acceptodds
Under review as a conference paper at ICLR 2027

One-Shot Calibration via Model Editing

Abstract

Model editing aims to correct errors in large, pretrained models by replacing model generations with user-provided target responses. In practice, however, user feedback often contains multiple plausible answers with varying uncertainty, forming a natural target distribution to not only correct but also calibrate model generation. Existing editors ignore this uncertainty and require a single target answer for each edit. We therefore propose SoftEdit, the first model editor to leverage uncertainty in user responses and perform one-shot calibration during editing. To generalize calibration to unseen samples, SoftEdit uses co-plausibility relationships between answers while preserving real distributional shifts, addressing the new distributional generality-locality tradeoff. Across three language models, three vision-language models, and two large datasets, SoftEdit achieves state-of-the-art performance, ultimately showing that uncertainty in user responses enables generalizable calibration of pretrained models beyond single-answer corrections.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Model Editing, Post-training Calibration, Large Language Models, Vision-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_q5e1OpenReview ↗
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, pk,t=(1−t)pK+tekp_{k,t}=(1-t)p_K+t e_k? 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 Lalign+LlocL_{\mathrm{align}}+L_{\mathrm{loc}} without the generality term? Does the proposed pipjp_i p_j 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

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:

  1. Full-sequence probabilities (B.2 Table 5 and below).

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.

EditorLLaVA AlignLLaVA SGen-ILLaVA SGen-TLLaVA SLocInstructBLIP AlignInstructBLIP SGen-IInstructBLIP SGen-TInstructBLIP SLocQwen3 AlignQwen3 SGen-IQwen3 SGen-TQwen3 SLoc
FT-Soft0.730.580.610.420.740.580.640.470.780.550.660.41
MEND-Soft0.110.130.110.130.130.160.130.160.060.070.060.06
GRACE-Soft0.900.380.530.190.870.480.590.470.900.280.450.53
BalancEdit-Soft0.220.190.170.270.230.250.200.320.200.250.170.43
ReasonEdit-Soft0.510.500.490.380.500.520.470.400.560.540.540.62
SoftEdit0.970.740.920.610.970.800.870.640.880.670.810.64
EditorLlama AlignLlama SGenLlama SLocMistral AlignMistral SGenMistral SLocQwen2.5 AlignQwen2.5 SGenQwen2.5 SLoc
FT-Soft0.780.530.360.730.440.300.780.500.31
MEND-Soft0.150.120.100.030.030.060.040.040.06
GRACE-Soft0.910.390.070.860.030.000.800.260.06
IKE-Soft0.070.070.100.080.080.070.340.300.35
AlphaEdit-Soft0.220.150.110.380.180.190.280.080.18
SoftEdit0.940.720.470.920.680.510.950.710.48
  1. Stochastic generation (B.3 Table 6 and below).

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.

ModelN allN multi-tokenN shared-prefixAlign (gated)Align (gated multi-token)Align (gated shared-prefix)Align (MC)Align (MC multi-token)Align (MC shared-prefix)
LLaVA-1.5-7B1000541130.9750.9700.8780.9730.9730.930
InstructBLIP-7B1000536110.9780.9730.8690.9750.9740.910
Qwen3-4B1000572120.9720.9670.8570.9640.9600.862
Llama-3.1-8B1000924470.9500.9480.8440.9390.9430.844
Mistral-7B1000847610.9280.9210.8080.9180.9240.820
Qwen2.5-7B1000934510.9370.9350.8190.9130.9170.817

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.

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.

  1. 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.

  2. 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.

  3. 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 t∈{0.0,0.3}t \in \{0.0, 0.3\} 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.

EditorLLaVA SGen-I (Perturb)LLaVA SGen-I (Variant)InstructBLIP SGen-I (Perturb)InstructBLIP SGen-I (Variant)Qwen3 SGen-I (Perturb)Qwen3 SGen-I (Variant)
FT-Soft0.6530.523 [0.520, 0.527]0.6240.550 [0.545, 0.554]0.6260.482 [0.478, 0.486]
MEND-Soft0.1310.173 [0.172, 0.173]0.1310.185 [0.185, 0.185]0.0620.079 [0.078, 0.079]
GRACE-Soft0.6640.333 [0.329, 0.337]0.5710.397 [0.392, 0.402]0.2770.296 [0.292, 0.300]
BalancEdit-Soft0.2260.296 [0.294, 0.299]0.2110.304 [0.301, 0.307]0.2060.308 [0.304, 0.312]
ReasonEdit-Soft0.5160.491 [0.490, 0.492]0.5190.535 [0.533, 0.536]0.5660.533 [0.527, 0.538]
SoftEdit0.9310.624 [0.619, 0.629]0.8810.756 [0.748, 0.764]0.7930.558 [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.

