acceptodds
Under review as a conference paper at ICLR 2027

TS-Edit: A Model Editing Framework for Adversarial Defense in Time-Series Foundation Models

Abstract

Time-series foundation models (TSFMs) report strong zero-shot forecasting performance, but where they store factual associations internally that drive a particular forecast, and whether the association can be corrected after a malicious attack are open questions. While these models are increasingly deployed in industrial settings such as power grids, traffic and finance, they remain vulnerable to careful backdoor injections inducing catastrophic operational hazards. To address this, we introduce TS-Edit, a novel model editing framework for adversarial defense in TSFMs. We extend model interpretability approaches built for language models such as causal tracing, ROME, sequential ROME and MEMIT on decoder-only transformer models (TimesFM 2.5, Toto 2.0) and evaluate the framework on popular time-series benchmarks. We use masked white-box attacks (FGSM, BIM and PGD) to corrupt a single segment of the context window; causal tracing then localizes the damage across various MLP layers and a closed-form rank-one update to the identified feed-forward weight then repairs the corrupted forecast. For instance, a single edit after a PGD attack on the UCI Electricity data reduces the corrupted-forecast NRMSE by while preserving the model's behavior on clean inputs. The same localize-then-edit procedure, applied per attack, also repairs BIM and FGSM by and at negligible drift on clean-input predictions (specificity) on the edited model. Subsequent iterative editing exposes a restoration-versus-specificity trade-off. Sequential ROME gives the most consistent restoration across attacks, whereas MEMIT attains both the lowest error and the lowest specificity under the strongest attack, PGD. In every case the clean-input drift remains negligible. We further verify the robustness of the model using cross quarter attacks with random starts and varying attack types. Together, the results demonstrate that both the key-value memory interpretation of feed-forward layers and their localized editability can extend beyond language models. The experiments also demonstrate that attack salience within TSFMs can be localized, diagnosed, and repaired through surgical edits to individual weights.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Foundation Models, Model Editing, Adversarial Robustness, Causal Tracing, Mechanistic Interpretability, Forecast Repair

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_TvsJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TS-Edit, a framework for repairing adversarially corrupted forecasts in time-series foundation models. It adapts causal tracing to identify influential feed-forward layers and applies ROME, Sequential ROME, or MEMIT to restore the model’s clean-input forecast. Its main contribution is extending existing localization and model-editing techniques to continuous time-series forecasting.

Validation covers TimesFM 2.5 and Toto 2.0 on Electricity, ETTh1, and Traffic under segment-restricted FGSM, BIM, and PGD attacks. Experiments assess forecast restoration, clean-input prediction drift, attack budgets, and transfer to fresh attack initializations and other attack families. The authors report substantial restoration in several settings, although effectiveness and specificity vary across datasets and editing methods.

However, the results primarily demonstrate repair with access to clean reference inputs and forecasts, rather than a practical defense when those references are unavailable. The supplied manuscript also lacks comparisons with established defenses and sufficient evidence of generalization to unrelated inputs or robustness against adaptive attacks targeting the edited model. These limitations, together with the largely incremental methodological contribution, motivate my recommendation to reject.

Critical strengths
  1. A meaningful extension of model editing to time-series forecasting. The paper adapts causal tracing and localized weight editing to investigate adversarial forecast corruption. This provides a useful experimental direction for studying whether editing techniques developed for language models can apply to continuous numerical predictions.

  2. Evidence of forecast restoration across multiple settings. Experiments cover two model families, three datasets, and three attack methods, including limited transfer tests with frozen edits. These results support the feasibility of localized repair beyond a single illustrative example.

However, these strengths do not outweigh the reliance on clean reference inputs and forecasts, the lack of established defense baselines, and the limited evidence of broader robustness. My overall recommendation remains rejection.

Critical weaknesses
  1. The defense relies on clean reference information whose availability is not justified. Causal tracing uses clean-input activations, and editing explicitly targets the model’s clean-input forecast. In deployment, the corresponding uncorrupted context may be unavailable. The experiments therefore establish repair under privileged access to clean references, but do not adequately demonstrate a usable defense under realistic conditions. This gap substantially limits the practical significance of the central claim.

  2. The evaluation is too narrow to establish general robustness. Although three benchmark datasets are included, the experiments emphasize selected univariate channels and a limited window-level evaluation. This provides useful preliminary evidence, but is insufficient to support broad claims about foundation-model robustness across diverse series, temporal regimes, and multivariate dependencies. Moreover, the absence of established defense baselines and adaptive attacks against the edited model leaves both comparative effectiveness and robustness unresolved. A substantially broader evaluation is needed to justify acceptance.

