{"id":"5a12c1c5-e902-49c5-bc8a-ccffaeeaed3a","arxiv_id":"2511.14117","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"On ChaosNLI, POPQUORN, and a CIFAR-10H subset, training on full annotation distributions instead of majority labels preserves accuracy while better matching human uncertainty.","lead":"A Microsoft team compared training classifiers on the full distribution of human labels versus a single majority label across three vision/NLP datasets. Soft-label training matched accuracy while making models uncertain on exactly the samples where humans disagree, supporting the view that annotator disagreement is information, not noise.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-sample empirical label distributions with few annotators are noisy proxies; the paper's KL gains may be dominated by fitting sampling noise, so the central 'distributions as ground truth' claim remains untested.","rationale":"The reader's weakest assumption was that persistent disagreement reflects epistemic ambiguity rather than noise or bias; I agree that is the philosophical lynchpin. But my concern sharpens this further: even if the ambiguity is genuine, the per-sample empirical distribution used as both training target and evaluation metric is a finite-sample estimate of the population distribution. With 6.7 annotators per sample on POPQUORN, the target itself is extremely noisy, so the observed KL reduction and entropy-correlation gains are partly expected from fitting that noise. This is a distinct, testable weakness that does not depend on resolving the ambiguity-versus-noise debate. If a held-out-annotator evaluation shows the benefit persists, the paper's claim is substantially strengthened; if it does not, the central empirical contribution is an artifact. I also note the reported 32% average is inconsistent with the paper's own Table 4, which reduces confidence in the numerical claims. The paper does have genuine strengths: using ChaosNLI's 100 annotations as a reliability anchor, matching accuracy while lowering KL across three domains, and the conceptual reframing. Those justify keeping the current CONDITIONAL verdict, but the sampling-noise confound should be resolved before the stronger 'distributions are ground truth' conclusion is accepted.","tokens_in":6154,"tokens_out":5595,"duration_ms":63564,"concrete_test":"For POPQUORN (and ideally all datasets), randomly split each sample's annotators into two disjoint sets A and B. Train soft-label models using only the distribution from set A as targets, and hard-label models using the majority vote of set A; evaluate both against the held-out set B's distribution (and symmetrically train on B, evaluate on A). Compare the soft-vs-hard KL gap when evaluation uses the training distribution versus held-out annotator distributions. If the soft-label advantage largely disappears or reverses on held-out annotators, the reported epistemic-alignment benefits are inflated by fitting finite-sample noise rather than capturing stable population-level disagreement. A secondary check: recompute the 'average 32% KL reduction' from Table 4; the numbers in the table imply roughly 20%.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's normative claim is that the annotation distribution is the correct learning target, not a noisy estimate of it. But the actual training targets are per-sample empirical distributions. For POPQUORN (Table 1) there are only 6.7 annotators per sample; CIFAR-10H-Hard has 50 and ChaosNLI has 100. With n=6.7 categorical draws, the empirical distribution is a high-variance estimate of any population-level distribution. Even granting that annotator disagreement reflects genuine epistemic ambiguity, the observed per-sample proportions are noisy measurements of that population distribution. Training with KL(p_hat || q) and evaluating with KL(p_hat || q) therefore rewards reproducing sampling noise. This is not a purely philosophical quibble: the reported entropy-correlation gains are largest on ChaosNLI (100 annotators, +119%) and smallest on POPQUORN (6.7 annotators, +24%), exactly the pattern expected if the benefit is partly an artifact of having more reliable targets. Moreover, the headline '32% average KL reduction' is not reproducible from Table 4: the reductions are 13.1% (ChaosNLI), 31.9% (CIFAR-10H-Hard), and 16.0% (POPQUORN), averaging about 20%, not 32%. The central claim that soft-label training aligns with epistemic truth, rather than with finite-sample noise, remains unsupported unless the evaluation targets are reliable estimates of the population distribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that when annotators disagree on genuinely ambiguous tasks, the full annotation distribution is the correct learning target and should be used as soft labels, rather than being collapsed to a majority-vote hard label. It compares soft-label training (cross-entropy to the empirical annotation distribution) with hard-label training on three datasets (ChaosNLI, POPQUORN, CIFAR-10H-Hard), using a frozen-embedding MLP head. It reports that soft-label training reduces test KL divergence to the annotation distribution, increases the correlation between model and annotator entropy, maintains or improves accuracy, and slows overfitting. The main evidence is in Tables 3 and 4; the paper frames these results as supporting a normative claim about epistemic alignment rather than as a purely