acceptodds
Under review as a conference paper at ICLR 2027

MARS: Model-Adaptive Reusable Task-Specific Data Selection for Instruction Tuning

Abstract

Task-specific instruction data selection aims to select a compact subset that improves performance on a target task. Different base models may benefit from different subsets for the same target task, motivating model-specific adaptation. Existing methods that rely on model-dependent candidate signals typically require recomputing these signals over the full candidate pool when the base model changes, incurring substantial repeated computation across models. Directly reusing a previously selected subset avoids this cost but cannot account for differences across base models. We propose MARS, a model-adaptive and reusable framework that combines reusable candidate characterization with lightweight reference-based model calibration. A fixed external LLM scores each candidate across six ability dimensions to construct a reusable Candidate Ability Atlas. For each target task and base model, MARS combines model feedback from a small target reference set with reference ability distributions to derive model-specific ability allocation and reference profiles. The allocation determines ability-specific quotas, while the profiles guide within-bucket matching. Adapting to a new base model therefore requires evaluation only on the target reference set. Using only 5% of the candidate pool, MARS achieves higher average performance than all compared selection methods and full-data training on every base model across four base models and five target tasks. It also reduces the estimated cumulative selection cost by up to 95.6% compared with existing model-dependent methods. Ablation studies demonstrate the contributions of task-level ability demand, model-specific difficulty, and within-bucket matching. The code is available at https://anonymous.4open.science/r/MARS-0803/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Data Selection, Instruction Tuning, Model Adaptation

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_FETuOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies task-specific instruction data selection across different base models. Authors propose MARS, which uses a fixed external LLM annotator scores every example in the candidate pool once along 6 ability dimensions, together with a separate quality score. For each base model, response NLLs are computed on a small target reference set and converted into percentile ranks, which determine the per-bucket quotas, and the difficulty-weighted reference profiles. Candidates are assigned to ability buckets by their dominant ability and ranked within each bucket by the profile distance, with the quality score as the secondary key. Experiments cover 4 base models and 5 target tasks under a 5% selection budget.

Critical strengths
  • The problem setting is practically relevant. Since data selection outcomes are sensitive to the experimental setup, including the base model, repeating model-dependent selection for every new model is a real cost, and the division of labor between reusable candidate characterization and lightweight reference-based model calibration is easy to deploy.

  • The method is transparent and interpretable. Both the allocation and the profiles have closed forms.

  • The experiments span 4 model families under a unified training configuration, and the paper provides a selection-cost audit, the complete annotation prompt, and code.r

Critical weaknesses
  1. The novelty is limited, and closely related work is neither discussed nor compared. The core idea of MARS has been explored extensively, e.g. [1-6]. Most of these methods require no target benchmark data, and [6] as well as CDT (cited only as a capability characterization framework), additionally support capability-targeted selection. None of [1–6] is discussed or used as a baseline. Besides, a control that uses the same annotator with simple scoring or matching (e.g., asking gpt-5.6-luna to rate task relevance and taking the top B) is also missing, so it cannot be ruled out that much of the gain stems from the knowledge of the strong external annotator rather than from the proposed calibration design.

  2. Model adaptivity is the main claim of the paper, but the supporting evidence is weak and partly contradicts the motivation. While the MMLU shifts in Fig4 differ across models, the BBH and TyDiQA shifts are nearly identical across all four, suggesting that the NLL difficulty ranking is largely shared across models. The "w/o model calibration" variant in Table 8 uses no model information and therefore appears to yield a single subset shared by all four models, requiring one annotation pass and no GPU time; MARS improves over it by only 0.68 points on Gemma-3-4B and 1.30 points on OLMo-2-7B, without variance estimates.

Transfer of MARS's own subsets is evaluated only between Llama-3.1 and OLMo-2, whereas Table 2 compares MARS with recipient-specific calibration against LESS/NAIT subsets selected on Llama-3.1-8B, changing the method and the calibration model simultaneously. However, the baselines themselves do not support authors' premise: LESS subsets transferred from Llama-3.1-8B outperform LESS subsets selected natively on every recipient model, and the same holds for NAIT on OLMo-2-7B and Gemma-3-4B. This contradiction with the core motivation is not addressed.

  1. The setting relies on a known target benchmark and gold-response references, and the ability dimensions appear entangled with the evaluated tasks. Although this setting has precedent, the limitation should be noted: The references come from the same benchmarks as the test sets, and open-ended tasks without gold responses are not supported.