  1. 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.

ModelThresholdAlignSGenSLocAccGenLoc
LLaVA-1.5-7B1st0.970.730.520.990.741.00
10th0.970.770.560.990.900.90
50th0.970.850.610.990.950.67
Tuned0.970.850.610.990.950.90
InstructBLIP-7B1st0.970.790.601.000.870.85
10th0.970.790.621.000.950.81
50th0.970.810.601.000.970.60
Tuned0.980.850.641.000.990.78
Qwen3-VL-4B1st0.890.680.620.940.671.00
10th0.890.730.640.940.880.95
50th0.890.740.650.940.900.79
Tuned0.890.750.650.940.900.95
Llama-3.1-8B1st0.950.140.070.980.170.81
10th0.940.540.300.980.690.75
50th0.940.700.440.980.910.70
Tuned0.950.720.470.980.940.76
Mistral-7B1st0.930.060.080.980.100.80
10th0.930.530.420.980.710.77
50th0.930.680.510.980.920.61
Tuned0.930.700.520.980.950.69
Qwen2.5-7B1st0.930.040.240.980.080.80
10th0.930.620.400.980.810.79
50th0.930.710.510.980.920.71
Tuned0.930.730.520.980.940.77
  1. 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.

ModelVariantAlignSGenSLocAccGenLoc
LLaVA-1.5-7BFrobenius0.970.840.600.990.950.90
Unweighted0.960.830.600.980.940.91
Weighted (SoftEdit)0.970.850.610.990.950.92
InstructBLIP-7BFrobenius0.950.830.620.960.980.79
Unweighted0.960.850.630.990.990.78
Weighted0.980.850.641.000.990.78
Qwen3-VL-4BFrobenius0.870.750.650.930.920.93
Unweighted0.800.720.600.880.870.93
Weighted0.890.750.650.940.900.95
Llama-3.1-8BFrobenius0.930.700.400.970.920.74
Unweighted0.920.690.400.980.920.74
Weighted0.950.720.470.980.940.76
Mistral-7BFrobenius0.900.720.480.960.910.69
Unweighted0.900.700.410.960.910.68
Weighted0.930.700.520.980.950.69
Qwen2.5-7BFrobenius0.910.730.480.950.930.77
Unweighted0.910.730.470.950.930.77
Weighted0.930.730.520.980.940.77
Reviewer_uiqsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SoftEdit, which extends one-shot model editing from single-answer correction to calibration of multiple plausible answers. It retrieves a row-sparse update to the output layer and combines alignment, generalization, and locality losses; experiments on three LLMs and three VLMs report improvements on VQAv2 and ProtoQA, including sequential editing.

Critical strengths
  1. Problem formulation. The distinction between semantic similarity and distributional equivalence is well motivated. Extending model editing to multi-answer human feedback is useful.

  2. Experimental coverage. The paper evaluates six backbones, sequential editing, and several ablations. The answer-conditioned locality construction is simple and interesting.

Critical weaknesses
  1. Validity of calibration evaluation. Align, SGen, and SLoc rely on first-token probabilities with a binary continuation gate, not verified probabilities of complete stochastic generations (Sec. 4.1; App. A.1). In addition, target distributions for latent-interpolated SGen-I/SLoc-I images are assigned without independent human re-annotation (App. A.2–A.3).

  2. Limited methodological novelty. ICE [1] already uses distributional editing targets, APP/PEAK [2] studies neighboring-answer preservation, and CGM [3] studies conditional distributional calibration; these works are not discussed. Figure 4 supports the usefulness of the generalization loss but not the necessity of its particular weighted Jacobian form.

