acceptodds
Under review as a conference paper at ICLR 2027

Storage Is Not Strategy: State-Conditioned Support Control for LLM Unlearning

Abstract

Many localized large language model (LLM) unlearning methods select a small parameter subset from a localization signal and keep it fixed during optimization. The parameters most associated with a target, however, need not be the best ones to update, and candidate interventions can change value as optimization proceeds. In a controlled experiment, a storage-localization score reaches an area under the receiver operating characteristic curve (AUROC) of , yet storage identity agrees with the better intervention on only targets, while low-rank adaptation (LoRA) wins . We introduce Intervention Score, which ranks editable groups by the predicted effect of the actual unlearning update while accounting for collateral damage, and use it to form the static intervention-value baseline (). We then introduce selective dynamic intervention re-ranking (), which revisits that subset only when a calibrated probe justifies the comparison. On the Natural-TOFU dataset, our method has positive descriptive margins in comparisons between methods and objectives, although several are near zero. On the LACUNA localization-precision benchmark, our mean terminal utility is higher in all six negative preference optimization (NPO) and SimNPO comparisons: NPO margins range from to , and SimNPO margins range from to . The gradient-difference (GradDiff) objective reveals substantial field dependence. Relative to , the primary four-field GradDiff evaluation has six wins, six ties, and no losses, with mean and median paired gains of and . The evidence supports separating localization, initial intervention selection, and checkpoint-dependent support revision.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Model Unlearning, Knowledge Localization, Parameter-Efficient Unlearning, Intervention Value, Dynamic Support Selection, Sparse Model Intervention

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ScPLOpenReview ↗
Rating
2(+1)
Confidence
4(+1)
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies which parameters should be updated during LLM unlearning and whether this choice should change during training. It proposes Intervention Score for initial parameter selection and DIR-R for selectively comparing alternative continuations at a later checkpoint. Evaluation includes a controlled fact-injection experiment, Natural-TOFU, and LACUNA.

Critical strengths
  1. The paper raises an interesting and important question: the parameters that store certain knowledge may not be the best parameters to modify during unlearning. This distinction between knowledge localization and intervention effectiveness is meaningful. It challenges a common assumption in localized unlearning and could help explain why accurate localization does not always lead to better forgetting.
  2. The paper attempts to study parameter selection from multiple perspectives. The controlled experiment provides an interesting observation: despite a localization AUROC of 0.981, the identified storage locations do not consistently lead to better unlearning outcomes. The authors also consider different unlearning objectives and include a static version of their method to isolate the effect of dynamic selection. These are useful experimental directions, although the current evidence is not sufficiently convincing.
Critical weaknesses
  1. The experimental comparisons are questionable. In Natural-TOFU, the reported 19/20 positive comparisons are not obtained from direct experiments. Instead, Eq. (22) combines results from different evaluation sets. This even changes the ordering against GRAIL under SimNPO. Such constructed comparisons cannot replace direct, paired experiments. Moreover, the LACUNA experiments select only three baselines per objective from a larger pool, without a sufficiently clear justification. Different experimental panels are also difficult to reconcile. For example, the NPO score of DIR-R is 6.194 in Table 14 but 1.328 in Table 15. The different settings and their relationship need to be much clearer.

  2. The improvement from DIR-R is very limited. Appendix C.7 reports gains of only 0.0024 and 0.0082 under NPO and SimNPO, despite increasing runtime by 33% and 52%, respectively. It is hard to see the practical value of such small improvements. More importantly, five static baselines outperform DIR-R in the primary four-field GradDiff evaluation. The main text instead emphasizes a three-field analysis that excludes Birth City after observing the results. Although this is disclosed, it makes the overall experimental evidence less convincing.

  3. The methodology has several important issues. Eq. (10) defines Intervention Score using a ratio of gradient-based effects, but its motivation is weak. The score may be sensitive to the scaling of different losses and to a near-zero denominator. No convincing analysis or ablation supports this particular formulation. Furthermore, DIR-R does not actually recompute parameter importance at later checkpoints. It only selects among candidate subsets constructed from the initial ranking. It is therefore unclear whether the improvement comes from state-dependent adaptation or simply from evaluating multiple candidates. In Eq. (15), candidate continuations are selected using the terminal score JJ, which is also used for performance reporting. The paper does not clearly explain whether selection and final evaluation use independent data. This raises concerns about evaluation bias.

  4. The model selection is outdated and insufficient. Natural-TOFU relies on Llama 3 and Gemma 2, both released in 2024. Given that Gemma 4 and Qwen3.5 are now available, experiments on more recent models would be much more meaningful.

  5. The presentation quality is very poor. The entire paper contains no figures or analysis plots, not even a basic illustration of the proposed framework. Instead, the authors repeatedly list numerical results in the main text, often reporting values to six decimal places without meaningful interpretation. Many paragraphs simply repeat information already available in tables. This makes the paper unnecessarily long and difficult to follow. For a method built around dynamic parameter selection, the absence of any visualization of how parameter importance or selected subsets change during training is especially concerning. Overall, the manuscript reads more like a collection of experiment logs than a well-organized research paper.

