acceptodds
Under review as a conference paper at ICLR 2027

AccentMatch: Adaptive Competence Control via External Network Teaching

Abstract

In this paper, we propose AccentMatch, a unified framework that distills the reasoning capabilities of Large Language Models (LLMs) into efficient Semi-Supervised Learning (SSL) students. Unlike standard distillation, AccentMatch introduces a dynamic teacher annealing mechanism based on training dynamics. We utilize an LLM as a teacher to generate high-quality initial pseudo-labels. We then fuse these with the student’s own predictions, dynamically adjusting the weight of the LLM guidance based on the student's learning status of unlabeled examples. This ensures the student relies on the LLM pseudo-label when it has a low learning status for an unlabeled example and transitions to self-training as it stabilizes. We show that our approach outperforms both state-of-the-art SSL baselines and few-shot LLMs across six text classification benchmarks, setting a new standard for efficient, high-accuracy classification in low-resource regimes.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
semi-supervised learning, teacher annealing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nQaCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AccentMatch combines offline LLM-generated pseudo-labels with a student model’s predictions for semi-supervised text classification. Its main contribution is an instance-specific weighting mechanism based on an exponential moving average of prediction margins: the student receives more weight as its predictions become stable. The method builds on SoftMatch and is evaluated on six datasets with 20 and 50 labels per class. The paper reports improvements across most settings, supported by extensive ablations, multiple teacher and student models, and additional ten-run significance tests.

Critical strengths

A practical approach to controlling external supervision. The method addresses a meaningful problem: an LLM can provide useful early supervision but should not permanently constrain a student that becomes more accurate. Instance-specific weighting is a sensible solution, and Tables 2 and 4 show benefits over equal mixing, agreement filtering, fixed weights, and abrupt teacher removal.

Broad empirical support. The evaluation covers six datasets, several teachers, two student backbones, and extensions to other SSL methods. Appendix H reports significant improvements over the stronger of SoftMatch and VerifyMatch in 10 of 12 settings. This breadth makes the findings more convincing than a result limited to one benchmark or model.

Critical weaknesses

The experimental protocol is not sufficiently specified. RCV1 and GoEmotions support multiple labels per example, whereas the method uses a single argmax class and Appendix M explicitly requests exactly one label. The paper does not explain whether examples were filtered, labels were collapsed, or the method was adapted for multi-label prediction. This matters for interpreting the reported F1 scores and baseline comparisons. The exact Qwen checkpoint, dataset sizes and splits, and validation/model-selection procedure are also missing.

The justification for the particular gating rule remains incomplete. Proposition 3.1 establishes optimality for an oracle hard gate in a randomized prediction setting, assuming the teacher–student error crossover is already known. It does not establish optimality for sigmoid-weighted training targets. Moreover, the binary ablation switches at APM = 0, while the reported empirical crossovers occur around S = 0.6–0.7. Thus, that comparison does not cleanly isolate the benefit of smooth mixing from the choice of switching threshold. The experiments support adaptive fusion, but the specific gate is less well justified.

Questions for authors
  1. How are multiple labels handled for RCV1 and GoEmotions? Please specify preprocessing, treatment of Neutral, resulting dataset sizes, and whether every baseline uses the same formulation.

  2. Which exact Qwen checkpoint and training/model-selection settings were used? How are APM values initialized, updated across classes, and used during the ten-epoch warmup?

  3. Does continuous gating still outperform a hard gate whose threshold is selected on validation data? This would help distinguish the benefit of smooth weighting from a poorly located switching threshold.

4 comments

We thank the reviewer for the valuable and constructive feedback and suggestions that help improve the quality of our paper. We address the concerns below and we will incorporate them in our paper.

W1 & Q1 & Q2:

Thank you for pointing out these omissions. We clarify the exact preprocessing protocols, split sizes, and model checkpoints below, and will incorporate all of these details into Section 4.2 and Appendix A.

