{"id":"83a21065-63d0-4d8f-a3fb-1b256a5b8db0","arxiv_id":"2506.11347","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Evidential Alignment improves worst-group accuracy by upweighting a biased model's high-uncertainty errors and retraining the last layer with a calibration set, without group annotations.","lead":"This paper proposes Evidential Alignment, a two-stage method that uses uncertainty from a biased classifier to reweight its own mistakes, then retrains the last layer, improving worst-group accuracy without group labels. A generalist might care because it targets the common failure where models rely on spurious correlations, such as identifying camels by desert backgrounds, using only model confidence.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reweighting rule in Eq. 14 assumes misclassified minority samples carry high evidential uncertainty, but spurious errors are often confidently wrong (low u(x)), and the claimed uncertainty-group correlation is never reported quantitatively.","rationale":"The reader's weakest_assumption correctly identifies the proxy assumption: high uncertainty on misclassified samples is treated as a reliable indicator of minority-group membership. My stress-test sharpens this into a concrete failure mode: the sign of the correlation matters. Because u(x) is inversely related to total evidence, a confidently wrong spurious prediction yields a low weight, so the method could systematically downweight the exact minority errors that JTT-style methods upweight. The paper's own text in §5.7 admits the key quantitative evidence is missing ('Quantitative analysis showed correlations...' without reporting them), and no ablation tests the contribution of the u(x) multiplier versus a constant. The theoretical results do not close this gap: Theorem 4.1 is a standard ELBO derivation for evidential learning, and Theorem 4.2 is a generic PAC-Bayes bound that assumes a group weighting scheme rather than deriving one from uncertainty. These issues do not invalidate the empirical results, which include code, hyperparameter tables, and consistent improvements across benchmarks. They do, however, make the central claim conditional on a quantitative validation that is currently absent. The reader's CONDITIONAL verdict is therefore appropriate, and my read does not change it. I would not escalate to REJECT because the method is reproducible in principle and the missing evidence is obtainable with a straightforward experiment.","tokens_in":18471,"tokens_out":5756,"duration_ms":70264,"concrete_test":"On the same calibration split used in the paper, compute the AUROC of u(x) for predicting minority-group membership among samples misclassified by f_theta1 (use group labels only for evaluation, not training). Then compare WGA of Evidential Alignment against a variant that sets u(x)=1 for all misclassified samples (constant JTT-style weight) with identical class balancing, beta, and model selection. If the AUROC is at or below 0.5, or if the constant-weight variant matches or exceeds Evidential Alignment's WGA, the evidential uncertainty multiplier is not the driver of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism rests on Eq. 14: misclassified samples are upweighted by u(x) = K/S(x), where S(x) is the total Dirichlet evidence. High-confidence errors have large S(x) and therefore low u(x), so if the biased model confidently predicts the spurious class for a minority sample, that sample is downweighted rather than upweighted. The method only works if the second-order risk minimization (in particular the KL regularization) systematically converts confident spurious errors into high-uncertainty predictions for minority samples. This is the load-bearing assumption, and it is not established: §5.7 reports only qualitative GradCAM and t-SNE evidence plus an unreported 'quantitative analysis showed correlations' between uncertainty and group labels, with no numbers, no AUROC, and no conditioning on misclassification. The ablation in Figure 4(c) does not isolate the role of u(x); it varies class balancing and regularization, not the uncertainty multiplier. Theorem 4.2 is a generic PAC-Bayes bound for an arbitrary group weighting {w_g} and does not show that the uncertainty-based sample weights approximate such a weighting. If the required conditional correlation is absent or negative, Evidential Alignment could underperform a simple constant-weight reweighting of all misclassified samples, because the u(x) factor would suppress the very errors that need correction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Evidential Alignment, a two-stage annotation-free method for improving worst-group accuracy under spurious correlations. Stage 1 trains (last-layer) models with a second-order risk objective that adds a KL regularizer to