Questions for authors
  1. In Eq. (15), are the data used for candidate selection completely separate from those used for final evaluation? Can the authors clarify this protocol and report results on an independent test set?

  2. Why are the Natural-TOFU comparisons constructed by combining results from different evaluation sets rather than directly evaluating all methods under identical settings? Can the authors provide direct comparisons?

  3. How were the baselines selected for the LACUNA experiments? Why are only three baselines reported for each objective? Can the authors explain the differences between Tables 14 and 15 and provide complete comparisons under consistent settings?

  4. What justifies the specific formulation of Intervention Score in Eq. (10)? How sensitive is it to loss scaling and the denominator? Can the authors also provide precise definitions of all diagnostic losses and evaluation metrics?

  5. Can the authors demonstrate that DIR-R offers meaningful improvements over simpler alternatives, such as random support replacement or compute-matched candidate search? Does its advantage remain when accounting for the total number of modified parameters and additional computation?

  6. Can the authors provide results on more recent model families, particularly Qwen3.5 or Gemma 4, and include figures illustrating the framework and the actual changes in parameter selection during training?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ScPL

W1 — Direct comparisons

The concern applies to the combined Natural-TOFU estimate; direct comparisons support full DIR-R on LACUNA and Intervention Score on Natural-TOFU. All six LACUNA mean gaps favor full DIR-R, with field, seed, objective, budget and evaluator held fixed in each pair (R10). Natural-TOFU's 19/20 adds a revision gain from other queries, shifting SimNPO vs. GRAIL from −0.0044 to +0.0007 J. Q2 gives the direct selector comparisons.

W2 — Effect size and cost

The four-field ranking is primary, and the gate's practical benefit is lower search cost. From the same starting columns, revision adds +0.1646 J (6 wins, 6 ties; median +0.0025) at 1.45× the fixed-support updates versus 16× for exhaustive search. Five selectors have higher four-field means. Without Birth City, DIR-R leads Projected Causal by +0.0144 J on the original seeds; the ordering reverses on fresh seeds. R5 gives the smaller NPO and SimNPO gains and their run times.

W3 — Score and controller

DIR-R compares continuations from the current checkpoint, so it can change the edited columns without rescoring them. The initial score rewards predicted forgetting, deducts the larger positive retain or protected effect, and scales by neutral change. Candidates share batches and randomness. Q5 tests ranked swaps against random search; Q4 explains the ratio.

W4 — Newer models

We ran a small-scale experiment on the newer Qwen family. On Qwen3.5-4B, we use one Natural-TOFU author, three seeds, 25 steps and 5,998,080 editable scalars per method. Intervention Score reaches J = 0.0703 under NPO, beating WAGLE, Storage and Activation by +0.0299, +0.0195 and +0.0156 J (3/3 each). The order reverses under SimNPO (rebuttal Table R14). We plan to run the same comparison on Gemma 4. If the runs finish in time, we hope to post the results before the November 18 AoE rebuttal deadline.

W5 — Presentation

The revision plots full DIR-R's gains and explains what drives the Natural-TOFU selector gains. Figure 1 labels paired wins; S1/S2 separate utility from collateral damage. Original tables display four decimals. Q6 reports the actual swaps.

Q1 — Independent test

The new Qwen test scores on held-out queries; the GradDiff search comparison scores on its selection queries. With frozen thresholds, we ran four fresh seeds per setting at 3M and 6M editable scalars. Locator queries select the branch; separate forget, retain and protected queries score both continuations. Revision yielded 27 ties, two gains and three declines in 32 held-out pairs. R13 reports the five GradDiff selection scores.

Q2 — Natural-TOFU controls

The direct Natural-TOFU comparison favors Intervention Score over WAGLE and MemFlex when we vary only the ranking. R7 pairs runs by model, author and seed, with the same loss, down-projection space, 25 steps and about 6M editable scalars. The gaps are +0.1344/+0.2926 J over WAGLE (NPO/SimNPO, 94/120 wins each) and +0.0928/+0.2051 J over MemFlex (92/120 each). Lower collateral damage accounts for these gains and the gains against Storage (S2).

Q3 — LACUNA controls

The full-method panel tests the strongest screened baselines for each objective; the ablation tests the contribution of revision alone. We locked the top three by four-field mean J per objective, then used fresh seeds. Table 14 gives +0.4308 J vs. Projected Causal (8/12) under NPO and +0.5034 J vs. Activation-Down (12/12) under SimNPO (R10). Both arms in Table 15 start from the same columns (G1).

Q4 — Diagnostics and scaling

The denominator downweights beneficial edits that also shift neutral answers; separate loss rescaling can alter the ranking. A common positive scale c changes only the effective floor to ε/c. We use mean gold-token NLL and ε=10⁻⁸; the floor matters near zero neutral effect. See Revision Change Summary for the loss definitions and overlap.

Q5 — Random search and cost

Birth City drives the mean difference between DIR-R and random search at equal update counts and parameter coverage. From the same step-160 checkpoint, both searches use 80 or 200 updates per pair; stage times differ by less than 0.55%. The full-panel mean difference in selection score is −0.5530 J. Excluding Birth City gives DIR-R +0.0100 J across the two remaining pairs (one win, one loss; R13).