  2. Despite the literature and cognitive-science motivation, the 6 dimensions map almost one-to-one onto the 5 evaluated tasks (Semantic Retrieval - MMLU, Rule - BBH, Quantitative Reasoning - GSM8K, Code / Procedural Reasoning - MBPP, and Context Maintenance - TyDiQA). How the method performs on more numerous or more complex benchmarks that do not align with this taxonomy?

  3. Statistical reliability and experimental scope are insufficient, and the comparison with base models is sidestepped. No training seeds or variance are reported, which is critical in data selection, where gains of 1.01–1.29 points over the strongest baseline may fall within run-to-run noise. Only a single 5% selection budget is evaluated, although SFT outcomes can differ substantially between low- and high-budget regimes, and the quota rounding and bucket-deficit filling in MARS may also behave differently across budgets. Furthermore, Qwen2.5-7B Base averages 70.96, higher than every fine-tuned result including MARS. Why does fine-tuning generally hurt in this case?

  4. The efficiency advantage is a bit overstated. Full-pool computation is not eliminated but shifted to the external annotator. And the relationship between the 0.12-hour GPU-stage elapsed time and the 0.34 GPU-hours in Appendix B.5 should be clarified. Similar to W1, the cheapest model-dependent baseline is also missing, and the selection costs of MoNA and TSDS are not reported.

[1] DataMan: Data Manager for Pre-training Large Language Models.

[2] Meta-rater: A Multi-dimensional Data Selection Method for Pre-training Language Models.

[3] QuRating: Selecting High-Quality Data for Training Language Models.

[4] What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning.

[5] Improving Data Efficiency via Curating LLM-Driven Rating Systems.

[6] Harvesting Rich, Scalable and Transferable Multi-Modal Data for Instruction Fine-Tuning.

Questions for authors

Please see weaknesses.

16 comments
AuthorsOpenReview ↗

Novelty and Relation to Prior Work (W1, Part 1/2)

We thank the reviewer for recognizing the practical relevance, transparency, and reproducibility of MARS. Your comments have helped us sharpen our methodological contributions and motivated additional controlled experiments.

1. Research Motivation and Key Contributions

We acknowledge the important foundations established by prior work on multidimensional data evaluation and capability-guided selection. We will expand Section 2.2 to discuss these studies and clarify the methodological distinctions.

  • DataMan, Meta-rater, and QuRating [1–3] investigate multidimensional data evaluation, primarily for pretraining. Meta-rater additionally uses proxy-model validation losses to learn combinations of rating dimensions.
  • DEITA and DS² [4,5] study instruction-data quality, complexity, diversity, and rating calibration.
  • mmSSR and CDT [6,7] are particularly relevant to capability-targeted selection. mmSSR groups data by specified abilities and interaction styles, while CDT matches candidates to cognition, domain, and task labels derived from target validation examples.

MARS investigates a distinct question: how can a reusable characterization of instruction data support effective selection for different recipient models using only lightweight model-specific feedback?

Our central insight is that model-specific data selection does not necessarily require model-specific candidate representations. Model dependence can instead be introduced when deciding how to prioritize reusable candidate information.

MARS realizes this insight through two complementary mechanisms:

  1. Model-adaptive ability allocation: Reference ability compositions and recipient-model NLLs jointly determine how the selection budget is distributed across abilities.
  2. Difficulty-weighted reference-profile matching: The same model feedback determines which candidate instances are prioritized within each ability bucket.

These mechanisms introduce recipient feedback at both the allocation and instance-matching levels, allowing different models to select different subsets from the same candidate representation.

The key methodological contribution is decoupling reusable candidate characterization from recipient-specific selection decisions. The Candidate Ability Atlas is constructed once and reused across target tasks and models. Adaptation to a new recipient requires reference-set NLL evaluation, without repeating model-dependent feature extraction over the entire candidate pool.

This provides a concrete way to combine cross-model reusability, model-specific adaptation, and low selection overhead within a unified framework.

2. Distinction from Existing Capability-Guided Selection

Capability-based descriptions alone do not determine how to allocate a fixed data budget for a particular recipient model, nor which candidates within an ability group are most appropriate for that model.