Multilabel Filtering & Label Spaces:

Both RCV1 and GoEmotions are evaluated strictly as single-label multiclass tasks:

  • GoEmotions: We preserve the official train and test splits and remove examples with multiple labels within each split. We also keep the neutral class, resulting in 28 classes: 27 emotions plus neutral.

  • RCV1: While the full RCV1 hierarchy contains 103 topic codes (referenced in Section 5 and Table 6), we restrict target annotations to the 82 leaf topic codes and keep only documents assigned to exactly one leaf code, discarding documents with multiple or zero leaf codes. We construct a custom split from the official 781,265-document test partition, assigning the first 200,000 documents chronologically (in ascending document-ID order) to training and the remaining 581,265 documents to testing, before applying this filtering. This yields 113,336 single-leaf training documents and 355,306 single-leaf test documents. We will update the class-count references in the text from 103 to 82 to reflect the leaf-code target space. Thank you again for raising this point.

Splits and Class-Wise Sampling:

  • For GoEmotions, IMDB, TREC-6, we use the standard benchmark train and test partitions. For Amazon Review and Yelp, we utilize the standard SSL text classification benchmark partition introduced in USB (Wang et al., 2022). RCV1 is discussed above.

  • Across all six datasets, we randomly sample k in {20, 50} labeled training examples per class without replacement. If a tail class in RCV1 or GoEmotions contains fewer than k examples after filtering, we include all available examples for that class in the labeled set L, which makes |L| slightly smaller than k * C. All remaining training examples form the unlabeled set U.

  • Baseline Parity: All baselines (supervised, LLM, and SSL) were run from scratch on these exact splits using identical evaluation metrics (macro-F1 for RCV1 and GoEmotions; accuracy for the remaining four datasets).

Reference: Yidong Wang et al., "USB: A Unified Semi-Supervised Learning Benchmark for Classification", NeurIPS 2022 Datasets and Benchmarks Track.

Final Dataset Split Sizes:

We provide the complete statistics across all six benchmarks below. We will also create a GitHub page for the project and make available our code and the exact splits for all datasets used in the paper for reproducibility purposes.

DatasetClasses (C)TestLabeled, 20/classLabeled, 50/classUnlabeled
IMDB225,0004010024,900
RCV182355,3061,4993,449109,887
GoEmotions284,5905601,38934,919
TREC-665001203005,152
Amazon Review5650,000100250249,750
Yelp Review550,000100250249,750

Main LLM Teacher Checkpoint:

Our main teacher is Qwen3-8B, specifically the Qwen/Qwen3-8B checkpoint, described in Yang et al. (2025). We utilize nucleus sampling (temperature = 0.7, top_p = 0.8) and a maximum generation length of 256 new tokens.

Reference: Yang et al. (2025). Qwen3 Technical Report. arXiv:2505.09388.

APM Warmup

For each unlabeled example, we maintain an APM value for every class, initializing each EMA with that class's first observed pseudo-margin. At each subsequent update, we compute the pseudo-margins and update the APMs for all classes (not just for the currently predicted class), using beta = 0.997.

The first ten epochs accumulate APM statistics without using them for adaptive teacher/student weighting, using an equal mixture (S = 0.5) of the teacher and student predictions while modulating the unsupervised loss with the confidence weight lambda(.) from Eq. 1. After warmup, we retrieve the accumulated APM of the student's current predicted class and apply the sigmoid to obtain its fusion weight S. All class-specific APMs continue to be updated throughout training, preserving their histories when the predicted class changes. We will clarify these details in the revised manuscript.

W2 & Q3:

Thank you for highlighting this distinction. We agree that Proposition 3.1 provides a conditional motivation for competence-based selection; it does not establish the optimality of the proposed sigmoid fusion. We will revise the theoretical discussion to make this scope explicit.