Q6 — Actual swaps

Qwen extends model coverage, and DIR-R changes which columns it edits at a fixed per-step budget. At step 160, DIR-R swaps 11,154 of 111,538 columns on Birth City, Email Address and Phone Number at seed 4099. The other two retain the original columns. Swapping keeps the number edited per step fixed and increases the total distinct scalars reached by about 10%. Figure 1 shows utility gains; these counts show actual support changes.

AuthorsOpenReview ↗

A Gentle Follow-up on Our Rebuttal — Reviewer ScPL

Dear Reviewer ScPL,

Thank you again for your detailed review and for identifying several important concerns regarding the experimental evidence and presentation of our work.

As the author–reviewer discussion period approaches its conclusion, we would greatly appreciate any further feedback on our rebuttal and revised manuscript.

We have made substantial efforts to address your concerns by clarifying the constructed Natural-TOFU comparisons, reporting direct paired results, distinguishing initial intervention selection from checkpoint-dependent revision, and providing more transparent computational-cost accounting.

We have also added results on a newer Qwen model, evaluated branch selection on held-out queries, compared DIR-R with simpler search alternatives, and revised the presentation with additional figures and clearer explanations.

Importantly, we have explicitly reported the negative and near-zero findings, including cases where simpler alternatives outperform DIR-R, to ensure that our conclusions remain supported by the available evidence.

We sincerely hope these revisions clarify the contributions and limitations of our method. We would be grateful to know whether they address any of your principal concerns, or whether important issues remain unresolved.

Thank you very much for your time and consideration.

Best regards,

Authors of Paper 29979

Thank you for the response. I appreciate the additional experiments and clearer LACUNA comparisons, which address some of my concerns. However, the 19/20 result on Natural-TOFU still combines estimates from different query sets. DIR-R’s gains over fixed support remain small relative to its added cost, and it shows no clear advantage over compute-matched random search. I have raised my score to 2, but I still lean toward rejection. I encourage the authors to revise the paper in light of the reviewers’ feedback.

The controller captures a meaningful share of the available revision gain, and its benefit is concentrated in the runs that trigger revision. Thank you for reconsidering our submission. We would like to address this point directly and specify how the feedback from all three reviewers has shaped the revision.

The NPO and SimNPO increments recover a substantial fraction of the improvement available from the tested local revisions. At the same decision checkpoint, evaluating every candidate in the frozen family yields mean gains of only +0.008024 and +0.014868 J over the fixed continuation. DIR-R obtains +0.002433 and +0.008228 J, capturing 30.32% and 55.34% of those available gains. This explains the scale of the increments: even the best branch in this tested family offers limited additional improvement. The measured runtime increases are 33% and 52%; we will present these rows as limited revision opportunities together with their explicit costs.

The GradDiff ablation shows why the distribution of gains also matters. In the primary four-field evaluation, six of twelve runs trigger revision. Those six have a mean gain of +0.3292 J and median gain of +0.3650 J; the other six keep the fixed continuation. Across all twelve, the mean is +0.1646 J and the median is +0.0025 J. The gate captures 58.26% of the available single-checkpoint gain at 1.45× the fixed trajectory's normalized update count. The gain is concentrated in the triggered runs, which we will show alongside the all-run mean and median. These are continuation-selection utility results.

The full-method baseline gains provide separate evidence for the overall approach. On LACUNA, the six direct comparisons improve mean utility by +0.4308–0.8479 J under NPO and +0.5034–0.5708 J under SimNPO, with a majority of paired wins in every comparison. The equal-cost random-search control addresses the replacement ranking and gives no overall advantage for DIR-R; we retain that result alongside the gate analysis.

We have already incorporated feedback from all three reviewers into the manuscript:

  • zAGL: Section 5.6 now provides fixed-interface evidence for Intervention Score through the direct Natural-TOFU comparisons in Table S1 and the collateral-damage analysis in Table S2. This addresses the request for evidence supporting initial intervention selection.
  • NaAs: The added discussion in §5.6 distinguishes the full DIR-R comparisons, fixed-selector comparisons and same-start controller ablation. This addresses the question of what the different experimental panels measure.
  • ScPL: Section 5.6 adds direct Natural-TOFU comparisons, a Qwen3.5-4B initial-selection experiment and Figure 1 showing paired full-method gains and win counts, addressing your requests for direct evidence, newer-model coverage and clearer presentation.

We will definitely incorporate the remaining feedback from all three reviewers into the next manuscript revision. In particular:

  1. For zAGL, place the statistical intervals for the controlled result and the scope of the fallback rule beside the corresponding claims.
  2. For NaAs, make the baseline-screening and final-evaluation protocols explicit beside the comparison panels, and discuss each controller increment together with its available gain and cost.
  3. For ScPL, lead with direct paired comparisons, label the original 19/20 as a cross-query estimate, and identify branch-selection versus held-out scoring beside each relevant result.

We hope this clarifies both why some increments are small and where revision produces a meaningful benefit. We would be grateful if you could consider the direct evidence and these concrete revision commitments in reassessing the paper.

