acceptodds
Under review as a conference paper at ICLR 2027

Controlling Model Updates at the Granularity of Output Permissions

Abstract

Model updating adapts a deployed model to revised training annotations. Such updates can also change outputs beyond their intended scope. For example, when relation annotations are added to a document, only the scores of entity pairs that gain relations are permitted to change. The scores of the remaining pairs are protected outputs that should stay unchanged. Existing methods control an update either for the whole model through compatible training or for each input through edit-scope routing. Preservation constraints instead restrict non-target distributions within an edit prompt or enforce specified output properties. However, a decision made for a whole input applies equally to the pairs permitted to change and to the pairs that must be protected. In this paper, we ask what shared controls cost and when that cost disappears. We compare document-scope switching with full redeployment for annotation revisions from DocRED to Re-DocRED. Switching only documents gaining relations retains 99.98% of full-redeployment root mean square (RMS) change in protected relation margins. This cost arises because a shared control scales the permitted and protected changes within each group by the same factor. For fixed updates with locally linear gains on permitted outputs, we derive in closed form the minimum mean squared protected change of each grouping, which reaches zero exactly when some group with useful slope has no protected response. Building on this analysis, we propose authorization-granular endpoint routing (AGER), which estimates whether each class score or entity pair may change and scales its update accordingly. On ImageNet, AGER reduces the mean squared protected change by a median of 97.4% relative to the strongest uniformly scaled update that meets the target gain. On document relation extraction, AGER retains 0.376 and 0.436 of the protected RMS change of retrained DREEAM and ATLOP while changing micro-F1 by −1.4 and −2.2 points. These results indicate that model updates should be controlled at the granularity of output permissions.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
model editing, post-deployment model updates, update regression, edit locality, constrained fine-tuning, document-level relation extraction, selective updates

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_koumOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies model updates when outputs within one input have different permissions. It defines authorization units and proves a closed-form floor for shared controls (Theorem 1), plus a sequential depletion result (Theorem 2). It proposes AGER, which estimates per-unit permissions from base-model scores, routes response at authorization granularity, and selects interface and multiplier under simultaneous utility lower bound and protected-response upper bound (Theorem 3). Experiments on ImageNet and DocRE compare fine routing with scalar, external, and native baselines at matched utility.

Critical strengths
  1. The paper’s main conceptual contribution is its formulation of model updates at the granularity of output permissions. It identifies a problem that editing and unlearning work usually ignores: when outputs inside one input have different permissions, a shared update control couples intended and protected changes, and this coupling has a closed-form floor. The authorization-unit framing and Theorem 1 give a clean structural result, and the refinement monotonicity is a useful design principle.

  2. A second contribution is AGER's utility-qualified selection with a finite-sample guarantee and abstention. If the assumptions hold, this links diagnosis to a deployable method rather than a post-hoc metric.

Critical weaknesses
  1. One critical weakness is that the theory is proved under assumptions that are much narrower than the claim. Theorem 1 requires locally linear utility and a fixed update direction, Theorem 3 requires a fixed candidate family and independent bounded observations, and the method treats authorization masks as given. But the paper's own Appendix admits ImageNet uses only a frozen backbone with 32 fitted coefficients and DocRE binds two specific checkpoints, so global nonlinear full-model finetuning is not covered. Permission uncertainty is also explicitly out of scope: Appendix also says a false writable label can authorize harmful change, yet no robust selector is given. So the headline "authorization-granular routing" is validated only in a narrow endpoint space, and the structural floor may not transfer to real model updates. This is consequential because the whole contribution rests on this theory.

  2. Annother weakness is readability: the paper has no conceptual figure for motivation or methodology, and all figures (Figures 1, 2, 4–7) are experimental data plots. A reader cannot see the permission-coupling problem or the AGER pipeline in one place, which makes a dense formal paper harder to follow.