Per your suggestion, we evaluated the binary hard-switching variant across a full grid of switching thresholds, where the model uses only the LLM pseudo-label when the stability score is below the threshold and switches completely to the student's pseudo-label once it reaches or exceeds the threshold. As shown in the table below, tuning the switching threshold improves over extreme values (0.1 or 0.9) as well as the unshifted APM = 0 baseline (tau = 0.5), with the best hard-gate performance occurring at 0.4 and 0.6 across settings (specifically, 0.6 achieves 61.3 on RCV1-20, 28.5 on GoEmotions-20, and 28.4 on GoEmotions-50, while 0.4 reaches 60.9 on RCV1-50).

Crucially, however, even the best-performing hard threshold in every setting consistently underperforms continuous AccentMatch across all datasets and label regimes (e.g., 61.3 vs. 63.2 on RCV1-20 and 28.5 vs. 30.2 on GoEmotions-20). These results show that continuous AccentMatch outperforms any fixed hard gate across the entire threshold grid, including one selected via validation tuning. We will include these results in the paper.

Method / Hard Threshold (tau)RCV1 (20)RCV1 (50)GoEmotions (20)GoEmotions (50)
Hard Gate (tau = 0.1)59.559.926.827.2
Hard Gate (tau = 0.3)60.160.227.127.6
Hard Gate (tau = 0.4)60.860.927.928.3
Hard Gate (tau = 0.5, APM = 0)59.859.727.428.1
Hard Gate (tau = 0.6)61.360.728.528.4
Hard Gate (tau = 0.7)61.060.528.328.1
Hard Gate (tau = 0.8)60.460.327.828.0
Hard Gate (tau = 0.9)59.860.027.327.5
AccentMatch (Continuous Sigmoid)63.261.530.230.6

Thank you for the clarifications and additional experiments. They address my main concerns, and I will raise my score accordingly.

Thank you very much for your response and for raising your score. We appreciate your time, effort, and constructive feedback, which have helped improve the quality of the paper.

Reviewer_ReWXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AccentMatch, a semi-supervised learning framework that combines pseudo-labels generated by a llm teacher with predictions from a smaller student model. The method introduces an instance-specific teacher annealing mechanism that dynamically adjusts the contributions of teacher and student predictions based on the student's learning status. Specifically, it employs the APM to estimate prediction stability and uses a sigmoid-based weighting function to combine the two sources of supervision.

Critical strengths
  1. The paper addresses confirmation bias and unreliable pseudo-labels in low-resource semi-supervised learning.
  2. The paper provides experimental comparisons across multiple datasets and model configurations.
Critical weaknesses
  1. AccentMatch primarily combines LLM-generated pseudo-labels, APM, and confidence-based adaptive weighting, all of which are closely related to existing semi-supervised learning techniques. The proposed sigmoid-based teacher annealing mechanism appears to be a relatively straightforward extension of existing pseudo-label fusion strategies. The paper should better clarify its methodological novelty over prior work such as MarginMatch and other LLM-guided SSL approaches.
  2. The theoretical analysis assumes that APM reliably reflects student competence and that teacher–student reliability follows a crossover pattern. However, these assumptions are not guaranteed in practice. Moreover, the optimality result applies to a hard selection rule rather than the proposed sigmoid-based fusion mechanism. Consequently, the theoretical analysis does not directly establish the effectiveness or optimality of the actual training objective.
  3. The experiments focus primarily on six classification benchmarks with 20 or 50 labeled examples per class. Although the paper evaluates different teacher and student architectures, it remains unclear whether the proposed method remains effective with larger labeled datasets or more complex tasks. Experiments covering additional supervision regimes and broader NLP tasks would strengthen its generalizability.
  4. The proposed framework relies heavily on APM to estimate student reliability, but the experiments do not sufficiently establish whether APM is more effective than alternative uncertainty or competence measures. Comparisons with predictive entropy, confidence calibration, prediction consistency, or uncertainty-based teacher–student weighting would help isolate the contribution of the proposed competence estimation mechanism.
  5. AccentMatch assumes that LLM-generated pseudo-labels provide reliable guidance when the student has low competence. However, LLM teachers may produce systematically incorrect or biased predictions, particularly on specialized domains. Although different teacher models are evaluated, controlled experiments with varying pseudo-label noise levels would better characterize the method's robustness and potential failure cases.
Questions for authors

See the weakness.

