acceptodds
Under review as a conference paper at ICLR 2027

Structured Auxiliary Supervision for Continuous Imbalanced Financial Sentiment Analysis

Abstract

Financial sentiment regression with large language models (LLMs) trains on neutral-dominated data in which the extreme scores that drive financial decisions are rare, so fine-tuned models pull their predictions toward the dense center, a failure mode we term neutral collapse. Existing remedies reweight samples by label density or classify over a partition fixed before training, leaving the category structure of the continuous label space unlearned, and an auxiliary classification task added to supply it interferes with regression and converges at a different rate, so no fixed task weight serves both. We propose Structured Auxiliary Supervision (SAS), a plug-and-play framework that makes the label partition a trained parameter and schedules its supervision by the convergence trajectory of each task. Differentiable Structured Partitioning (DSP) learns thresholds that split the sentiment space into balanced auxiliary categories through differentiable soft assignments, giving the sparse extremes their own supervision targets. Trajectory-based Convergence Balancing (TCB) sets the task weights from windowed convergence velocities, shifting gradient toward the slower-converging task, and weights the auxiliary categories by a softmax over their losses. The two components are coupled in one loss and one backward pass, and the auxiliary head is discarded after training, so SAS adds no inference cost to the backbone. On three financial datasets and three backbone scales, fixed-weight multi-task fine-tuning (Fixed-MTL) is worse than Single-task regression in all nine settings, whereas SAS attains the lowest MSE in all nine, 10.5% below Fixed-MTL, a reduction that stays between 9.9% and 11.6% across a 24-fold increase in backbone size. On RoBERTa-base it has the lowest MSE among twelve baselines under identical protocols, and on public financial and cross-domain benchmarks it lowers MSE by 9.4% to 10.7%; DSP alone yields 52% of the joint MSE gain and TCB alone 30%. On NEU/RoBERTa-base, SAS compresses the extreme-to-neutral MSE ratio from 3.85 to 1.95, and its advantage widens as imbalance grows: under 20 label skew, its MSE rises by 25% versus 80% for Single-task regression and 92% for Fixed-MTL.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
imbalanced regression, financial sentiment analysis, auxiliary classification, differentiable label partitioning, multi-task loss balancing, LLM fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dqmSOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the severe imbalance problem in continuous regression, using financial sentiment regression as a representative testbed, and identifies two coupled challenges: at the data level, labels are concentrated in the neutral region, causing predictions to collapse toward neutrality during MSE fine-tuning (neutral collapse); at the task level, the auxiliary classification task interferes with regression, and the two tasks converge at asymmetric rates, making it impossible to optimize both simultaneously with fixed task weights.

The authors propose the SAS framework: DSP parameterizes partition thresholds in the continuous label space as trainable variables, jointly learning the auxiliary classification partitions through soft assignment and three regularization terms; TCB dynamically adjusts task weights based on windowed convergence speed derived from EMA loss trajectories, and reweights auxiliary classes according to class-specific losses. Both components are jointly optimized within a single loss function and one backward pass, with the auxiliary head discarded after training, incurring no additional inference cost.

Experiments cover 3 private financial, 3 public financial, and 4 cross-domain datasets, involving 7 backbones and 12 baselines, including main experiments, ablation studies, per-region MSE analysis, imbalance sweep, boundary condition evaluation, and cost analysis.

Critical strengths
  1. The problem motivation is clear, and both challenges are directly supported by empirical evidence.
  2. The method design is reasonable and the engineering is complete.
  3. The experimental coverage is comprehensive.
  4. The ablation and boundary analysis are honest.
  5. The cross-domain validation design is appropriate.
Critical weaknesses

Scope misalignment and missing downstream evaluation.

The abstract and introduction heavily emphasize financial sentiment analysis as the core motivation and application scenario. Because financial sentiment analysis typically serves as an upstream component of quantitative trading systems in practice, readers naturally expect the subsequent experiments to include downstream evaluations—such as trading simulations, backtests across stock markets, or the translation of MSE reductions into trading win rates and profitability. However, the paper only reports standard ML regression metrics (MSE, MAE, Pearson r) and auxiliary classification metrics, without any downstream trading or economic evaluation. Furthermore, the Limitations section fails to explicitly clarify this boundary.

The claimed MSE improvement is marginal, statistically questionable, and comes at a high cost.

The authors highlight a "10.5% MSE reduction," but this number is measured against a weak baseline (Fixed-MTL). If you compare it to the Single-task model, the gain shrinks to only about 6.3%. More critically, when compared to the strongest competing method (DWA), the difference is just 0.0002 (0.0193 vs. 0.0195). The authors mention in §5.1 that the multi-seed standard deviation is up to 0.0008 MSE. This means the gap between SAS and DWA is 4 times smaller than the seed noise itself, making it highly likely to be statistically insignificant. Although the authors state they use Holm-Bonferroni tests "where reported," the main tables contain no significance markers (e.g., p-values or asterisks) for these comparisons. Thus, readers cannot verify if SAS truly outperforms DWA on regression, or if the gain is just random seed variance.

  1. On top of that, Table 27 shows SAS requires 35.7% more training time than DWA (38h vs. 28h). Spending nearly 1.5 times the training cost for an MSE drop that cannot be statistically distinguished from seed noise is a very poor trade-off. The paper needs to provide a clearer cost-accuracy Pareto analysis or explicitly justify why this extra overhead is worth it.
Questions for authors

Same to Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer dqmS

Dear Reviewer dqmS:

Thank you for recognizing the clear motivation, the method design, and the comprehensive experiments. As your questions are the same as the weaknesses, we respond to the weaknesses W1 to W3 below. We have responded point by point to your concerns regarding downstream evaluation, significance tests, and cost report. In the marked copy appended to the revised PDF (pages 47–94), the revisions made in response to your comments are tagged with @Reviewer dqmS. We hope our responses resolve your concerns and look forward to further discussion.


Summarized W1: Add downstream trading evaluation and clarify the Limitations

A1: Following your suggestion, we further evaluate the downstream financial decisions with a daily trading simulation on EUR/USD, where the average predicted sentiment of each day's texts in the held-out test split of NEU sets a long, short, or flat position for the next day, on metrics (including hit rate, annualized return, and Sharpe ratio). Table R1 shows that SAS reaches the highest hit rate, annualized return, and Sharpe ratio of the four methods in this simulation, as its Sharpe ratio reaches 0.66 against 0.47 for DWA, the strongest of the three baselines.

Table R1: Downstream trading simulation on EUR/USD.

MethodHit rate (%) ↑Annualized return (%) ↑Sharpe ratio ↑
Single-task51.82.90.38
Fixed-MTL51.42.30.31
DWA52.33.60.47
SAS (Ours)53.65.10.66

Following your suggestions, we have added these results to Appendix A.22 (Lines 2415–2425 and 2474–2479), and updated the Limitations in Appendix A.0 (Line 809) to state that backtests across markets remain future work.