  3. The contribution of causal localization is insufficiently isolated. The paper combines existing tracing and editing methods, but the supplied manuscript does not establish whether selecting layers through causal tracing improves repair over random or fixed-layer selection under comparable editing budgets. Without this ablation, it is difficult to determine whether the central “localize-then-edit” design is necessary or whether similar improvements follow from ordinary localized optimization.

  4. The backdoor motivation is not matched by the evaluated threat model. The abstract motivates the framework through malicious backdoor injections, whereas the experiments evaluate inference-time input perturbations using FGSM, BIM, and PGD. These experiments do not establish removal of a learned backdoor or suppression of trigger-induced behavior. This mismatch leaves an important part of the security motivation unsupported and weakens the claimed scope of the contribution.

Questions for authors

Clarification: Could the authors clarify whether TS-Edit requires access to the uncorrupted counterpart of each attacked context and its clean forecast at repair time? This distinction is central to assessing its practical validity as a defense.

1 comment

We thank the reviewer for the detailed assessment and for recognizing the relevance of extending model editing to time-series forecasting. We address the main concerns below.

Clarification question / W1: Does TS-Edit require the uncorrupted context and its clean forecast at repair time?

Yes, the clean reference is required to construct the edit. Our intended setting is offline repair using trusted historical data, followed by evaluation of the frozen edited model on new attacked inputs. The clean counterpart is not provided at that later inference step. We will make the distinction explicit and avoid presenting this as an online repair method that works without a clean reference.

Our frozen-edit experiments evaluate exactly this setting: each edit is fitted once on a single attacked window and then applied, without refitting or a clean reference, to new attacked inputs. The edit transfers well to new attacks on the fitted window, removing over 90% of the forecast error for unseen PGD seeds, BIM, and other PGD budgets. Transfer to new inputs is more limited. On 10 other windows of the same client and 11 other clients, ROME and SeqROME remove 48–66% of the error at the median, with some cases below 20%, and MEMIT removes only 16–23%. MEMIT also fails to transfer when a different segment of the fitted window is attacked. The full table is in our response to Reviewer 1 (W1/Q1), and per-case results are in Appendix D.5.

W3: Is causal localization necessary compared with random or fixed-layer selection?

Across the six model and dataset settings, edits in the final layers lose effectiveness in every case, and on Toto they leave the forecast worse than the attack. Layer choice therefore matters. The traced layer always falls in the early and middle region where editing works, repairs more than the average alternative layer in all six settings, and in four of the six it is either the best layer or within one seed standard deviation of it. Because tracing selects L0 in five of six settings, these results do not separate tracing from a fixed early-layer rule, and we present tracing as a practical heuristic for finding the editable region rather than as a necessary step. The full results are in our response to Reviewer 1 (W1).

W2: Narrow evaluation, multivariate dependencies, limited windows, missing baselines, and adaptive attacks.

The goal of the present study is to establish and characterize the behavior of localized model editing in time-series forecasting. A broader evaluation, including multivariate dependencies, adaptive attacks, and competing defense methods, is an important next step.

To address unseen windows and series directly, we added frozen-edit experiments in which a single edit, fitted on one window, is evaluated without refitting on 10 other windows of the same client and on 11 other clients over the same period. Transfer to these inputs is partial and varies across cases; see our response to Reviewer 1 (W1/Q1) and Appendix [D.5]. This analysis is separate from the population study, where an edit is fitted for each window individually.

W4: The backdoor motivation does not match the evaluated threat model.

We agree. The backdoor reference was meant as broader security context, but the abstract wording suggested that TS-Edit addresses backdoor attacks, which we do not evaluate. Our experiments study inference-time evasion attacks (FGSM, BIM, and PGD perturbations of the input context), and we have revised the abstract and related work to reflect this threat model. Please see our response to Reviewer 2 (Q4) for details.

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

The paper explores whether model editing can be used to repair adversarial failures in time-series foundation models. It proposes TS-Edit, which localizes attack effects through causal tracing and applies ROME-style weight edits to recover predictions. On Electricity, ETTh1 and Traffic, the edits remove most of the attack-induced error, with small drift in the forecast on the clean input. The authors conclude that the key-value memory view of feed-forward layers carries over to forecasting models.

