acceptodds
Under review as a conference paper at ICLR 2027

Evidence Inheritance in Model Updates: Performance Continuity Is Not Enough

Abstract

Deployment teams commonly release updated models after data refreshes, continual training, or maintenance, while continuing to cite behavioral evidence from an earlier version. We study evidence inheritance: whether diagnostic evidence established for a reference model M0 remains empirically representative of an updated model Mt. Identification uses a three-level control chain. Level 1 is a magnitude sanity check: positive-scale reparameterizations match the weight-space or representation-space displacement of a claim-breaking update while the prediction function stays fixed, so near-zero diagnostic transfer loss (DTL) is the arithmetic consequence of that symmetry and only closes the false positive that any movement of weights or features would invalidate the diagnostic. Level 2 is the identification. Reliance migrates onto a third channel the inherited diagnostic never probes: the function changes and weight-space displacement exceeds the claim-breaking update, yet the inherited claim remains intact. Level 3 transfers that sanity check, and the claim-breaking phenomenon, to stacked MNIST-CIFAR inputs on a convolutional network. Prediction churn and claim-aligned functional drift (FD) reverse rank between a saturated synthetic task and stacked image inputs. Thus performance continuity is insufficient for automatically inheriting historical diagnostic evidence: continuity must be assessed relative to the historical claim being transferred.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Learning, Model Evaluation, Model Updates, Evidence Inheritance, Model Reliability, Concept Drift, Model Monitoring

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VUwhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether behavioral evidence validated for one model version still holds for an update with unchanged accuracy, which it calls evidence inheritance. It proposes a claim-aligned drift measure (FD) and uses controlled toy experiments (small MLPs and an MNIST–CIFAR CNN) to show that an update can keep accuracy stable yet break an old claim. Weight change alone doesn’t predict this: one update moves the weights further and still leaves the claim intact. Its conclusion is that continuity must be judged per claim, not by performance.

Critical strengths

S1\textbf{S1} The conceptual problem is novel. Prior work on prediction churn, backward compatibility and shift detection compares outputs or test results. This paper instead asks whether a specific historical claim qq, validated for a reference model M0M_0, still holds for an updated model MtM_t.

S2\textbf{S2} The formulation is clear and easy to follow. It cleanly separates performance continuity, ∣P(Mt)−P(M0)∣≤δP|P(M_t) - P(M_0)| \le \delta_P, the ground-truth consequence, DTLt=R^D2(Mt)−R^D2(M0){\rm DTL_{\it t}} = \hat R_{D_2}(M_t) - \hat R_{D_2}(M_0), and claim-aligned drift, FDt(q){\rm FD_{\it t}^{(\it q)}}. It also distinguishes evidence inheritance from evidence refresh.

Critical weaknesses

W1\textbf{W1} The experiments are small in scale, as the authors use tiny MLPs on 2–3-dimensional synthetic inputs and a three-layer CNN on an MNIST/CIFAR mosaic, with hand-designed updates and a single claim type. Whether the phenomenon arises in realistic updates such as fine-tuning or compression remains untested.

W2\textbf{W2} The headline evidence is partly driven by construction. Because the Level 1 arms are exact reparameterizations, DTL=0{\rm DTL} = 0 is guaranteed, so the significant gaps in Tables 3 and 5 are largely arithmetic. Although FD and DTL are measured on independent samples and are not equivalent by definition, the hand-designed alignment among the migration, the intervention, and the claim may favor FD. Its usefulness under less controlled updates thus remains unestablished, and the more informative Level 2 control is limited to a synthetic MLP.

Questions for authors

Q1\textbf{Q1} How does FD compare with simply rerunning the frozen diagnostic, i.e. computing R^D1(Mt)\hat R_{D_1}(M_t) directly? Since mapping FD to DTL already requires DTL labels, in what practical setting would FD be preferable?

Q2\textbf{Q2} Can the Level 2 orthogonal control be extended beyond the synthetic MLP, for instance to the stacked-image setting of Section 4.3 or to a realistic update chain on a standard spurious-correlation benchmark?