[1] Qi, S., Yang, B., Jiang, K., Wang, X., Li, J., Zhong, Y., Yang, Y., & Zheng, Z. (2025). In-context editing: Learning knowledge from self-induced distributions. International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2025/hash/c0ff9e52e94ae331bc0f2d28be06a9ca-Abstract-Conference.html

[2] Ma, J.-Y., Ling, Z.-H., Zhang, N., & Gu, J.-C. (2024). Neighboring perturbations of knowledge editing on large language models. Proceedings of Machine Learning Research, 235, 33839–33854. https://proceedings.mlr.press/v235/ma24h.html

[3] Smith, H., Diamant, N. L., & Trippe, B. L. (2026). Calibrating generative models to distributional constraints. Proceedings of Machine Learning Research, 306, 113976–114023. https://proceedings.mlr.press/v306/smith26a.html

Questions for authors
  1. Could the authors compare gated-JS results with Monte Carlo full-generation distributions on a subset, including multi-token or shared-prefix answers?

  2. Could the assigned SGen-I/SLoc-I distributions be checked using fresh human annotations of interpolated images?

  3. How is SoftEdit distinct from [1–3], and does the weighted Jacobian regularizer outperform simpler unweighted smoothing?

  4. What is the end-to-end inference cost including the Qwen2.5-1.5B paraphrase judge and growing codebook? Did soft/ranked baselines use the same probability information and comparable tuning budgets?

  5. Which loss was implemented: Lalign+λgenLgen+λlocLlocL_{align}+\lambda_{gen}L_{gen}+\lambda_{loc}L_{loc} (main text) or Lalign+λgen(Lgen+λlocLloc)L_{align}+\lambda_{gen}(L_{gen}+\lambda_{loc}L_{loc}) (App. B.1)?

  6. How robust are the results to less-filtered VQAv2/ProtoQA data, including multi-word answers and more than three answer clusters?

6 comments

Thank you for the very constructive feedback and for recognizing that our new problem definition is well-motivated and useful, as well as the breadth of our experimental evaluation. Our paper has been largely improved based on your suggestions.

Q1. Compare fast gated readout distribution with Monte Carlo full-generation distributions on a subset, including multi-token or shared-prefix answers.

This is a great suggestion! It validates our proposed gated readout approach as a fast approximation of full-generation distribution. We have added a direct comparison of gated readout and Monte Carlo (MC) sampling distributions (100 stochastic full generations per edit) against human target distributions, using 1,000 edits per backbone model. Results are shown below (and in Appendix B.3 Table 6).

ModelN allN multi-tokenN shared-prefixAlign (gated)Align (gated multi-token)Align (gated shared-prefix)Align (MC)Align (MC multi-token)Align (MC shared-prefix)
LLaVA-1.5-7B1000541130.9750.9700.8780.9730.9730.930
InstructBLIP-7B1000536110.9780.9730.8690.9750.9740.910
Qwen3-4B1000572120.9720.9670.8570.9640.9600.862
Llama-3.1-8B1000924470.9500.9480.8440.9390.9430.844
Mistral-7B1000847610.9280.9210.8080.9180.9240.820
Qwen2.5-7B1000934510.9370.9350.8190.9130.9170.817

Gated and MC alignment scores closely agree across all six models, with absolute differences of 0.002–0.024 overall and 0.001–0.018 for multi-token answers. For shared-first-token answers (small subsets of 11-61 edits per backbone), both approaches show lower alignment scores. We thus also acknowledged this 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.

Q2. Human evaluation of target distributions on SGen-I/SLoc-I generated images.

Thank you and reviewer q5e1 for this suggestion. Collecting multi-answer distributions from multiple human annotators for every generated image would be highly time- and resource-intensive. Thus, we address this concern as follows.

(1) 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) and 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.

(2) SGen-I: We include interpolated image variants (Variant) to evaluate more challenging generalizations beyond photometric perturbations (Perturb). To reduce reliance on a single generated image, we evaluate multiple interpolation strengths t∈{0.0,0.3}t \in \{0.0, 0.3\} (~4 images per edit) and report mean [min, max] separately from 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.