Critical strengths
  1. It frames the problem in an interesting way: where adversarial corruption is localized inside a time-series foundation model, and whether it can be edited out.
  2. Experiments are extensive and also honestly report the failures.
Critical weaknesses
  1. The main limitation is that the technical novelty is modest, as the approach primarily transfers existing LLM editing techniques to the time-series domain.

  2. The paper draws strong mechanistic conclusions from causal tracing results, but high restoration scores only indicate influential layers rather than demonstrating that attack-related information is actually stored in those layers. As a result, the evidence for the claimed TSFM memory mechanism remains limited.

  3. The specificity evaluation is relatively narrow, as it measures prediction changes on a limited set of clean inputs rather than assessing the edited model on broader held-out data. Consequently, it remains unclear whether the edits preserve forecasting behavior more generally.

  4. The empirical evaluation lacks comparison with simple repair baselines (e.g., LoRA updates, lightweight fine-tuning, or input denoising). As a result, the practical benefit of the proposed causal-editing framework over simpler alternatives remains unclear.

  5. some inconsistencies: The text says segment Q2 was selected, while Figures 3 and 10 are labeled Q3.

Questions for authors
  1. How would TS-Edit be used when the clean input and clean forecast are unknown? Does an edit fitted on one window improve forecasts on other attacked windows or series? 2.What happens if ROME edits a random layer or the lowest-ranked layer? Does the repair still work?
  2. Can you measure specificity on many held-out clean windows, and explain the ETTh1 drift?
  3. How does TS-Edit compare with input purification, segment imputation and LoRA fine-tuning toward the same target?
  4. Can you normalize e relative to each series’ scale, and test the backdoor setting the abstract motivates?
2 comments

We thank the reviewer for the thoughtful feedback and for recognizing the value of investigating model editing in this setting. We address the questions below.

W1: The technical novelty is limited because the methods come from language-model editing.

We agree that ROME, SeqROME, and MEMIT are existing methods. Our contribution is in adapting and evaluating them for numerical forecasting, where there are no discrete subject tokens or vocabulary-level factual targets. TS-Edit treats a selected segment of the input context as the subject and uses the model's clean forecast as the editing target. The experiments then ask whether these edits can repair the forecast after adversarial perturbation.

The implementation also required architecture-specific choices for TimesFM and Toto, including activation scaling, the construction of editing keys, and handling Toto's projection scaling. These issues are discussed in the submitted appendix.

W2: High causal-trace restoration shows influence, not that a fact is stored in a particular layer.

We see this in our layer sweeps. If a model has n layers, we look at the first half, m = n/2 (10 of 20 layers in TimesFM, 12 of 24 in Toto). In every one of the six settings, the layer with the highest trace restoration and the layer with the highest R_edit are both in this first half. Between 66% and 94% of the positive trace restoration also sits there, and edits there repair more on average than edits in the second half (0.891 against 0.586 on Toto Electricity, 0.914 against 0.481 on Toto ETTh1). So the layers that matter causally and the layers that can be edited are found in the same part of the network.

That match does not carry down to single layers, though. The peak trace layer and the peak R_edit layer are the same layer only on Toto Electricity. On TimesFM ETTh1, layer 2 has the second-highest trace score, but editing it removes none of the attack (R_edit 0.00). We agree with the reviewer that tracing does not show the attack is stored in one particular layer. What it does show is the region where the attack's effect can be both influenced and edited, and we will word the paper that way. The full layer-sweep results are in our response to Reviewer 1 (W1).

W5: Q2 versus Q3 inconsistency.

Thank you for catching this. The selected segment is Q2, and we have corrected the inconsistent labels in Figures 3 and 10.

Q1: How can the method be used without the clean input? Does the edit transfer?

The trusted clean reference is needed when constructing an edit, not when evaluating a frozen edited model on a new input. We will clarify this offline repair setting and its limitations.

To test this, we fit each edit once on a single attacked window, freeze the weights, and evaluate the frozen model on attacks and inputs not seen during fitting, with no clean reference at test time. On the fitted window, the frozen edit removes over 90% of the forecast error for new PGD seeds, BIM, and PGD budgets from half to twice the fitting budget. On other windows and other clients, transfer is partial: ROME and SeqROME remove about half to two-thirds of the error at the median, with wide variation across cases, and MEMIT transfers least. The full table is in our response to Reviewer 1 (W1/Q1), and per-case results are in Appendix [D.5].

Q2: What happens when ROME edits a random layer or the lowest-ranked layer?