Summarized W2 (1): Headline gain against Single-task and the strongest baseline

A2 (1): We respectfully clarify that Fixed-MTL is the headline reference because it shares the backbone and the auxiliary head of SAS and differs in DSP and TCB alone, which isolates the contribution of the two components. To report the gain against the other references as well, Table R2 lists the MSE reduction of SAS over Fixed-MTL, Single-task, and the strongest baseline (i.e., DWA on NEU and FXE, and FAMO on INV) on the original validation split of NEU, FXE, and INV in Table 29 (Lines 2053–2075).

Table R2: MSE reduction of SAS over three references.

ReferenceNEUFXEINV
Fixed-MTL0.0216 → 0.0193 (−10.6%)0.0225 → 0.0203 (−9.8%)0.0220 → 0.0188 (−14.5%)
Single-task0.0206 → 0.0193 (−6.3%)0.0217 → 0.0203 (−6.5%)0.0199 → 0.0188 (−5.5%)
Strongest baseline0.0195 → 0.0193 (−1.0%)0.0206 → 0.0203 (−1.5%)0.0196 → 0.0188 (−4.1%)

Following your suggestions, we have revised Appendix A.12.4 (Lines 2039–2042) to report these validation reductions, the conclusion (Lines 478–485) to report the corresponding test-split reductions over Single-task and the strongest of all compared methods, and Table 3 (Lines 383–392) to add the change against the best other baseline on the other financial and cross-domain datasets.


Summarized W2 (2): Significance of the gain over DWA within seed noise

A2 (2): To check whether SAS truly outperforms DWA beyond seed variance, we report the per-seed MSE of SAS and the closest of the twelve baselines (i.e., DWA on NEU, and FAMO on FXE and INV) on a held-out test split of NEU, FXE, and INV (i.e., selecting on 10% of the training data and reporting the original validation split once as the test split), together with a two-sided paired t-test with Holm correction over the three datasets. The methods share the split and initialization of each seed, and the paired test compares their MSE within each seed. Table R3 shows that SAS has a lower MSE than the closest of the twelve baselines in each of the five seeds on the three datasets.

Table R3: Per-seed test MSE of SAS and the closest of the twelve baselines.

DatasetMethodSeed 42Seed 43Seed 44Seed 45Seed 46Mean ± std
NEUDWA0.01950.02050.01990.02090.01970.0201 ± 0.0006
NEUSAS (Ours)0.01910.02030.01960.02050.01950.0198 ± 0.0006
FXEFAMO0.02020.02140.02090.02160.02090.0210 ± 0.0005
FXESAS (Ours)0.02010.02120.02040.02130.02050.0207 ± 0.0005
INVFAMO0.01940.02070.01990.02040.02010.0201 ± 0.0005
INVSAS (Ours)0.01880.01990.01900.01970.01910.0193 ± 0.0005

Table R4 shows that the gain over the closest of the twelve baselines is significant after Holm correction on the three datasets, as the corrected p-value is 0.0051 on NEU.

(continued)

Table R4: Paired t-test of SAS against the closest of the twelve baselines.

DatasetMean differenceCohen's dHolm-corrected p
NEU0.00033.000.0051
FXE0.00031.900.0132
INV0.00085.060.0010

Following your suggestions, we have added these p-values to Appendix A.12.4 (Lines 1995–1997, 2035–2037 and 2107–2122).


Summarized W3: Cost-accuracy trade-off against DWA

A3: We respectfully clarify that the extra training time of SAS is a one-time cost, and SAS keeps the inference cost of the original backbone, since the auxiliary head is discarded after training. To show the cost-accuracy trade-off, we combine the training time and memory in Table 33 with the MSE on NEU in Table 29. Table R5 shows that SAS lies on the cost-accuracy Pareto front with the lowest MSE, together with Single-task and DWA, where the 10 extra hours over DWA are spent once in training, and the gain over DWA on NEU is significant on the test split as shown in Table R4.

Table R5: Training time (one seed, summed over NEU, FXE, and INV), memory, and validation MSE on NEU.

MethodTraining time (h) ↓Memory (GB) ↓MSE ↓Pareto front
Single-task24.022.10.0206✅
Fixed-MTL27.022.40.0216❌
DWA28.022.50.0195✅
LDS29.022.60.0201❌
FAMO32.023.00.0198❌
SAS (Ours)38.023.80.0193✅

Following your suggestions, we have revised Appendix A.13 (Lines 2128–2130 and 2174–2181) to present this Pareto analysis.


Best Wishes,

The Authors of Paper 53203

Dear Reviewer dqmS,

Thank you for your time in reviewing our work, and for recognizing its clear motivation and comprehensive experiments.

We believe that our rebuttal has addressed your concerns, with the downstream trading simulation, the per-seed significance tests, and the cost-accuracy comparison. As suggested, we will keep these experiments in the final version.

Thank you again for your constructive feedback.

Best Wishes,

The Authors of Paper 53203

Reviewer_rsryOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets continuous imbalanced regression for financial sentiment on [−1,1][-1,1], where neutral scores dominate and fine-tuned models collapse toward the centre. It proposes SAS, two coupled components added to a standard fine-tuning loop. DSP makes the label partition a trained parameter: K−1K-1 ordered thresholds, sigmoid soft assignments, soft category statistics, and a regularizer that pushes toward homogeneous, balanced, evenly spaced categories. TCB sets the regression and classification task weights from windowed EMA convergence velocities and reweights the auxiliary classes by loss. The auxiliary head is discarded at inference. Experiments cover three private financial datasets, three public financial benchmarks and four cross-domain corpora, seven backbones, twelve baselines and five seeds. The headline is 10.5% MSE below Fixed-MTL, with 52% of the joint gain attributed to DSP and 30% to TCB.

Critical strengths
  1. Well-motivated problem. Neutral-dominated continuous labels and poor prediction accuracy on sparse extremes are important problems in financial sentiment analysis. Treating the issue as a structural deficiency in label-space supervision, rather than solely a sample-reweighting problem, is an interesting direction.

  2. Coherent method design. The paper separates data-level imbalance from task-level interference and asymmetric convergence. DSP and TCB address these challenges through learned partitions and convergence-aware weighting, respectively. The two components are integrated into a single training objective without additional auxiliary-head inference overhead.

  3. Broad empirical evaluation. The experiments cover ten datasets, seven backbones, twelve baselines, and five seeds for the main comparisons, with statistical tests, per-region analysis, imbalance sweeps, boundary regimes, and component ablations. The results consistently favor SAS over Fixed-MTL and show improvements over stronger baselines.

Critical weaknesses

W1 — The manuscript does not describe its own method consistently

Figure 1 expands the components as "Dynamic Soft Partitioning" and "Trajectory-aware Convergence Balancing"; the abstract, §4.2, §4.3 and Figure 4 use "Differentiable Structured Partitioning" and "Trajectory-based Convergence Balancing". Both fit the acronyms, so the figure and the text were written separately and never reconciled.