MARS addresses both decisions through reference-calibrated allocation and continuous profile matching. Its contribution therefore extends beyond assigning capability labels or matching task categories: it provides a model-conditioned decision mechanism operating on reusable functional representations.

We further test this distinction against direct scoring, ability distribution matching, and CDT using controlled comparisons, reported in the following comment.

References: [1] DataMan (ICLR 2025); [2] Meta-rater (ACL 2025); [3] QuRating (ICML 2024); [4] DEITA (ICLR 2024); [5] DS² (ICLR 2025); [6] mmSSR (2025); [7] CDT (Findings of EMNLP 2025).

AuthorsOpenReview ↗

Same-Annotator and Capability-Targeted Comparisons (W1, Part 2/2)

3. Does the Strong External Annotator Explain the Gains?

To isolate the contribution of our selection design, we introduce three additional baselines:

  • Quality Top-B: Selects candidates using quality scores from the same LLM annotator as MARS.
  • Task Relevance Top-B: Uses the same annotator to score candidate relevance to the target task.
  • CDT: Performs capability-targeted selection through cognition–domain–task label matching.

We also compare with Ability Distribution Matching and w/o Model Calibration, which reuse MARS's ability annotations without its complete model-calibrated selection mechanism. All methods use the same candidate pool, 5% budget, and downstream training configuration.

Table R1: Same-annotator and capability-targeted comparisons (5% budget). CDT and MARS multi-seed results are reported as mean ± sample SD over three matched training seeds (42, 43, 44). Direct-scoring baselines are single-run results.

(a) Llama-3.1-8B

MethodMMLUBBHTyDiQAGSM8KMBPPAvg.
Quality Top-B61.0461.2072.4753.3743.6058.34
Task Relevance Top-B64.4163.5272.4663.6148.6062.52
CDT (3 seeds)62.87 ± 0.7963.27 ± 0.5671.66 ± 0.3464.34 ± 0.3147.80 ± 0.5361.99 ± 0.15
MARS (seed 42)64.8364.7273.2066.1149.6063.69
MARS (3 seeds)64.72 ± 0.1764.01 ± 1.3972.76 ± 0.3865.53 ± 0.5048.40 ± 1.0463.09 ± 0.57

(b) Gemma-3-4B-PT

MethodMMLUBBHTyDiQAGSM8KMBPPAvg.
Quality Top-B57.1653.5271.3244.4324.0050.09
Task Relevance Top-B57.9655.9373.8250.4944.6056.56
CDT (3 seeds)56.91 ± 0.6656.05 ± 0.5473.96 ± 0.4649.91 ± 0.3145.87 ± 0.4656.54 ± 0.19
MARS (seed 42)58.4356.3074.4451.8646.4057.49
MARS (3 seeds)58.32 ± 0.3056.91 ± 0.6074.74 ± 0.4952.14 ± 0.8346.47 ± 0.3157.72 ± 0.27

Direct-scoring comparisons. Task Relevance Top-B substantially outperforms Quality Top-B, confirming that task relevance is a meaningful selection signal. Nevertheless, under the original single-run setting, MARS further improves over Task Relevance Top-B by 1.17 and 0.93 points on Llama and Gemma, respectively, outperforming it on all ten individual task comparisons.

Across all four evaluated models, MARS also exceeds Quality Top-B in all 20 model–task settings.

Ability-matching controls. Ability Distribution Matching achieves 59.79 and 55.87 on Llama and Gemma, respectively; w/o Model Calibration achieves 60.11 and 56.81. MARS reaches 63.69 and 57.49 under the same single-run evaluation, showing that ability annotations alone do not reproduce the full selection gains.

Multi-seed CDT comparison. Across three matched training seeds, MARS improves the five-task average from 61.99 ± 0.15 to 63.09 ± 0.57 on Llama (+1.10) and from 56.54 ± 0.19 to 57.72 ± 0.27 on Gemma (+1.18). All six model–seed comparisons show positive average gains.

Together, these results establish that MARS provides additional value beyond strong annotator knowledge, direct task relevance, and existing capability matching. The gains support the contribution of reference-calibrated ability allocation and instance-level matching to effective data selection.

AuthorsOpenReview ↗

Recipient-Specific Calibration Beyond Direct Subset Transfer (W2, Part 1/2)

We thank the reviewer for highlighting the strong cross-model transfer results. We agree that these findings deserve closer examination and provide an important baseline for evaluating model adaptivity.