EditorLLaVA SGen-I (Perturb)LLaVA SGen-I (Variant)InstructBLIP SGen-I (Perturb)InstructBLIP SGen-I (Variant)Qwen3 SGen-I (Perturb)Qwen3 SGen-I (Variant)
FT-Soft0.6530.523 [0.520, 0.527]0.6240.550 [0.545, 0.554]0.6260.482 [0.478, 0.486]
MEND-Soft0.1310.173 [0.172, 0.173]0.1310.185 [0.185, 0.185]0.0620.079 [0.078, 0.079]
GRACE-Soft0.6640.333 [0.329, 0.337]0.5710.397 [0.392, 0.402]0.2770.296 [0.292, 0.300]
BalancEdit-Soft0.2260.296 [0.294, 0.299]0.2110.304 [0.301, 0.307]0.2060.308 [0.304, 0.312]
ReasonEdit-Soft0.5160.491 [0.490, 0.492]0.5190.535 [0.533, 0.536]0.5660.533 [0.527, 0.538]
SoftEdit0.9310.624 [0.619, 0.629]0.8810.756 [0.748, 0.764]0.7930.558 [0.551, 0.565]

(3) Human validation on SGen-I: Collecting multi-answer distributions from multiple independent annotators (e.g., through crowdsourcing) for each generated image is particularly time- and resource-intensive. Thus, given the difficulty of estimating numerical target distributions without crowdsourcing, we conducted a simplified human validation in which two authors independently evaluated interpolated images from 50 randomly sampled edits. We found that 96% of images support intended answers with rankings consistent with the target distributions, with inter-annotator agreement of 98%. To further address your concern, in the limitation, we have explicitly acknowledged the potential evaluation variance in this part of SGen-I and noted that crowd-source validation could provide more reliable ground-truth distributions given sufficient time and resources for future studies.

(4) Overall clarification: Interpolated images were introduced to provide a more challenging evaluation of SGen-I beyond photometric perturbations. They constitute only part of one VLM editing metric among the 14 metrics in VLM and LLM editing evaluations. Our results on original images, photometric perturbations, and text variations in VLM editing, as well as all evaluations in LLM editing, independently provide extensive experimental evidence supporting SoftEdit's superior performance in the novel one-shot calibration problem.

Q3. Compare methodological novelty with three related works.

Thank you for suggesting these related works! We have discussed them in the Related Work section due to their relevancy. We further clarify how they differ from the novel one-shot calibration task introduced by SoftEdit in the table below.

What distributionWhy distribution is usedAlign with target ranking/distributionSoft generalitySoft locality
ICE (2025)Model-induced contextual distributionImprove single-fact editing and reduce overfittingXXX
APP/PEAK (2024)Model-predicted probabilitiesMaintain probability margins to preserve existing answersXXX
CGM (2026)Aggregate model generative distributionMatch specified generation statistics (e.g., demographic proportions)XXX
SoftEdit (Ours)Per-edit human answer distributionCalibrate multiple plausible answers' distribution/ranking√√√

Q5. Typo.

We have fixed the typo in the main text into Lalign+λgen(Lgen+λlocLloc)L_{align}+\lambda_{gen}(L_{gen}+\lambda_{loc}L_{loc}), consistent with implementation. Thank you for catching this!

Q3-Q6. Additional Ablations.

Thanks to your suggestion, we have added the following ablations.

(1) Weighted graph-Laplacian regularization (Q3).

To evaluate varying designs of LgenL_{gen}, 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.

ModelVariantAlignSGenSLocAccGenLoc
LLaVA-1.5-7BFrobenius0.970.840.600.990.950.90
Unweighted0.960.830.600.980.940.91
Weighted (SoftEdit)0.970.850.610.990.950.92
InstructBLIP-7BFrobenius0.950.830.620.960.980.79
Unweighted0.960.850.630.990.990.78
Weighted (SoftEdit)0.980.850.641.000.990.78
Qwen3-VL-4BFrobenius0.870.750.650.930.920.93
Unweighted0.800.720.600.880.870.93
Weighted (SoftEdit)0.890.750.650.940.900.95
Llama-3.1-8BFrobenius0.930.700.400.970.920.74
Unweighted0.920.690.400.980.920.74
Weighted (SoftEdit)0.950.720.470.980.940.76
Mistral-7BFrobenius0.900.720.480.960.910.69
Unweighted0.900.700.410.960.910.68
Weighted (SoftEdit)0.930.700.520.980.950.69
Qwen2.5-7BFrobenius0.910.730.480.950.930.77
Unweighted0.910.730.470.950.930.77
Weighted (SoftEdit)0.930.730.520.980.940.77

(2) Retrieval cost at inference (Q4).