Questions for authors
  1. Theorem 1 assumes locally linear utility and a fixed update direction. Your ImageNet setting uses a frozen backbone with 32 coefficients and DocRE binds two checkpoints. Can you show the floor and refinement result still hold under full-model nonlinear finetuning, or is this a scope limitation of the theory?

  2. Appendix I.1.1 says a false writable label can authorize harmful change, and the selector treats permissions as fixed. Does Theorem 3 survive noisy or adversarial authorization masks? If not, how should a deployment decide when the permission input is untrustworthy?

  3. The main reduction (97.4% median on ImageNet) is measured in a relateively narrow endpoint space. Would the advatage persist under a wider update space or a different backbone, or is it an artifact of the 32-dimensional construction?

1 comment

Thank you for your constructive comments. We clarify the distinction between update generation and output routing: refinement remains valid for completed nonlinear-model updates. To address policy uncertainty, we also derive selection over a declared family of plausible permissions, with simultaneous validity whenever the true policy belongs to that family.

Q1. Which results apply to nonlinear model updates?

For completed full-model updates, refinement remains valid with nonlinear utility. The closed-form floor additionally requires linear utility along a fixed checkpoint response. Write that response as d=Fθ+Δ−Fθd=F_{\theta+\Delta}-F_\theta. Broadcasting a route maximum uu adds exactly ∑j∈P(u2−rj2)dj2≥0\sum_{j\in P}(u^2-r_j^2)d_j^2\ge0 protected squared change, regardless of how the checkpoints were trained. This identity therefore applies to nonlinear retraining, as in our DocRE comparisons.

The refinement proof follows by copying every coarse route to its finer child cells, producing the identical endpoint and utility. Consequently, the minimum protected response over finer routes cannot increase at the same utility constraint. Under linear utility, the closed-form value t2/(mHG)t^2/(mH_G) further supplies the exact optimum, its attaining strengths and the condition for zero protected response.

Q2. What happens when permissions are noisy or untrustworthy?

Theorem 3 tolerates noisy route-fitting labels when support permissions are correct and its bounded, independent-support conditions hold. A route trained from noisy labels is still a fixed candidate before support evaluation. In our fitting-label permutation comparison, the full family selects the unchanged raw-confidence route, showing how evaluation selects a useful alternative to the damaged learned route.

For uncertainty in the permission policy itself, we derive a robust extension over LL plausible policies fixed before support evaluation. Replace log⁡(4M/δ)\log(4M/\delta) in the confidence radii by log⁡(4ML/δ)\log(4ML/\delta), require the utility lower bound to reach the target under every policy, and minimize the largest protected-response upper bound. A union bound over all candidate–policy pairs gives simultaneous utility validity and a response gap of 2ϵR2\epsilon_R against uniformly utility-qualified candidates. Hence the guarantee covers the true policy whenever it belongs to this declared family. If the evidence cannot justify a policy family, retaining the base endpoint avoids authorizing an unsupported change. This distinguishes uncertainty about which outputs may change from error in predicting a supplied permission.

Q3. Is the gain an artifact of 32 fitted coefficients?

The effect also appears when routing completed full-model DocRE updates and in comparisons against million-parameter ImageNet heads. For DREEAM, even the optimal document-level strength allocation has a protected RMS floor 1.94 times the pair route's at matched utility on the 160 design documents. Separately, the 91 documents excluded from development give pair/native RMS ratios of 0.376 [0.356, 0.398] for DREEAM and 0.438 [0.416, 0.460] for ATLOP. Thus two language-model backbones exhibit the effect on a natural annotation revision.

For ImageNet, the existing native baselines train ResNet-50, DenseNet-121 and MobileNetV3-Large heads with approximately 1–2 million parameters. The route/native protected RMS ratios are 0.08, 0.09 and 0.42, respectively, with higher point-estimate utility for the route in all three. To isolate preservation within an identical update space, we additionally fitted coarse baselines with a direct protected-output MSE penalty and the same available permission labels. At exact calibration NLL targets of 0.01/0.02, fine routing reduces held-out protected MSE by 92.4–98.1%. The coarse baselines yield higher held-out NLL gains, so this comparison quantifies the protection–utility trade-off at a common calibration target. It complements the wider native-head and fixed-checkpoint evidence.