1. Why Is Recipient-Specific Calibration Valuable When Direct Transfer Works?

As the reviewer observes, LESS subsets selected using Llama transfer better than their recipient-specific counterparts on all three recipient models, while NAIT shows similar behavior on OLMo and Gemma. These findings establish that recomputing selection signals for each model does not automatically improve performance. They also highlight the importance of comparing against strong transferred subsets.

This motivates a more precise question for MARS:

Given that a fixed subset can already transfer effectively, can lightweight recipient-model feedback further improve selection while preserving reusable candidate representations?

To answer this, we extend the original Llama–OLMo experiment (Appendix C.2) to all three recipient models. We hold the Candidate Ability Atlas, target reference sets, selection algorithm, data budget, and recipient training configurations fixed, changing only which model provides the reference NLLs.

Table R2: Controlled comparison of calibration sources (5% budget; single-run results).

Recipient ModelCalibration SourceBBHTyDiQAFive-Task Avg.
Qwen2.5-7BLlama63.8971.0869.14
Recipient66.1172.2270.64
OLMo-2-7BLlama48.6163.4752.85
Recipient49.4465.9254.50
Gemma-3-4B-PTLlama55.0971.5954.94
Recipient56.3074.4457.49

Recipient-specific calibration improves the five-task average by +1.50, +1.65, and +2.55 points on Qwen, OLMo, and Gemma, respectively, with improvements in 14 of the 15 individual task comparisons.

The comparison also clarifies the relationship with strong transfer baselines. With Llama-based calibration, MARS scores 69.14, 52.85, and 54.94 on the three recipients, below the strongest LESS/NAIT transfer results of 69.50, 53.52, and 55.51. After recipient-specific calibration, MARS achieves 70.64, 54.50, and 57.49, exceeding the strongest transferred baseline by +1.14, +0.98, and +1.98 points, respectively.

This establishes the incremental value of recipient feedback under a controlled selection procedure, even when direct subset transfer is already effective.

2. Why Are the Allocation Shifts Similar in Figure 4?

We agree that Figure 4 shows similar ability-allocation patterns across models on BBH and TyDiQA. Such patterns are consistent with shared task-level ability demands.

Importantly, Figure 4 visualizes allocation proportions, while the complete MARS calibration also adjusts difficulty-weighted reference profiles within each ability bucket. Similar aggregate allocations can therefore lead to different within-bucket rankings and selected subsets.

The controlled results in Table R2 directly demonstrate the value of recipient-specific calibration on these two tasks: all six BBH and TyDiQA comparisons improve, with gains ranging from +0.83 to +2.85 points, despite their similar allocation patterns.

These findings show that shared task-level ability demands and recipient-specific selection gains can coexist. MARS combines global ability allocation with local reference-profile matching, allowing reusable candidate representations to respond effectively to each recipient model's reference difficulty information.

We will clarify this distinction in the discussion of Figure 4 and include the full calibration-source comparison in the revised manuscript.

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

This paper proposes MARS (Model-Adaptive Reusable Selection), a three-stage targeted data selection framework for instruction tuning that offline-annotates candidate instructions with a 6-dimensional capability profile and quality score, calibrates target-task capability quotas using a small reference set and the target model's response NLL percentiles, and selects candidates via bucketed Euclidean distance to calibrated centroids. Experiments across four base models and five benchmarks compare 5% selected subsets against full-pool SFT and existing selection baselines.