This is a valuable additional result to report! We have measured retrieval-based inference latency on an A100 GPU (Appendix D.4). The lightweight Qwen2.5-1.5B paraphrase judge introduces an overhead of 16.1 ± 1.0 ms/query. Including key encoding and lookup, total inference time increases only from 31.1 to 31.2 ms/query as the codebook grows from 100 to 4,000 edits, with nearest-neighbor retrieval accounting for only 0.04–0.20 ms/query. Compatible baselines. All soft/ranked baselines use the same human-provided probability/ranking information. Learning-based editors use the same distributional alignment objective as SoftEdit. We preserve the original editing mechanisms of prior editors. In addition, we devoted extra tuning effort to GRACE's retrieval radius using 100 edits for each backbone model to improve its generality on the new dataset. We also used a batch size of 10 instead of 1 for AlphaEdit in sequential editing to improve its performance and ensure competitive and fair comparisons (Appendix D.3).

(3) Robustness on less-filtered data (Q6).

We clarify that, after preprocessing, both datasets used in our experiments include multi-token answers, and their sizes are comparable to or larger than single-answer datasets used in most prior VLM and LLM model editing studies, providing a sufficiently large and consistent benchmark for comparing editing methods. Based on your suggestion, we agree it’s valuable to test robustness under less filtering, so we have added a sensitivity analysis of SoftEdit and adapted soft editors on 1,000 unfiltered edits from each of VQAv2 and ProtoQA (Appendix D.5, Table 12). For VQAv2, we use previously excluded raw data samples; for ProtoQA, we retain all available answer clusters rather than only the top three. SoftEdit consistently achieves the best one-shot calibration performances across all six models, using unfiltered samples in both datasets.

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

In this paper, authors claim that existing models editors assume a single correct answer while in real life there’re scenarios that several plausible answers could exist. Based on this motivation, they propose soft model editing. Given the query and a target distribution over multiple plausible answers, based on annotators vote. They design their model editing method to achieve the following goals: 1). Reproduce the distribution on the query. 2). The plausible answer distribution should be carried over to similar queries with the same data distribution, and 3) leave similar queries with different data distribution unaffected. Specifically, the proposed method SoftEdit is a retrieval-based editor, which learns a small update to the output-layer rows of the answer tokens, via three objectives: a distribution-matching loss, a regularizer that keeps the answer probabilities changing together near the edit, and a locality loss on synthetic prompts that state the answer. Experimental results on three VLMs across various benchmarks demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The evaluation is broad. It covers 6 backbones across 2 modalities, batch and sequential editing, and ranked feedback, with soft-label versions of all baselines.
  2. The proposed editing method is cheap, as it only edits the output layer.
Critical weaknesses
  1. The term “calibration” is a bit different from what it normally stands for. In this paper, it means the model distribution should match the annotator vote distribution, and no standard calibration metric is reported, i.e. ECE, Brier, etc.
  2. Based on the motivation example, the client is expected to be presented of several answers with their likelihood at one output. But in this paper, the method calibrates the single output prob using the sampling frequencies, which only generate one answer. The users are not able to know the likelihood behind the model output, since it only has the access to what model generated.
  3. The proposed retrieval method relies on two components that the baselines lack: an auto-tuned threshold and an LLM paraphrase judge (relegated to App. B.2 and omitted from the main text). Crucially, the threshold is tuned on paraphrases and image augmentations similar to the SGen probes, yet neither component is ablated.
Questions for authors

Some scenario might better fit the introduced method, like survey simulation, which requires the overall output distribution match the real human distribution.

4 comments

Thank you for the very constructive suggestions and for recognizing the efficient methodology and broad evaluation of our work. We have used your feedback to improve our work in many ways!

W1. Clarify how our work relates to conventional calibration and conventional calibration metrics (e.g., ECE).

In our work, we use "calibration" in its general sense of aligning predicted probabilities with empirical outcome frequencies. The key distinction between conventional calibration and our one-shot local calibration lies in their granularity: conventional calibration evaluates agreement between predicted confidence and empirical correctness frequencies across samples, whereas our method aligns predicted answer probabilities with empirical human response uncertainty for an individual sample. Complementing our problem definition, we have elaborated on this connection to conventional calibration in Section 4.1 and Appendix B.4.