Controller ablation. Available gain is the best improvement in the tested candidate family at the same frozen checkpoint. Capture fractions use unrounded values.

ObjectiveDIR-R gainAvailable gainCapturedCost vs fixed continuation
NPO+0.0024+0.008030.32%1.33× measured time
SimNPO+0.0082+0.014955.34%1.52× measured time
GradDiff, four fields+0.1646+0.282558.26%1.45× normalized updates
Reviewer_NaAsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to investigate whether parameters associated with target knowledge are the parameters to update during the forgetting process in an LLM, and whether the selected subset of parameters should be dynamically adjusted throughout the optimization process or not. The authors proposed the Intervention Score, which ranks parameter sets based on the predictive performance of actual forgetting updates. This paper separates knowledge localization from intervention selection, and proposes an initial parameter selection method based on objective conditions. In this way, they propose a support correction strategy based on checkpoints.

Critical strengths

1.This paper proposes a well-researched question supported by a controlled experiment. 2.The initial parameter selection approach shows some results, with transparent reporting of limitations. 3.STATIC-IV performs well, suggesting that accounting for the specific unlearning objective may be useful for parameter selection.

Critical weaknesses
  1. Tables 3 and 14 compare DIR-R against only three selected baselines for each objective, but Table 7 includes ten methods. Although the paper states that the baseline identities were fixed before final evaluation, the criteria used to select these particular baselines are not sufficiently clear.
  2. Some reported results also differ substantially across tables. For example, Projected Causal achieves an NPO mean utility of 6.53 in Table 7, compared with 5.76 in Tables 3 and 14, where DIR-R achieves 6.19. Although this difference may reflect evaluation settings, the relationship between these results is not explained clearly enough. In Table 8, five external static methods achieve higher mean utility than DIR-R. It is difficult to show improvements.
  3. The GradDiff results can vary across evaluation settings. Although the four-field evaluation reports a mean gain, the median paired gain is only 0.0025, and the mean gain decreases to 0.01 in the three replications. Are the additional improvements large enough to justify the approach?
  4. While the paper explains that the threshold is from development data, it is not clear whether the terminal utility used to select the final continuation is evaluated on data separate from the final test set.
Questions for authors

1。A clearer explanation of the evaluation subsets and protocols is needed. Could the authors clarify the criteria used to select the three baselines for each objective in Tables 3 and 14? Also, why does Projected Causal achieve an NPO mean utility in Table 7 that is far from that shown in Tables 3 and 14? 2.Are the terminal utility scores used to select the final continuation in Equation 15 computed on data separate from the final test set? Given the relatively small gains over STATIC-IV, the authors need to show whether DIR-R provides a meaningful advantage over simpler alternatives under the same settings.

2 comments
AuthorsOpenReview ↗

Response to Reviewer NaAs

W1 — Criteria for the three selected baselines in Tables 3 / 14

We selected the baselines by mean utility on screening seeds for each objective before running the final evaluation. We ranked by mean J at b = 1 over four fields and seeds 42, 100 and 2026. The NPO shortlist is Projected Causal, MemFlex and Activation exact; SimNPO uses MemFlex, FILA-relative Fisher and Activation-Down. The earlier fixed-selector panel illustrates why the lists differ: Projected Causal ranks first under NPO and eighth under SimNPO. Both final evaluations use all four fields and fresh seeds 31415, 27182 and 16180 (R10). Fresh seeds separate baseline screening scores from the reported DIR-R margins.

W2 — Results differ across tables; five static methods beat DIR-R in Table 8

The earlier selector score does not overturn the direct NPO win over Projected Causal; under GradDiff the ranking is mixed. At the same objective, budget and evaluator, full DIR-R beats Projected Causal 6.1939 to 5.7632 J (+0.4308 J; 8/12 wins) under NPO. Table 7’s 6.5384 J comes from the earlier fixed-selector panel. The four-field GradDiff result remains primary: five comparators have higher means, while DIR-R beats WAGLE and AP-MLP by +1.6523 J (8/12) and +3.8577 J (10/12), respectively (R4). The controller ablation in R5 starts both arms from the same columns. For example, DIR-R wins 8/12 runs against GRAIL, yet GRAIL has the higher mean because the size of the losses matters as well as the number of wins.

W3 — GradDiff gains vary across settings

The GradDiff gains do not justify routine search; the practical benefit of the gate is to cut the cost when the alternative is exhaustive search. Starting from the same columns, revision adds +0.1646 J on average (6 wins, 6 ties; median +0.0025). Six triggers bring the cost to 1.45× the fixed-support update steps, about 9% of the 16× multi-checkpoint search. The three-field test on new seeds gains +0.0100 J over nine pairs. R5 distinguishes update counts from measured time. In the three-field test, the gate triggers in only 2/9 runs; the other seven keep the fixed continuation.

W4 — Is the terminal utility used for selection separate from the final test set?