4 comments

We thank the reviewer for the valuable and constructive feedback and suggestions that help improve the quality of our paper. We address the concerns below and we will incorporate them in our paper.

W1

Thank you for your comment. Our methodological contribution is a new competence-guided, instance-specific teacher-annealing mechanism that determines how authority is transferred from the generalist LLM teacher to the task-specific-tuned SSL student, throughout the entire training, using a continuous dynamic handover: the LLM pseudo-label receives greater weight for an example on which the student remains unstable, while the student progressively takes over only when its predictions across training iterations become consistently reliable for that example.

  • Comparison to MarginMatch. MarginMatch consists of a single SSL model and uses accumulated pseudo-margins (APMs) to identify potentially unreliable pseudo-labels and to filter out their contribution to training. Thus, MarginMatch relies on its own predictions, making these predictions susceptible to overconfident errors, which can reinforce themselves during training, leading to confirmation bias and error accumulation (as we show in Figure 1 and our results in Table 1). In contrast, AccentMatch consists of two independent models (LLM teacher and SSL student) and uses the APM signal to determine how to fuse the pseudo-label from the LLM with that from the SSL for the same example. When historical support for the student’s prediction is weak, the fused output (pseudo-label) gives greater weight to the external LLM teacher; as that support increases, the student receives greater weight and can depart from the teacher’s prediction. Thus, APM in AccentMatch controls the composition of the output pseudo-label, extending its use beyond filtering student-generated supervision.
  • Comparison to confidence-based weighting. The confidence weight for an unlabeled example determines how strongly that example contributes to the loss, whereas our score S determines the teacher-student composition of their pseudo-labels. When LLM and SSL predictions disagree, AccentMatch redistributes target probability mass between classes; scaling the loss alone does not provide this functionality.
  • Comparison to LLM-guided SSL. VerifyMatch uses teacher-student agreement and adaptive confidence thresholds to identify potentially mislabeled pseudo-labeled examples, which are denoised through MixUp with labeled examples. AccentMatch instead directly fuses teacher and student pseudo-labels for each example using its accumulated prediction history (APM). Its fusion coefficient can differ across examples at the same training step and can increase or decrease as that history changes.

As we can see from Table 1 from the paper, AccentMatch outperforms all these methods.

W2

Thank you for highlighting this distinction. We agree that Proposition 3.1 provides a conditional motivation for competence-based selection; it does not establish the optimality of the proposed sigmoid fusion. We will revise the theoretical discussion to make this scope explicit.

We also make the following observations:

First, the relationship between APM and student competence, and the existence of a teacher-student reliability crossover, are assumptions of the analysis rather than properties guaranteed by the construction of APM. Figure 2b and Appendix K provide empirical support for the intended relationship in the evaluated settings: student performance improves with increasing stability, and the student eventually outperforms the teacher in higher-stability regions. However, these observations do not establish the assumed relationship for every backbone or training stage. We will explicitly acknowledge this limitation in the paper.

Second, the proposition concerns the conditional prediction risk of selecting between fixed teacher and student predictions. Under the stated assumptions, its optimal solution is an oracle hard selector at the crossover threshold. This result motivates adapting supervision to student competence, but does not determine the sigmoid’s functional form, midpoint, or scale, nor does the risk guarantee automatically extend to a student learned through mixed-target training. We will describe the sigmoid as a practical, smooth weighting mechanism motivated by the analysis, without attributing the oracle selector’s optimality guarantee to it. We will revise the proposition’s title and the discussion to distinguish the conditional selection result, the practical design of the sigmoid gate, and the empirical evaluation of the training procedure. Thank you for your comment.

W3:

Thank you for the suggestion. We performed experiments with 100 and 200 labeled examples per class and we show the results comparing AccentMatch with SoftMatch and FlexMatch in the table below. Results are reported as mean ± standard deviation over three runs.