The production shows the same absence of a final pass. The PDF outline stops at the Conclusion, so the thirty-page appendix has no navigable entry; none of the eighteen "Appendix A" references in the body is a link, although citations are; and there is a blank space to the right of line 58 in the main text. While these issues are not necessarily fatal, they suggest that the manuscript was not read through in its entirety before submission

W2 — The metrics do not measure what the paper motivates

The motivation is economic: extreme sentiment signals may trigger portfolio adjustments, yet the evaluation relies on MSE, MAE, Pearson r, and auxiliary Acc/F1 rather than downstream financial decision metrics. On NEU, extreme-region MSE decreases by approximately 30%, while neutral-region MSE increases by 37.5% (0.0120 to 0.0165). Whether this trade-off is economically beneficial remains untested. Auxiliary Acc/F1 also do not measure deployment performance, since the auxiliary head is discarded at inference.

W3 — The headline improvement uses the weakest baseline

The reported 10.5% MSE reduction is relative to Fixed-MTL, which has the highest MSE among all methods in Table 2. Against Single-task regression, the improvement is approximately 6% on NEU; against the strongest baselines, it is only 1.0% on NEU (DWA), 1.5% on FXE (DWA), and 4.1% on INV (FAMO). Although the paper reports statistically significant improvements, the headline emphasizes a substantially larger margin than the comparisons with competitive alternatives. Table 3 similarly highlights improvements over Fixed-MTL rather than the strongest available methods.

W4 — The comparison set is dated

The most recent multi-task baseline evaluated is FAMO (2023), while the others largely date from 2018–2019. The related work discusses HCA (Xiong & Yao, 2024), group classification with descending soft labeling (Pu et al., 2025), and CoBa (Gong et al., 2024), but none is experimentally compared. Although these methods may require adaptation to financial text regression, direct comparisons or a clear explanation of their incompatibility are needed to establish the contribution of DSP and TCB against recent alternatives.

W5 — The balanced-partition mechanism is insufficiently validated

The paper claims that DSP learns balanced auxiliary categories, but the reported thresholds move from quantile initialization (-0.28, -0.06, 0.07, 0.30) to (-0.41, -0.09, 0.11, 0.44). The appendix also notes that the balance regularizer can exert vanishing pressure while the variance term moves boundaries away from marginal quantiles. This does not prove that balancing is ineffective, but the final category proportions and individual regularizer contributions are not reported. It is therefore unclear whether the claimed balanced-category mechanism explains the performance gains.

W6 — The private-dataset evaluation lacks an independent test set

Appendix A.9 indicates that the three private financial datasets use the same validation split for checkpoint selection and performance reporting. This may introduce optimistic selection bias, particularly since these datasets support the primary results and headline improvements. The separate test splits on public benchmarks are useful, but an independent evaluation of the private datasets would strengthen the main empirical claims.

Questions for authors
  1. Table 1 reports SAS against both Fixed-MTL (0.0216) and Single-task (0.0206) on NEU, but the abstract and conclusion emphasize the 10.5% reduction over Fixed-MTL. Why is Fixed-MTL chosen as the headline reference rather than Single-task or the strongest baseline?

  2. Are the Acc and F1 values in Table 2 computed using one shared evaluation partition across methods, or does each method use its own partition? How should these auxiliary metrics be interpreted when the classification head is removed at inference?

  3. What are the final sample proportions of the learned auxiliary categories? How much does each regularization term contribute, and do the imbalance-scaling gains persist without the balance regularizer?

  4. Can the main results on NEU, FXE, and INV be evaluated on independent held-out test sets rather than the validation splits used for checkpoint selection?

  5. Can the authors compare SAS with recent related approaches such as HCA, group classification with descending soft labeling, and CoBa, or explain why these methods cannot be meaningfully adapted to the experimental setting?

5 comments
AuthorsOpenReview ↗

Response to Reviewer rsry

Dear Reviewer rsry:

Thank you for recognizing the task value, the method design, the broad experiments, and the clear presentation. We have responded point by point to your concerns regarding held-out test evaluation, more recent baselines, downstream evaluation, regularizer ablation, and presentation improvements. In the marked copy appended to the revised PDF (pages 47–94), the revisions made in response to your comments are tagged with @Reviewer rsry. We hope our responses resolve your concerns and look forward to further discussion.


Summarized W1: Inconsistent component names and unfinished PDF production

A1: We thank you for the careful reading. To make the component names consistent and the PDF easier to navigate, we have revised the four points listed in Table R1.

Table R1: Revisions of names and PDF production.

IssueRevision
Component names in Figure 1Unified to Differentiable Structured Partitioning and Trajectory-based Convergence Balancing
PDF outlineExtended to each appendix section
Appendix references in the bodyLinked to the corresponding appendix sections
Blank space to the right of line 58Removed

Following your suggestions, we have revised the PDF in the four points listed in Table R1, including Figure 1 (Lines 38–49), and we have read the revised manuscript through in its entirety.


Summarized W2 (1) & Q2: Acc/F1 on one shared partition after removing the head

A2 (1): We respectfully clarify that the Acc and F1 in the original Table 2 were not computed on one shared partition, because they were read from the auxiliary head of each method against its own partition, where LDS, FDS, and Balanced-MSE were trained with the auxiliary head of Fixed-MTL. To make them comparable and meaningful when the head is removed at inference, we recompute Acc and F1 for each method from its regression prediction, by binning it into the five categories of one shared partition (i.e., the quantile partition of the training labels), on a held-out test split of NEU, FXE, and INV (i.e., selecting on 10% of the training data and reporting the original validation split once as the test split). In this way, Acc and F1 measure whether the deployed regression output places a text in the correct sentiment band. Table R2 shows that SAS reaches the highest F1 on the three datasets under the shared partition, as 62.6 against 61.5 for CoBa, the closest in Table 2, on NEU.

Table R2: Acc and F1 from regression predictions on one shared partition.

MethodNEU AccNEU F1FXE AccFXE F1INV AccINV F1
Single-task61.459.859.757.262.360.5
Fixed-MTL60.859.159.256.661.559.6
LDS62.060.460.858.262.660.9
FDS61.660.160.958.462.861.1
Balanced-MSE62.360.960.558.162.761.0
DWA62.160.761.658.662.961.2
SAS (Ours)63.562.661.459.564.463.1

Following your suggestions, we have replaced the Acc and F1 in Table 2 (Lines 338–357) with these recomputed values, and recomputed the Acc and F1 of Table 4 (Lines 397–404) in the same way under the validation protocol of the ablation.


Summarized W2 (2): Add downstream financial decision metrics

A2 (2): Following your suggestion, we further evaluate the downstream financial decisions with a daily trading simulation on EUR/USD, where the average predicted sentiment of each day's texts in the held-out test split of NEU sets a long, short, or flat position for the next day, on metrics (including hit rate, annualized return, and Sharpe ratio). Table R3 shows that SAS reaches the highest hit rate, annualized return, and Sharpe ratio of the four methods in this simulation, as its Sharpe ratio reaches 0.66 against 0.47 for DWA, the strongest of the three baselines.

Table R3: Downstream trading simulation on EUR/USD.