W2. Make the permission-coupling problem and pipeline easier to follow

AGER separates writable and protected controls before selecting a useful update. A two-output example makes this mechanism explicit: if only the first output may change, a shared strength applies to both, whereas separate controls can open the first and preserve the second. The pipeline is supplied permissions → route fitting → fixed route/scale candidates → separate utility and protection supports → utility-qualified selection or no update. The utility constraint comes first because suppressing every change would also suppress the requested improvement. The protected-response comparison then identifies the least disruptive useful candidate.

Reviewer_RgtVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how control granularity affects the protection cost of model updates when output permissions within the same input are different. The author gives the minimum protected mean square change of shared control grouping and its reachable construction under fixed response and linear utility, and proposes AGER to scale updates unit by unit with predicted class/pair permissions. The limited candidate selector first requires the utility lower bound to meet the standard, and then minimizes the upper bound of the protection loss. ImageNet and DocRED→Re-DocRED experiments show that fine-grained control can significantly reduce overall protected endpoint variation compared to sharing strength.

Critical strengths
  1. Clear problem formalization. By decoupling permissions, responses, and routing—and requiring positive utility as a prerequisite for applying protection—the framework avoids the trivial solution of simply disabling all updates. DocRE provides a natural test case for scenarios involving mixed permissions for the same input.
  2. Direct comparison of granularity effects. In the DocRE setting, the checkpoint response is held constant; even so, the optimal document-level RMS floor remains 1.94 times higher than that achieved via pair-based routing. Similarly, when the direction is fixed for ImageNet, the median reduction in mean squared change remains 96.95%—closely aligning with the 97.4% figure reported in the main results (see Appendices D and E).
Critical weaknesses
  1. The DocRE main panel contains development data. Table 7 divides the 443 documents into 160 for design, 96 for fitting, 96 for support, and 91 that were not used in development. The pooled results for the 443 documents (Tables 4/8) cannot substitute for independent validation. While RMS gains remain clear on the independent set of 91 documents, Table 9 shows that DREEAM fails to meet all micro-F1/retained-recall criteria, and ATLOP fails to meet all task criteria on both the full and independent panels. A distinction should be made between "failure to establish non-inferiority under uncertainty" and "confirmed performance falling beyond the tolerable loss threshold."
  2. Overall improvements in protection do not extend to all protected subgroups. The RMS for "pure protected-positive" instances increased from 4.204 (under full redeploy) to 4.486, yielding a ratio of 1.067 [1.041, 1.093] (Table 4; Appendix D). The base-confidence routing tends to assign higher weight to retained positive instances, with the overall gains driven primarily by a large number of "revised-negative" pairs. While this result has been disclosed, the current average-based objective fails to achieve stratified protection.
  3. The link between endpoint metrics and behavioral preservation remains to be established. Constancy in a "protected logit" does not guarantee that the corresponding softmax probability—or the final prediction—remains unchanged. Consequently, the "zero-response" property of the oracle applies only to the specified endpoints. The ImageNet selection process also employs truncated NLL and normalized truncated squared responses; metrics concerning untruncated utility, tail-end responses, and actual shifts in decision-making should be presented more explicitly.
  4. Comparisons with external methods do not yet strictly align across all conditions. SERAC and AlphaEdit serve as relevant baselines; however, the use of discrete-scale grids can lead some methods to "overshoot"—for instance, DenseNet shows a native utility/target of 2.80, compared to a routing value of 1.62 (Table 13). While fixed-direction comparisons support the efficacy of granularity, they are insufficient to attribute the entire performance gap to granularity alone. Further analysis is required, incorporating baselines that utilize identical supervision privileges, apply direct penalties to the loss of protected endpoints, or employ distillation.
  5. Conclusions regarding broad-scale deployment extend beyond the current scope of coverage. In ImageNet, the backbone remains fixed, and the update space is constrained; the "natural revision" setting is essentially a transformation for Document-Level Relation Extraction (DocRE). Auxiliary tasks rely on constructed responses, while CounterFact employs isolated, single-instance edits with known masks. The fixed slope and linear quality growth posited in Theorem 2 are based on conditional modeling and have not been empirically verified against actual, continuous update trajectories; furthermore, the observed trends in Table 24 are not consistent.
Questions for authors

