. A computationally motivated implementation with substantial comparative evaluation.
Restricting edits to selected rows of the final vocabulary projection layer permits reuse of frozen hidden states during optimization, providing a concrete computational motivation for the design. The evaluation spans six backbones and includes sequential editing, ranked feedback, and component ablations. Importantly, several baselines are adapted to use multi-answer target distributions rather than being compared only in their original single-answer form. These experiments provide promising evidence that the proposed method improves the reported soft generality–locality tradeoff.
Critical weaknesses
W1. The fast answer-level probability readout does not establish calibration of complete stochastic responses.
Section 3.1 and Appendix A.1 define the goal in terms of generating each complete answer with its target probability. However, the evaluated readout is the first-token probability multiplied by a binary continuation gate. This gate checks whether all continuation tokens are argmax predictions under teacher forcing. This differs from the autoregressive sequence probability,
\prod_{\ell=2}^{L_k} P_\theta(t_{k,\ell}\mid x,t_{k,<\ell}).$$ An argmax continuation need not have probability close to one. For example, a first-token probability of $0.5$ followed by an argmax continuation with probability $0.6$ gives a sequence probability of $0.3$, assuming certain termination. However, the gated readout assigns $0.5$. Conversely, a failed gate sets the estimated probability to zero even when the answer has nonzero generation probability. Although the continuation loss encourages correct answer completion, the paper does not verify whether the gated readout accurately approximates the actual probability of generating the full answer, particularly for unseen queries and across different editing methods. Furthermore, Appendix A.1 merges answers sharing the same first token, so errors in their relative probabilities are not measured. Since Align, SGen, and SLoc all use this readout, this is a central limitation of the evidence for full-answer calibration. ## W2. The target distributions for generated image variants are not independently validated. For generated SGen-I examples, Appendix A.2.1 constructs images using target-distribution-weighted combinations of answer-specific VAE latent embeddings. For SLoc-I, it interpolates toward an answer-specific image and assigns the target distribution $p_{k,t}=(1-t)p_K+t e_k$. However, interpolation coefficients in image latent space do not establish the corresponding human answer probabilities. Even when the endpoints are recognizable, human judgments about intermediate images need not vary linearly with the interpolation coefficient. Appendix A.3 validates the intended answer and absence of watermark text in answer-specific images before interpolation. It does not validate the distributions assigned to intermediate images. Consequently, these results show some agreement with a constructed target distribution, but do not establish preservation of genuine human distributional shifts. This concern specifically affects the generated/interpolated image evaluations and should not be interpreted as invalidating the original-input or text-based results. ## W3. The ablations do not sufficiently isolate the contribution of graph-Laplacian regularization. The graph-Laplacian regularization is a reasonable local smoothness mechanism, but the evidence does not isolate what is gained from its particular pairwise co-plausibility graph. Figure 4 compares $L_{\mathrm{align}}$, $L_{\mathrm{align}}+L_{\mathrm{gen}}$, and the complete objective, but does not report $L_{\mathrm{align}}+L_{\mathrm{loc}}$. Thus, it shows that the generality term helps without the locality term, but does not establish its incremental benefit once the synthetic text-only locality samples are included. There is also no reported comparison against a simpler Jacobian penalty or uniform graph weights under the same editing and retrieval setup. This matters because $\sum_{i,j} p_i p_j \mid\mid J_i-J_j \mid\mid^2$ is a probability-weighted gradient-difference penalty. These existing ablations therefore support regularization as a useful ingredient, but are less conclusive about the paper's mechanism. ## W4. The contribution of gated retrieval and the paraphrase judge remains unresolved. Appendix B.2 introduces an automatically calibrated retrieval rejection threshold and a Qwen2.5-1.5B-Instruct paraphrase judge, both of which determine whether an edit is applied. These components may influence generality and locality independently of the learned weight perturbation, but the paper does not isolate their contributions to the improvements over existing editors through ablations or retrieval-conditioned evaluation. Furthermore, Figure 3 reports seconds per edit, but does not separately quantify the additional per-query inference latency of the paraphrase judge. Appendix D acknowledges retrieval and storage overhead, but does not quantify these costs. Clarifying these effects would help assess both the source of SoftEdit's improvements and its practical efficiency.Questions for authors
Q1. Clarification and validation of the probability being calibrated.