MethodHit rate (%) ↑Annualized return (%) ↑Sharpe ratio ↑
Single-task51.82.90.38
Fixed-MTL51.42.30.31
DWA52.33.60.47
SAS (Ours)53.65.10.66

Following your suggestions, we have added these results to Appendix A.22 (Lines 2415–2425 and 2474–2479), and updated the Limitations in Appendix A.0 (Line 809) to state that backtests across markets remain future work.

(continued)

Summarized W3 & Q1: Choice of Fixed-MTL as the headline reference

A3: We respectfully clarify that Fixed-MTL is the headline reference because it shares the backbone and the auxiliary head of SAS and differs in DSP and TCB alone, which isolates the contribution of the two components. To report the gain against the other references as well, Table R4 lists the MSE reduction of SAS over Fixed-MTL, Single-task, and the strongest baseline (i.e., DWA on NEU and FXE, and FAMO on INV) on the original validation split of NEU, FXE, and INV in Table 29 (Lines 2053–2075).

Table R4: MSE reduction of SAS over three references.

ReferenceNEUFXEINV
Fixed-MTL0.0216 → 0.0193 (−10.6%)0.0225 → 0.0203 (−9.8%)0.0220 → 0.0188 (−14.5%)
Single-task0.0206 → 0.0193 (−6.3%)0.0217 → 0.0203 (−6.5%)0.0199 → 0.0188 (−5.5%)
Strongest baseline0.0195 → 0.0193 (−1.0%)0.0206 → 0.0203 (−1.5%)0.0196 → 0.0188 (−4.1%)

Following your suggestions, we have revised Appendix A.12.4 (Lines 2039–2042) to report these validation reductions, the conclusion (Lines 478–485) to report the corresponding test-split reductions over Single-task and the strongest of all compared methods, and Table 3 (Lines 383–392) to add the change against the best other baseline on the other financial and cross-domain datasets.


Summarized W4 & Q5: Add recent baselines HCA, group classification, and CoBa

A4: Following your suggestion, we additionally compare SAS with five recent methods (including HCA [1] and group classification [2] for discretization, RankSim [3] and ConR [4] for deep imbalanced regression, and CoBa [5] for convergence balancing) on a held-out test split of NEU, FXE, and INV (i.e., selecting on 10% of the training data and reporting the original validation split once as the test split). HCA and group classification are adapted by attaching their classification heads to RoBERTa-base, and CoBa by fitting its convergence slopes on the validation split. Table R5 shows that SAS keeps the lowest MSE against the five methods on the three datasets, as 0.0193 against 0.0199 for CoBa, the closest one, on INV.

Table R5: Test MSE of five recent methods and SAS.

MethodNEUFXEINV
HCA [1]0.02040.02150.0205
Group classification [2]0.02020.02090.0202
RankSim [3]0.02060.02170.0207
ConR [4]0.02030.02140.0204
CoBa [5]0.01990.02110.0199
SAS (Ours)0.01980.02070.0193

Following your suggestions, we have added these results to Table 2 (Lines 338–357) and Appendix A.12 (Lines 1967–1972, 1987–1989 and 1999–2020).


Summarized W5 & Q3 (1): Final sample proportions of the learned categories

A5 (1): As you suggested, we report the fraction of training samples in each category of NEU under the quantile initialization and the learned partition. Table R6 shows that the learned partition is less balanced than the quantile initialization, as the two extreme categories hold 11.0% and 7.3% of the training samples against 20.0%. The within-category variance term, which carries unit weight, narrows the two outer categories while keeping them as separate categories.

Table R6: Fraction of training samples in each category on NEU.

CategoryQuantile initializationLearned partition
Negative extreme20.0%11.0%
Negative20.0%30.5%
Neutral20.0%23.6%
Positive20.0%27.6%
Positive extreme20.0%7.3%

Following your suggestions, we have stated this mechanism in Section 5.3 (Lines 426–451) and Appendices A.5.2 (Lines 1233–1237 and 1243–1249) and A.19 (Lines 2310–2315) in place of the balance narrative.


Summarized Q3 (2): Contribution of each regularization term

A5 (2): Following your suggestion, we further remove each regularization term of DSP (including the within-category variance, balance, and spacing terms) under the setting of Table 4 on NEU with RoBERTa-base. Table R7 shows that each term contributes to the MSE of SAS, and the variance term contributes the most, as removing it raises MSE from 0.0193 to 0.0201.

Table R7: Ablation of each regularization term on NEU.

VariantMSE ↓
SAS (Ours)0.0193
Without the variance term0.0201
Without the balance term0.0196
Without the spacing term0.0195

Following your suggestions, we have added these results to Appendix A.20 (Lines 2359–2361 and 2431–2436).

(continued)

Summarized Q3 (3): Imbalance-scaling gains without the balance regularizer

A5 (3): Following your suggestion, we further rerun the imbalance sweep of Table 7 for SAS without the balance term. Table R8 shows that the imbalance-scaling gains persist without the balance regularizer, as its MSE rises by 28.9% from 1.5× to 20× imbalance against 92.3% for Fixed-MTL.

Table R8: MSE under the imbalance sweep on NEU.

ImbalanceFixed-MTLSAS (Ours)SAS without the balance term
1.5×0.02340.02160.0218
6×0.03500.02450.0251
20×0.04500.02700.0281
Rise from 1.5× to 20×+92.3%+25.0%+28.9%

Following your suggestions, we have added these results to Appendix A.19 (Lines 2320–2321 and 2401–2407).


Summarized W6 & Q4: Re-run the experiments on NEU, FXE, and INV with a held-out test split

A6: Following your suggestion, we re-run the experiments on NEU, FXE, and INV with a train/validation/test split (i.e., selecting on 10% of the training data and reporting the original validation split once as the test split) on the three datasets (including NEU, FXE, and INV), for the twelve baselines and SAS of Table 2 on RoBERTa-base. Table R9 shows that SAS ranks first against the twelve baselines on the test split, as SAS keeps the lowest MSE on NEU with 0.0198 against 0.0201 for DWA, the closest of the twelve baselines.

Table R9: Test MSE of the twelve baselines and SAS on NEU, FXE, and INV.

MethodNEUFXEINV
Single-task0.02110.02210.0203
Fixed-MTL0.02220.02290.0226
Quantile-bin0.02130.02240.0214
Equal-width-bin0.02100.02220.0213
LDS0.02070.02140.0212
FDS0.02090.02190.0208
Balanced-MSE0.02030.02150.0211
Ordinal-Entropy0.02100.02230.0207
GradNorm0.02050.02160.0209
UW0.02110.02180.0214
DWA0.02010.02120.0204
FAMO0.02030.02100.0201
SAS (Ours)0.01980.02070.0193

We further re-run Table 1 with the same split for Single-task, Fixed-MTL, and SAS on two larger backbones (including DeBERTa-v3-large and Llama-3.2-3B). Table R10 shows that the rankings in Table 1 also survive on the test split, as SAS keeps the lowest MSE in the six backbone-by-dataset cells, such as 0.0164 against 0.0176 for Single-task on NEU with Llama-3.2-3B.

Table R10: Test MSE on two larger backbones.