Dataset (metric)Labels per classFlexMatchSoftMatchAccentMatch
Amazon Review (accuracy)10057.6 ± 1.3758.9 ± 2.4160.8 ± 1.45
Amazon Review (accuracy)20059.1 ± 2.1759.7 ± 2.3361.1 ± 1.42
Yelp (accuracy)10052.7 ± 2.3158.6 ± 1.5460.8 ± 1.57
Yelp (accuracy)20057.5 ± 1.7661.2 ± 1.6762.7 ± 1.41

As results show, AccentMatch consistently outperforms both SoftMatch and FlexMatch on Amazon Review and Yelp in both the 100- and 200-labels-per-class setups, achieving higher mean scores across all four setups. This shows that our method continues to outperform strong baselines using larger labeled sets. We will add these results in the paper.

W4:

Thank you for this suggestion. To isolate the contribution of APM, we compare it with exponential moving averages of prediction entropy (normalized as 1 - H(p)/log(C) so higher values reflect higher certainty) and maximum softmax confidence as alternative instance-level student reliability measures within the exact same gating framework. The table below reports macro-F1 on RCV1 and GoEmotions for both 20 and 50 labels per class.

DatasetLabels/classAvg. EntropyAvg. ConfidenceAccentMatch (APM)
RCV12062.162.063.2
RCV15060.460.161.5
GoEmotions2025.127.230.2
GoEmotions5028.528.530.6

As shown in the table, APM consistently outperforms both alternatives across all four settings. Compared with the stronger alternative in each setting, APM improves macro-F1 by 1.1 points on RCV1 and by 2.1 to 3.0 points on GoEmotions. Unlike softmax confidence or entropy, which can remain high even when the student switches its argmax prediction across epochs, APM tracks class-specific logit margins that turn negative whenever predictions oscillate, providing a more discriminative signal of true student stability. We will include these comparisons in the revised manuscript.

W5:

To evaluate AccentMatch's robustness to noisy teacher supervision, we conducted controlled noise-injection experiments on IMDB and RCV1 (20 and 50 labels/class) by introducing 20% and 40% noise in the teacher’s labels. Precisely, because our unlabeled set is constructed by masking labels from the training set, we have access to the ground-truth labels beforehand; to strictly control the teacher's exact error rate, we start from these ground-truth labels and randomly flip 20% and 40% of correct labels to an incorrect class chosen uniformly at random from the remaining classes.

As shown in the table below, at 20% synthetic noise, AccentMatch not only outperforms both FlexMatch and SoftMatch across all settings, but on RCV1 it actually surpasses the default Qwen teacher (reaching 66.4 and 67.3 F1 vs. 63.2 and 61.5 for Default). This occurs because RCV1 is a challenging 82-class dataset where the real Qwen teacher has a natural error rate above 40% (58.6 F1 in Table 1), meaning a synthetic teacher with a 20% error rate (80% accuracy) provides substantially cleaner supervision than the actual LLM, whereas on IMDB the real Qwen teacher (~14.2% error rate, 85.8% accuracy) is cleaner than the 20% noisy oracle. Even under 40% synthetic noise, which approximates or slightly exceeds the teacher's natural error rate, AccentMatch degrades gracefully rather than collapsing: it remains on par with SoftMatch on IMDB (85.6 vs. 85.5 at 20/cls) and only dips slightly below SoftMatch on RCV1 (58.5 vs. 59.3 at 20/cls; 59.4 vs. 60.2 at 50/cls) while still outperforming FlexMatch (56.3 and 58.9). We will include these results and discussion in the paper.

MethodIMDB (20/cls)IMDB (50/cls)RCV1 (20/cls)RCV1 (50/cls)
FlexMatch84.186.356.358.9
SoftMatch85.588.159.360.2
AccentMatch (Default)87.789.863.261.5
AccentMatch (η=20\eta = 20\\% Noise)86.889.166.467.3
AccentMatch (η=40\eta = 40\\% Noise)85.688.058.559.4

I appreciate the authors’ response and their careful answers to my questions. I will adjust my score in the reviewer-AC discussion round.

