acceptodds
Under review as a conference paper at ICLR 2027

D²S: Dual-Perspective Dynamic Data Selection for Supervised Fine-Tuning

Abstract

Effective data selection is crucial for supervised fine-tuning under limited training budgets. Existing data selection methods can be broadly categorized into data-centric and model-centric approaches. Data-centric methods rely on reusable static scores but do not directly capture the evolving learning needs of the model, whereas model-centric methods reflect the current learning needs but incur substantial costs when repeatedly evaluating large data pools. To address these issues, we introduce Dual-perspective Dynamic Data Selection (DS), a framework that first performs coarse selection from the data perspective and then refines the candidates from the model perspective. The resulting model-perspective feedback is used to guide subsequent data-perspective selection, forming an iterative selection process which adapts to the evolving learning needs of the model while restricting expensive model evaluation to a small candidate set. To provide fine-grained model-perspective feedback, we further introduce Model-aware Scoring (MaS), which distinguishes between important and unimportant response tokens based on the model's prediction probabilities, applies different evaluation criteria to each group, and aggregates the resulting scores into a sample-level score. Experiments on three data pools and two base models show that our framework achieves an average relative improvement of 3.6% over the best baseline and 16.4% over full-data training, while reducing average runtime by 6.1% and 66.2%, respectively.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language model, supervised fine-tuning, data selection

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_UN5iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents D2S, a data selection framework that simultaneously considers static data-centric scoring and dynamic model-centric scoring. Specifically, it precomputed multiple static metrics (quality, difficulty, and diversity) and adaptively weights them with a reuse penalty to shortlist a candidate set. The candidate set is then re-ranked using model-aware scoring, which first distinguishes important and unimportant tokens and then scores them separately. Top-rank candidates are then selected to fine-tune the model for one epoch. The MaS rankings are also used update category allocation budgets and to re-optimize data-perspective attribute weights for subsequent rounds. The quantitative evaluation reports performance improvements and runtime savings over full-data training and existing data selection baselines.

Critical strengths
  • The motivation is clear, and the paper is overall easy to read.
  • The evaluation is extensive, covering different instruction-tuning pools, base models, downstream tasks, and baseline methods.
Critical weaknesses
  1. The paper frames the combination of static data-centric scoring and dynamic model-centric scoring as an unexplored area in instruction tuning. However, this hybrid philosophy has been explored in several work, including MoDS [1], LEAD [2], and UDS [3]. The paper fails to distinguish itself from these work in both concept and evaluation.

  2. Without the dual-perspective framing, the contributions narrow down to a token-partitioning heuristic and a Spearman-based weight grid search. However, the paper's own ablations in Table 4 show that the closed-loop feedback provides only a marginal +0.9 point gain, and a simple iterative random-candidate control (w/ random selection) already reaches 55.9. The feedback mechanism is interesting but underexplored.

  3. The evaluation is confounded by different training strategy. D2S uses an iterative, multi-round loop, while all baseline methods uses a static single-stage configuration. The ablation study in Table 4 reveals that an iterative control (i.e., w/ random selection) already reaches an average score of 55.9 on Alpaca-GPT4 with LLaMA-3.1-8B, outperforming every external baseline in Table 1. Therefore, the observed performance advantage may be driven by the iterative multi-round training paradigm and schedule resets rather than the proposed dual-perspective selection algorithm.

[1] Qianlong Du, Chengqing Zong, and Jiajun Zhang. 2023. MoDS: Model-oriented data selection for instruction tuning. arXiv preprint arXiv:2311.15653.

[2] Xiaotian Lin, Yanlin Qi, Yizhang Zhu, Themis Palpanas, Chengliang Chai, Nan Tang, and Yuyu Luo. 2026. LEAD: Iterative data selection for efficient LLM instruction tuning. Proceedings of the VLDB Endowment (PVLDB), 19(3): 426–439.

[3] Heming Zou, Yixiu Mao, Yun Qu, Qi Wang, and Xiangyang Ji. 2026. Utility-diversity aware online batch selection for LLM supervised fine-tuning. In Proceedings of the 43rd International Conference on Machine Learning (ICML).

Questions for authors
  • If strong dynamic baselines (e.g., UDS) and a random selection baseline are trained using the exact same multi-round training schedule as D2S, how much of the performance advantage of D2S remains?

  • How does D2S conceptually and empirically differ from work such as MoDS, LEAD, and UDS?