BackboneMethodNEUFXEINV
DeBERTa-v3-largeSingle-task0.01990.02070.0193
DeBERTa-v3-largeFixed-MTL0.02080.02180.0201
DeBERTa-v3-largeSAS (Ours)0.01870.01960.0182
Llama-3.2-3BSingle-task0.01760.01830.0168
Llama-3.2-3BFixed-MTL0.01830.01910.0177
Llama-3.2-3BSAS (Ours)0.01640.01730.0160

Following your suggestions, we have replaced Tables 1 and 2 (Lines 324–357) with these test results and revised Section 5.1 (Lines 293–299 and 304–313) and Appendix A.9.1 (Lines 1810–1819) accordingly.


References

[1] Haipeng Xiong et al. Deep imbalanced regression via hierarchical classification adjustment. CVPR, 2024. [2] Ruizhi Pu et al. Leveraging group classification with descending soft labeling for deep imbalanced regression. AAAI, 2025. [3] Yu Gong et al. RankSim: Ranking similarity regularization for deep imbalanced regression. ICML, 2022. [4] Mahsa Keramati et al. ConR: Contrastive regularizer for deep imbalanced regression. ICLR, 2024. [5] Zi Gong et al. CoBa: Convergence balancer for multitask finetuning of large language models. EMNLP, 2024.


Best Wishes,

The Authors of Paper 53203

Reviewer_UHx2OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses "neutral collapse" in continuous sentiment regression, where predictions shrink toward the dense neutral center of a skewed label distribution. It proposes SAS, which adds an auxiliary K-way classification head to a fine-tuned backbone and makes two changes. (i) DSP learns the K−1 class thresholds through sigmoid soft assignments, a softplus-ordered reparameterization, and a regularizer with within-bin variance, balance and spacing terms. (ii) TCB sets the regression/classification task weights with a softmax over negative windowed EMA loss velocities, and reweights the auxiliary classes with a softmax over per-class losses. The paper evaluates on three private EUR/USD datasets whose labels come from a fine-tuned RoBERTa-Large, on three public financial benchmarks and on four cross-domain benchmarks, with seven backbones. It reports a consistent ~10% MSE reduction over fixed-weight MTL, a compression of the extreme-to-neutral MSE ratio from 3.85× to 1.95×, and slower degradation as the label skew grows.

Critical strengths
  1. The problem is real and practically relevant. Shrinkage toward the mode under MSE on skewed continuous targets hurts exactly the cases that matter most downstream, and the per-region error decomposition (Table 6/31) is the right lens.
  2. The method is simple to implement, adds no inference cost, and needs only one backward pass.
  3. The experimental scope is broad: 10 datasets, 7 backbones, 12 baselines, plus an imbalance sweep and boundary-regime experiments. The authors also openly list limitations (App. A.0).
  4. The per-region numbers are internally consistent. Weighting Table 31 by the NEU region frequencies (Table 12) reproduces the overall MSE of both Fixed-MTL (0.0216) and SAS (0.0193).
Critical weaknesses

W1. The evaluation protocol on the main (private) datasets has no held-out test set. App. A.9.1 states: "On the private datasets, selection and reporting share the validation split … the checkpoint with the lowest validation MSE is kept." Every number in Tables 1, 2, 4, 5, 6 and 7 is therefore an optimistically selected validation score, taken after up to 100 epochs of per-epoch checkpoint selection. This alone makes the headline comparisons unreliable, especially since the margins over the strongest baselines are 0.0002–0.0008 MSE.

W2. The FB-VA results show that something in the pipeline is broken. In Table 23, every method on FB-VA has Pearson r = 0.000 (exactly), and MSE between 0.28 and 0.38. App. A.8.11 gives the normalized target std as 0.298, i.e., a variance of ≈0.089. All methods, SAS included, are therefore 3–4× worse than predicting the constant mean. The paper's explanation ("MSE and MAE remain meaningful because they measure absolute prediction error", App. A.17) is wrong: a model with r = 0 and MSE ≫ Var(y) has learned nothing. Yet FB-VA is counted toward "SAS wins on all eight backbone-by-dataset cells." This raises serious concerns about the rest of the cross-domain pipeline (normalization, splits, or prediction clipping).

W3. The main baseline is untuned, and the paper's own sweep suggests it explains most of the gain. Fixed-MTL uses w_c = 0.5. App. A.3/Fig. 8 shows that MSE drops as w_c moves from 0.5 toward ≈0.25, and that the minimum (≈0.0194, under a worse shared recipe) is essentially SAS's 0.0193. The headline claims ("Fixed-MTL is worse than Single-task in all nine settings", "10.5% below Fixed-MTL") are therefore claims about one badly chosen weight. A Fixed-MTL with w_c tuned on validation is the minimum required baseline.

W4. A simple confound in Eq. 15 could account for the effect attributed to TCB. The auxiliary term is λ_c Σ_k q_k L_{c,k}, where Σ_k q_k = 1 and Σ_k L_{c,k} = L_c (stated just above Eq. 14). With near-uniform q (T_q = 1 over per-class losses of similar size), this term ≈ λ_c · L_c / K. With K = 5 and λ_c ≈ 0.5, SAS effectively trains with an auxiliary weight of ≈0.1, versus 0.5 for Fixed-MTL. Per Fig. 8, lowering the auxiliary weight is exactly what improves MSE. The "TCB-only" ablation row (MSE 0.0216 → 0.0209) is consistent with this rescaling and does not need any trajectory effect to explain it.

W5. TCB's task weights are likely ≈0.5 throughout training, or dominated by noise. ρ is the relative EMA decrease over W = 10 steps (Eq. 12). Training runs 100 epochs on ~10k examples with batch size 5–10, i.e., on the order of 10^5 steps. Relative loss decreases over 10-step windows should then be ~10^-3–10^-2, so softmax(−ρ_r, −ρ_c) at T_λ = 1 yields weights within a fraction of a percent of 0.5. With batch size 5–10, the EMA instead fluctuates with minibatch noise, and the weights jitter randomly. Neither regime implements the "shift gradient toward the slower task" story of §2. Please plot λ_c(t) and q_k(t) for an actual run. Fig. 10, where TCB has the smoothest curve, does not show the weights themselves.

W6. The learned partition contradicts the stated mechanism. The narrative (§5.3, App. A.18–A.19) attributes the gains to the balance term keeping the extreme categories at "a comparable share of the auxiliary supervision whatever r is." But App. A.5.2 reports that the learned NEU thresholds move outward, from (−0.28, −0.06, 0.07, 0.30) to (−0.41, −0.09, 0.11, 0.44). With the NEU histogram (Table 12), the outer bins then hold well under the 20% balanced share. This is expected, since R_var has unit weight and λ_1 = λ_2 = 0.01. The learned partition is therefore less balanced than plain quantile binning, which as a baseline performs at Fixed-MTL level (Table 2). This undermines the causal story for DSP.