Our main evaluation uses Jensen–Shannon (JS) distance to directly measure the discrepancy between model-predicted and human-provided answer distributions for individual samples. This directly evaluates our objective and follows prior work on modeling human response uncertainty (Lee et al., 2023; Wu et al., 2026; Inoshita et al., 2026). But based on your suggestion, we further evaluate whether locally calibrated models produce answer probabilities consistent with empirical human response frequencies across samples. We report aggregate calibration diagnostics over 1,000 edits per model, including calibration diagrams and expected calibration error (ECE) (Fig. 7 and table below). Across all six models, ECE is low (0.007–0.025), demonstrating close aggregate agreement with human response frequencies across samples, consistent with the local distribution-alignment objective.

ModelECE ↓
LLaVA-1.5-7B0.008
InstructBLIP-7B0.007
Qwen3-VL-4B0.008
Llama-3.1-8B0.017
Mistral-7B0.025
Qwen2.5-7B0.021

W2. Can SoftEdit edit a model to generate multiple answers with their probabilities?

This is a great question that helps us better describe SoftEdit’s utility! SoftEdit directly and efficiently supports generating multiple (K) complete answers with their estimated probabilities, rather than only a single answer. Specifically, by obtaining the top-K first-token probabilities (calibrated toward the corresponding answer probabilities) and greedily completing each answer, users can efficiently access multiple complete answers with their estimated probabilities in a single formatted output using simple templates. We have detailed this approach in Appendix B.1 of the revised paper.

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.

Q. Survey simulation as another motivating scenario.

Thank you for this well-aligned motivating example. We have strengthened the motivation in the Introduction by adding concrete application scenarios as follows: “Correct and calibrated generations are important for faithful simulations, such as LLM-based survey simulations (Hu et al., 2026; Cao et al., 2025; Grief-Albert et al., 2026), and reliable decision support in high-stakes domains, such as medical language model question answering and differential diagnoses (Singhal et al., 2025; McDuff et al., 2025).”

W3. Additional ablations: (1) Retrieval threshold. (2) LLM paraphrase judge.

Thank you for suggesting these additional ablations! We have first detailed both components in the main text's method section, taking advantage of the additional space, and added ablation experiments to evaluate their contributions in Appendix D.4, extending the primary ablation studies in the main text.

(1) 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.

ModelThresholdAlignSGenSLocAccGenLoc
LLaVA-1.5-7B1st0.970.730.520.990.741.00
10th0.970.770.560.990.900.90
50th0.970.850.610.990.950.67
Tuned0.970.850.610.990.950.90
InstructBLIP-7B1st0.970.790.601.000.870.85
10th0.970.790.621.000.950.81
50th0.970.810.601.000.970.60
Tuned0.980.850.641.000.990.78
Qwen3-VL-4B1st0.890.680.620.940.671.00
10th0.890.730.640.940.880.95
50th0.890.740.650.940.900.79
Tuned0.890.750.650.940.900.95
Llama-3.1-8B1st0.950.140.070.980.170.81
10th0.940.540.300.980.690.75
50th0.940.700.440.980.910.70
Tuned0.950.720.470.980.940.76
Mistral-7B1st0.930.060.080.980.100.80
10th0.930.530.420.980.710.77
50th0.930.680.510.980.920.61
Tuned0.930.700.520.980.950.69
Qwen2.5-7B1st0.930.040.240.980.080.80
10th0.930.620.400.980.810.79
50th0.930.710.510.980.920.71
Tuned0.930.730.520.980.940.77

We also clarify that the retrieval percentile is tuned using only the first 100 edits with similar augmentation procedures, then held fixed for evaluation on the vast majority of subsequent, unseen edits. Across 4,157 - 4,625 VLM edits and 7,207 - 8,326 LLM edits per backbone, 97.6% - 97.8% of VLM edits and 98.6% - 98.8% of LLM edits were unseen during threshold tuning, so our evaluation reflects generalization on unseen edits. Moreover, the generality evaluation includes generated semantically related queries beyond photometric perturbations, where SoftEdit consistently outperforms other baselines.

(2) Retrieval with vs. without the paraphrase judge.