6 comments
AuthorsOpenReview ↗

Author Response (1/3)

Thank you for carefully reviewing our paper and for your valuable suggestions. In response to your concerns, we provide the following replies. The corresponding revisions are highlighted in blue in the revised paper.


Response to W1 and Q2: How does D2S conceptually and empirically differ from work such as MoDS, LEAD, and UDS?

Thank you for your comment. The novelty of D²S lies in the following three aspects:

  • Systematic dual-perspective modeling. D²S is a systematic modeling framework that characterizes data supply and model demand as two complementary perspectives. The data perspective jointly considers data quality, difficulty, and diversity to characterize the intrinsic value of candidate data, while the model perspective evaluates data utility based on the model's current learning needs.
  • Dynamic coordination through iterative feedback. D²S connects the two perspectives through iterative feedback, enabling their coordination across multiple training rounds. This allows data selection to adapt to the model's evolving learning needs while reducing model-based evaluation overhead.
  • Fine-grained model-aware scoring. We introduce a model-aware scoring mechanism MaS that accounts for differences in token importance to provide fine-grained assessments of sample utility, supporting both model-perspective selection and data-perspective weight adaptation.

Although MoDS, LEAD, and UDS incorporate certain data-related or model-related signals, they do not systematically model the data and model perspectives or explicitly capture their interactions as D²S does. Specifically, the differences are as follows.

Specific differences from MoDS, LEAD, and UDS.

  • MoDS [1] considers both data quality and model needs through a sequential selection process: it first filters data based on quality and then selects data according to model needs in a subsequent stage. In contrast, D²S coordinates the two perspectives through iterative feedback across multiple training rounds, allowing data selection to continuously adapt to the model's evolving learning needs.
  • LEAD [2] groups data into clusters and iteratively selects the highest-scoring cluster, from which high-utility samples are selected for training. It then uses model feedback to update the selected cluster's score for subsequent selection. However, it does not explicitly incorporate data quality attributes into candidate selection, potentially allowing challenging but noisy samples to be selected. In contrast, D²S jointly considers data quality, difficulty, and diversity to guide high-quality candidate selection and reduce this risk. Furthermore, D²S introduces MaS to account for differences in token importance, enabling fine-grained model-aware assessments of sample utility and supporting both model-perspective selection and data-perspective weight adaptation.
  • UDS [3] combines data diversity and model utility for online data selection. However, it does not explicitly incorporate data-quality attributes to guide candidate selection and still requires expensive model-based evaluation of every candidate batch. In contrast, D²S jointly models data quality, difficulty, and diversity to construct a smaller and high-quality candidate pool before model-based evaluation, thereby reducing evaluation overhead while maintaining the ability to select data according to the model's learning needs.

Empirical comparison with MoDS, LEAD, and UDS.

In the original paper, we compared D²S with UDS. We have extended this comparison to include MoDS and LEAD, using LLaMA-3.1-8B on three datasets under the same experimental setting. As shown in Table 1, D²S achieves the highest average on all three datasets, with average improvements of 2.8, 3.1, and 3.2 points over MoDS, LEAD, and UDS, respectively. These results provide empirical evidence of the effectiveness of D²S, demonstrating the benefits of systematically modeling and dynamically coordinating the data and model perspectives for SFT data selection.

Table 1: Comparison with related methods using LLaMA-3.1-8B on Alpaca-GPT4, WizardLM, and CoT

DatasetMethodBBHGSM8KMMLUTQAHEMBPPAvg.
Alpaca-GPT4MoDS62.351.564.256.944.646.854.4
LEAD63.150.264.457.744.546.054.3
UDS62.551.564.757.643.947.054.5
D²S66.458.665.558.245.151.057.5
WizardLMMoDS64.955.164.853.345.045.054.7
LEAD64.052.364.255.144.545.654.3
UDS63.954.464.454.145.749.855.4
D²S65.557.765.554.347.053.257.2
CoTMoDS60.152.362.643.539.046.250.6
LEAD62.150.261.941.439.645.650.1
UDS53.052.558.944.036.047.848.7
D²S62.854.264.145.843.350.253.4

AuthorsOpenReview ↗

Author Response (2/3)

Response to W3 and Q1: If strong dynamic baselines (e.g., UDS) and a random selection baseline are trained using the exact same multi-round training schedule as D2S, how much of the performance advantage of D2S remains?

Thank you for pointing this out. D²S outperforms strong baselines and pure random selection using the same multi-round training schedule. In the following content, we will clarify the experimental settings and present comparisons with other iterative selection methods and pure random selection.