Editing the lowest-ranked layer is a poor choice on Toto for every dataset, where it leaves the forecast worse than the attack (R_edit between -0.67 and -2.41). On TimesFM, the lowest-ranked layer repairs less than the traced layer on Electricity and ETTh1 (0.56 and 0.29), but not on Traffic (0.79). A randomly chosen layer from the early or middle part of the network often performs about as well as the traced layer, while a choice from the final layers reduces repair in every setting. The full table is in our response to Reviewer 1 (W1).

W3 / Q3: Specificity on held-out windows and ETTh1 drift.

We agree that a small parameter update does not guarantee small behavioral change, and that specificity must be measured on clean inputs beyond the edited window. On held-out windows and clients, the frozen edit leaves clean forecasts nearly unchanged (median drift under 1% of the series range).

The higher drift on ETTh1 arises because the edit there is fitted to a weak attack. At ε = 0.2, ETTh1 is the weakest-attacked dataset: the attack displaces the forecast by only 0.14 NRMSE (TimesFM) and 0.35 (Toto), compared with 0.47–1.77 on Electricity and Traffic. Because the attacked forecast is already close to the clean one, the edit is fitted to a very small correction, and the resulting weight change is less specific to the attacked input, so it carries over more to clean inputs. When the attack is substantial, as on Electricity and Traffic, the edit targets a clear displacement and clean drift stays negligible. Editing is therefore most appropriate when the attack causes material damage, which is also when repair is needed.

We have revised the abstract to state this precisely: "Clean-input drift remains small when the model is edited on strong attacks, and is largest on ETTh1, where the attack is weakest." Please see our response to Reviewer 1 (W2/Q2) for the full numbers.

W4 / Q3: Why not compare against LoRA, fine-tuning, input purification, or segment imputation?

These are relevant alternatives, and we agree that direct comparisons would make it easier to assess TS-Edit as a defense. The current work focuses on a narrower question, namely whether existing model editing techniques can repair disrupted forecasts through targeted parameter changes in TSFMs. It does not establish that editing is more effective or more efficient than the alternatives.

Q4: The abstract motivates backdoor attacks, but the experiments do not evaluate them.

We thank the reviewer for pointing out this mismatch. The reference to backdoor attacks was intended to provide broader security context for studying vulnerabilities in time-series foundation models. However, we recognize that the wording of the abstract suggests that TS-Edit addresses backdoor injections directly, which is not what we evaluate. Our experiments focus on inference-time evasion attacks, where FGSM, BIM, and PGD perturb the input context without introducing a learned backdoor trigger. We have therefore revised the abstract and related work to reflect the evaluated threat model more accurately. We do not claim that TS-Edit has been evaluated against backdoor attacks.

Reviewer_Hj57OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether an adversarial attack on a time-series foundation model can be traced to a few layers and then undone by editing those layers, the way ROME and MEMIT edit facts in language models. The method, TS-Edit, attacks one quarter of the input history with FGSM, BIM or PGD. It then uses causal tracing to find the MLP layer whose clean activations best restore the forecast, and applies a rank-one update to that layer so that the attacked input again produces the clean forecast. Two variants repeat the edit at the same layer or spread it over three layers. To check side effects, the paper compares the edited and original models on the same window without the attack.

The experiments use TimesFM 2.5 and Toto 2.0 on one series each from Electricity, ETTh1 and Traffic. On Electricity under PGD, one edit cuts the forecast error from 0.889 to 0.023 and barely changes the clean forecast. The gains are smaller on ETTh1. The traced layer differs a lot across series, from the first layer to the 19th, but the edit reduces the error in every case. An edit fitted once also works on new PGD runs and on BIM and FGSM, as long as the attack is on the same window and segment. The paper concludes that layer-level localization and editing carry over from language models to time-series models, and presents TS-Edit as an adversarial defense.

Critical strengths

The paper is, to my knowledge, the first to apply causal tracing and ROME or MEMIT style editing to time-series foundation models, and it makes them work on two quite different architectures. Along the way it shows some useful facts about how these models react to attacks. Attacking different parts of the history has very different effects, with the oldest quarter barely mattering. ETTh1 is much harder to attack than Electricity. On two of the three datasets, the layers that restore the forecast are the first few. This is the main thing I learned from the paper, and the code could help others study how these models use their input, even apart from the defense framing.