a Dirichlet-evidential classification loss, yielding per-sample uncertainty estimates. Stage 2 uses a calibration set to reweight the loss for samples misclassified by the stage-1 model, weighting them by their evidential uncertainty u(x) = K/S(x), and retrains the last layer with class-balanced sampling and a proximal regularization. The authors validate the method on Colored MNIST, Waterbirds, CelebA, CheXpert, MultiNLI, and CivilComments, reporting improvements over ERM and several annotation-free baselines, and provide ablations for the KL annealing step, the weight regularization strength, and class balancing. They also present a theoretical analysis consisting of an ELBO derivation for the second-order objective and a PAC-Bayes bound for a generic reweighted empirical risk.","tokens_in":18720,"tokens_out":4720,"duration_ms":48401,"significance":"If the central mechanism is validated, Evidential Alignment would be a valuable addition to the annotation-free group-robustness toolbox: it is last-layer-only, computationally light, and shows consistent worst-group accuracy gains across vision and language benchmarks, with code released. The paper also includes useful backbone experiments (ViT) and calibration-set-size sensitivity. However, the paper's main conceptual contribution is the claim that evidential uncertainty identifies minority-group samples that a spurious-correlation-biased model misclassifies, and this claim is not quantitatively established. The theoretical results are generic and do not connect the uncertainty weights to group reweighting. The empirical gains are encouraging but do not by themselves verify the mechanism, so the paper currently overstates its theoretical and mechanistic support.","major_comments":[{"comment":"The reweighting rule upweights misclassified samples by u(x) = K/S(x), but a confidently wrong prediction (large S(x), small u(x)) receives a weight less than 1, i.e., it is downweighted relative to correctly classified samples. The method therefore works only if second-order risk minimization systematically assigns high uncertainty to exactly the minority-group samples that the biased model misclassifies. This is the load-bearing assumption of the paper, and it is not established. Section 5.7 reports only qualitative GradCAM and t-SNE evidence plus an unreported 'quantitative analysis showed correlations' without any numbers, AUROC, or conditioning on misclassification. Please report a quantitative measure (e.g., AUROC or correlation) of how well u(x) separates minority from majority groups, ideally restricted to the misclassified subset, and show that the result is not driven by a trivial class-imbalance artifact.","section":"§4.2, Eq. (14)"},{"comment":"The ablation in Figure 4(c) varies class balancing (CB) and the regularizations (Reg) but does not isolate the role of the uncertainty multiplier u(x) in Eq. (14). A critical missing baseline is a constant-weight reweighting that upweights all misclassified samples equally (u(x) = 1 for mistakes), which would show whether the uncertainty scaling contributes anything beyond JTT-style upweighting of errors. Separately, the theory in Section 4.4 does not fill this gap: Theorem 4.1 is a standard ELBO identity for evidential learning, and Theorem 4.2 is a PAC-Bayes bound for an arbitrary group weighting {w_g} that never references u(x) or the sample-level reweighting in Eq. (14). As stated, these theorems hold for any weighting scheme and therefore do not justify the claim that uncertainty-based reweighting identifies or suppresses spurious correlations. Please either add a theoretical statement that connects u(x) to group weights or soften the claim that the method is theoretically justified.","section":"§5.5, Figure 4(c); §4.4, Theorems 4.1 and 4.2"},{"comment":"The empirical comparison is broad, but the main text's framing that the method 'achieves the best WGA' should be more carefully scoped. On CelebA, CnC (which uses validation group labels) achieves a higher WGA (88.8) than the proposed method (84.4), and on CivilComments, DFR† with validation group labels achieves 81.8 vs. 80.2 for the proposed method. The text acknowledges these cases, but the abstract and introduction claim without qualification that the method 'significantly improves group robustness' and 'identifies and suppresses spurious correlations.' Please clarify in the comparisons and the abstract that the claimed advantage is among fully annotation-free methods, not against methods with validation set group labels.","section":"§5.4, Tables 2–4"}],"minor_comments":[{"comment":"The CelebA description contains a typo: 'feamale' should be 'female'.","section":"§5.1"},{"comment":"In the ablation description, 'improves group robustness to some extend' should read 'to some extent'.","section":"§5.5"},{"comment":"The