What stochastic decoding policy is the method intended to calibrate? On a representative subset, could the authors compare the gated readout with complete-answer probabilities (including termination) or empirical answer frequencies under the stated decoder? Also, how frequently do answers share the same first token, and does evaluating them separately change the main conclusions?
Q2. Validation of the image-based target distributions
What evidence supports the target distributions assigned to blended and interpolated images, ? Could the authors validate these distributions through independent human annotations on a representative subset, beyond endpoint recognizability? Separating SGen-I results for photometric perturbations and generated image variants would also help clarify the effect of synthetic target distributions.
Q3. Controlled validation of graph-Laplacian regularization
With the editing and retrieval setup held fixed, how does the full method compare with without the generality term? Does the proposed weighting outperform a simpler Jacobian penalty or uniform graph weights under comparable tuning? Such comparisons would help establish the contribution of the pairwise co-plausibility graph beyond generic smoothing.
Q4. Clarification of retrieval's contribution and cost
How much do the automatically calibrated retrieval rejection threshold and paraphrase judge contribute to the reported SGen, SLoc, and Loc results? Could the authors provide an ablation or retrieval-conditioned analysis to separate retrieval effects from those of the learned weight perturbation? Additionally, what is the per-query inference latency introduced by the paraphrase judge?
8 comments
Q2. Human evaluation of target distributions on SGen-I/SLoc-I generated images.
Thank you and reviewer q5e1 for this suggestion. In our one-shot calibration setting, independently collecting multi-answer distributions from multiple human annotators for every generated image would be highly time- and resource-intensive. Thus, we address this concern through the following clarifications and improvements.
-
We first clarify that image interpolation was originally only used in parts of two VLM evaluation metrics (SGen-I and SLoc-I). Our results on original images, photometric perturbations, and text variations in VLM editing, as well as all evaluations in LLM editing, independently demonstrate SoftEdit's superior one-shot calibration performance without relying on interpolated images.
-
SLoc-I: We originally included interpolated images with approximate target distributions to make soft locality evaluation more challenging. However, answer-specific images are sufficient to evaluate whether editors preserve distributional shifts. We therefore remove intermediate interpolation and evaluate only answer-specific images with one-hot target distributions, following prior VLM editing works (Cheng et al., 2023; Zhang et al., 2025; Guo et al., 2025; Qiu et al., 2026) but with iterative validation using GPT-4o (Appendix A.3). Under the updated SLoc-I evaluation, SoftEdit consistently outperforms all baselines across three VLMs in preserving distributional shifts.
-
SGen-I: We include generated ambiguous images to evaluate more challenging generalizations beyond photometric perturbations. To reduce reliance on a single generated image, we evaluate multiple interpolation strengths and report mean [min, max] separately from SGen-I on photometric perturbations below (and in Appendix D.5 Table 11). Notably, SoftEdit demonstrates superior SGen-I than other adapted editors on both photometric perturbations and the more challenging variants.
| Editor | LLaVA SGen-I (Perturb) | LLaVA SGen-I (Variant) | InstructBLIP SGen-I (Perturb) | InstructBLIP SGen-I (Variant) | Qwen3 SGen-I (Perturb) | Qwen3 SGen-I (Variant) |
|---|---|---|---|---|---|---|
| FT-Soft | 0.653 | 0.523 [0.520, 0.527] | 0.624 | 0.550 [0.545, 0.554] | 0.626 | 0.482 [0.478, 0.486] |
| MEND-Soft | 0.131 | 0.173 [0.172, 0.173] | 0.131 | 0.185 [0.185, 0.185] | 0.062 | 0.079 [0.078, 0.079] |
| GRACE-Soft | 0.664 | 0.333 [0.329, 0.337] | 0.571 | 0.397 [0.392, 0.402] | 0.277 | 0.296 [0.292, 0.300] |
| BalancEdit-Soft | 0.226 | 0.296 [0.294, 0.299] | 0.211 | 0.304 [0.301, 0.307] | 0.206 | 0.308 [0.304, 0.312] |
| ReasonEdit-Soft | 0.516 | 0.491 [0.490, 0.492] | 0.519 | 0.535 [0.533, 0.536] | 0.566 | 0.533 [0.527, 0.538] |
| SoftEdit | 0.931 | 0.624 [0.619, 0.629] | 0.881 | 0.756 [0.748, 0.764] | 0.793 | 0.558 [0.551, 0.565] |
We have also explicitly acknowledged in the Limitations section the potential evaluation variance introduced by approximated target distributions for SGen-I images. Collecting multi-answer responses from multiple independent human annotators for every generated image would be highly time- and resource-intensive, but could provide more reliable ground-truth distributions.
