{"id":"2125764c-4fe8-4340-a389-4a2ddd295692","arxiv_id":"2504.15432","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Fine-tuning BERT-class models on LLM-generated labels consistently produces less accurate, more unstable classifiers that plateau early, and common fixes only partially help.","lead":"This paper shows that using AI-generated labels to train smaller text classifiers degrades accuracy, increases run-to-run instability, and causes early learning plateaus. It is a cautionary empirical study for anyone building low-cost text classifiers from LLM annotations, especially in high-stakes settings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's IMDB variance (σacc: gold 0.28 vs 3B 0.16 and 7B 0.20) contradicts the abstract's 'increased instability' claim, undermining a central finding.","rationale":"The reader's weakest assumption focuses on the premature-plateau claim being confounded by the 5000-point training cap. That is a valid and load-bearing concern, and I agree that the plateau conclusion lacks statistical support and data beyond the cap. However, I identified a more direct internal inconsistency: the paper's own variance data in Table 1 contradict the abstract's blanket statement of 'increased instability across training runs' on IMDB, the simplest task. Since the paper's methods explicitly equate stability with variance, this is a concrete factual contradiction rather than an absence of evidence. The plateau concern is also serious, but it hinges on missing external evidence; the instability issue is visible in the paper's own reported numbers. Both concerns point to the need for revision, but the verdict remains CONDITIONAL because the core performance-degradation finding is solid and the instability might still hold on harder datasets (Ecommerce, Manifestos, Toxic) despite the IMDB counterexample. The authors should qualify or correct the instability claim and provide the raw data for verification. I therefore recommend keeping the reader's CONDITIONAL verdict unchanged, while noting that the instability inconsistency is the more immediate correctness risk.","tokens_in":9858,"tokens_out":8596,"duration_ms":73428,"concrete_test":"Recompute per-run accuracy and macro-F1 for IMDB and Manifestos from the raw experimental logs (or from the five individual runs if code/data are released) and run Levene's test for equality of variances between the gold and 3B/7B conditions. If the IMDB variance is significantly lower for synthetic labels, the 'increased instability' claim must be qualified to harder datasets or to individual-level stability metrics. Also inspect the five Manifestos 3B runs to verify whether σacc=0.00 is accurate; if any run differs, the table entry is erroneous.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim includes 'increased instability across training runs' (Abstract, Section 3). Its own methods (§2.2) define stability via the variance of accuracy and macro-F1, reporting these as σacc and σf1 in Table 1. On IMDB, the gold condition has σacc=0.28 and σf1=0.29, while the 3B condition has σacc=0.16 and σf1=0.16, and the 7B condition has σacc=0.20 and σf1=0.20. Thus, synthetic labels reduce variance on this dataset, directly contradicting the universal claim of increased instability. Additionally, the Manifestos 3B row reports σacc=0.00, which is suspicious: five runs with identical accuracy despite αK=52.72 and puc=50 imply exact cancellation of prediction flips, which is unlikely and may indicate a data error. This internal inconsistency means the instability finding is not robust across the reported experiments; the paper must either restrict the claim to specific datasets/measures or reconcile the variance data with the individual-level measures (αK, puc). If the authors intended 'instability' to refer only to αK/puc, that conflicts with their own methodological emphasis on variance. This undermines a headline contribution of the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper empirically studies the use of LLM-generated labels (Qwen2.5-Instruct 3B/7B) to fine-tune RoBERTa-base classifiers on four text classification datasets (IMDB, ECommerce, Manifestos, Toxic). It reports that models trained on synthetic labels suffer accuracy/F1 degradation, increased prediction instability as measured by Krippendorff's alpha and the proportion of unchanged predictions, and premature performance plateaus, relative to models trained on gold labels. The paper proposes a bias-variance decomposition to interpret the degradation, and evaluates two mitigation strategies: entropy-based filtering and consistency ensembles. It concludes that these mitigations only partially recover the gap and do not fully resolve the instability and plateau issues.","tokens_in":10116,"tokens_out":7271,"duration_ms":58310,"significance":"If correct, the paper provides a useful empirical cautionary study that goes beyond simple accuracy comparisons, and its comparison