notation around Eq. (19) is inconsistent: the displayed WGA is defined with 1[f_theta(x)=y], but the following sentence says '1[f_theta(x) ≠ y] is the 0-1 loss'; please correct the sentence to say that the 0-1 loss is 1[f_theta(x) ≠ y] and that WGA uses the corresponding accuracy.","section":"§5.2, Eq. (19)"},{"comment":"The text in Section 4.2 states that samples are drawn in a class-balanced manner during evidential calibration, but the pseudocode in Algorithm 1 does not include the class-balanced sampling step; please add it for reproducibility.","section":"Algorithm 1, §4.2"},{"comment":"Reference [6] appears to be mis-attributed: the cited work 'Subjective Logic: A Formalism for Reasoning Under Uncertainty' is a book by Audun Jøsang, and the author list and URL should be checked.","section":"References"},{"comment":"Several baseline entries (e.g., ERM, LfF, BPA, GEORGE) are reported without standard deviations; please add them or state that they are taken from prior papers with single runs.","section":"Tables 2–3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid empirical core and the code release is valuable, but the central mechanistic claim—that evidential uncertainty is a reliable proxy for minority-group membership, particularly for misclassified samples—is not quantitatively supported. The theory section is generic and does not connect to the proposed weighting scheme. I would encourage the editor to require the authors to provide the missing quantitative correlation analysis and an ablation that isolates the u(x) multiplier; these additions are feasible within the paper's scope and would substantially strengthen the contribution. If the authors cannot provide such evidence, the claimed conceptual novelty over simpler reweighting baselines such as BAM would be weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the paper does something genuinely new and empirically useful—it replaces deterministic bias-amplified auxiliaries with evidential uncertainty from the biased model itself, and uses that uncertainty to reweight the model's own mistakes in a last-layer retraining. The results are competitive with methods that use more supervision. But the central mechanism is not actually established, and the theory in Section 4.4 does not carry the weight the paper puts on it. I'd send it out, but with a request for one or two specific additions.\n\nWhat's new: BAM and LfF use a separate deterministic model to identify biased samples; this paper uses the same model's evidential uncertainty, computed from a Dirichlet head, as a weight on misclassified samples. That framing is new relative to the cited work, and the empirical sweep is wider than most: six datasets, vision and text, with ablations. On Waterbirds (92.2 WGA) and CivilComments (80.2), they match or beat DFR and SELF, which use validation group labels; on CelebA they are mid-pack but competitive. Code and hyperparameters are provided. That's real.\n\nSoft spots, in order of importance:\n\n1. The reweighting rule (Eq. 14) assumes the biased model is uncertain on the minority groups it gets wrong. That is the load-bearing claim, and the paper does not supply quantitative evidence for it. Section 5.7 mentions 'quantitative analysis showed correlations' but gives no numbers, no AUROC, no conditioning on misclassification. t-SNE and GradCAM are illustrative, not proof. Without this, the method could be upweighting noise.\n\n2. The theory is generic. Theorem 4.1 is just the ELBO for evidential learning; Theorem 4.2 is a PAC-Bayes bound for any group weights. Neither shows that uncertainty weights approximate a worst-group weighting. This is not a fatal issue if the empirical story is solid, but the abstract overclaims.\n\n3. The ablation does not isolate the uncertainty multiplier. Figure 4(c) varies class balancing and regularization, but there is no comparison to a constant-weight version (e.g., upweight all misclassified samples equally). Such an ablation would tell us whether u(x) is the thing that matters.\n\nNone of these is a killer by itself; together they mean the paper needs revision, not rejection. The empirical core is sound enough that a careful referee could push it to the point where the mechanism is either verified or honestly reframed.