See weakness.

2 comments

Thank you for your thorough review. We added same-supervision ImageNet preservation-loss baselines and measured predictions: fine routing retains substantial score protection while improving utility accuracy over the base model with few protected decision changes.

W1. Independent DocRE evidence and non-inferiority

The independent 91-document comparison confirms large response reductions. Its task intervals establish neither non-inferiority nor losses beyond the stated tolerances. The separate design/fitting/support sets contain 160/96/96 documents. Below are 95% document-bootstrap intervals. RMS is pair/native, and task differences are pair minus native in percentage points.

ModelProtected RMS ratioMicro-F1Recall concern
DREEAM0.376 [0.356, 0.398]−1.5 [−2.6, −0.4]Retained: −1.0 [−2.2, +0.2]
ATLOP0.438 [0.416, 0.460]−2.1 [−3.1, −1.1]Added: −3.3 [−5.5, −1.3]

As the table shows, independent response reductions remain large. The micro-F1 intervals straddle the −2.5-point tolerance, and the recall intervals likewise straddle their respective −2/−5-point tolerances. These comparisons distinguish measured task costs from evidence of costs exceeding the stated tolerances. The 443-document panel separately describes the full deployment population.

W2. Pure protected-positive pairs and subgroup protection

Retained-positive protection can be imposed separately from average protection. On the 443-document panel, fixed-route scaling reaches full-redeployment RMS at multiplier 1.217 while retaining 93.7% of the support margin gain at multiplier 1.299. Since responses scale linearly, the corresponding overall RMS ratio is approximately 0.35.

Base-confidence routing gives retained positives larger strengths, explaining their 1.067-times-native RMS at the utility-matching multiplier. A subgroup requirement can instead enter selection as an additional response upper-bound constraint, using simultaneous bounds over predefined groups. This makes subgroup protection explicit in the objective.

W3. Untruncated utility, probabilities, decisions and tails

Utility top-1 accuracy improves while protected prediction changes remain rare in the new ImageNet comparison. We evaluate the full 1000-class softmax, untruncated NLL and unnormalized responses. Across six model–target settings, utility-set accuracy increases by 0.4–1.1 points. On each 1000-image protected set, there are 0–2 prediction changes and zero correct-to-wrong changes. The separate 2500-image sentinel set has 0–2 correct-to-wrong changes per setting.

The tail measurements distinguish mean protection from worst-coordinate movement. AGER's protected absolute-response 99th percentile is 0.002–0.018 logits, with maxima of 0.10–0.36. Maxima exceed the direct-loss baseline on ResNet/DenseNet, separating mean reduction from tail dominance. Likewise, the oracle's zero-response statement concerns specified endpoints, with probabilities and decisions evaluated separately.

W4. Equal supervision and direct protected-output loss

Fine routing reduces protected MSE by 92.4–98.1% against direct preservation-loss baselines at matched calibration NLL targets. All methods use the same 32-dimensional space and permission labels. Identity and max-union baselines minimize cross-entropy plus protected-coordinate MSE, with five fixed penalty weights. Penalty/route selection uses protected support only. Continuous scaling removes grid overshoot at calibration targets 0.01 and 0.02.