Q3 - Q4. Additional ablations.
We have added the following ablations based on your suggestions in Appendix D.4, extending the primary ablations studies in the main text.
- Tuned retrieval threshold vs. fixed.
The retrieval threshold balances the conventional semantic generality–locality tradeoff, similar to prior retrieval-based model editors. Per your suggestion, we have added a comparison of our tuned threshold with fixed percentiles (1st, 10th, and 50th) below (and in Appendix D.4 Table 7). We find that our auto-tuned threshold improves generality, soft generality, and soft locality while maintaining high locality on unseen samples.
| Model | Threshold | Align | SGen | SLoc | Acc | Gen | Loc |
|---|---|---|---|---|---|---|---|
| LLaVA-1.5-7B | 1st | 0.97 | 0.73 | 0.52 | 0.99 | 0.74 | 1.00 |
| 10th | 0.97 | 0.77 | 0.56 | 0.99 | 0.90 | 0.90 | |
| 50th | 0.97 | 0.85 | 0.61 | 0.99 | 0.95 | 0.67 | |
| Tuned | 0.97 | 0.85 | 0.61 | 0.99 | 0.95 | 0.90 | |
| InstructBLIP-7B | 1st | 0.97 | 0.79 | 0.60 | 1.00 | 0.87 | 0.85 |
| 10th | 0.97 | 0.79 | 0.62 | 1.00 | 0.95 | 0.81 | |
| 50th | 0.97 | 0.81 | 0.60 | 1.00 | 0.97 | 0.60 | |
| Tuned | 0.98 | 0.85 | 0.64 | 1.00 | 0.99 | 0.78 | |
| Qwen3-VL-4B | 1st | 0.89 | 0.68 | 0.62 | 0.94 | 0.67 | 1.00 |
| 10th | 0.89 | 0.73 | 0.64 | 0.94 | 0.88 | 0.95 | |
| 50th | 0.89 | 0.74 | 0.65 | 0.94 | 0.90 | 0.79 | |
| Tuned | 0.89 | 0.75 | 0.65 | 0.94 | 0.90 | 0.95 | |
| Llama-3.1-8B | 1st | 0.95 | 0.14 | 0.07 | 0.98 | 0.17 | 0.81 |
| 10th | 0.94 | 0.54 | 0.30 | 0.98 | 0.69 | 0.75 | |
| 50th | 0.94 | 0.70 | 0.44 | 0.98 | 0.91 | 0.70 | |
| Tuned | 0.95 | 0.72 | 0.47 | 0.98 | 0.94 | 0.76 | |
| Mistral-7B | 1st | 0.93 | 0.06 | 0.08 | 0.98 | 0.10 | 0.80 |
| 10th | 0.93 | 0.53 | 0.42 | 0.98 | 0.71 | 0.77 | |
| 50th | 0.93 | 0.68 | 0.51 | 0.98 | 0.92 | 0.61 | |
| Tuned | 0.93 | 0.70 | 0.52 | 0.98 | 0.95 | 0.69 | |
| Qwen2.5-7B | 1st | 0.93 | 0.04 | 0.24 | 0.98 | 0.08 | 0.80 |
| 10th | 0.93 | 0.62 | 0.40 | 0.98 | 0.81 | 0.79 | |
| 50th | 0.93 | 0.71 | 0.51 | 0.98 | 0.92 | 0.71 | |
| Tuned | 0.93 | 0.73 | 0.52 | 0.98 | 0.94 | 0.77 |
- Weighted graph-Laplacian regularization.