Critical strengths
  1. Practical amortized pipeline: Decoupling offline LLM capability tagging from lightweight per-model forward-pass calibration avoids expensive full-pool gradient computation (e.g., LESS) when adapting to new base models.
  2. Broad evaluation: The method is evaluated across four base model families and five diverse downstream benchmarks.
Critical weaknesses
  1. Severe performance degradation After SFT (Full < Base and Selected < Base) undermines the experimental setup: In Table 1, full-data SFT (Full) degrades performance relative to the untuned Base model in 13 out of 20 model–task settings, lowering the 5-task average on Qwen-2.5-7B (68.25 vs. 70.96) and OLMo-2-7B (52.15 vs. 53.61), and hurting MBPP across all four models (by up to -7.80 points). Even after data selection, every method including MARS (70.64) underperforms the untuned Base model (70.96) on Qwen-2.5-7B (with GSM8K dropping -5.00 points under MARS). Because the setup pairs an outdated 2022–2023 candidate pool with strong 2024–2025 base models and runs full-parameter SFT for 2 epochs with identical hyperparameters (Appendix B.3)—meaning Full takes 20× more update steps (~4,230 vs. ~211 steps) and suffers heavier catastrophic forgetting, even losing to Random (5%)—selecting 5% of the data primarily mitigates training damage rather than improving capability over Base.
  2. Fragility and incomplete coverage of the ablation study (Table 3 and Table 8): Because MARS is primarily an empirical integration of existing components, its claims rely heavily on rigorous ablation validation, yet the ablation study in Table 3 is fragile in several ways.
  • First, full component-level ablations in Table 3 are conducted on only a single model (Llama-3.1-8B); when w/o model calibration is evaluated on other models in Table 8 (Appendix C), the average gain on Gemma-3-4B shrinks to just +0.68 points (54.47 vs. 53.79) and model calibration actually degrades TyDiQA (74.44 vs. 74.83).
  • Second, more than half of the average gap between MARS (63.69) and the ablated variants (60.11–61.93) in Table 3 is driven by a single task (GSM8K: 66.11 vs. 56.33–62.93), whereas the intermediate ablations (w/o allocation difficulty at 61.84, w/o profile weighting at 61.93, and w/o allocation demand at 61.78) differ by at most 0.15 points without reported standard deviations across random seeds, and every variant in Table 3 (including full MARS at 64.83) still underperforms the untuned Base model on MMLU (65.64).
  • Finally, although the quality score q(z) is presented as a core atlas component, Eq. 13 places -q(z) after the continuous Euclidean distance d(z) as a lexicographic tie-breaker, and no ablation on q(z) is provided.

References

  • Mengzhou Xia et al., LESS: Selecting Influential Data for Targeted Instruction Tuning, ICML 2024.
Questions for authors
  1. How do the authors explain that Full underperforms Base in 13/20 settings and even MARS (70.64) underperforms Base (70.96) on Qwen-2.5-7B, and how does Full perform when trained for an equal training budget or with early stopping / LoRA?
  2. Given the lexicographic key in Eq. 13 where d(z) is a continuous Euclidean distance, how often does q(z) actually affect selection, and why are the allocation shifts in Figure 4 nearly identical across all four base models?
7 comments
AuthorsOpenReview ↗

Performance Degradation After SFT and Training-Budget Controls (W1/Q1)

We thank the reviewer for recognizing MARS's practical design and broad evaluation. Your concerns motivated us to further examine the effects of training length and fine-tuning strategy through additional matched-step and LoRA experiments.

1. Understanding Performance Degradation After SFT

The candidate pool (FLAN v2, CoT, Dolly, and OASST1) contains heterogeneous general-purpose instruction data, whose training objectives may differ from those of specific downstream benchmarks. For stronger pretrained models, full-data SFT may introduce task-irrelevant or potentially detrimental updates that interfere with previously acquired capabilities.

Similar behavior has been reported in LESS [1] using the same candidate pool. Full-data fine-tuning reduces MMLU from 55.3 to 54.5 on Llama-2-13B, and reduces MMLU and BBH from 62.4 to 60.4 and from 54.9 to 53.0, respectively, on Mistral-7B. LESS also discusses the potential impact of task-irrelevant or harmful training examples, particularly for stronger base models.

Importantly, MARS improves over the untuned Base in 17 of the 20 model–task settings and outperforms Random selection in all 20 settings under the same 5% budget. Its five-task average also exceeds Base on Llama, OLMo, and Gemma. On Qwen, the 0.32-point average decline is concentrated in GSM8K (83.62 → 78.62), while the other four tasks improve.

2. Matched Optimization Steps and LoRA Controls

To directly examine whether our gains arise primarily from reduced training length, we introduce two controls:

  • Matched-step Full: We sample from the complete candidate pool and match the number of optimizer steps used by MARS, while keeping the optimization configuration consistent.
  • LoRA fine-tuning: We compare Full, Random 5%, and MARS 5% using identical LoRA settings to evaluate whether the selection advantage persists under parameter-efficient fine-tuning.