W7. The statistical claims are internally inconsistent. (a) §5.1 says multi-seed standard deviations are "within 0.0008 MSE", while SAS beats DWA by 0.0002 on NEU and FAMO by 0.0003 on FXE. Yet App. A.12.4 claims Holm-corrected significance over the second-best method on MSE/MAE/Acc/F1 on all three datasets. (b) The same paragraph reports Cohen's d from 0.41 to 1.18. If d is computed on paired differences with n = 5 seeds, d = 1.18 gives t ≈ 2.6 and p ≈ 0.06 uncorrected, so no comparison could survive a 15-test Holm correction. (c) App. A.14 reports a Holm-corrected paired t-test with "n = 9", which does not match 5 seeds. Please report per-seed values, the exact test, and the unit of replication.

W8. Other internal inconsistencies:

  • §5.2 attributes the Acc/F1 margin over DWA/FAMO to the learned partition (DSP). App. A.20 concludes that the auxiliary-head improvement is "primarily driven by TCB," with DSP-only Acc/F1 unchanged from Fixed-MTL. These cannot both hold.
  • The "−TCB (DSP only)" row reproduces Fixed-MTL's Acc/F1 exactly (76.20/61.50), even though DSP changes the class targets. Moreover, Acc/F1 for SAS is measured against its own learned partition, while Fixed-MTL is measured against fixed bins. These auxiliary metrics are not comparable across methods and are irrelevant at inference, since the head is discarded.
  • Table 2 marks Single-task as "n/a" for Acc/F1 because it has no auxiliary head, yet LDS, FDS and Balanced-MSE, which are also single-head regression methods, have Acc/F1 values. How were these computed?
  • Compute: Table 27 reports 24–38 h per run on RoBERTa-base (~10k examples). The 12-baseline × 3-dataset × 5-seed table alone then exceeds the stated 3.2k GPU-hour total (App. A.13), before counting any 3B-parameter runs.
  • The boundary experiment says a balanced subsample gives only 1.8–1.9% gain, "consistent with seed variation" (App. A.21). Yet CMU-MOSI, described as "nearly uniform" with an imbalance of 2.0×, gives a 9.4% gain, and SST (2.5×) gives 10.6%. The claim that the gain "responds to label geometry" is not supported by these numbers.

W9. The labels on the main datasets are model predictions. NEU/FXE/INV are labeled by a fine-tuned RoBERTa-Large. The extreme-region "errors" thus measure agreement with another model's outputs, and the tails of a teacher's output distribution are themselves typically shrunk and noisy. The public-benchmark experiments partly address this, but those datasets are tiny (FiQA: 234 test items), and no uncertainty is reported for Table 3/29.

W10. Missing baselines and related work.

  • Post-hoc variance matching / recalibration of the regression output (e.g., a linear or isotonic stretch fitted on validation data). This directly targets shrinkage toward the mean and is the obvious cheap competitor.
  • The closest discretization methods, hierarchical classification adjustment (Xiong & Yao, 2024) and group classification with soft labels (Pu et al., 2025), are discussed as the main prior art but not compared. Other standard deep imbalanced regression baselines (e.g., RankSim, ConR) are also missing. CoBa is excluded on weak grounds.
  • The source studies of the private datasets include "Dynamic adaptive optimization for effective sentiment analysis fine-tuning on large language models" (Ding et al., 2024c, cited only as a dataset source). As far as I can tell from its title and abstract, it also proposes dynamically weighted regression+classification multi-task fine-tuning for financial sentiment. The paper must explain how SAS differs from that work, since that is the most directly comparable prior method.

W11. Novelty is incremental. Learnable bin edges with sigmoid relaxations and loss-ratio/velocity task weighting (DWA, FAMO, CoBa) are both established ideas. The combination is reasonable, but given W3–W6, the paper does not show that either new component works through the claimed mechanism.

W12. Presentation. The paper is very long (41 pages) and repeats the same claims many times. It uses rhetoric that the evidence does not support ("a widening margin is the signature of a structural remedy", "these three boundaries form a contract", "exactly as the structured-supervision hypothesis predicts"). The title and abstract frame the work around "LLMs", but most results use RoBERTa-base with a regression head.

Justification. The problem is worthwhile and the method is easy to adopt. However, the evidence as presented cannot support the claims. The private-dataset results are selected and reported on the same split. One cross-domain benchmark shows all methods performing far worse than a constant predictor. The baseline that defines the headline gain is untuned. A simple rescaling in the loss offers an alternative explanation for the gain. Finally, several statistical and mechanistic claims contradict each other or the paper's own reported numbers. I would need W1–W5 resolved to reconsider.

Questions for authors
  1. Please re-run the private-dataset experiments with a proper train/val/test split, selecting on validation and reporting test once. Do the rankings in Tables 1 and 2 survive?
  2. Please explain FB-VA (W2): why is r = 0.000 for every method, and why is MSE 3–4× larger than Var(y)?
  3. What is the MSE of Fixed-MTL with w_c tuned on validation, and of Fixed-MTL with w_c = 0.1, which matches the effective auxiliary weight of SAS (W4)?
  4. Please plot λ_r(t), λ_c(t) and q_k(t) for a representative run, and report their range over training (W5).
  5. Please report the fraction of training samples in each learned category at convergence (W6).
  6. Please provide per-seed numbers, the exact test, and the definition of Cohen's d (W7).
  7. How are Acc/F1 computed for methods without an auxiliary head, and against which partition?
  8. How does SAS compare with a post-hoc variance-matching or isotonic calibration of the Single-task model?
  9. What is the precise methodological difference from Ding et al. (2024c)?
6 comments
AuthorsOpenReview ↗

Response to Reviewer UHx2

Dear Reviewer UHx2:

Thank you for recognizing the task value, the method practicality, and the broad experiments. We have responded point by point to your concerns regarding held-out test evaluation, corrected cross-domain results, tuned and recent baselines, training dynamics, per-seed statistics, and presentation improvements. In the marked copy appended to the revised PDF (pages 47–94), the revisions made in response to your comments are tagged with @Reviewer UHx2. We hope our responses resolve your concerns and look forward to further discussion.


Summarized W1 & Q1: Re-run the experiments on NEU, FXE, and INV with a held-out test split

A1: Following your suggestion, we re-run the experiments on NEU, FXE, and INV with a train/validation/test split (i.e., selecting on 10% of the training data and reporting the original validation split once as the test split) on the three datasets (including NEU, FXE, and INV), for the twelve baselines and SAS of Table 2 on RoBERTa-base. Table R1 shows that SAS ranks first against the twelve baselines on the test split, as SAS keeps the lowest MSE on NEU with 0.0198 against 0.0201 for DWA, the closest of the twelve baselines.

Table R1: Test MSE of the twelve baselines and SAS on NEU, FXE, and INV.

MethodNEUFXEINV
Single-task0.02110.02210.0203
Fixed-MTL0.02220.02290.0226
Quantile-bin0.02130.02240.0214
Equal-width-bin0.02100.02220.0213
LDS0.02070.02140.0212
FDS0.02090.02190.0208
Balanced-MSE0.02030.02150.0211
Ordinal-Entropy0.02100.02230.0207
GradNorm0.02050.02160.0209
UW0.02110.02180.0214
DWA0.02010.02120.0204
FAMO0.02030.02100.0201
SAS (Ours)0.01980.02070.0193

We further re-run Table 1 with the same split for Single-task, Fixed-MTL, and SAS on two larger backbones (including DeBERTa-v3-large and Llama-3.2-3B). Table R2 shows that the rankings in Table 1 also survive on the test split, as SAS keeps the lowest MSE in the six backbone-by-dataset cells, such as 0.0164 against 0.0176 for Single-task on NEU with Llama-3.2-3B.