practical trick.","tokens_in":6545,"tokens_out":7009,"duration_ms":68304,"significance":"If the empirical claims were robust, this would be a useful conceptual contribution: it would reframe soft-label training as the epistemically aligned objective for subjectively labeled data and would provide a concrete evaluation criterion (KL to annotation distributions). The paper's strengths are its use of three heterogeneous datasets, significance testing over 10 seeds, and clear reporting of per-dataset results. However, the central quantitative claim is not reproducible from the tables, and the key interpretative premise—that observed empirical distributions are ground truth rather than noisy estimates—is not actually tested by the experiments. The paper is therefore not ready for acceptance in its current form, but the issues are addressable with additional analysis and revised claims.","major_comments":[{"comment":"The headline '32% average KL reduction' is not supported by Table 4. The relative reductions are (0.367–0.319)/0.367 ≈ 13.1% for ChaosNLI, (0.596–0.406)/0.596 ≈ 31.9% for CIFAR-10H-Hard, and (0.430–0.361)/0.430 ≈ 16.0% for POPQUORN. The unweighted mean is about 20%, and pooling by hard-label KL gives about 22%; only CIFAR-10H-Hard reaches 32%. The abstract and §4.2 should be corrected, and per-dataset relative reductions should be reported instead of a single unsupported average.","section":"Abstract; §4.2, Table 4"},{"comment":"The central claim that the annotation distribution is the ground truth, not a noisy estimate, is not tested. The targets used for both training and evaluation are per-sample empirical distributions p_hat. For POPQUORN, Table 1 gives 6.7 annotators per sample, so p_hat is a high-variance estimate of any population-level distribution. Furthermore, soft-label training minimizes cross-entropy to p_hat, which is D_KL(p_hat||q)+H(p_hat), so reporting lower test KL to the same p_hat is partly a restatement of the objective. The noise concern is supported by Table 4: entropy-correlation gains are largest on ChaosNLI (100 annotators, +119%), intermediate on CIFAR-10H-Hard (50, +40%), and smallest on POPQUORN (6.7, +24%)—exactly the pattern expected if the benefit partly reflects the reliability of the estimated targets. To support the normative claim, the authors should model annotation sampling","section":"§1.1, §2, Eq. (1)"},{"comment":"The text states that 'all other training procedures, model architectures, and optimization settings are identical across both conditions,' but Table 2 reports different learning rates, batch sizes, epochs, weight decays, and schedulers for soft and hard models on every dataset. This means the accuracy/KL comparisons and, in particular, the overfitting claim based on Figure 4 are not controlled for optimization configuration. The authors should provide a matched-hyperparameter ablation (or report results over the full grid) to show that the qualitative conclusions are not an artifact of selecting different configurations per condition.","section":"§3.3, Table 2"},{"comment":"The paper itself notes that the effect 'strengthens with more annotations' (ChaosNLI +119%) and is smallest on POPQUORN (+24%). This observation is treated as supporting evidence, but it directly undercuts the Section 1.1 claim that the observed distribution is the ground truth rather than a noisy estimate: if the distribution were the target concept, model quality should not depend so strongly on how many annotators were averaged. The authors need to address this confound explicitly, for example by showing that the entropy-correlation gain persists when the number of annotations per sample is held fixed.","section":"§4.3, §5"}],"minor_comments":[{"comment":"The entries '1000.59' and '500.24' appear to be missing a separator between the annotation count and the mean entropy (should be '100 0.59' and '50 0.24'). Please fix the formatting.","section":"Table 1"},{"comment":"The text describes the entropy correlation as 'Pearson correlation coefficient ... averaged across all test samples,' but a Pearson correlation is typically computed over a set of points. Clarify whether Table 4 reports a single correlation over the whole test set, and if so how (or whether) it is averaged over seeds.","section":"§4.3, Table 4"},{"comment":"The use of OpenAI's proprietary Text Embeddings 3 Large as a frozen embedding model makes the exact features irreproducible. Releasing the extracted feature cache or using an open embedding model would improve reproducibility, which is especially important given the small datasets.","section":"§3.2, §3.3"},{"comment":"The validation-loss curves compare raw cross-entropy between soft and hard targets, but soft-label targets contain an additional entropy term H(p) that differs per dataset. The curves would be more informative if plotted as KL divergence to a common reference, or if the constant shift were acknowledged.","section":"Figure 4"},{"comment":"For CIFAR-10H-Hard and POPQUORN, 'statistically indistinguishable' is inferred from failure to reject at α=0.05; an equivalence test or confidence interval on the accuracy difference would be a more appropriate way to support the 'no degradation' claim.","section":"§4.2, Table 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take: this is a respectable empirical paper that makes a philosophical point — when annotators disagree for reasons tied to real ambiguity, the distribution is the target — and then checks what happens if you train on that distribution. The results are mostly as expected: soft-label models match accuracy and report lower KL to annotation distributions. I think the paper deserves referee time, but the abstract overstates one number and the epistemic ground-truth claim is argued, not proven.