The new Qwen experiment separates branch selection from final evaluation; the GradDiff search comparison reports scores on its selection queries. After freezing thresholds, we ran four seeds per objective and probe policy at 3M and 6M editable scalars. Eq. (15) chooses a branch on the locator queries. The final evaluator scores that branch and the unchanged continuation on separate forget, retain and protected queries, including real-author and world-fact questions. The five LACUNA GradDiff comparisons used the same forget and retain queries to choose the continuation and report J (R6). The Qwen held-out evaluation yields 27 ties, two improvements and three declines across 32 pairs.

Q1 — Evaluation subsets and protocols; baseline selection; Projected Causal discrepancy

The higher Projected Causal score comes from the earlier fixed-selector panel; the lower score is its result on the fresh seeds used to compare full DIR-R. Table 7 reports 6.5384 J. Tables 3A/14 give 5.7632 J for Projected Causal and 6.1939 J for DIR-R over the same 12 test pairs. The pattern extends beyond that comparator: against NPO Activation exact, full DIR-R gains +0.8479 J, with 9/12 wins.

Q2 — Eq. 15 data separation; meaningful advantage over simpler alternatives

The Qwen test uses held-out scoring, and the separate GradDiff comparison does not show that DIR-R beats either simpler policy. From each of five shared checkpoints, the fixed swap uses 40 updates at a size chosen on development data. Random search and DIR-R share the early branches and differ in the replacement columns for larger swaps. They use 80 or 200 updates each, with stage times within 0.55%. DIR-R trails the fixed swap by 0.0555 J and random search by 0.5530 J on branch-selection scores (R6). One Birth City case accounts for nearly all of the mean gap to random search.

Tables R4 and R13 are in Revision Change Summary; R10 is in Thanks for All Reviewers.

Table R5. Controller ablation from the same initial support

LACUNA settingPairsMean ΔJCost relative to STATIC-IV reference
NPO12+0.00241.33× measured wall time
SimNPO12+0.00821.52× measured wall time
GradDiff, four fields12+0.16461.45× normalized steps
GradDiff, three fields on new seeds9+0.0100—

Table R6. GradDiff checkpoint policies: DIR-R minus each alternative (five cases; branch-selection scores)

AlternativeDIR-R − alternative (J)DIR-R W/T/L
Development-fixed swap−0.05552/1/2
Random search−0.55301/2/2
AuthorsOpenReview ↗

Discussion Phase Approaching Closure — We Welcome Reviewer NaAs's Feedback

Dear Reviewer NaAs,

As the author–reviewer discussion period approaches its conclusion, we would greatly appreciate any further feedback on our detailed rebuttal and revised manuscript.

We have carefully addressed your concerns regarding baseline selection, the differences between experimental panels, the practical magnitude of DIR-R's improvements, and the separation between branch-selection and final-evaluation data.

In particular, we clarified the screening and fresh-seed evaluation protocols, distinguished full-method comparisons from controller ablations, and reported additional experiments with held-out evaluation and comparisons against simpler search alternatives.

We have also explicitly acknowledged the settings in which DIR-R provides limited improvements or does not outperform simpler alternatives.

We would be grateful to know whether these clarifications and additional results have addressed your principal concerns, or whether any important questions remain.

Thank you again for your time and constructive suggestions.

Best regards,

Authors of Paper 29979

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

The paper argues that parameter localization and intervention selection are distinct decisions in localized LLM unlearning. (i) A controlled experiment in Qwen2.5-7B with synthetic facts injected through known interfaces shows a storage-localization score reaching AUROC 0.981 while agreeing with the better of two interventions on only 17/36 targets (LoRA wins 35/36). (ii) The authors propose Intervention Score, s_int(g) = (e_{g,F} − max{e_{g,R}, e_{g,P}, 0}) / (|e_{g,N}| + ε), a first-order gradient-alignment proxy computed at θ0, and its fixed-subset realization STATIC-IV. (iii) Selective DIR-R re-ranks a frozen local family A = {0, 0.01, 0.025, 0.05, 0.10} at a checkpoint only when a short probe exceeds a calibrated threshold. Evaluation covers a derived Natural-TOFU setting (Llama 3 8B, Gemma 2 2B; NPO, SimNPO) and the LACUNA benchmark (OLMo 3 7B; NPO, SimNPO, GradDiff over four PII fields).

Critical strengths

S1. The decomposition into (a) localization, (b) initial support selection, (c) checkpoint- dependent revision is conceptually clean, and the formalization in Sec. 3 (support as an intervention variable, V_0 and V_t as distinct objectives) is the right abstraction. Separating Eqs. 5 and 7 is a genuine clarification over "localize then freeze."

S2. The controlled experiment is the paper's best evidence. Holding objective, batches, steps, seeds, and ~6M editable scalars fixed while varying only the interface is a well-designed comparison, and the 0.981 AUROC vs 17/36 agreement gap is a striking, concrete demonstration that localization accuracy does not transfer to intervention efficacy.