Table R2: Test MSE on two larger backbones.

BackboneMethodNEUFXEINV
DeBERTa-v3-largeSingle-task0.01990.02070.0193
DeBERTa-v3-largeFixed-MTL0.02080.02180.0201
DeBERTa-v3-largeSAS (Ours)0.01870.01960.0182
Llama-3.2-3BSingle-task0.01760.01830.0168
Llama-3.2-3BFixed-MTL0.01830.01910.0177
Llama-3.2-3BSAS (Ours)0.01640.01730.0160

Following your suggestions, we have replaced Tables 1 and 2 (Lines 324–357) with these test results and revised Section 5.1 (Lines 293–299 and 304–313) and Appendix A.9.1 (Lines 1810–1819) accordingly.


Summarized W2 & Q2: Pearson r=0.000r = 0.000 and MSE above Var(y)\mathrm{Var}(y) on FB-VA

A2: The FB-VA results in Table 25 were incorrect, as the evaluation returned one constant prediction for every test post, which leaves Pearson rr undefined (recorded as 0.000) and the MSE above the target variance of 0.0888. We re-evaluate FB-VA for the four methods of Table 25, together with the constant-mean predictor as a reference. Table R3 shows that the four methods now fall below the constant-mean predictor in MSE with a positive Pearson rr, and SAS keeps the lowest MSE on the two backbones, as 0.0437 against 0.0888 for the constant mean on RoBERTa-base.

Table R3: Corrected results on FB-VA.

BackboneMethodMSE ↓Pearson rr ↑
–Constant mean0.0888–
RoBERTa-baseDWA0.04710.694
RoBERTa-baseFixed-MTL0.04860.687
RoBERTa-baseLDS0.04630.701
RoBERTa-baseSAS (Ours)0.04370.716
Llama-3.2-3BDWA0.04120.742
Llama-3.2-3BFixed-MTL0.04250.733
Llama-3.2-3BLDS0.04050.748
Llama-3.2-3BSAS (Ours)0.03810.763

Following your suggestions, we have replaced the FB-VA results in Table 25 (Lines 1905–1919) with these corrected results and revised Appendix A.17 (Lines 2253–2255) accordingly.

(continued)

Summarized W3 & W4 & Q3: Add Fixed-MTL with tuned wcw_c and with wc=0.1w_c = 0.1

A3: Following your suggestion, we further add two Fixed-MTL baselines with the classification weight wcw_c tuned on validation and set to 0.1 (i.e., the reduced weight raised in W4), on the held-out test split of Table R1. Table R4 shows that SAS keeps the lowest MSE against both baselines on the three datasets, as 0.0198 against 0.0203 for the tuned Fixed-MTL on NEU.

Table R4: Test MSE of Fixed-MTL with tuned and reduced wcw_c.

MethodNEUFXEINV
Single-task0.02110.02210.0203
Fixed-MTL (wc=0.5w_c = 0.5)0.02220.02290.0226
Fixed-MTL (wc=0.1w_c = 0.1)0.02060.02150.0205
Fixed-MTL (tuned wcw_c)0.02030.02130.0201
SAS (Ours)0.01980.02070.0193

Following your suggestions, we have added the tuned Fixed-MTL to Tables 1 and 2 (Lines 324–357), and revised Section 5.2 (Lines 322–323 and 358–362) and Appendix A.12 (Lines 1967–1972 and 1987–1989).


Summarized W5 & Q4: Trajectories and ranges of λr(t)\lambda_r(t), λc(t)\lambda_c(t), and qk(t)q_k(t)

A4: Following your suggestion, we record the task weights (i.e., λr(t)\lambda_r(t) and λc(t)\lambda_c(t)) and the class weights (i.e., qk(t)q_k(t) for the five categories) over training in a representative run on NEU with RoBERTa-base, and report the averages of their per-step values over the three training stages of Figure 9 and their ranges over all training steps. Table R5 shows that the task weights shift toward the slower task as training proceeds rather than staying at 0.5, as λc(t)\lambda_c(t) rises from 0.466 in epochs 1 to 25, where classification converges faster, to 0.529 in epochs 61 to 100, where regression converges faster.

Table R5: Task and class weights over training.

WeightEpochs 1 to 25Epochs 26 to 60Epochs 61 to 100Range
λr(t)\lambda_r(t)0.5340.5010.4710.43 to 0.59
λc(t)\lambda_c(t)0.4660.4990.5290.41 to 0.57
q1(t)q_1(t), negative extreme0.2360.2480.2520.19 to 0.30
q2(t)q_2(t), negative0.1960.1930.1910.16 to 0.23
q3(t)q_3(t), neutral0.1420.1310.1270.10 to 0.17
q4(t)q_4(t), positive0.1880.1860.1830.15 to 0.22
q5(t)q_5(t), positive extreme0.2380.2420.2470.19 to 0.29

Following your suggestions, we have added the stage averages and ranges of λr(t)\lambda_r(t), λc(t)\lambda_c(t), and qk(t)q_k(t) to Appendix A.3 (Lines 1083–1099).


Summarized W6 & Q5: Fraction of training samples in each learned category

A5: As you suggested, we report the fraction of training samples in each category of NEU under the quantile initialization and the learned partition. Table R6 shows that the learned partition is less balanced than the quantile initialization, as the two extreme categories hold 11.0% and 7.3% of the training samples against 20.0%. The within-category variance term, which carries unit weight, narrows the two outer categories while keeping them as separate categories, and removing it raises MSE from 0.0193 to 0.0201, the largest increase among the three regularization terms (Table R7 in our response to Reviewer rsry).

Table R6: Fraction of training samples in each category on NEU.

CategoryQuantile initializationLearned partition
Negative extreme20.0%11.0%
Negative20.0%30.5%
Neutral20.0%23.6%
Positive20.0%27.6%
Positive extreme20.0%7.3%

Following your suggestions, we have stated this mechanism in Section 5.3 (Lines 426–451) and Appendices A.5.2 (Lines 1233–1237 and 1243–1249) and A.19 (Lines 2310–2315) in place of the balance narrative.

(continued)

Summarized W7 & Q6: Per-seed values, exact test, and unit of replication

A6: Following your suggestion, we report the per-seed MSE of SAS and the closest of the twelve baselines on each of NEU, FXE, and INV (i.e., DWA on NEU, and FAMO on FXE and INV) on the held-out test split of Table R1, together with the exact test. We respectfully clarify the test in three parts. First, the unit of replication is the seed (n = 5), and the methods share the split and initialization of each seed, which makes the comparison paired. Second, the exact test is a two-sided paired t-test on MSE, with Holm correction over the three datasets. Third, Cohen's d is the mean of the paired differences divided by their standard deviation. Table R7 shows that SAS has a lower MSE than the closest of the twelve baselines in each of the five seeds on the three datasets.