Critical weaknesses

Weakness 1: The defense needs the answer it is supposed to recover. Causal tracing uses the clean input's activations, and the edit is fitted so that the attacked input reproduces the clean forecast. A defender who has the clean input can simply forecast from it, and a defender who does not cannot run TS-Edit at all. Each edit is also fitted to a single attacked window and reverted afterwards, and the only transfer test keeps the same window and segment. Because the edit is optimized directly toward a known target, success is expected at almost any layer, and the paper never compares with editing a random or untraced layer. Its own results point the same way: the traced layer ranges from the first to the 19th across series, yet the edit works everywhere. This weighs most in my assessment, because the defense and the localization are the paper's two main claims.

Weakness 2: The paper barely tests whether the edit is surgical. The hard part of model editing is changing one behavior without disturbing others, but the paper only checks the same window's clean input. Where it does measure, the result is weak: on ETTh1 the edit shifts the clean forecast by more than the error it leaves on the attacked input, about 0.033 against 0.02. There are no tests on other windows, other series or overall forecast accuracy, and no comparison with simple alternatives that use the same knowledge, such as filling in the attacked segment or fine-tuning on the same example. The main results also come from one window per dataset, and the ten seeds only rerun PGD. Without these, even the narrower claim that the edits are precise is not supported.

Questions for authors
  1. In what deployment could TS-Edit run without the clean input? Could the edit target come from something a defender actually has, such as a forecast made after filling in the suspect segment?
  2. How much does the edit change forecasts on other windows, on other series, and in overall accuracy? How does the ETTh1 result fit the abstract's claim that clean-input drift is negligible in every case?
2 comments

We thank the reviewer for the careful review and for recognizing the potential value of applying model editing methods to time-series foundation models. We address the concerns below and will clarify the relevant assumptions in the revised manuscript.

W1 / Q1: Does TS-Edit need the clean input at deployment time? If the defender has it, why not just forecast from it?

TS-Edit requires a trusted clean reference when the edit is constructed. It does not require the clean counterpart of each input when the edited model is later used for inference. Our intended setting is an offline repair procedure in which the model owner uses historical data to identify an attack-sensitive behavior and construct an edit. The edited weights are then held fixed for subsequent evaluation.

To test whether one edit transfers, we fit each editor once on a single attacked window (TimesFM 2.5, Electricity client MT_253, PGD ε = 0.2 on context segment Q2), freeze the weights, and evaluate the frozen model on attacks it never saw. Each test attack is crafted against the unedited model, and the edited model receives only the attacked input; no clean reference is used at test time. We report the percentage of the attack's forecast error (NRMSE against the clean forecast) that the frozen edit removes.

Test conditionCaseCasesROMESeqROME
Same window, new PGD seedsmedian (min–max)598.1 (96.9–98.2)98.3 (97.4–98.5)
Same window, other attacked segmentQ0189.193.8
Q1196.495.8
Q3192.491.6
Whole context179.072.0
Same window, other attack / budgetBIM, ε = 0.2191.892.2
PGD, ε = 0.1195.495.8
PGD, ε = 0.3198.899.0
PGD, ε = 0.4198.999.1
Other windows of the same clientmedian (min–max)1048.2 (17.4–92.7)65.5 (34.5–89.0)
Other clients, same time periodmedian (min–max)1147.7 (15.5–69.9)61.9 (26.2–89.6)

% of the attack's forecast error removed by the frozen edit (100% = clean forecast fully restored). Each case is averaged over 3 independent fits. FGSM (ε = 0.2) and client MT_164 are omitted because the attack barely changes the forecast (adversarial NRMSE < 0.05).

On the fitted window, the frozen edit transfers strongly to new attacks: both editors remove over 90% of the error for unseen PGD seeds, BIM, and PGD budgets from half to twice the fitting budget. When a different segment of the same window is attacked, ROME and SeqROME still remove at least 72% of the error. On new inputs, transfer is partial and varies widely: on other windows and other clients, both editors remove about half to two-thirds of the error at the median, with individual cases ranging from under 20% to about 90%. Per-case results for every seed, segment, window, client, and attack are shown in Appendix D.5 (Tables 12–16 and Figures 16–17).

MEMIT gives comparably strong repair on the fitted window in early runs, but its transfer to other segments and inputs is more variable; we are still characterizing it and do not report it here.

These results support using a frozen edit against new attacks on the inputs it was calibrated on; they do not show that one edit reliably repairs arbitrary new windows or series.

W1: Would editing a random or untraced layer work just as well?

We applied the same ROME edit (same target and optimization) at every MLP layer of both models on the same attacked inputs (5 PGD seeds). R_edit is the fraction of the attack's forecast displacement removed: 1 = clean forecast restored, negative = worse than the attack.