Critical weaknesses

W2 (major). The controlled experiment motivates a conclusion the paper does not pursue. LoRA beats the storage interface 35/36. The natural inference is "use a low-rank adapter," not "use Intervention Score." Intervention Score is never evaluated against LoRA in that controlled setting — all subsequent experiments live in the MLP down-projection column geometry (~6M scalars). The motivational arc therefore terminates in an untested step. Additionally, 17/36 = 47.2% is indistinguishable from chance, and the reported correlation of −0.2302 (n = 36) is not significant (t ≈ −1.38, p ≈ 0.18); it should not be used as evidence of a negative relationship without an interval. W3 (major). The DIR-R "safety guarantee" is a tautology. Sec. 5.5 states that "when all branches complete, terminal utility cannot decrease relative to STATIC-IV." This is true only because ρ = 0 ∈ A and selection takes an argmax — it is the definition of the procedure, not a property of it. It also does not apply when the gate does not fire, which is the majority of runs. There is no statement about the gate's error: Table 11 shows trigger precision 1.000 on calibration but the gate misses Birth City seed 311 and Driver's License seed 997, whose available gains are +0.52 and +0.22.

Questions for authors

Q1. Table 3/14 vs Table 7: why is the NPO Projected Causal baseline 5.763156 in the main comparison but 6.538359 in App. C.2, and why is NPO DIR-R 6.193908 in Table 14 but 1.328368 in Table 15? Please give a run-level ledger mapping every table cell to run IDs, seeds, fields, and evaluation set.

Q2. Did the baselines in Table 3 Panel A receive the same DIR-R revision as "Ours"? If not, the reported margins combine the static and dynamic components and do not isolate the effect the paper claims to measure.

1 comment
AuthorsOpenReview ↗

Response to Reviewer zAGL

W2 — Controlled experiment motivates LoRA, not Intervention Score; 17/36 is chance; r = −0.23 not significant

Fresh LoRA's win answers which interface worked better in the controlled test; our direct comparisons show that Intervention Score improves on storage-based selection within a fixed interface. In Natural-TOFU, we change only the column ranking and keep the MLP down-projection, objective, update rule and budget fixed. STATIC-IV beats Storage by +0.0308 J under NPO (69/114 wins, 19 authors) and +0.1571 J under SimNPO (91/120 wins, 20 authors) by reducing collateral damage (S2). It also beats Storage by +0.0195 J (3/3 seeds) under NPO on Qwen3.5-4B, and AP-MLP by +0.9852 J (44/48 wins) in LACUNA's fixed-selector test. This isolates the benefit of using predicted editing effects to select columns.

The controlled study motivates that change in ranking. Storage identity can be recovered at 0.9815 AUROC, yet using the known storage interface gives the better edit on only 17/36 targets. Fresh LoRA wins on 35/36. Both NPO updates use the same batches, steps, seeds and about 6M editable scalars; the endpoint measures the change in the NLL gap between the gold and best perturbed answers. Even when the storage label is known, it gives little guidance about which update will work better in this setting.

We agree that 17/36 is consistent with chance (nominal 95% interval [0.304, 0.645]) and that the correlation does not establish a negative relationship (r=−0.2302r=-0.2302; Fisher 95% interval [−0.519, +0.106], treating targets as independent). The argument only requires that accurate localization can fail as a decision rule. Lee et al. test edits to an injected MLP region against a disjoint region; Boglioni et al. show that precise localization can help resist relearning. Those results concern different editing choices and endpoints.

W3 — DIR-R "safety guarantee" is a tautology; no statement about gate error

The fallback protects the selection score by construction; the empirical result is that the gate captures most of the available gain while skipping searches. When the gate stays closed, the original columns are retained, and when it opens, the unchanged branch is still eligible. All five calibration triggers found a better branch, but Birth City seed 311 and Driver’s License seed 997 missed gains of +0.5200 J and +0.2200 J. In the final four-field panel we froze the gate decisions, then scored the skipped candidates. All 12 runs had an improving candidate. Searching 6/12 recovered 1.9750 of 3.3900 J (58.26%), leaving 1.4150 J unused (R2). The threshold, 0.02125, was frozen using four fields and calibration seeds 73, 311 and 997.

Q1 — Table 3/14 vs. Table 7 vs. Table 15; run-level ledger

The scores are from different panels, and the relevant test of each claim is the comparison within its own panel. Table 7 gives Projected Causal 6.5384 J on the earlier fixed-selector grid. Tables 3A/14 use four fields and seeds 31415, 27182 and 16180: Projected Causal scores 5.7632 J and full DIR-R scores 6.1939 J. Table 15 instead starts both arms from the Intervention Score columns; enabling revision reaches 1.3284 J (R3). In Tables 3A/14, both methods use the same terminal evaluator in each pair; DIR-R leads Projected Causal by +0.4308 J, with 8/12 wins.

Q2 — Did Panel A baselines receive the same DIR-R revision?

The external baselines did not receive DIR-R; Panel A compares complete methods, and the separate ablation isolates the effect of revision. The Panel A differences therefore include both the initial ranking and revision. In Table 15, both arms start from the same Intervention Score columns; STATIC-IV keeps them fixed and DIR-R can swap them. Allowing revision adds +0.0024 J under NPO and +0.0082 J under SimNPO in 12 pairs per objective. This estimates the controller’s contribution for its own starting ranking. With the controller off, Intervention Score leads WAGLE by +0.1344 J under NPO and +0.2926 J under SimNPO, winning 94/120 pairs under each objective.