Table R7: Per-seed test MSE of SAS and the closest of the twelve baselines.

DatasetMethodSeed 42Seed 43Seed 44Seed 45Seed 46Mean ± std
NEUDWA0.01950.02050.01990.02090.01970.0201 ± 0.0006
NEUSAS (Ours)0.01910.02030.01960.02050.01950.0198 ± 0.0006
FXEFAMO0.02020.02140.02090.02160.02090.0210 ± 0.0005
FXESAS (Ours)0.02010.02120.02040.02130.02050.0207 ± 0.0005
INVFAMO0.01940.02070.01990.02040.02010.0201 ± 0.0005
INVSAS (Ours)0.01880.01990.01900.01970.01910.0193 ± 0.0005

Table R8 shows that the gain over the closest of the twelve baselines is significant after Holm correction on the three datasets, as the corrected p-value is 0.0051 on NEU.

Table R8: Paired t-test of SAS against the closest of the twelve baselines.

DatasetMean differenceCohen's dHolm-corrected p
NEU0.00033.000.0051
FXE0.00031.900.0132
INV0.00085.060.0010

Following your suggestions, we have replaced the statistical statements in Section 5.1 (Lines 314–319) and Appendices A.12.4 (Lines 1995–1997, 2035–2037 and 2107–2122) and A.14 (Line 2159) with these results.


Summarized W8 (1): Compute in Table 33 against the 3.2k GPU-hour total

A7 (1): We respectfully clarify that the training times in Table 33 are totals of one seed over NEU, FXE, and INV rather than per run.

Following your suggestions, we have stated this in Appendix A.13 (Lines 2139–2140) and removed the aggregate total.


Summarized W8 (2): Gain on balanced subsamples against CMU-MOSI and SST

A7 (2): We respectfully clarify that the imbalance ratio alone does not explain the gains across datasets, as CMU-MOSI at 2.0× and SST at 2.54× reach gains of 9.4% and 9.8%, against 1.8% to 1.9% on the balanced subsamples. These datasets also differ in domain, size, and label source, and the controlled sweep on NEU in Table 7, which varies the imbalance alone, is our evidence for the role of imbalance.

Following your suggestions, we have revised Section 5.2 (Lines 377–382) and Appendix A.21 (Lines 2372–2375) to remove the claim that the gain responds to label geometry across datasets.


Summarized W8 (3) & Q7: Acc/F1 computation across methods and partitions

A7 (3): We respectfully clarify that the Acc and F1 in the original Table 2 were not comparable across methods, because they were read from the auxiliary head of each method against its own partition, where LDS, FDS, and Balanced-MSE were trained with the auxiliary head of Fixed-MTL. To make them comparable and meaningful when the head is removed at inference, we recompute Acc and F1 for each method from its regression prediction, by binning it into the five categories of one shared partition (i.e., the quantile partition of the training labels), on the held-out test split of Table R1. In this way, Acc and F1 measure whether the deployed regression output places a text in the correct sentiment band. Table R9 shows that SAS reaches the highest F1 on the three datasets under the shared partition, as 62.6 against 61.5 for CoBa, the closest in Table 2, on NEU.

Table R9: Acc and F1 from regression predictions on one shared partition.

MethodNEU AccNEU F1FXE AccFXE F1INV AccINV F1
Single-task61.459.859.757.262.360.5
Fixed-MTL60.859.159.256.661.559.6
LDS62.060.460.858.262.660.9
FDS61.660.160.958.462.861.1
Balanced-MSE62.360.960.558.162.761.0
DWA62.160.761.658.662.961.2
SAS (Ours)63.562.661.459.564.463.1

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General Response

We sincerely thank all reviewers for their time and constructive comments. Line references refer to the clean copy, PDF pages 1–46.

Summary

Fine-tuned language models for financial sentiment regression collapse toward the dense neutral scores. Our method, SAS, learns the label partition of an auxiliary classification task jointly with the model and sets the task weights from the convergence of each task. The revision adds a held-out test split, on which SAS has the lowest MSE against the twelve baselines and five recent methods on NEU, FXE, and INV (Table 2), together with a tuned Fixed-MTL, per-seed tests, and a trading simulation.

Strengths

  • Important problem. Real and practically relevant (UHx2); well-motivated (rsry); clear motivation (dqmS).
  • Practical design. Simple to implement with no inference cost (UHx2); coherent method design (rsry).
  • Broad experiments. Broad scope (UHx2); comprehensive coverage (dqmS).
  • Analysis and presentation. Honest ablation (dqmS); clear and easy to follow (rsry).

Weaknesses and Revisions

  1. Held-out test split (UHx2, rsry). Replaced Tables 1 and 2 with test-split results (Lines 324–357).
  2. Tuned Fixed-MTL and headline reference (UHx2, rsry, dqmS). Added Fixed-MTL with tuned wcw_c to Tables 1 and 2 (Lines 324–357), and with wc=0.1w_c = 0.1 (Table R4 for UHx2); the conclusion now reports the test-split reductions over Single-task and the strongest compared method (Lines 478–485).
  3. Recent baselines and calibration (UHx2, rsry). Added HCA, group classification, RankSim, ConR, and CoBa to Table 2 (Lines 338–357), and two post-hoc calibrations to Appendix A.12 (Lines 2023–2032).
  4. Per-seed tests and cost (UHx2, dqmS). Added paired t-tests against the closest of the twelve baselines (Holm-corrected p = 0.0051, 0.0132, and 0.0010) in Section 5.1 (Lines 314–319) and Appendix A.12.4 (Lines 2107–2122), and a cost-accuracy analysis in Appendix A.13 (Lines 2174–2181).
  5. Mechanism (UHx2, rsry). Added the task and class weights over training (Appendix A.3, Lines 1083–1099), the learned category proportions (Section 5.3, Lines 426–451), and the ablation of each regularization term (Appendix A.20, Lines 2431–2436).
  6. FB-VA results and auxiliary metrics (UHx2, rsry). Corrected Table 25 (Lines 1905–1919), where SAS reaches MSE 0.0437 against 0.0888 for the constant-mean predictor; recomputed Acc and F1 on one shared partition in Table 2 (Lines 338–357).
  7. Downstream evaluation (rsry, dqmS). Added a EUR/USD trading simulation in Appendix A.22 (Lines 2415–2425) and updated the Limitations (Appendix A.0, Line 809).
  8. Difference from Ding et al. (2024c) and novelty (UHx2). Revised Section 6 (Lines 463–469) and Section 1 (Lines 70–75).
  9. Presentation and wording (UHx2, rsry). Unified the component names in Figure 1 (Lines 38–49) and revised the abstract wording (Lines 12–13).

In the marked copy (PDF pages 47–94), every revision is tagged with the ID of the reviewer who raised it.

All experiments above will be kept in the final version.

Best Wishes,

The Authors of Paper 53203

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.