Clarification of UDS’s iterative selection procedure.

It is worth clarifying that UDS and D²S adopt fundamentally different iterative frameworks. UDS selects a small subset from each candidate batch (e.g., 4 samples) and repeats this process thousands of times. In contrast, D²S selects 3,000 training examples per round and trains for only a few rounds. This substantial difference in selection granularity and iteration frequency makes it difficult to directly adapt UDS to D²S's multi-round training schedule.

Clarification of the “w/ random selection” ablation.

The “w/ random selection” ablation in Table 4 only ablates the data perspective candidate selection, while the model perspective still keeps the MaS scores to select the final training samples. The “w/ random selection” ablation therefore includes model-aware selection and should not be interpreted as the effect of multi-round training alone.

Comparison with selection methods under multi-round training.

We reproduced UDS following its official implementation of iterative data selection and reported the results in Tables 1, 2, and 7 of the original paper. To extend our comparison to more multi-round methods, we compared D²S with the strongest dynamic baseline LEAD [2], using the same multi-round training schedule for both methods.

As shown in Table 2, D²S achieves an average score of 57.5, outperforming LEAD (54.3) and UDS (54.5) by 3.2 and 3.0 points, respectively. These results suggest that D²S’s gains come not only from multi-round training but also from dual-perspective data selection.

Table 2: Comparison with LEAD and UDS using LLaMA-3.1-8B on Alpaca-GPT4

MethodBBHGSM8KMMLUTQAHumanEvalMBPPAvg.
LEAD63.1‾\underline{63.1}50.264.457.7‾\underline{57.7}44.5‾\underline{44.5}46.054.3
UDS62.551.5‾\underline{51.5}64.7‾\underline{64.7}57.643.947.0‾\underline{47.0}54.5‾\underline{54.5}
D²S66.458.665.558.245.151.057.5

Comparison with pure random selection.

To further examine whether multi-round training alone accounts for the gains, we additionally conducted an experiment that selects the final training samples entirely at random in each round. As shown in Table 3, this variant achieves an average score of 52.9, compared with 57.5 for D²S, a difference of 4.6 points. This result further supports the contribution of our dual-perspective data selection framework.

Table 3: Comparison between pure random multi-round training and D²S

MethodBBHGSM8KMMLUTQAHumanEvalMBPPAvg.
Pure random62.2‾\underline{62.2}50.8‾\underline{50.8}64.5‾\underline{64.5}54.2‾\underline{54.2}40.2‾\underline{40.2}45.2‾\underline{45.2}52.9‾\underline{52.9}
D²S66.458.665.558.245.151.057.5

AuthorsOpenReview ↗

Author Response (3/3)

Response to W2: The variability of feedback mechanism.

Thank you for your suggestion. To assess the variability of D²S and the no-feedback variant, we conducted three independent training runs with different random seeds using LLaMA-3.1-8B + Alpaca-GPT4.

As shown in Table 4, the full D²S method achieves higher mean performance on all six benchmarks, with an average improvement of 1.2 ± 0.3 points, showing that the feedback mechanism provides a consistent performance advantage.

Table 4: Comparison between the full D²S method and the variant without feedback

MethodBBHGSM8KMMLUTruthfulQAHumanEvalMBPPAvg.
w/o feedback64.7±0.6‾\underline{64.7 \pm 0.6}55.7±0.4‾\underline{55.7 \pm 0.4}65.5±0.3‾\underline{65.5 \pm 0.3}56.8±0.6‾\underline{56.8 \pm 0.6}44.9±1.5‾\underline{44.9 \pm 1.5}49.5±1.0‾\underline{49.5 \pm 1.0}56.2±0.4‾\underline{56.2 \pm 0.4}
D²S66.5 ± 0.158.0 ± 0.665.8 ± 0.358.6 ± 0.445.5 ± 1.350.1 ± 1.057.4 ± 0.1
Paired improvement+1.8 ± 0.7+2.3 ± 0.8+0.3 ± 0.2+1.8 ± 1.0+0.6 ± 1.6+0.7 ± 0.1+1.2 ± 0.3

Thank you again for your valuable suggestions for improving the paper. We hope the explanations and additional experiments above address your concerns.


References

[1] Qianlong Du, Chengqing Zong, and Jiajun Zhang. 2023. MoDS: Model-oriented Data Selection for Instruction Tuning. arXiv preprint arXiv:2311.15653.