Table R2 is in Revision Change Summary; R7 is in Thanks for All Reviewers.

Table R1. Original controlled NPO results

MeasurementResult
Blind storage-interface discriminationAUROC 0.9815
Choose the known storage interface17/36 better actions
Always choose fresh LoRA35/36 better actions

Table R3. NPO panel identities

Paper tableComparator JDIR-R J
7 · historical static selectorsProjected Causal 6.5384—
3A/14 · fresh-seed full policyProjected Causal 5.76326.1939
15 · controller ablationSTATIC-IV 1.32591.3284

Comments on the paper

2 comments
AuthorsOpenReview ↗

Thanks for All Reviewers

Thank you to the reviewers and area chair for the careful reading. Our method, DIR-R, combines Intervention Score for initial selection with selective checkpoint revision. The full method is compared with external baselines. In the controller ablation, STATIC-IV keeps the initially selected columns fixed. The revised main text spans 10 pages; all changes are marked in blue.

In the 25-step Natural-TOFU initial-selection comparison, fixed Intervention Score beats WAGLE by +0.0762 / +0.3486 J on Gemma 2 2B and +0.1927 / +0.2367 J on Llama 3 8B (NPO / SimNPO). Against WAGLE, MemFlex and Storage, all six mean gaps stay positive when we drop any one author's runs or lower the damage weight to 0.6; at that weight the smallest gaps are +0.0053 J for NPO and +0.0650 J for SimNPO.

[G1] Cross-table consistency and run-level ledger

Tables 3A/14 compare full DIR-R with external baselines; Table 15 is a controller ablation. In the 12 LACUNA NPO pairs, full DIR-R scores 6.1939 J vs. 5.7632 J for Projected Causal, a gain of +0.4308 J. Table 7's 6.5384 J refers to the earlier fixed-selector panel. In Table 15 both arms start from the Intervention Score columns and STATIC-IV keeps them fixed. This isolates the contribution of revision: +0.0024 J under NPO (1.3284 vs. 1.3259).

[G2] Selection vs. evaluation data in Eq. 15

For the five four-arm GradDiff comparisons in R13, we used the same forget and retain queries to choose the continuation under Eq. (15) and to report its J. R13 therefore reports branch-selection scores. The gate threshold was set on development data and held fixed for these runs.

The new Qwen3.5-4B checkpoint test uses separate queries. We froze the thresholds before four evaluation seeds per setting. At 3M and 6M editable scalars, we selected branches on locator queries and then scored the chosen continuation on held-out queries.

[G3] Baseline selection criteria for Tables 3 / 14

The screening rule ranks fixed selectors by mean J separately for NPO and SimNPO, at b=1b=1 across four fields and seeds 42, 100 and 2026. The full-policy panel compares three selectors per objective on all four fields, using seeds 31415, 27182 and 16180. Table R10 gives the six comparisons, with 12 field–seed pairs each.

[G4] Magnitude and practical value of the DIR-R gain

The four-field GradDiff controller ablation gains +0.1646 J on average, with six wins, six ties and a median of +0.0025 J from the same initial support. In the retrospective audit it captures 1.9750 of the 3.3900 J available at the frozen checkpoint. R5 distinguishes measured time from normalized updates: the 1× denominator is the fixed-support update trajectory. ScPL Q5/R13 directly compares DIR-R with a fixed swap and random search.

Table labels beginning with R refer to tables in these responses; Table 14 without the R is in the paper.

In the Qwen NPO comparison, the forget NLL increase exceeds WAGLE's by 0.0436 and Storage's by 0.0332. That gain outweighs the additional collateral damage.

Table R10. LACUNA full DIR-R vs. external baselines (12 field–seed pairs per row; positive ΔJ favors DIR-R)

ObjectiveBaselineΔJWins
NPOProjected Causal+0.43088
NPOMemFlex+0.691210
NPOActivation exact+0.84799
SimNPOMemFlex+0.505511
SimNPOFILA-relative Fisher+0.57088
SimNPOActivation-Down+0.503412

Table R14. Qwen3.5-4B initial selector comparison: mean J (one Natural-TOFU author, three seeds per objective, 5,998,080 editable scalars, 25 steps)

SelectorNPOSimNPO
Random−1.15920.0026
Activation0.05470.1217
WAGLE0.04040.1905
Storage0.05080.2076
STATIC-IV0.07030.0363

Table R7. Initial selection: fixed Intervention Score ΔJ vs. baseline

SelectorNPOSimNPO
Random+0.0028+0.0018
Activation+0.0390+0.1496
WAGLE+0.1344+0.2926
Storage+0.0308+0.1571
MemFlex+0.0928+0.2051
FILA-relative Fisher+0.0014−0.0059
AP-MLP+0.0278+0.0930
GRAIL+0.0002−0.0044
Activation-Down+0.0376+0.1441
Projected Causal+0.1089+0.1267

Table R11. Original LACUNA fixed-selector comparison (Paper Table 13)

SelectorSTATIC-IV − selector (J)Run wins95% interval
FILA-relative Fisher+0.703736/48[+0.3765, +1.0244]
AP-MLP+0.985244/48[+0.7097, +1.2726]
GRAIL+0.590039/48[+0.3185, +0.8568]
Activation-Down+0.310137/48[+0.0491, +0.5454]
Projected Causal+0.252835/48[−0.1109, +0.6112]
AuthorsOpenReview ↗

Revision Change Summary

Revisions are marked in blue. The additional results appear in §5.6, at the end of Experiments.