Below, NLL entries are AGER/direct-loss held-out gains at target 0.01. The last column spans both targets.

BackboneNLL gain (×10−2\times10^{-2})Protected MSE reduction
ResNet-500.95 / 1.2192.4–92.6%
DenseNet-1210.97 / 1.1095.6–95.8%
MobileNetV3-L1.13 / 1.7298.0–98.1%

As the table shows, fine routing retains much smaller protected response against explicit preservation penalties. The baselines produce larger held-out utility, making this a protection–utility comparison after exact calibration matching. This controls supervision and loss design separately from the original SERAC/AlphaEdit comparisons.

W5. Nonlinear updates and sequential interpretation

The nonlinear-update evidence concerns completed checkpoint transitions, while Theorem 2 characterizes a conditional sequential mechanism. DocRE and actual GPT-2 XL edits test the exact response identity after weight changes. Auxiliary constructed responses test additional endpoint geometries. Theorem 2 assumes fixed useful slopes, linearly increasing protected mass and bounded endpoint width. Table 24 separates realized trajectories from that model, so differing slopes describe different growth regimes. The practical implication is to measure mixed permissions and useful gains when selecting control granularity.

Thank you for your point-by-point response. As I previously noted, MEMIT and AlphaEdit are essential baselines for this scenario; the baselines you have selected are insufficient to demonstrate the contribution of your method, especially when viewed from the perspective of late 2026. Furthermore, since you have not addressed several specific, actionable points I raised, I am maintaining my original score.

Reviewer_Jq47OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Authorization-Granular Endpoint Routing (AGER), a framework for controlling model updates at the level of individual output permissions. The authors theoretically analyze the limitations of coarse-grained update controls and demonstrate that finer-grained routing can reduce unintended changes to protected outputs. Experiments on ImageNet classification and document-level relation extraction show substantial reductions in protected output changes while maintaining a required level of update utility.

Critical strengths

Well-motivated problem formulation. The paper addresses an important challenge in model updating: preserving unaffected outputs while incorporating new information. The output-level permission perspective provides a more fine-grained formulation than conventional input-level update controls.

Theoretical analysis and extensive experiments. The paper provides formal results characterizing the limitations of coarse-grained controls and evaluates the proposed approach across multiple architectures and tasks, supported by comparisons with existing methods.

Critical weaknesses

Limited novelty of the theoretical insights. The central conclusion that finer-grained controls can better preserve protected outputs is relatively intuitive. Since finer-grained routing enlarges the feasible solution space, some theoretical results appear to formalize expected properties rather than establish fundamentally new insights.

Strong assumptions about output permissions. The framework assumes that writable and protected outputs can be meaningfully distinguished. However, in realistic model updates, changes may propagate across semantically related outputs, making such permission boundaries difficult to define or estimate reliably.

Limited connection between the proposed metric and actual model behavior. The primary evaluation focuses on reducing changes in protected logits or relation margins. However, smaller numerical changes do not necessarily translate into better prediction consistency or more reliable downstream behavior.

Performance degradation despite improved output preservation. On document-level relation extraction, AGER reduces protected RMS substantially but decreases micro-F1 by 1.4 and 2.2 percentage points for DREEAM and ATLOP, respectively. This raises concerns about whether the preservation benefits justify the loss in actual task performance.

Insufficient validation on modern large-scale models. The main experiments primarily involve conventional image classifiers and document-level relation extraction models. Although additional GPT-2 XL experiments are included, the effectiveness of AGER in contemporary large language models, complex knowledge editing, and realistic sequential update scenarios remains insufficiently demonstrated.

Questions for authors

NA

1 comment

Thank you for your feedback. We clarify the exact shared-control cost and its attaining allocation, and test routing with preservation-loss baselines and prediction measurements. On ImageNet, these show score protection, utility accuracy gains over the base model, and few protected prediction changes.

W1. What does the theory add beyond a larger feasible set?