Model / datasetTraced layer, R_editBest layer, R_editOther layers, R_edit (mean ± sd across layers)Final layer
TimesFM / ElectricityL0, 0.977 ± 0.005L9, 0.980 ± 0.0040.875 ± 0.2360.00
TimesFM / ETTh1L0, 0.639 ± 0.002L1, 0.742 ± 0.0050.465 ± 0.2090.14
TimesFM / TrafficL0, 0.779 ± 0.034L3, 0.811 ± 0.0370.767 ± 0.0670.52
Toto / ElectricityL0, 0.937 ± 0.001L0, 0.937 ± 0.0010.730 ± 0.539-1.71
Toto / ETTh1L0, 0.744 ± 0.037L7, 0.945 ± 0.0020.695 ± 0.694-2.41
Toto / TrafficL6, 0.937 ± 0.015L1, 0.950 ± 0.0100.852 ± 0.338-0.67

Traced/best: mean ± sd over seeds. Other layers: mean ± sd across non-traced layers.

Editing a random layer does not work just as well. Repair depends strongly on the layer (across-layer sd 0.21–0.69 in five of six settings), and final-layer edits fail consistently; on Toto they leave the forecast worse than the attack (R_edit as low as -2.41). The traced layer is above the average alternative in all six settings, never among the failing final layers, and stable across seeds (sd ≤ 0.037). On Toto Electricity it is the best of all 24 layers. On TimesFM Electricity, TimesFM Traffic and Toto Traffic, its gap to the best layer (0.003, 0.032 and 0.013) is smaller than its own seed standard deviation (0.005, 0.034 and 0.015). In four of six settings it is thus indistinguishable from the best layer an exhaustive search would find, without fitting an edit at each layer.

On ETTh1, tracing falls short of the best layer. On TimesFM, the traced L0 (0.639) is well above the average alternative (0.465) but trails L1 (0.742), and repair is not monotonic in depth (layers 2–4: 0.00–0.44, although layer 2 has the second-highest trace score). On Toto ETTh1, the traced L0 (0.744 ± 0.037) is above the mean of the other layers (0.695) but below their median (0.881) and below the best layer, L7 (0.945). Both gaps are well beyond seed variation, and we will report them as limitations of tracing as a layer selector.

Tracing selected L0 in five of six settings, so a fixed early-layer rule would have performed similarly here; the contribution of tracing is that it finds this editable region from the model's own activations, without prior knowledge of the architecture. Per-layer results: Figure 18.

W2 / Q2: Is the edit surgical? What happens to clean inputs from other windows and series, including ETTh1?

We agree that a small parameter update does not by itself guarantee small behavioral change, and that specificity must be measured on clean inputs not used for the edit.

Held-out clean inputs. In the frozen-edit experiments (TimesFM 2.5, Electricity), we applied each edit to the clean contexts of 10 other windows and 11 other clients. The median clean drift is 0.0055–0.0091 NRMSE for ROME and SeqROME, i.e. under 1% of the series range, so the edit leaves clean forecasts on unseen inputs essentially unchanged.

Why ETTh1 drift is higher. The higher drift on ETTh1 arises because the edit there is fitted to a weak attack. At the shared absolute budget ε = 0.2, ETTh1 is by far the weakest-attacked dataset: the median attack displaces the forecast by only 0.14 NRMSE (TimesFM) and 0.35 (Toto), compared with 0.47–1.77 on Electricity and Traffic. Because the attacked forecast is already close to the clean one, the edit is fitted to a very small correction, and the resulting weight change is less specific to the attacked input, so it carries over more to clean inputs (median specificity 0.045 / 0.025, versus ≤ 0.0006 on Electricity and Traffic). When the attack is substantial, the edit targets a clear displacement and clean drift stays negligible.

Editing is therefore most appropriate when an attack causes material forecast damage, which is also when repair is needed. We have revised the abstract accordingly, replacing "In every case the clean-input drift remains negligible" with: "Clean-input drift remains small when the model is edited on strong attacks, and is largest on ETTh1, where the attack is weakest."

W2: Comparison with alternatives that have access to the same clean reference, including segment imputation and fine-tuning.

Our question is whether targeted model editing can repair an adversarially disrupted forecast in a pretrained TSFM. Purification and imputation act on the observations, while editing changes the parameters; the approaches are complementary. Direct comparisons would be useful but are not part of the present evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.