To assess the contribution of the lightweight paraphrase judge to SoftEdit’s performance, we compare SoftEdit with and without the judge in the ablation study below (and in Appendix D.4 Table 8), and found that it improves conventional locality by 0.22–0.34 for LLMs while maintaining comparable performance on other metrics. For VLMs, its contribution is smaller, improving locality by 0.03 for InstructBLIP while maintaining comparable performance on the other two models.

ModelParaphrase judgeAlignSGenSLocAccGenLoc
LLaVA-1.5-7BNo judge0.980.850.611.000.960.93
With judge0.970.850.610.990.950.92
InstructBLIP-7BNo judge0.970.850.641.000.990.75
With judge0.980.850.641.000.990.78
Qwen3-VL-4BNo judge0.890.750.650.950.910.95
With judge0.890.750.650.940.900.95
Llama-3.1-8BNo judge0.960.730.471.000.940.54
With judge0.950.720.470.980.940.76
Mistral-7BNo judge0.940.710.520.990.950.35
With judge0.930.700.520.980.950.69
Qwen2.5-7BNo judge0.940.730.500.980.950.54
With judge0.930.730.520.980.940.77

Dear reviewer,

Thank you for your time and effort in reviewing our paper. We hope our additional experiments, clarifications, and improvements in the revised paper have addressed your concerns.

We would be grateful to hear your thoughts and welcome any remaining questions before the discussion period ends.

Thank you again for your valuable feedback!

Dear Reviewer,

As we approach the very end of the rebuttal period, we sincerely hope to hear back from you.

Thank you again for your time and valuable feedback!

Comments on the paper

1 comment

Dear AC,

Thank you for overseeing the review process and facilitating the discussion. We are looking forward to hearing back from two reviewers on our rebuttal. Meanwhile, we summarize the strengths recognized by reviewers and our improvements addressing the main concerns below.

Strengths recognized by reviewers

  1. Our new problem formulation is meaningful and well-motivated, extending model editing beyond single-answer correctness to one-shot calibration using multi-answer human feedback.
  2. Our new soft generality–locality tradeoff distinguishes distributional equivalence from conventional semantic similarity, making the problem formulation more informative.
  3. Our experimental evaluation is broad, covering six LLMs and VLMs, sequential editing, ranked feedback, and soft-label adaptations of existing baselines.
  4. Our method SoftEdit is computationally efficient through lightweight output-layer updates.
  5. Our presentation was considered clear and easy to follow.

Main concerns and how we address them

We are very grateful to all reviewers for their time and constructive feedback. Based on the suggestions, we have improved our paper as follows.

  1. Clarify calibrated generation enabled by SoftEdit.

Thanks to all reviewers, we have improved the description in problem definition (Section 3.2) and dedicated Appendix B to detailing three approaches to calibrated generation enabled by SoftEdit: fast deterministic readout of multiple answers with their probabilities in B.1, estimating autoregressive full-sequence probabilities given user-specified answers in B.2, and stochastic sampling of complete answers with Monte Carlo sampling validation in B.3. We also discussed the relationship to conventional calibration in B.4.

  1. For SGen-I evaluation in VLM editing, validate approximated target distributions for interpolated images.

Following prior VLM editing works, we use generated images for part of the evaluation (SGen-I). Thanks to reviewers uiqs and q5e1, we have (1) added human validation on a subset of generated and interpolated images for SGen-I; (2) separately reported SGen-I on photometric perturbations and more challenging interpolated image variants across multiple strengths (Appendix D.5), showing SoftEdit outperforms baselines in both settings; and (3) acknowledged in the limitation the approximated target distributions for interpolated images and clarified that interpolated images are intended to make a small part of our evaluation more challenging (half of one metric). Extensive evaluations in both LLM and VLM editing independently support our contributions and conclusions.

  1. Additional ablations.

Thanks to all reviewers, besides the four primary design ablations in Section 4.4, we added these additional ablations and sensitivity analyses in Appendix D.4: (1) Auto-tuned vs. fixed retrieval thresholds, showing improved generality–locality balance. (2) Probability-weighted vs. unweighted and Frobenius regularization, showing the best overall performance with probability weighting. (3) With vs. without the LLM paraphrase judge, showing improved LLM locality. (4) Retrieval inference cost, showing low retrieval overhead as the codebook grows.

We sincerely appreciate the reviewers' constructive feedback, which has helped us improve our paper in valuable ways. Thank you again for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.