Thank you very much for your response and for your willingness to adjust your score during the discussion phase. We appreciate your time, effort, and thoughtful feedback, which have helped strengthen our paper.

Reviewer_3C6FOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AccentMatch combines an external LLM teacher with a semi-supervised text classifier. It tracks an exponential moving average of per-example, per-class pseudo-margins, then maps the history for the student's current predicted class through a sigmoid to mix student and teacher pseudo-labels. Teacher labels are generated once before training, so the method does not require repeated LLM inference during optimization. Experiments cover six datasets, several SSL backbones, alternative teacher/student models, and extensive ablations. The practical idea is simple and useful, but the task definitions for two datasets and the calibration of the competence gate need clarification.

Critical strengths
  1. The method is practical: teacher calls are made once, and the training-time modification is small. The paper compares against equal weighting, teacher-student agreement filtering, a linear schedule, fixed weights, a hard switch, and a no-teacher variant. These controls make it plausible that the adaptive combination contributes beyond simply adding LLM labels. The results cover six datasets rather than a single favorable case.

  2. The experimental appendix addresses several common questions well. It includes alternative SSL methods and teacher/student backbones, EMA sensitivity, teacher confidence filtering, and ten-run paired tests. Appendix H reports significant gains in 10 of 12 settings, while RCV1 and Amazon with 50 labels per class are not significant. The repeated runs help distinguish consistent gains from differences that may be within run-to-run variation.

Critical weaknesses
  1. The label space and preprocessing are not sufficiently specified. Algorithm 1 and Appendix M describe single-class predictions, categorical targets, and a teacher instructed to choose exactly one class. RCV1 and GoEmotions are natively multilabel datasets, so the conversion or filtering policy is essential for interpreting these experiments. It is unclear whether multilabel examples are removed, reduced to one label, or handled by a different loss. The paper also lists 27 GoEmotions classes in the main text and 28 in Appendix F. Different choices would change both the difficulty and the meaning of macro-F1. The main teacher is identified as Qwen without a precise checkpoint, and the final split sizes are not given. These are material reproducibility issues, rather than minor omissions.

  2. The gate's calibration across models needs more evidence. Proposition 3.1 motivates a threshold under an assumed teacher/student risk crossover; it does not directly characterize training with mixed targets. The authors acknowledge the smooth approximation and provide relevant fixed-weight, beta=0, and crossover analyses. My remaining concern is the choice of an unshifted sigmoid: the empirical crossover in Figure 2b is near S=0.6 or 0.7, while equal weighting occurs at S=0.5. Rescaling logits can also change the weights without changing class predictions. These observations do not invalidate the method, but a scale-and-offset sensitivity test would clarify how reliably the gate transfers across backbones.

Questions for authors
  1. Please specify the exact RCV1 and GoEmotions datasets, train/validation/test sizes, handling of multilabel examples, treatment of neutral, and class-wise label sampling. Were all baselines rerun on precisely these processed splits with the same metrics? Please resolve the 27-versus-28 class discrepancy.

  2. Which Qwen checkpoint generated the main pseudo-labels, with what decoding settings? How were demonstrations selected and truncated for the 103-class RCV1 prompt under the 4,096-token limit?

  3. How sensitive are the results to a simple positive rescaling of student logits before computing margins, or to the sigmoid's slope and offset? This would show whether changing backbones requires recalibrating the gate.

  4. Appendix I reports cross-domain performance for AccentMatch. Can you include SoftMatch and a fixed teacher/student mixture under the same transfer protocol? Without those references, the absolute transfer scores do not establish that adaptive gating improves cross-domain robustness.

3 comments

We thank the reviewer for the valuable and constructive feedback and suggestions that help improve the quality of our paper. We address the concerns below and we will incorporate them in our paper.

W1 & Q1

Thank you for pointing out these omissions. We clarify the exact preprocessing protocols, split sizes, and model checkpoints below, and will incorporate all of these details into Section 4.2 and Appendix A.