  • Table S1 (zAGL W2, ScPL W1/Q2). We pair existing Natural-TOFU runs for WAGLE, MemFlex, and Storage under the same 25-step update procedure and budget. STATIC-IV beats WAGLE by +0.1344 J under NPO and +0.2926 J under SimNPO, with 94/120 wins per objective: 52/60 on Gemma 2 2B and 42/60 on Llama 3 8B. These pairs measure the initial selector. The original 19/20 counts positive margins after adding the DIR-R gain measured on a separate set of queries to the initial-selector gap. For SimNPO vs. GRAIL, that changes the gap from −0.0044 J to +0.0007 J.
  • Table S2 (zAGL W2, ScPL W1/Q4). On the same runs, STATIC-IV reduces mean collateral damage against all three baselines under both objectives. Against WAGLE the reductions are 0.1880 under NPO and 0.4177 under SimNPO. All six mean utility gaps remain positive after omitting any one author or rescoring with damage weight 0.6.
  • Figure 1 (ScPL W1/W5/Q3). The bar chart shows the original paired results from Table 14: full DIR-R gains +0.4308 to +0.8479 J under NPO and +0.5034 to +0.5708 J under SimNPO against the selected baselines. Bar labels give wins out of 12 field–seed pairs.
  • Newer backbone (ScPL W4). Section 5.6 adds a Qwen3.5-4B NPO test of initial selection on one Natural-TOFU author. At 25 steps, fixed Intervention Score beats WAGLE, Storage and Activation in 3/3 seed pairs each. Rebuttal Table R14 reports both objectives.
  • LACUNA comparisons (NaAs W1/Q1, ScPL W3/Q3). Section 5.6 distinguishes baseline comparisons from controller ablation. Table 14 compares full DIR-R with external baselines: +0.4308 to +0.8479 J for NPO and +0.5034 to +0.5708 J for SimNPO, with 12 pairs per comparison. Table 15 holds the initial columns fixed across arms to isolate revision: +0.0024 J under NPO and +0.0082 J under SimNPO.
  • Presentation (ScPL W5). The original tables are shown to four decimal places. Their underlying values and table numbers are unchanged. Section 7 summarizes the evidence in one paragraph.

Supporting evidence for the replies

Score and metrics (ScPL Q4). For group gg, let eg,qe_{g,q} be the predicted change per unit step in mean gold-token NLL on the forget, retain, protected or neutral set. We rank groups by (eg,F−max⁡(0,eg,R,eg,P))/(∣eg,N∣+ϵ)(e_{g,F}-\max(0,e_{g,R},e_{g,P}))/(|e_{g,N}|+\epsilon). The utility JJ subtracts the largest positive collateral NLL increase from the observed forget NLL increase. Collateral queries cover retain, real-author and world-fact answers on Natural-TOFU and retain answers in the LACUNA selector comparison. Under GradDiff, overlap with WAGLE is at most 6.20%, versus at least 32.76% between baseline pairs at the tested cutoffs. This checks the combined ranking rule.

Search control (ScPL Q5, NaAs Q2). DIR-R and random search share probes and the 0% and 1% branches; larger swaps use different replacement columns. The fixed swap uses a rate chosen on development data and 40 updates. R13 reports branch-selection scores, including the full sample and both Birth City exclusions.

Table R2. Retrospective final four-field GradDiff gate audit

QuantityResult
Cells with an improving candidate12/12
Gate hits / misses6 / 6
Recall within the tested candidates50.00%
Available / captured / missed gain (J)3.3900 / 1.9750 / 1.4150
Fraction captured58.26%

Table R4. Four-field GradDiff: full DIR-R vs. other selectors (12 pairs per row)

BaselineMean JDIR-R − baseline (J)DIR-R W/T/L
Projected Causal109.7329−7.10464/0/8
Activation-Down105.3194−2.69106/0/6
FILA104.5629−1.93467/0/5
GRAIL103.8154−1.18718/0/4
MemFlex103.1044−0.47607/0/5
DIR-R102.6284——
Storage102.2273+0.40108/0/4
WAGLE100.9761+1.65238/0/4
AP-MLP98.7706+3.857710/0/2

Table R13. GradDiff branch-selection ΔJ: DIR-R minus each alternative

Field / seedvs. fixed swapvs. random search
Birth City / 40990.0000−2.7850
Birth City / 8191−0.31000.0000
Birth City / 16381−0.30250.0000
Email Address / 4099+0.0200−0.0150
Phone Number / 4099+0.3150+0.0350
Full mean (5 pairs)−0.0555−0.5530
Mean without Birth City seed 4099 (4 pairs)−0.0694+0.0050
Mean without Birth City (2 pairs)+0.1675+0.0100

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.