against gold labels avoids circularity. The theoretical interpretation uses a standard Hessian-based bias-variance decomposition without fitted free parameters, and the paper gives detailed prompt and training protocols. However, the significance is limited by the absence of significance tests, the small number of runs (five), the borrowed Figure 1, and several internal inconsistencies in the reported stability and mitigation results, which currently weaken the headline claims.","major_comments":[{"comment":"The universal claim of \"increased instability\" is contradicted by the variance measures on IMDB, where gold σacc=0.28 and σf1=0.29 are higher than both synthetic conditions (3B: 0.16/0.16; 7B: 0.20/0.20). Since Section 2.2 defines stability via these variances, the paper must either restrict the instability claim to the αK/puc measures or explain why variance is not the operative stability measure on this dataset.","section":"Section 2.2 and Table 1"},{"comment":"σacc=0.00 with five runs is implausible given αK=52.72 and puc=50, which imply substantial prediction flips. Please verify the data and report the actual run-level values; if correct, explain how zero accuracy variance coexists with only 50% prediction overlap.","section":"Table 1, Manifestos 3B row"},{"comment":"The premature plateau claim lacks support. Figure 1 is borrowed from Csanady et al. (2024) and shows \"LlamBERT\" rather than the RoBERTa models used here, and Figure 2 has no error bars or significance tests. Because training is capped at 5,000 examples (Appendix A), the plateau may be an artifact of the cap; provide data beyond the cap or a statistical test (e.g., comparing learning-curve slopes) to show synthetic labels plateau before gold labels.","section":"Section 3, Figures 1–2"},{"comment":"The introduction claims that \"neither strategy stabilizes training variance,\" but the consistency ensemble reduces σacc and σf1 on all four datasets (e.g., Toxic σacc from 5.31 to 0.69; IMDB from 0.20 to 0.02). Please correct this misstatement or qualify it with respect to the gold baseline.","section":"Section 1 vs Table 2"},{"comment":"The reported µf1=83.45 is 10 points below µacc=93.46 on a balanced binary task and is far below the synthetic-label µf1=92.94, which would imply the ensemble drastically reduces macro-F1. If this is a typo (e.g., 93.45), fix it; otherwise, the claim that the ensemble \"seems to work\" is not supported on this dataset.","section":"Table 2, IMDB Ensemble row"},{"comment":"The abstract's claim that mitigations \"recover only 60–75% of the gold-label performance gap\" is not supported by the table: on Manifestos the ensemble recovers roughly (81.58−71.51)/(83.56−71.51) ≈ 84% of the accuracy gap. Please either recompute the range or adjust the claim.","section":"Section 1 and Table 2"}],"minor_comments":[{"comment":"The caption says \"Plot from Csanady et al. (2024)\" and the figure legend refers to \"LlamBERT,\" not the RoBERTa models used in this paper; clarify the provenance and relation to the present experiments, or replace with the paper's own data.","section":"Figure 1 caption"},{"comment":"\"All numbers are scaled up by 100\" is ambiguous for αK and puc, which are already percentages; state the units for each metric explicitly.","section":"Table 1 caption"},{"comment":"The approximation in the bias-variance decomposition is stated without justification; cite the Hessian-based expansion in Heskes (1998) and note its validity conditions.","section":"Section 3.1, Eq. (1)"},{"comment":"The sentence \"entropy-based filtering is theoretical more appealing simple alternative simple uncertainty estimation heuristics\" is ungrammatical and should be rewritten for clarity.","section":"Section 5, first paragraph"},{"comment":"\"as representatives LLMs in their respective weight classes\" should be \"as representative LLMs\".","section":"Section 2.1"},{"comment":"The statement \"shrink the size of test datasets with <= 2000\" is unclear; specify how many samples were used for each test set.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The IMDB ensemble F1 value (83.45) and the Manifestos 3B σacc=0.00 both look like possible data errors; please have the authors verify them and, if possible, release run-level results. The paper would also benefit from significance testing or at least an explicit acknowledgment that five runs limit inferential strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing worth knowing about this paper is that it measures something most of the LLM-annotation literature doesn't: run-to-run instability of the fine-tuned classifier, using Krippendorff's alpha and the proportion of unchanged predictions across five training runs per condition, on four datasets. That is a real, useful addition. The authors also deserve credit for testing mitigation heuristics (entropy filtering, consistency ensembles) and reporting that these only partially close the gap.