[2] Xiaotian Lin, Yanlin Qi, Yizhang Zhu, Themis Palpanas, Chengliang Chai, Nan Tang, and Yuyu Luo. 2025. LEAD: iterative data selection for efficient LLM instruction tuning. Proceedings of the VLDB Endowment.

[3] Heming Zou, Yixiu Mao, Yun Qu, Qi Wang, and Xiangyang Ji. 2026. Utility-Diversity Aware Online Batch Selection for LLM Supervised Fine-tuning. Forty-third International Conference on Machine Learning.

[4] Yanjun Fu, Faisal Hamman, and Sanghamitra Dutta. 2025. T-SHIRT: Token-selective hierarchical data selection for instruction tuning. Advances in Neural Information Processing Systems.


Reviewer_sfqgOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The work proposes a new balanced data selection method for SFT, comibing the benefit from two scoring methods: data-centric (cheap but static) and model-centric (reflecting model need but expensive).

The proposed D2S method performs 3-4 rounds of iterative selection, and each round contains 3 steps:

  1. Coarse selection: use attribute-weighted static scores + category proportions to select 4500 candidates
  2. Use model-aware score (MaS) to further select a subset of 3000 and train the target model on it for 1 epoch
  3. Use MaS score to update attribute weights + category proportions for next round's coarse selection

Tested on 3 data pool (Alpaca-GPT4, WizardLM, CoT) x 2 base models (LLaMA-3.1-8B, Qwen2.5-7B), D2S method outperforms the baseline ADG by 3.6% and full-data training by 16.4%, while reducing training time by 41.9% and 66.2%, respectively.

Critical strengths
  • The method is well motivated and clearly explained. Coarsely selecting candidates with cheap static scores and allowing the current model to pick the final subset directly addresses the cost of model-centric selection.
  • The ablation supports this claim. Without coarse selection, runtime rises from 84.5 / 118.7 / 151.1 to 138.4 / 335.1 / 418.8 minutes (Table 3), and the average score on Alpaca-GPT4 drops from 57.5 to 54.5 (Table 4). This comparison is within the authors' own pipeline, so it is not affected by the concern below.
Critical weaknesses

My main concern is how much of D2S's performance gain comes from author's training pipeline rather than from the method itself.

  • Author's pipeline raises scores. In Table 1, author's own run of ADG achieves avg scores of 54.1 / 55.4 / 51.2 on the 3 datasets, which is +3.3 / 3.7 / 6.1 points above the 50.85 / 51.73 / 45.11 as reported in ADG's paper Table 11.
  • However, all other baselines in Table 1 (except UDS) have the exact same numbers as in ADG's Table 11 after rounding. So these are almost certainly not run in author's training pipeline. Therefore, comparing them directly with D2S is unfair.
  • On Qwen (Table 2), even the ADG row matches ADG's Table 12, and D2S leads it by only 1.1-2.7 points, so the Qwen results do not clearly show that D2S "generalizes across different data pools and base models" (Sec 4.2).
Questions for authors

Which baseline rows in Tables 1, 2 and 7 did you run yourselves, and under what setting? If the other rows were not run in your pipeline, please either explain why comparing them with D2S is still fair, or report "All Data" on the 3 datasets of Table 1, run with the same training pipeline as your ADG and UDS runs. Either could raise my rating.

5 comments
AuthorsOpenReview ↗

Author Response (1/3)

Thank you for carefully reviewing our paper and for your valuable suggestions. In response to your concerns, we provide the following replies. The corresponding revisions are highlighted in blue in the revised paper.


Response to W1 and Q1: Which baseline rows in Tables 1, 2 and 7 did you run yourselves, and under what setting?

Many thanks for your constructive comments. In the original paper, we reproduced the strongest baselines, including ADG and UDS, and cited the results of the other weaker baselines from the ADG paper[1]. We explain the rationale for this reporting results and present our updated results of baselines in the following content.

Rationale for the reporting results.

For the LLaMA-3.1-8B results in Table 1 of the original paper, we used the selected data officially released by ADG and trained for three epochs with LLaMA-Factory, using the same training and evaluation hyperparameters as ADG. However, our reproduced ADG results were higher than those reported in its paper. Our investigation found that the original ADG experiments did not use the official LLaMA template during SFT of the LLaMA models, which may lead to incorrectly formatted answers and thus an underestimation of model performance. To avoid underestimating the baseline, we reported our higher reproduced results in Table 1. For the other weaker baselines, we cited the results from the ADG paper, which were compared under unified settings and substantially underperformed ADG, providing useful references without affecting our main conclusions.