Multilabel Filtering & Label Spaces:

Both RCV1 and GoEmotions are evaluated strictly as single-label multiclass tasks:

  • GoEmotions: We preserve the official train and test splits and remove examples with multiple labels within each split. We also keep the neutral class, resulting in 28 classes: 27 emotions plus neutral.

  • RCV1: While the full RCV1 hierarchy contains 103 topic codes (referenced in Section 5 and Table 6), we restrict target annotations to the 82 leaf topic codes and keep only documents assigned to exactly one leaf code, discarding documents with multiple or zero leaf codes. We construct a custom split from the official 781,265-document test partition, assigning the first 200,000 documents chronologically (in ascending document-ID order) to training and the remaining 581,265 documents to testing, before applying this filtering. This yields 113,336 single-leaf training documents and 355,306 single-leaf test documents. We will update the class-count references in the text from 103 to 82 to reflect the leaf-code target space. Thank you again for raising this point.

Splits and Class-Wise Sampling:

  • For GoEmotions, IMDB, TREC-6, we use the standard benchmark train and test partitions. For Amazon Review and Yelp, we utilize the standard SSL text classification benchmark partition introduced in USB (Wang et al., 2022). RCV1 is discussed above.

  • Across all six datasets, we randomly sample k in {20, 50} labeled training examples per class without replacement. If a tail class in RCV1 or GoEmotions contains fewer than k examples after filtering, we include all available examples for that class in the labeled set L, which makes |L| slightly smaller than k * C. All remaining training examples form the unlabeled set U.

  • Baseline Parity: All baselines (supervised, LLM, and SSL) were run from scratch on these exact splits using identical evaluation metrics (macro-F1 for RCV1 and GoEmotions; accuracy for the remaining four datasets).

Reference: Yidong Wang et al., "USB: A Unified Semi-Supervised Learning Benchmark for Classification", NeurIPS 2022 Datasets and Benchmarks Track.

Final Dataset Split Sizes:

We provide the complete statistics across all six benchmarks below. We will also create a GitHub page for the project and make available our code and the exact splits for all datasets used in the paper for reproducibility purposes.

DatasetClasses (C)TestLabeled, 20/classLabeled, 50/classUnlabeled
IMDB225,0004010024,900
RCV182355,3061,4993,449109,887
GoEmotions284,5905601,38934,919
TREC-665001203005,152
Amazon Review5650,000100250249,750
Yelp Review550,000100250249,750

Main LLM Teacher Checkpoint:

Our main teacher is Qwen3-8B, specifically the Qwen/Qwen3-8B checkpoint, described in Yang et al. (2025). We utilize nucleus sampling (temperature = 0.7, top_p = 0.8) and a maximum generation length of 256 new tokens.

Reference: Yang et al. (2025). Qwen3 Technical Report. arXiv:2505.09388.

W2 & Q3:

Thank you for raising the concern about the choice of an unshifted sigmoid. The unshifted sigmoid uses zero accumulated margin as a natural reference: it separates positive from negative historical support for the student’s current pseudo-label. However, this reference need not coincide with the point where the student becomes more reliable than the teacher. Accordingly, Figure 2b supports the usefulness of APM for ordering examples by student reliability, but does not establish that equal weighting at S=0.5 is optimal. We agree that the observed crossovers require testing a shifted gate. We also agree that positive logit rescaling preserves class predictions while changing margin magnitudes and hence the fusion weights. Temporal averaging stabilizes the margin signal but does not remove this scale dependence. Our results across BERT and RoBERTa provide evidence that the default gate transfers to these backbones, but do not isolate its sensitivity to rescaling. To address this directly, we evaluate a shifted sigmoid Sτ,b(x)=σ(APM⁡(x)−bτ),S_{\tau,b}(x)=\sigma\left(\frac{\operatorname{APM}(x)-b}{\tau}\right),