We have added an ablation comparing our probability-weighted graph-Laplacian regularization with unweighted and Frobenius-norm alternatives below (and in Appendix D.4, Table 9). We find that the weighted formulation achieves the best overall performance across calibration and editing metrics, while the alternatives remain competitive but exhibit lower soft generality and soft locality, with variations across models.
| Model | Variant | Align | SGen | SLoc | Acc | Gen | Loc |
|---|---|---|---|---|---|---|---|
| LLaVA-1.5-7B | Frobenius | 0.97 | 0.84 | 0.60 | 0.99 | 0.95 | 0.90 |
| Unweighted | 0.96 | 0.83 | 0.60 | 0.98 | 0.94 | 0.91 | |
| Weighted (SoftEdit) | 0.97 | 0.85 | 0.61 | 0.99 | 0.95 | 0.92 | |
| InstructBLIP-7B | Frobenius | 0.95 | 0.83 | 0.62 | 0.96 | 0.98 | 0.79 |
| Unweighted | 0.96 | 0.85 | 0.63 | 0.99 | 0.99 | 0.78 | |
| Weighted | 0.98 | 0.85 | 0.64 | 1.00 | 0.99 | 0.78 | |
| Qwen3-VL-4B | Frobenius | 0.87 | 0.75 | 0.65 | 0.93 | 0.92 | 0.93 |
| Unweighted | 0.80 | 0.72 | 0.60 | 0.88 | 0.87 | 0.93 | |
| Weighted | 0.89 | 0.75 | 0.65 | 0.94 | 0.90 | 0.95 | |
| Llama-3.1-8B | Frobenius | 0.93 | 0.70 | 0.40 | 0.97 | 0.92 | 0.74 |
| Unweighted | 0.92 | 0.69 | 0.40 | 0.98 | 0.92 | 0.74 | |
| Weighted | 0.95 | 0.72 | 0.47 | 0.98 | 0.94 | 0.76 | |
| Mistral-7B | Frobenius | 0.90 | 0.72 | 0.48 | 0.96 | 0.91 | 0.69 |
| Unweighted | 0.90 | 0.70 | 0.41 | 0.96 | 0.91 | 0.68 | |
| Weighted | 0.93 | 0.70 | 0.52 | 0.98 | 0.95 | 0.69 | |
| Qwen2.5-7B | Frobenius | 0.91 | 0.73 | 0.48 | 0.95 | 0.93 | 0.77 |
| Unweighted | 0.91 | 0.73 | 0.47 | 0.95 | 0.93 | 0.77 | |
| Weighted | 0.93 | 0.73 | 0.52 | 0.98 | 0.94 | 0.77 |
Thank you for finding our new problem definition meaningful, well-motivated, and finding our method substantially outperforming other baseline editors in broad evaluation and experiments. Your detailed feedback has helped us improve our paper in many ways.
Q1.1 Clarification of probability being calibrated.
Thank you for this great question! We clarify that SoftEdit already encourages calibration of autoregressive full-sequence token-product probabilities: it aligns first-token probabilities with the human target distribution while maximizing the likelihood of each answer's remaining tokens, encouraging their conditional probabilities toward one. This factorization directly optimizes the full-sequence probability of each answer to approach its target probability.
Q1.2 Validate fast deterministic gated readout probabilities.
To validate the deterministic gated readout as a fast approximation of full-sequence probabilities, both through direct full-sequence probability computation (when users specify complete answers) and stochastic generation (when users use Monte Carlo sampling), we have added the following experiments:
We evaluate all editors using exact autoregressive full-sequence probabilities, including on unseen generality and locality samples. SoftEdit's results closely match its fast readout results (median absolute score difference: 0.01, range: 0.00–0.04) and consistently outperform other editors across all six models.
Using 100 stochastic generations per edit across 1,000 edits per model, we find close agreement between gated readout and empirical generation frequencies (ECE: 0.002–0.033), with differences in target-distribution alignment of only 0.002–0.024.
We also added the following limitation: “As the first study of one-shot calibration via model editing, we adopt first-token-based probability estimation following prior work (Lee et al., 2023; Ma et al., 2025). Shared-first-token answers occur in only 4.65–5.06% of ProtoQA samples and 0.62–0.63% of VQAv2 samples. To better distinguish answers sharing the same first token, future work could explore more flexible token-level probabilistic factorizations.”
Thanks to you and all reviewers, we have dedicated Appendix B to detailing three approaches to calibrated answer generation enabled by SoftEdit: (1) fast deterministic readout of multiple answers with their probabilities in B.1, (2) estimating autoregressive full-sequence probabilities given user-specified answers in B.2, and (3) stochastic sampling of complete answers in B.3.