For the Qwen2.5-7B results in Tables 2 and 7 of the original paper, we also trained and evaluated ADG. Our reproduced results were slightly lower than those reported in the ADG paper. Similarly, to avoid underestimating the baseline, we retained the higher results from the ADG paper in the main tables. The results for the other weaker baselines were also cited from the ADG paper, where they were compared under unified settings and therefore provided a useful reference.

Updated results of baselines.

To make the comparison more rigorous, we trained and evaluated all baselines and updated the corresponding results in Tables 1, 2, and 7 of the revised paper. The full results are shown in Table 1 and 2.

With LLaMA-3.1-8B, our reproduced results are generally higher than those reported in the ADG paper, which is related to the use of a more appropriate LLaMA template. After the update, the six-benchmark average scores of the other weaker baselines remain below ADG, and D²S retains the best average performance on all three datasets. With Qwen2.5-7B, the reproduced results are generally close to the originally reported average performance, and D²S also remains advantages on all three datasets. Therefore, updating the baseline results does not change our main conclusion regarding the advantage of D²S, but does make our paper more convincing.


Thank you again for your valuable suggestions for improving the paper. We hope the explanations and additional experiments above address your concerns.


References

[1] Bo Li, Mingda Wang, Shikun Zhang, and Wei Ye. 2026. Instruction data selection via answer divergence. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).


AuthorsOpenReview ↗

Author Response (2/3)

Table 1: Full experimental results of LLaMA-3.1-8B on three datasets

DatasetMethodBBHGSM8KMMLUTQAHumanEvalMBPPAvg.
Alpaca-GPT4Full Data62.846.963.453.840.844.652.1
Random61.748.664.655.040.245.052.5
Rethinking61.946.863.355.741.546.452.6
SCAR62.9‾\underline{62.9}51.563.657.039.642.852.9
MIG62.250.663.958.2‾\underline{58.2}34.246.252.5
ZIP61.948.263.556.544.5‾\underline{44.5}45.453.3
IFD61.449.564.157.643.347.053.8
SuperFiltering61.649.164.158.440.247.4‾\underline{47.4}53.5
SelectIT*62.447.663.255.139.043.451.8
ADG62.052.5‾\underline{52.5}64.456.642.147.4‾\underline{47.4}54.2
UDS62.551.564.7‾\underline{64.7}57.643.947.054.5‾\underline{54.5}
D²S66.458.665.558.2‾\underline{58.2}45.151.057.5
WizardLMFull Data61.153.963.049.144.745.652.9
Random62.054.563.649.543.346.253.2
Rethinking62.852.063.452.745.149.054.2
SCAR63.252.863.453.645.144.053.7
MIG63.254.063.653.146.346.054.4
ZIP62.053.263.654.7‾\underline{54.7}44.546.654.1
IFD64.151.063.356.043.347.054.1
SuperFiltering64.5‾\underline{64.5}48.763.554.7‾\underline{54.7}44.546.653.7
SelectIT*63.152.763.652.946.4‾\underline{46.4}50.0‾\underline{50.0}54.8
ADG64.255.4‾\underline{55.4}65.3‾\underline{65.3}53.544.549.655.4‾\underline{55.4}
UDS63.954.464.454.145.749.855.4‾\underline{55.4}
D²S65.557.765.554.347.053.257.2
CoTFull Data58.352.360.842.539.843.649.5
Random59.945.661.544.137.845.449.0
Rethinking45.444.163.1‾\underline{63.1}45.5‾\underline{45.5}37.842.046.3
SCAR54.547.259.244.436.043.647.5
MIG60.553.761.843.940.243.250.6
ZIP39.750.961.241.741.5‾\underline{41.5}44.646.6
IFD59.944.161.142.738.445.848.7
SuperFiltering60.745.362.742.337.245.048.9
SelectIT*58.244.860.544.338.443.648.3
ADG61.5‾\underline{61.5}53.8‾\underline{53.8}61.244.440.945.251.2‾\underline{51.2}
UDS53.052.558.944.036.047.8‾\underline{47.8}48.7
D²S62.854.264.145.843.350.253.4
AuthorsOpenReview ↗

Author Response (3/3)