\n\nWhat's new: the specific comparisons on ChaosNLI, POPQUORN, CIFAR-10H-Hard (using a curated hard subset), with frozen embeddings and a light MLP head, appear not to be in the prior literature. The paper is well written, cites the relevant label-distribution-learning and knowledge-distillation work, and is transparent about its setup. The entropy-correlation result is a genuinely less circular measure than KL, and it improves on all three datasets.\n\nWhere it's soft: First, the abstract's '32% average' KL reduction does not match Table 4. The numbers in that table give roughly 13% (ChaosNLI), 32% (CIFAR-10H-Hard), 16% (POPQUORN), averaging around 20%. So the headline is off, and the 'p < 10^-4' for the combined average isn't supported by the per-dataset t-tests shown. That's a reporting problem, not a fatal one. Second, the stress-test noise concern has teeth. POPQUORN has 6.7 annotators per sample; the empirical distribution is a noisy estimate of whatever 'truth' is. Training and evaluating against the same noisy target means part of the KL gain could be fitting sampling noise. The pattern they report — the entropy-correlation gain is largest on ChaosNLI (100 annotators, +119%) and smallest on POPQUORN (6.7, +24%) — is consistent with that reading. It's also consistent with genuine ambiguity being measured more precisely with more annotators, so I wouldn't call this a refutation. But the paper does not confront it: they never model annotation noise or show the result is robust to it. Third, the per-condition hyperparameter search means each model uses its own best configuration; that's defensible, but it slightly blurs the causal attribution.\n\nBottom line: the central argument — that the annotation distribution is the right target on subjective tasks — is a position, not a theorem, and the paper's evidence supports it directionally. The reader who takes away 'soft labels preserve uncertainty without hurting accuracy' is getting a fair summary. The paper would benefit from fixing the 32% number, adding a noise analysis or at least a discussion of it, and ideally releasing code. Still, the core measurements are new, the writing is honest, and it's a useful empirical datapoint. I'd send it to a serious referee and I'd read it closely in my own work.","headline":"A clear, honest empirical paper on soft-label training; the headline KL number is off and the noise concern is real, but the core argument and most measurements hold up.","tokens_in":7012,"tokens_out":3807,"would_cite":true,"duration_ms":38316,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Soft-label training, which uses full annotation distributions instead of majority-vote labels, matches or exceeds hard-label accuracy while producing models whose uncertainty closely tracks human disagreement.","keywords":["soft-label training","annotation distributions","epistemic uncertainty","majority vote","KL divergence","entropy correlation","human uncertainty","ambiguous classification"],"falsifier":"Construct a dataset with known true class-conditional distributions, then simulate annotator disagreement that is pure input-independent noise rather than genuine ambiguity; if soft-label training on these noisy distributions does not reduce KL divergence to the true distributions or does not maintain accuracy relative to hard-label training, the benefit would be specific to genuine disagreement rather than distributional targets generally.","tokens_in":6049,"feed_emoji":"🎯","tokens_out":2620,"duration_ms":29578,"temperature":0.7,"pith_summary":"This paper argues that when annotators genuinely disagree about a sample, the distribution of their labels is the ground truth, not a noisy estimate to be collapsed into a single majority vote. Training on those full distributions, rather than on one-hot hard labels, preserves the ambiguity that humans actually experience. Across natural language inference, politeness classification, and image categorization, soft-label training matches or improves accuracy while reducing KL divergence to human annotations by 32% and strengthening the correlation between model and annotator entropy by 61%. If correct, this reframes annotation distributions from signals to be aggregated away into faithful targets that models should learn to reproduce. The practical stake is that AI systems deployed in human contexts should be uncertain in exactly the places where people are uncertain.","feed_headline":"Soft labels beat majority votes at capturing human uncertainty","feed_subtitle":"Training on full annotation distributions keeps accuracy and cuts KL divergence to human labels by 32%.","key_machinery":"The central mechanism is soft-label training with cross-entropy loss against the full human annotation distribution, compared against hard-label training with categorical cross-entropy on the majority-vote label. Two evaluation metrics carry the argument: KL divergence between model predictions and annotation distributions, which measures how closely the model reproduces human uncertainty, and the Pearson correlation between annotation