Reviewer_66j2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

After a model update, it is impractical to rerun every diagnostic that was measured on the previous version. This paper investigates whether, and how, a model's behavior on those diagnostics changes when a new version is trained. On small synthetic experiments and a toy image task, it shows that updates which keep accuracy flat can still break an old result, with the old test's error rising by about 20 percentage points. It further shows that weight distance and prediction flip rates cannot distinguish these updates from harmless ones, and proposes a label-free check that can.

Critical strengths

S1. The problem is real and under-studied. Model cards and audits are tied to one model version, and Table 7 shows how the question differs from prediction compatibility, shift monitoring and regression testing.

S2. Table 4 shows that updating the model to predict from an input channel the old test never perturbs (channel A to C) gives a larger normalized L2 weight distance from the original model than the harmful update (channel A to B; 0.566 vs 0.452) and the same prediction churn (0.0195), yet leaves the old claim intact (change in the old test's error, DTL, of −0.019 vs 0.231). Weight distance and churn therefore cannot tell the two updates apart, while the claim-specific measure FD_q can (0.021 vs 0.424).

Critical weaknesses

W1. The paper does not explain what FD_q offers over measuring DTL directly. FD_q is the change in the model's response to the old test's intervention, and DTL is the change in the old test's error under that same intervention. The only practical difference is that FD_q needs no labels, yet the paper mentions labels only in passing (Table 2 caption, Section 6, Table 6 caption) and never states this as the reason FD_q exists. Under the paper's own definition (Section 2.1), inheritance means attaching old evidence without a new measurement, but FD_q still requires running the old test's intervention on the new model, so what it saves is only the labeling of the diagnostic sample. The paper should state that estimating DTL without labels is the goal, and report how many labeled samples a direct DTL measurement needs to match the error of FD_q.

W2. Two relevant lines of work are not cited, and two cited references are wrong. Predicting a labeled change in error from unlabeled data is the problem studied by ATC (Garg et al. 2022), agreement on the line (Baek et al. 2022), disagreement-based generalization (Jiang et al. 2022) and DoC (Guillory et al. 2021). The paper cites none of these and does not compare FD_q with them under the old test's intervention. Work on what is preserved across model updates is also largely missing: apart from Yan et al. 2021, the paper does not cite BCT (Shen et al., CVPR 2020), FCT (Ramanujan et al., CVPR 2022), FastFill (Jaeckle et al., ICLR 2023) or MUSCLE (Echterhoff et al., 2024), which measures negative flips across LLM updates. These works define compatibility on representations or predictions, and the paper should explain what defining it per claim adds. Separately, two entries in the reference list are wrong: the Yan et al. 2021 entry has authors and pages that do not match the CVPR 2021 paper, and Dodge et al. 2020 is cited as a TACL paper that does not exist (details under Ethics).

W3. The results do not show when FD is better than simpler measures. Churn predicts DTL better than FD on the synthetic task, while FD is better on images (Section 6, Table 6). Combining the two helps on images but hurts on the synthetic task (Appendix A.5). In the Section 5 scope check, the median final DTL of 1.87 percentage points is below the 2.0 point median reference margin from Section 4, so it is unclear whether the old claim breaks there at all, and the 0.951 correlation between FD and DTL adds little.

W4. Both testbeds are synthetic, and accuracy never changes. The first testbed is a two-feature synthetic task with a small MLP (Section 3.2); the second places an MNIST digit above a CIFAR-10 image so that either half can be replaced independently (Section 4.3). In both, the authors choose which input the model uses before and after the update, so the harmful update is designed in. The analysis also keeps only updates that change accuracy by less than 2 percentage points (Table 8), and on the synthetic task accuracy stays between 98.8% and 99.2% (Figure 2a). The method is therefore never tested on updates that also change accuracy, or on claims measured with a biased benchmark. Real update sequences would test this, for example successive fine-tuned checkpoints evaluated for a Waterbirds or CelebA subgroup claim, or successive LLM versions evaluated on a bias benchmark.