Table 2: Full experimental results of Qwen2.5-7B on three datasets

DatasetMethodBBHGSM8KMMLUTQAHumanEvalMBPPAvg.
Alpaca-GPT4Full Data63.481.272.258.958.553.064.5
Random63.981.772.260.560.454.065.4
Rethinking62.681.772.260.4‾\underline{60.4}63.364.467.4
SCAR63.282.272.358.864.563.867.5
MIG62.482.172.359.159.064.866.6
ZIP62.082.372.4‾\underline{72.4}60.4‾\underline{60.4}62.165.867.5
IFD63.683.3‾\underline{83.3}72.057.563.365.067.5
SuperFiltering63.182.072.359.163.366.4‾\underline{66.4}67.7
SelectIT*63.582.672.859.862.165.067.6
ADG64.9‾\underline{64.9}81.772.359.766.164.868.2
UDS58.181.572.358.876.864.468.6‾\underline{68.6}
D²S67.986.372.4‾\underline{72.4}58.869.7‾\underline{69.7}67.070.4
WizardLMFull Data67.9‾\underline{67.9}80.472.5‾\underline{72.5}55.761.650.664.8
Random64.480.873.058.8‾\underline{58.8}59.856.065.5
Rethinking66.579.372.156.762.164.066.8
SCAR66.479.971.656.355.463.465.5
MIG67.680.572.355.261.563.466.7
ZIP66.779.072.056.956.662.465.6
IFD66.379.971.557.759.765.066.7
SuperFiltering66.279.871.457.159.063.266.1
SelectIT*66.877.071.955.956.662.065.0
ADG65.979.372.357.266.0‾\underline{66.0}61.267.0
UDS62.081.7‾\underline{81.7}72.359.166.0‾\underline{66.0}64.2‾\underline{64.2}67.6‾\underline{67.6}
D²S68.083.172.058.666.663.868.7
CoTFull Data63.1‾\underline{63.1}76.473.744.854.955.861.4
Random60.776.773.9‾\underline{73.9}44.955.556.861.4
Rethinking60.575.471.745.456.161.461.7
SCAR61.475.471.343.253.161.0‾\underline{61.0}60.9
MIG61.376.972.042.954.960.461.4
ZIP60.474.872.442.752.460.660.6
IFD61.578.371.746.3‾\underline{46.3}48.258.860.8
SuperFiltering61.576.071.645.352.459.861.1
SelectIT*61.376.271.843.341.560.459.1
ADG61.574.772.545.764.658.262.9‾\underline{62.9}
UDS57.173.071.645.368.357.662.1
D²S64.578.0‾\underline{78.0}75.147.267.0‾\underline{67.0}60.665.4
Reviewer_esYVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

D2SD^2S alternates between inexpensive candidate selection using quality, difficulty, and diversity scores, and a more expensive selection step using the current model. Its MaS score separates response tokens by how much the instruction changes their likelihood, then combines instruction dependence for the important tokens with fluency for the remainder. The model scores also update the coarse selector for the next round. Experiments cover three instruction pools, two base models, and six downstream benchmarks, with ablations of the scoring and feedback components and an analysis of runtime.

Critical strengths
  1. The method offers a practical connection between static data scores and changing model feedback. Table 3 includes preprocessing and shows a substantial saving over ADG and full-data training; the comparison with UDS is more modest and varies by pool. This is a relevant tradeoff for using dynamic selection in practice.
  2. The improvements are reasonably broad. D2SD^2S has the highest reported average in all six model/pool combinations. Table 4 also tests random candidates, removing coarse selection, fixed MaS, an alternative scorer, and removing feedback. These controls give useful evidence that the result is not solely due to one static quality filter.
Critical weaknesses
  1. The reported checkpoint and hyperparameter selection are not sufficiently separated from evaluation. Section 4.5 selects settings using the six-benchmark average on Alpaca/LLaMA and reports that the runs peak after three or four rounds. Appendix B says that all scores come from the same checkpoint, but does not identify a separate validation set or a pre-specified rule for choosing that checkpoint. This distinction matters because the extra rounds can reduce performance, and some improvements over the strongest baseline are around one or two average-score points. I cannot tell whether these are held-out results or results after selection on the reported benchmarks.
  2. The main comparison does not isolate data selection from the training schedule. D2SD^2S receives 9,000–12,000 sample presentations, whereas the selection baselines receive 30,000. Better performance with less training is useful, but the paper itself observes deterioration from additional SFT rounds, so the baseline schedule could account for part of the difference. A comparison against the strongest baseline at a matched number of updates or training tokens, using the same checkpoint-selection rule, would make the claimed selection advantage much more convincing. The within-framework ablations help, although the 0.9-point feedback gain is reported without variation across runs.
Questions for authors
  1. Which data and metric were used to choose the hyperparameters and the three or four-round checkpoint? Were the six reported evaluation splits consulted in those choices, and how were baseline checkpoints selected?
  2. Does the improvement over ADG or UDS persist when training tokens/updates and checkpoint selection are matched? A focused comparison on one existing model/pool setting would help resolve this concern.
  3. How variable are the full method and the no-feedback ablation across independent training/selection runs? This would help judge whether the feedback mechanism's reported 0.9-point contribution is reliable.
5 comments
AuthorsOpenReview ↗