\n\nThe central degradation claim holds. Table 1 clearly shows synthetic-label training underperforms gold, especially on tail classes, and the theoretical framing via Heskes's bias-variance decomposition is a reasonable, parameter-free way to explain plateaus. The citation pattern looks fine; no red flags.\n\nNow the soft spots, in order of importance.\n\nFirst, the abstract's 'increased instability across training runs' is too broad for the paper's own table. On IMDB, sigma_acc for gold is 0.28, whereas 3B gives 0.16 and 7B gives 0.20 — synthetic labels reduce variance on that dataset. The alpha_K and puc values do move in the 'more unstable' direction, but the variance-based instability claim is contradicted on IMDB. The paper should either qualify the claim by dataset/measure or explain why variance is the right indicator in some cases. The Manifestos 3B row reporting sigma_acc=0.00 with alpha_K=52.72 also looks suspicious.\n\nSecond, the plateau claim rests on weak footing. Figure 2 has no error bars or significance tests; Figure 1 is borrowed from Csanady et al. (2024), not the present experiments; and the 5000-point cap means there is no evidence that gold labels would not also plateau at that point. This needs either scaling data or a softer conclusion.\n\nThird, reproducibility: no code, no data, and the mitigation section doesn't specify which entropy percentile (5 or 25?) produced Table 2, nor how many ensemble runs were used. Also, Table 2's IMDB ensemble row reports macro-F1=83.45 with accuracy=93.46 on a balanced binary set — that gap looks like a data entry error.\n\nNone of this kills the core finding. The degradation and individual-level instability are documented honestly, and the limitations section is candid about what's missing. The paper is a usable empirical warning shot for anyone who uses LLM-generated labels to fine-tune BERT-class classifiers.\n\nBottom line: send it out for review, with a request for major revisions — qualification of the instability claims, more rigorous plateau analysis, and release of code/data. I'd bring it to the reading group.","headline":"Core finding is real, but the instability claim overreaches and the plateau evidence is thin.","tokens_in":10620,"tokens_out":6006,"would_cite":true,"duration_ms":48939,"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":"Fine-tuning BERT on LLM-generated labels degrades accuracy and, more importantly, makes training unstable and prone to early plateaus.","keywords":["LLM-generated labels","synthetic data","text classification","fine-tuning","RoBERTa","prediction instability","performance plateau","entropy filtering"],"falsifier":"Train gold-label and LLM-label RoBERTa models on the same four datasets with training sizes beyond 5000 (for example 10k and 20k) and compare the slopes of their learning curves; if the gold-label curve continues improving while the synthetic-label curve flattens, the plateau stands, and if both flatten together it does not.","tokens_in":9653,"feed_emoji":"⚠️","tokens_out":4372,"duration_ms":35280,"temperature":0.7,"pith_summary":"The paper sets out to show that using LLM-generated labels to fine-tune smaller classifiers such as RoBERTa costs more than accuracy: it makes training runs unpredictable and causes learning to plateau early. These effects appear across four text-classification datasets, are worst on complex or imbalanced tasks, and persist when the annotating LLM grows from 3B to 7B parameters. The authors offer a theoretical explanation based on the divergence between the LLM's label distribution and the true label distribution, and test simple mitigations that only partially close the gap. A sympathetic reader would take this as a caution that cheap LLM annotation is not a safe substitute for human labels in high-stakes applications.","feed_headline":"LLM labels make BERT training unstable and plateau early","feed_subtitle":"Run-to-run predictions flip and bigger LLMs don't help; filters recover only part of the gold-label gap.","key_machinery":"The central analytic object is the bias/variance-style decomposition of the classifier's expected KL error: $$\\text{Error}(\\hat P) \\approx \\mathrm{KL}(P\\,\\|\\,P_S) + \\mathbb{E}_{D_S}\\left[\\mathrm{KL}(P_S\\,\\|\\,\\hat P)\\right].