W5. The Level 1 result is already known. The Level 1 matched arms rescale ReLU units so that the network's outputs do not change (Appendix A.2), so DTL, FD and churn are zero for these arms while weight distance and representation distance change. That such function-preserving reparameterizations change weight-space measures without changing the function is established: Dinh et al. 2017 use the same positive rescaling of ReLU layers, and Entezari et al. 2022 study the related permutation symmetries. Neither is cited. The paper should cite them and present Level 1 as a sanity check that follows from this prior work rather than as a finding.

Questions for authors

Q1. Existing methods such as ATC and agreement on the line estimate a model’s error without labels. If they are applied to the old test’s perturbed inputs to predict DTL, does FD_q predict DTL more accurately than they do?

Q2. In the Section 5 check, the median DTL (1.87 percentage points) is just below the 2.0 point margin. In how many of the 20 runs does the old claim actually break?

Q3. How many labeled samples does a direct DTL measurement need to match the prediction error of FD_q?

Q4. Can you show one real sequence of model updates where accuracy stays within 2 percentage points but an old subgroup or robustness result no longer holds, and FD_q detects it?

Answers to Q1 and Q4 would change my score the most: the framing is worth publishing if FD_q beats existing label-free estimators on real updates.

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

This paper examines evidence inheritance—the problem of determining whether historical diagnostic evidence established for a reference model remain empirically representative of an updated model. The authors argue that performance continuity is insufficient to guarantee that historical diagnostic conclusions remain valid, as updated models can undergo substantial internal functional drift while maintaining stable aggregate performance. To analyze this, the authors introduce a three-level control chain (utilizing reparameterization symmetries and orthogonal controls), define a claim-aligned functional drift metric, and demonstrate across synthetic and image datasets that simple metrics like prediction churn can reverse effectiveness ranks depending on update regimes.

Critical strengths
  1. In machine learning operations (MLOps), deployment teams often practice "launch and iterate." When they update a model (e.g., via retraining or data refreshes), they check if the overall accuracy or performance is stable. If the new model is just as accurate as the old one, teams often assume it is safe to ship without re-running expensive safety audits. Two models can have identical overall accuracy while behaving completely differently under the hood. An updated model might fix old errors while introducing new failure modes, shifting its reliance to biased data channels, or violating safety constraints. The paper calls out this dangerous shortcut, proving that performance continuity does not equal behavioral continuity—meaning stable accuracy is never a guarantee that old safety tests still apply.

  2. The use of precise reparameterization symmetries (Level 1) and orthogonal controls (Level 2) elegantly resolves potential confounding variables, cleanly separating weight-space movement from actual claim-breaking functional migration.

Critical weaknesses
  1. The paper is poorly motivated and is difficult to understand. Instead of starting with a relatable, concrete engineering pain point—like a team deploying a model update, relying on an old safety audit, and accidentally shipping a regression—the introduction immediately bombards the reader with heavy, abstract terminology: evidence inheritance, diagnostic transfer loss (DTL), functional drift (FD), and three-level control chains all within the first few paragraphs. By burying the motivation behind dense formalism, the authors make it difficult for the reader to appreciate why performance continuity isn't enough until deep into the methodology section.

  2. The empirical evidence rests almost entirely on low-dimensional models and a simplified stacked MNIST-CIFAR setup. There is no demonstration that this methodology scales to high-dimensional, opaque foundation models or LLMs, where defining a clean, isolated intervention family for a nuanced historical claim is notoriously intractable. While the internal logic of the three-level control chain and its mathematical reparameterization symmetries are tight for low-dimensional synthetic models, the jump from controlled sandboxes to real-world, high-dimensional software lifecycles introduces major verification gaps

Questions for authors
  1. Can the authors rewrite the framing to ground "evidence inheritance" in a concrete engineering failure mode early in the introduction rather than opening with abstract notation?
  2. How can engineering teams practically scope, define, and freeze intervention families for complex generative or multi-modal models where historical claims are broad, qualitative, or open-ended?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.