Author Response (1/3)

Thank you for carefully reviewing our paper and for your valuable suggestions. In response to your concerns, we provide the following replies. The corresponding revisions are highlighted in blue in the revised paper.


Response to W1 and Q1: Which data and metric were used to choose the hyperparameters and the three or four-round checkpoint? Were the six reported evaluation splits consulted in those choices, and how were baseline checkpoints selected?

Thank you for your comments. We select our hyperparameters and checkpoints based on the average performance across the six benchmarks following the evaluation protocols adopted by existing data selection baselines, including ADG (ACL 2026) [1], SelectIT (NeurIPS 2024) [2], T-SHIRT (NeurIPS 2025) [3], LEAD (VLDB 2026) [4], and ssToken (ICLR 2026) [5]. As for the specific checkpoints used for the baselines, we adopt the optimal checkpoint settings reported in their respective papers.

In addition, to demonstrate the general applicability of our hyperparameters, we select hyperparameters in a single setting and transfer them to the others without separate selection for each setting. Specifically, we conduct hyperparameter experiments only with LLaMA-3.1-8B + Alpaca-GPT4 and apply the selected hyperparameters to the other data pools and base models. Similarly, for checkpoint selection, we determine the best training round for each data pool using LLaMA-3.1-8B and directly apply the selected rounds to the other base models.


AuthorsOpenReview ↗

Author Response (2/3)

Response to W2 and Q2: Does the improvement over ADG or UDS persist when training samples and checkpoint selection are matched? A focused comparison on one existing model/pool setting would help resolve this concern.

Thank you for pointing this out. D²S retains its advantage over ADG both under the same training sample budgets and under optimal checkpoint selection.

In our paper, D²S and the strongest baseline ADG use the same checkpoint-selection protocol (response to W1 and Q1), with the optimal checkpoint settings from their hyperparameter experiments. To further compare ADG and D²S under the same training sample budgets, we train ADG and D²S using LLaMA-3.1-8B on Alpaca-GPT4 under the same budget of 9,000 samples. As shown in Table 1, D²S outperforms all three ADG checkpoints, including its best checkpoint at the third epoch. With matched training sample budgets, the advantage of D²S further increases to 4.5 points, supporting the effectiveness of our data selection method rather than less training.

Table 1: Comparison between ADG and D²S under different training samples

MethodTraining stageTraining samplesBBHGSM8KMMLUTQAHEMBPPAvg.
ADG1 epoch9,00062.7‾\underline{62.7}50.464.456.7‾\underline{56.7}40.243.853.0
ADG2 epochs18,00062.352.8‾\underline{52.8}64.9‾\underline{64.9}56.642.7‾\underline{42.7}45.854.2‾\underline{54.2}
ADG3 epochs27,00062.052.564.456.642.147.4‾\underline{47.4}54.2‾\underline{54.2}
D²S--9,00066.458.665.558.245.151.057.5

AuthorsOpenReview ↗

Author Response (3/3)

Response to Q3: How variable are the full method and the no-feedback ablation across independent training runs?

Thank you for your suggestion. To assess the variability of D²S and the no-feedback variant, we conducted three independent training runs with different random seeds using LLaMA-3.1-8B + Alpaca-GPT4.

As shown in Table 2, the full D²S method achieves higher mean performance on all six benchmarks, with an average improvement of 1.2 ± 0.3 points, showing that the feedback mechanism provides a consistent performance advantage.