LoRA configuration:

HyperparameterSetting
Rank / Alpha / Dropout128 / 512 / 0.1
Target Modulesq_proj, k_proj, v_proj, o_proj
Learning Rate / Scheduler2e-5 / Cosine / 1% warmup
Global Batch / Max Length128 / 4096 tokens
Precision / Training SeedBF16 / 42

LoRA experiments use four GPUs with DeepSpeed ZeRO-2, and all LoRA variants share the same configuration.

Table R1: Training-budget and fine-tuning controls on Qwen2.5-7B.

SettingMMLUBBHTyDiQAGSM8KMBPPAvg.
Base74.1665.8370.0083.6261.2070.96
Full (2 epochs)73.4262.0471.2976.1258.4068.25
Full (Matched Steps)74.1564.1770.2077.0361.8069.47
MARS (2 epochs)74.2666.1172.2278.6262.0070.64
LoRA Full73.8961.3971.1775.1360.8068.48
LoRA Random 5%74.2163.0671.4976.5060.6069.17
LoRA MARS 5%74.2863.2471.9978.5462.0070.01

Training length indeed affects Full-data performance, but MARS retains a clear advantage after controlling for optimization steps. Reducing Full to the same number of optimizer steps improves its average from 68.25 to 69.47 (+1.22), yet MARS still achieves 70.64 (+1.17) and performs better on all five tasks.

The selection advantage also persists under LoRA. With the same 5% data budget and LoRA configuration, MARS achieves 70.01 compared with 69.17 for Random (+0.84), with improvements across all five tasks.

These experiments demonstrate that MARS's benefits extend beyond reducing training updates and persist under different fine-tuning strategies.

Reference

[1] Xia et al. LESS: Selecting Influential Data for Targeted Instruction Tuning. ICML 2024. https://proceedings.mlr.press/v235/xia24c.html

AuthorsOpenReview ↗

Robustness and Generalizability of Component Ablations (W2)

To address the reviewer's concerns about the limited model coverage, task-level concentration of gains, and missing variance estimates, we extend the component ablations to all four base models and repeat the experiments using three matched training seeds (42, 43, 44). The Candidate Ability Atlas, reference sets, selection budget, and training configurations are held fixed, with only the corresponding component removed.

1. Cross-Model Component Ablations

Table R2: Component ablations across four base models (5% budget, five-task average, mean ± sample SD over three seeds).

VariantLlama-3.1-8BQwen2.5-7BOLMo-2-7BGemma-3-4B
w/o model calibration60.04 ± 0.4768.34 ± 0.5553.13 ± 0.3156.94 ± 0.30
w/o allocation difficulty61.59 ± 0.5769.33 ± 0.6853.86 ± 0.3357.33 ± 0.38
w/o allocation demand61.45 ± 0.5369.15 ± 0.6653.78 ± 0.1857.21 ± 0.28
w/o profile weighting61.61 ± 0.6069.27 ± 0.7653.88 ± 0.2657.31 ± 0.20
w/o profile matching60.68 ± 0.5868.88 ± 0.3053.54 ± 0.3557.09 ± 0.39
MARS63.09 ± 0.5770.10 ± 0.5154.37 ± 0.1157.72 ± 0.27

MARS achieves the highest five-task average in all 20 model–component comparisons, with improvements in 97 of 100 individual model–task comparisons based on three-seed means.

Removing model calibration reduces average performance by 3.05, 1.76, 1.24, and 0.78 points on Llama, Qwen, OLMo, and Gemma, respectively. The allocation and profile components also contribute positively across all four models, supporting their generalizability beyond Llama.

2. Are the Ablation Gains Dominated by GSM8K?

We further recompute the improvements after excluding GSM8K, which accounts for a substantial portion of certain ablation gaps in the original Table 3.

Table R3: MARS improvements over each ablation, averaged across the four tasks excluding GSM8K.

Removed ComponentLlamaQwenOLMoGemma
Model calibration+1.69+1.73+1.16+0.81
Allocation difficulty+1.19+0.70+0.48+0.43
Allocation demand+1.30+0.96+0.58+0.43
Profile weighting+0.65+0.75+0.53+0.36
Profile matching+2.01+1.11+0.74+0.53

All 20 model–component comparisons remain positive even after excluding GSM8K. This provides direct evidence that the contributions of the calibration, allocation, and profile-matching components extend across multiple downstream tasks.