\n\nRecommendation: send to peer review, but flag the missing quantitative validation of Eq. 14 and the overclaimed theory. If the authors provide either an AUROC for uncertainty vs. group label conditioned on errors, or a constant-weight ablation, the contribution would be much clearer.","headline":"A useful annotation-free debiasing method with solid empirical results, but the load-bearing uncertainty assumption is unverified and the theory is generic.","tokens_in":19264,"tokens_out":3823,"would_cite":true,"duration_ms":42301,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Evidential Alignment claims that a model's own uncertainty can locate and suppress spurious correlations without group labels, improving worst-group accuracy across vision and language benchmarks.","keywords":["spurious correlations","group robustness","uncertainty quantification","evidential deep learning","worst-group accuracy","last-layer retraining","sample reweighting","Dirichlet distribution"],"falsifier":"On a split of Waterbirds or CelebA with known group labels, compute the AUROC of the stage-1 uncertainty $u(x)$ for detecting samples from the minority group among those the biased model misclassifies; if that AUROC is near 0.5, the reweighting rule is selecting noise rather than minority failures, and the method's stated mechanism is falsified even if some worst-group accuracy gain remains.","tokens_in":18226,"feed_emoji":"🎯","tokens_out":9220,"duration_ms":94216,"temperature":0.7,"pith_summary":"Evidential Alignment tries to prove that the epistemic uncertainty of a deliberately biased model is enough to debias it, with no group annotations. The authors convert the usual logit classifier into a Dirichlet-evidence model via second-order risk minimization, read out per-sample uncertainty, and retrain only the last layer by upweighting the samples the biased model misclassifies in proportion to their uncertainty. On six vision and language benchmarks, they report that this reweighting lifts worst-group accuracy substantially, often matching or beating methods that use group labels for model selection. The reason to care is practical: group labels are expensive, and existing annotation-free debiasing relies on deterministic auxiliary models, while this pipeline is a lightweight extension of ordinary ERM training.","feed_headline":"Uncertainty alone can locate spurious correlations","feed_subtitle":"Reweighting a biased model's unsure mistakes lifts worst-group accuracy with no group labels.","key_machinery":"The central object is the evidential Dirichlet distribution produced by second-order risk minimization. For a $K$-class problem the model outputs evidence $e_k(x)\\ge 0$ and forms concentration parameters $\\alpha_k(x)=e_k(x)+1$; the epistemic uncertainty readout is $u(x)=K/\\sum_k\\alpha_k(x)$, with high values marking samples about which the model has little evidence. The reweighting rule that carries the argument is $w(x,y)=\\mathbf{1}[f_{\\theta_1}(x)=y]+u(x)\\mathbf{1}[f_{\\theta_1}(x)\\neq y]$ on a calibration set, followed by last-layer retraining with class-balanced sampling and a proximity penalty $\\beta\\|\\theta_2-\\theta_1\\|_2^2$. The uncertainty readout selects which failures to amplify, and the last-layer retraining converts that selection into a debiased classifier.","core_discovery":"The central discovery, on the paper's own terms, is that uncertainty estimates from a spurious-correlation-biased model carry the signal needed to identify minority-group failures: high uncertainty on misclassified calibration samples clusters in minority groups, and reweighting by $u(x)=K/\\sum_k\\alpha_k(x)$ suppresses the spurious correlation while preserving core features. The authors support this with an ELBO argument for second-order risk minimization, a PAC-Bayes bound showing that low reweighted empirical risk controls worst-group risk, and empirical gains on Colored MNIST, Waterbirds, CelebA, CheXpert, MultiNLI, and CivilComments across ResNet-50, ViT, and BERT backbones.","pith_inferences":["Editorial inference: If uncertainty-driven reweighting is the active ingredient, the same two-stage recipe should transfer to continuous or multi-valued spurious attributes, where uncertainty readouts are aggregated over calibration samples rather than binarized by misclassification.","Editorial inference: The last-layer-only design leaves the feature extractor's spurious directions intact, so an extension that uses the uncertainty weights to fine-tune the full network or prune spurious embedding directions could yield larger gains, at the risk of disturbing the core features the paper says are preserved.","Editorial inference: The PAC-Bayes bound holds for any weighting scheme, so the paper's specific claim is that evidential uncertainty is a good proxy for the worst-group weight; comparing evidential weights with alternatives such as forgetting counts or gradient-norm outliers under the same bound would isolate that proxy's value.","Editorial inference: Because the method relies on the biased model's own errors, it likely inherits that model's failure modes; on datasets where majority-group samples are themselves hard or label noise is high, the uncertainty-reweighted retraining may need a denoising step."],"forward_implications":["Reported worst-group accuracy rises from 3.74% to 84.58% on Colored MNIST, from 