Table 2: Comparison between D²S and the variant without feedback

MethodBBHGSM8KMMLUTruthfulQAHumanEvalMBPPAvg.
w/o feedback64.7±0.6‾\underline{64.7 \pm 0.6}55.7±0.4‾\underline{55.7 \pm 0.4}65.5±0.3‾\underline{65.5 \pm 0.3}56.8±0.6‾\underline{56.8 \pm 0.6}44.9±1.5‾\underline{44.9 \pm 1.5}49.5±1.0‾\underline{49.5 \pm 1.0}56.2±0.4‾\underline{56.2 \pm 0.4}
D²S66.5 ± 0.158.0 ± 0.665.8 ± 0.358.6 ± 0.445.5 ± 1.350.1 ± 1.057.4 ± 0.1
Paired improvement+1.8 ± 0.7+2.3 ± 0.8+0.3 ± 0.2+1.8 ± 1.0+0.6 ± 1.6+0.7 ± 0.1+1.2 ± 0.3

Thank you again for your valuable suggestions for improving the paper. We hope the explanations and additional experiments above address your concerns.


References

[1] Bo Li, Mingda Wang, Shikun Zhang, and Wei Ye. 2026. Instruction data selection via answer divergence. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

[2] Liangxin Liu, Xuebo Liu, Derek F Wong, Dongfang Li, Ziyi Wang, Baotian Hu, and Min Zhang. 2024. Selectit: Selective instruction tuning for llms via uncertainty-aware self-reflection. Advances in Neural Information Processing Systems.

[3] Yanjun Fu, Faisal Hamman, and Sanghamitra Dutta. 2025. T-SHIRT: Token-selective hierarchical data selection for instruction tuning. Advances in Neural Information Processing Systems.

[4] Xiaotian Lin, Yanlin Qi, Yizhang Zhu, Themis Palpanas, Chengliang Chai, Nan Tang, and Yuyu Luo. 2025. LEAD: iterative data selection for efficient LLM instruction tuning. Proceedings of the VLDB Endowment.

[5] Xiaohan Qin, Xiaoxing Wang, Ning Liao, Cancheng Zhang, Xiangdong Zhang, Mingquan Feng, Jingzhi Wang, and Junchi Yan. 2026. ssToken: Self-modulated and Semantic-aware Token Selection for LLM Fine-tuning. The Fourteenth International Conference on Learning Representations.


Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Reviewers and Area Chair,

Thank you for your careful reviews and constructive feedback. We have carefully responded to the concerns raised by the reviewers through additional experiments, detailed analyses, methodological clarifications, and paper revisions. The key additions have been incorporated into the revised paper. As the discussion period approaches its end, we would like to summarize our main responses as follows:

  1. Baseline settings (Reviewer sfqg). To address the reviewer’s concerns, we clarified the rationale for the original reporting and provided updated baseline results. We have incorporated these updated results into the revised paper. Reviewer V8VQ acknowledged that our response addressed this concern and raised the score to 3.
  2. Training budgets (Reviewer esYV). We compared D²S with ADG at different training stages. D²S outperforms ADG under the same training sample budgets, addressing the reviewer’s concern about this question.
  3. Multi-round training (Reviewer UN5i). We compared D²S with LEAD, UDS, and pure random under multi-round training. D²S consistently outperforms these methods, supporting the contribution of dual-perspective data selection beyond multi-round training alone. Reviewer UN5i acknowledged that our response addressed this concern and the score is between 2 and 3.
  4. Feedback effectiveness and stability (Reviewers esYV and UN5i). To address the reviewers’ concerns about feedback variability, we conducted three independent training runs with different random seeds for D²S and its no-feedback variant. D²S achieves higher mean performance on all six benchmarks, supporting a consistent advantage from feedback. Reviewer UN5i acknowledged that our response addressed this concern and the score is between 2 and 3.
  5. Hyperparameter and checkpoint selection (Reviewers esYV and UN5i). Actually, most existing data selection baselines use benchmark results to guide hyperparameter choices, including ADG (ACL 2026), SelectIT (NeurIPS 2024), T-SHIRT (NeurIPS 2025), LEAD (VLDB 2026), and ssToken (ICLR 2026). We follow their evaluation protocols for a fair comparison. Furthermore, we select hyperparameters in a single setting and transfer them to the others, which demonstrates that the gains come from our method rather than reliance on specific hyperparameter selection.
  6. Methodological novelty (Reviewer UN5i). To clarify the differences from MoDS, LEAD, and UDS, we explained D²S’s systematic dual-perspective modeling, dynamic coordination through iterative feedback, and fine-grained model-aware scoring. Results across three datasets and two base models show that D²S outperforms these related methods, demonstrating its effectiveness. We have also included MoDS and LEAD as baselines in the revised paper (Tables 1, 2, and 7).

We hope these additional experiments and clarifications address the main concerns raised in the reviews. We would appreciate the reviewers and Area Chair considering this additional evidence in their final assessment.

Thank you very much for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.