$$ The first term, the irreducible approximation error from using the LLM's conditional label distribution, is what produces the plateau; the second explains why variance inflates where the LLM is off. On the mitigation side, the paper tests an entropy-ranked filtering function $F(x)$ that mixes gold labels into high-uncertainty inputs, and a consistency ensemble that replaces flip-prone predictions, recovering 60–75% of the performance gap.","core_discovery":"The paper claims that fine-tuning a smaller encoder-only text classifier (RoBERTa-base) on LLM-generated labels, rather than human gold labels, degrades more than accuracy: runs become less reproducible (Krippendorff's alpha drops, predictions flip across seeds), and learning curves flatten prematurely as training data grows. These effects intensify with task complexity and class imbalance, and scaling the annotator from 3B to 7B parameters does not remove the instability. The authors attribute the pattern to irreducible approximation error in the LLM's label distribution $P_S(Y|X)$ that no amount of training data can average away, with estimation error amplified wherever $P_S$ diverges from the true $P$.","pith_inferences":["If the plateau generalizes beyond the 5000-point cap, then acquiring more unlabeled data will not rescue a synthetic-label pipeline; the ceiling is set by annotator bias, not sample size.","The paper's own error decomposition suggests a testable hierarchy: label-only synthetic data should beat synthetic text-plus-label pairs for the same data budget, a comparison the paper notes but does not run.","Because entropy filtering failed worst on the Toxic dataset, LLM confidence scores may be least trustworthy exactly where the annotation decision is hardest; uncertainty-based fixes need calibration checks per dataset.","A practical extension would be to adopt instability metrics like Krippendorff's alpha and the proportion of unchanged predictions as standard diagnostics in any LLM-annotation pipeline."],"forward_implications":["On complex or imbalanced datasets, LLM labeling can cause large accuracy drops (for example 66% versus 96% on Ecommerce with the 3B annotator), and scaling to 7B does not reliably close them.","Training instability is not a small-model artifact: Krippendorff's alpha and unchanged-prediction rates drop even on simple sentiment data, so a single fine-tuning run with LLM labels is not a trustworthy estimate.","Entropy-based filtering and consistency ensembles recover only 60–75% of the gold-label performance gap and do not restore training stability, so post-hoc fixes are not a full remedy.","Minority classes are systematically under-learned with synthetic labels, a pattern the paper reads as a mild form of model collapse during synthetic-data training."],"supporting_citations":[{"why":"Supplies the LlamBERT pipeline and the figure the paper reuses to introduce the premature-plateau comparison.","marker":"Csan´ady et al. (2024)"},{"why":"Gives the model-collapse concept used to interpret systematic under-learning of minority classes.","marker":"Shumailov et al. (2024b)"},{"why":"Provides the bias/variance decomposition that yields the irreducible approximation error and amplified variance terms.","marker":"Heskes (1998)"},{"why":"Documents LLM limitations on subjective classification tasks, cited to explain why plateaus appear.","marker":"Li et al. (2023)"},{"why":"Earlier work claiming GPT-3 labels can nearly match human annotations; the paper's instability results challenge that parity assumption.","marker":"Wang et al. (2021)"},{"why":"Shows few-shot fine-tuning pathologies that the paper cites as evidence of systematic error propagation.","marker":"Chen et al. (2022)"},{"why":"Defines the RoBERTa-base encoder used as the classifier in every experiment.","marker":"Liu et al. (2019)"}],"fun_headline_variants":["LLM labels breed unstable BERT classifiers","Synthetic labels make BERT runs unreproducible","Filters can't fix LLM label noise in BERT","Bigger LLM annotators still destabilize BERT","LLM annotations: BERT classifiers at risk"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premature plateau result depends on the decision to cap every training set at 5000 data points; if the gold-label baseline also plateaus by that budget, the seeming plateau is an artifact of the cap rather than a failure mode of LLM labels.","fun_headline_variants_meta":{"raw":{"variants":["LLM labels breed unstable BERT classifiers","Synthetic labels make BERT runs unreproducible","Filters can't fix LLM label noise in BERT","Bigger LLM annotators still destabilize BERT","LLM annotations: BERT classifiers at risk"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1147,"prompt_tokens":842,"completion_tokens":305,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":230}},"tokens_in":458,"tokens_out":305,"duration_ms":2817,"temperature":1.0,"reasoning_tokens":230,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:26:14.465345+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train gold-label and LLM-label RoBERTa models on the same four datasets with training sizes beyond 5000 (for example 10k and 20k) and compare the slopes of their learning curves; if the gold-label curve continues improving while the synthetic-label curve flattens, the plateau stands, and if both flatten together it does not.","supporting_citations":[{"cited_title":"Pathologies of Pre-trained Language Models in Few-shot Fine-tuning","cited_arxiv_id":"2204.08039","evidence_quote":"Shows few-shot fine-tuning pathologies that the paper cites as evidence of systematic error propagation."}],"review_version":1}