72.6% to 92.2% on Waterbirds, from 47.2% to 84.4% on CelebA, and reaches 73.6% on CheXpert, 74.5% on MultiNLI, and 80.2% on CivilComments.","Because only the last layer is retrained, the added cost is a second-order last-layer pass plus a reweighted last-layer pass, the same big-O as DFR and much cheaper than JTT's full retraining.","The method needs no group annotations at any stage, including model selection, where the paper uses worst-class accuracy instead of worst-group accuracy.","The paper claims the evidential uncertainty estimates are reliable proxies for group membership, which would let the same scores support other group-aware objectives such as balanced sampling in fully annotation-free pipelines."],"supporting_citations":[{"why":"Supplies the Dirichlet-evidence formulation, including evidence $e_k$, concentration parameters $\\alpha_k=e_k+1$, and the uncertainty readout $u=K/\\sum\\alpha_k$, that the method builds its calibration on.","marker":"[33]"},{"why":"Shows that last-layer retraining on a debiased set suffices for group robustness; supplies the retraining backbone and the DFR baseline the paper compares against.","marker":"[18]"},{"why":"Introduces retraining on misclassified samples from a biased model, the strategy that the uncertainty-based reweighting rule refines.","marker":"[25]"},{"why":"Establishes the last-layer retraining protocol and the half-validation-set setup that Evidential Alignment adopts for its calibration stage.","marker":"[21]"},{"why":"The strongest annotation-free baseline; its bias-amplified reweighting is the deterministic alternative to uncertainty-based selection.","marker":"[24]"},{"why":"Defines worst-group accuracy and the group DRO objective, the problem formulation the paper optimizes without group labels.","marker":"[31]"},{"why":"Supplies the prior and posterior network view of second-order risk minimization that the method extends to evidential calibration.","marker":"[37]"},{"why":"Contributes the Colored MNIST benchmark and the invariant-risk framing used for the paper's synthetic experiment.","marker":"[2]"}],"fun_headline_variants":["Uncertainty maps pinpoint spurious correlations without labels","Reweight by uncertainty to debias models with no group info","Uncertainty signals reveal hidden group bias, boosting robustness","Evidential alignment kills spurious correlations via uncertainty","No group labels? Use uncertainty to lift worst-group accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method works only if high uncertainty on a misclassified sample reliably marks that sample as belonging to a minority group whose spurious correlation fails; if unsure mistakes are mostly ordinary noise or hard majority cases, the reweighting pushes the model the wrong way.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty maps pinpoint spurious correlations without labels","Reweight by uncertainty to debias models with no group info","Uncertainty signals reveal hidden group bias, boosting robustness","Evidential alignment kills spurious correlations via uncertainty","No group labels? Use uncertainty to lift worst-group accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3247,"prompt_tokens":911,"completion_tokens":2336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":2257}},"tokens_in":527,"tokens_out":2336,"duration_ms":15331,"temperature":1.0,"reasoning_tokens":2257,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:10:36.307658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a split of Waterbirds or CelebA with known group labels, compute the AUROC of the stage-1 uncertainty $u(x)$ for detecting samples from the minority group among those the biased model misclassifies; if that AUROC is near 0.5, the reweighting rule is selecting noise rather than minority failures, and the method's stated mechanism is falsified even if some worst-group accuracy gain remains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that last-layer retraining on a debiased set suffices for group robustness; supplies the retraining backbone and the DFR baseline the paper compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces retraining on misclassified samples from a biased model, the strategy that the uncertainty-based reweighting rule refines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the last-layer retraining protocol and the half-validation-set setup that Evidential Alignment adopts for its calibration stage."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The strongest annotation-free baseline; its bias-amplified reweighting is the deterministic alternative to uncertainty-based selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines worst-group accuracy and the group DRO objective, the problem formulation the paper optimizes without group labels."}],"review_version":1}