The closed-form result gives the best protection achievable by every control with a given grouping and linear utility. For useful slopes BgB_g and protected masses PgP_g, define HG=∑g:Bg>0Bg2/PgH_G=\sum_{g:B_g>0}B_g^2/P_g. At utility target tt, the minimum mean squared response is t2/(mHG)t^2/(mH_G), attained by strengths proportional to Bg/PgB_g/P_g. The floor vanishes exactly when a useful cell has zero protected mass. Thus the result identifies an unavoidable cost and its equality conditions.

The DocRE comparison makes this distinction concrete. On the 160 design documents, optimizing all document-level strengths gives a minimum protected MSE of 0.290, versus 0.077 for pair routing at the same utility. Consequently, even the optimal document-level RMS remains 1.94 times the pair route's. This quantifies a region that no document-wide allocation can attain.

W2. How are permissions defined when semantic changes propagate?

The supplied revision permits related relation scores within an entity pair to change jointly. Under strictly aligned DocRED/Re-DocRED annotations, a directed pair becomes writable when a relation is added. All its relation margins may then change, while pairs with unchanged annotations are protected. Thus the permission unit matches the revision event rather than an isolated label.

AGER estimates the supplied permissions from base scores. For uncertain policies, we derive a finite-family extension: fix LL plausible policies before support evaluation, require the utility lower bound under each, and minimize the largest response upper bound. Replacing MM by MLML in the confidence radii gives simultaneous validity for any true policy in that family, with no update when no candidate qualifies.

W3. Do smaller score changes correspond to measured behavior?

The measured ImageNet behavior shows useful accuracy gains with few protected prediction changes. We evaluate full 1000-class NLL, softmax changes and top-1 decisions. Against coarse updates trained with protected-output MSE and the same supervision, AGER reduces held-out protected MSE by 92.4–98.1% across three backbones. Calibration NLL gains are exactly matched at 0.01/0.02. The baselines have larger held-out gains, so these points describe a protection–utility trade-off.

Specifically, utility-set top-1 accuracy rises by 0.4–1.1 points, and each 1000-image protected set has 0–2 prediction changes, all among already incorrect predictions. The separate 2500-image sentinel set has 0–2 correct-to-wrong changes. These direct measurements connect the response objective to observed decisions.

W4. What protection is gained for the DocRE micro-F1 cost?

The independent comparison quantifies the trade-off: DREEAM/ATLOP retain 37.6%/43.8% of native RMS with micro-F1 differences of −1.5/−2.1 points. On these 91 documents excluded from development, the RMS ratios have 95% intervals [35.6%, 39.8%]/[41.6%, 46.0%]. The result is consistent with the full-panel −1.4/−2.2-point costs and supports choosing an update according to its required protection and utility.

The DREEAM full-panel breakdown explains the trade-off: added-relation recall decreases by 2.2 points, while retained recall changes by −0.3 points [−0.8, +0.1]. Smaller route strengths install fewer new relations while suppressing unintended changes. A utility constraint makes this choice explicit. For pure protected-positive pairs, fixed-route scaling also identifies native-level RMS at 93.7% of matched support utility, providing a concrete subgroup-constrained operating point.

W5. How does the evidence relate to large models and sequential editing?

The large-model evidence tests the routing mechanism after actual GPT-2 XL weight edits, and the sequential result gives a conditional prediction as protection requirements grow. The exact response identity is architecture-independent. Across 100 isolated ROME edits and 100 constrained-fine-tuning edits, input-scope routing retains median protected-response fractions of 99.7%/73.2%. The known prompt–token mask preserves writable change with zero protected endpoint response, isolating the mechanism after actual weight changes.

For sequential updates, the theorem predicts how a shared-control floor changes under fixed useful slopes and linearly growing protected mass. These are quantities to measure as updates accumulate. The empirical evidence concerns specified update transitions, while the theorem explains how growing protection requirements change the control problem.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.