entropy and model prediction entropy, which measures whether the model is uncertain on the same samples humans are. To isolate the effect of label type, both conditions use frozen pretrained embeddings (OpenAI Text Embeddings 3 Large for text, DINOv2 Small for","core_discovery":"The central claim is that the annotation distribution itself is the correct learning target for data with genuine epistemic ambiguity. The authors demonstrate this by training identical models on either soft labels (the full distribution over annotator votes) or hard labels (the collapsed majority vote) across three datasets: ChaosNLI, POPQUORN, and a curated CIFAR-10H subset. Soft-label training achieves significantly lower KL divergence to human annotations on all three datasets (p < 0.001, average 32% improvement), produces predictions whose per-sample entropy correlates 61% more strongly with annotator entropy, and maintains or improves standard accuracy—most notably raising ChaosNLI acc","pith_inferences":["Inference: A natural testable extension would be to synthetic datasets where true conditional distributions are known and annotation disagreement is generated independently of the input; the claim predicts soft-label training should recover the true distribution better than hard-label training, while pure labeling noise should not confer this advantage.","Inference: The entropy-correlation result suggests that soft-label models may also yield better downstream calibration and selective prediction, since a model that knows where it is uncertain can defer or abstain more reliably—though the paper does not test these downstream tasks directly.","Inference: The paper's framing implies a caution for demographic labeling bias: if annotator disagreement partly reflects systematic demographic differences rather than irreducible ambiguity, then treating the observed distribution as ground truth could entrench those biases, so the normative claim likely depends on disagreement being genuinely epistemic.","Inference: The few-annotator regime, such as POPQUORN's 6.7 annotations per sample, introduces sampling noise into the observed distributions; modeling that noise explicitly could either strengthen the argument (if noise is accounted for) or bound the benefit in low-annotation settings, a direction the paper does not explore."],"forward_implications":["Accuracy-only evaluation penalizes honest uncertainty: a model that confidently predicts the majority class on a 60-40 split may score higher accuracy but worse epistemic alignment, so evaluation should include distributional metrics on ambiguous tasks.","Investing in multiple annotations per sample yields measurable returns in model uncertainty alignment, with the effect strengthening as the number of annotators grows, from 6.7 per sample in POPQUORN to 100 in ChaosNLI.","Soft-label training resists overfitting: validation loss continues to improve for more epochs with distributional targets, suggesting a regularizing effect that could benefit small-data and transfer regimes.","On tasks with inherent subjectivity, annotation distributions should be treated as ground truth rather than as noisy measurements, which repositions multi-annotator data collection as a way to capture epistemic uncertainty rather than merely to estimate a single correct label.","The benefit generalizes across domains: the observed improvements hold in both NLP (natural language inference, politeness) and vision (image categorization), indicating a domain-independent effect of training on distributional targets."],"fun_headline_variants":["Soft-label training preserves accuracy, cuts human-label divergence 32%","Soft labels match hard-label accuracy while capturing human uncertainty","Train on annotator distributions: uncertainty where humans disagree","Soft-label training: 32% closer to human judgment, same accuracy"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The argument rests on persistent annotator disagreement being genuine epistemic ambiguity rather than annotator noise, task artifact, or demographic labeling bias, and on the observed annotation distributions being reliable enough to train against even with as few as 6.7 annotators per sample.","fun_headline_variants_meta":{"raw":{"variants":["Soft-label training preserves accuracy, cuts human-label divergence 32%","Soft labels match hard-label accuracy while capturing human uncertainty","Train on annotator distributions: uncertainty where humans disagree","Soft-label training: 32% closer to human judgment, same accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001173,"raw_usage":{"total_tokens":4662,"prompt_tokens":696,"completion_tokens":3966,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":3896}},"tokens_in":440,"tokens_out":3966,"duration_ms":26123,"temperature":1.0,"reasoning_tokens":3896,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:39:28.289993+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a dataset with known true class-conditional distributions, then simulate annotator disagreement that is pure input-independent noise rather than genuine ambiguity; if soft-label training on these noisy distributions does not reduce KL divergence to the true distributions or does not maintain accuracy relative to hard-label training, the benefit would be specific to genuine disagreement rather than distributional targets generally.","supporting_citations":[],"review_version":1}