where >0 controls the transition’s sharpness and b controls its midpoint; the original gate corresponds to (,b)=(1,0). We test τ∈0.5,1,1.5\tau\in\\{0.5,1,1.5\\} and b∈−0.847,−0.405,0,0.405,0.847b\in\\{-0.847, -0.405, 0, 0.405, 0.847\\}, using the same grid across BERT and RoBERTa on RCV1 and Amazon Review. The nonzero offsets −0.847,−0.405,0.405,0.847\\{-0.847, -0.405, 0.405, 0.847\\} correspond to S=0.3,0.4,0.6,0.7S = \\{0.3, 0.4, 0.6, 0.7\\}, directly probing the reviewer’s observation. We report downstream performance below:

Backbone / Offset (b)RCV1 20 labels/class (tau = 0.5 / 1.0 / 1.5)Amazon Review 20 labels/class (tau = 0.5 / 1.0 / 1.5)
BERT: b = -0.845 (S = 0.3)61.3 / 61.6 / 61.452.0 / 52.2 / 52.1
BERT: b = -0.405 (S = 0.4)62.2 / 62.5 / 62.352.6 / 52.8 / 52.7
BERT: b = 0.000 (S = 0.5, Default)62.8 / 63.2 / 62.952.9 / 53.2 / 53.0
BERT: b = 0.405 (S = 0.6)63.1 / 63.4 / 63.153.1 / 53.3 / 53.1
BERT: b = 0.847 (S = 0.7)62.7 / 62.9 / 62.653.0 / 53.1 / 52.9
RoBERTa: b = -0.845 (S = 0.3)61.6 / 61.9 / 61.752.2 / 52.5 / 52.3
RoBERTa: b = -0.405 (S = 0.4)62.5 / 62.8 / 62.652.8 / 53.1 / 52.9
RoBERTa: b = 0.000 (S = 0.5, Default)63.1 / 63.5 / 63.253.2 / 53.5 / 53.3
RoBERTa: b = 0.405 (S = 0.6)63.3 / 63.6 / 63.453.4 / 53.6 / 53.4
RoBERTa: b = 0.847 (S = 0.7)62.8 / 63.1 / 62.953.2 / 53.4 / 53.2

These results confirm that our default parameter-free gate (=1.0,b=0) lies within a broad, flat optimum and transfers reliably across BERT and RoBERTa without requiring per-backbone or per-dataset recalibration.

Q2:

We use Qwen3-8B (Qwen/Qwen3-8B; Yang et al., 2025) with nucleus sampling (temperature = 0.7, top_p = 0.8) and a maximum generation length of 256 new tokens. For datasets with large label spaces such as RCV1, where including demonstrations for all classes would exceed the 4,096-token input budget, we randomly sample demonstrations from the labeled set L until the 4,096-token budget is exhausted. We will clarify these details in the revised manuscript.

Q4:

Thank you for this suggestion. We agree that the absolute transfer scores alone do not isolate the benefit of adaptive gating. We therefore include SoftMatch and a fixed teacher/student mixture (S=0.5, corresponding to AccentMatch_naive in Table 2) under the same transfer protocol. The table below reports accuracy for both in-domain and cross-domain evaluation.

Source → TargetLabels/classSoftMatchFixed mixture (S=0.5)AccentMatch
Amazon → Amazon2051.752.553.2
Amazon → Amazon5053.754.054.6
Yelp → Amazon2035.236.236.4
Yelp → Amazon5039.639.740.4
Amazon → Yelp2033.734.635.4
Amazon → Yelp5040.140.441.2
Yelp → Yelp2050.450.752.7
Yelp → Yelp5055.956.356.9

As we can see from the table, AccentMatch consistently outperforms both baselines in both transfer directions and label settings, improving cross-domain accuracy by 0.8–1.7 percentage points over SoftMatch and 0.2–0.8 points over the fixed mixture. We will add these comparisons to Appendix I.

Thank you again for your time and effort in reviewing our paper. Please let us know if our responses have addressed your concerns, or if we can provide any additional information.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.