3. Interpreting the Similar Ablation Scores

The reviewer notes that three intermediate ablations in the original Table 3 differ by at most 0.15 points. The relevant comparison for evaluating each component is its performance difference from the complete MARS framework.

In the original results, removing allocation difficulty, allocation demand, or profile weighting reduces the average by 1.85, 1.91, and 1.76 points, respectively. The new cross-model and multi-seed experiments further confirm positive average contributions from each component.

We will include the complete per-task, per-seed results in the revised paper to make the component contributions and training variability transparent.

Together, these results substantially strengthen the ablation evidence: the key components contribute across model families, their benefits extend beyond GSM8K, and the observed improvements persist in three-seed averages.

AuthorsOpenReview ↗

Effect of the Quality Score on Data Selection (Q2, Part 1/2)

1. Why Does the Quality Score Affect Selection?

We appreciate the reviewer's observation that the quality score q(z)q(z) serves as a secondary criterion in the lexicographic ranking. Despite using Euclidean distance, MARS produces exact distance ties because its ability representations are derived from discrete scores.

Specifically, the LLM annotator assigns integer scores from 0 to 100 across six ability dimensions. After normalization, different candidates can have identical ability distributions c(z)c(z), resulting in identical distances to the same reference profile.

For example, the raw ability vectors (20,40,0,0,0,0)(20,40,0,0,0,0) and (40,80,0,0,0,0)(40,80,0,0,0,0) produce the same normalized representation. Consequently, quality-based tie-breaking can meaningfully affect which candidates are selected, particularly when tied candidates occur at the selection boundary.

2. Direct Ablation of Quality-Based Tie-Breaking

To quantify its actual contribution, we keep the Candidate Ability Atlas, ability buckets, quotas, reference profiles, and deterministic hash unchanged. We only replace the lexicographic ranking key:

(d(z),−q(z),ρ(z))⟶(d(z),ρ(z)).(d(z),-q(z),\rho(z))\quad\longrightarrow\quad(d(z),\rho(z)).

We then measure both the changes in the selected subsets and the resulting downstream performance.

Table R4: Effect of quality-based tie-breaking across four base models (5% budget, 13,534 examples per task).

Base ModelAvg. Replaced ExamplesAvg. Replacement Ratew/o Quality Avg.MARS Avg.Gain
Llama-3.1-8B1,2269.06%63.4263.69+0.27
Qwen2.5-7B8326.15%69.7570.64+0.89
OLMo-2-7B1,1728.66%54.0254.50+0.48
Gemma-3-4B9136.75%56.9757.49+0.52

The number of replaced examples is averaged across five target tasks and rounded down. The replacement rate is computed as (∣Sq∖Sw/o quality∣/B)(|S_q\setminus S_{\mathrm{w/o\ quality}}|/B), averaged across tasks. Downstream performance corresponds to the original single-seed evaluation.

Quality-based tie-breaking changes 6.15%–9.06% of the selected examples and improves five-task average performance by 0.27–0.89 points across the four models.

These results directly establish that the quality score has a measurable effect on both subset composition and downstream performance, even though it is used only as a secondary ranking criterion.

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

MARS selects instruction-tuning data using reusable candidate descriptions. An external LLM assigns six ability scores to each example. The model being fine-tuned scores a small target reference set; these losses determine ability quotas and profiles for selecting candidates. This avoids running each new model over the full candidate pool. At a 5% data budget, MARS reports the highest five-task average among the compared selectors on four base models. Ablations and cross-model subset transfer support the value of model-specific selection.

Critical strengths
  • Reusable annotations make adaptation inexpensive. Each new model scores only the small reference set instead of the full candidate pool. The reported selection cost includes the one-time annotation charge.
  • The controls support useful model-specific selection. Experiments cover four model families and five tasks. Within-MARS ablations and cross-model transfer show that difficulty weighting and the recipient model's own feedback improve results.
Critical weaknesses
  • My main concern is limited significance. Skill-guided selection using model feedback has precedents in Skill-It and STAT. What remains unclear is why these six abilities are the appropriate representation. Removing or merging abilities supports this particular design, but does not explain its advantage over other representations. The choice therefore still feels somewhat ad hoc, and the paper provides limited insight beyond showing that the recipe works.
  • The practical gains do not convincingly offset the limited advance. Average margins over the strongest selectors are about 1.0–1.3 points, without reported run counts or variation. The largest cost saving is against repeated LESS selection; savings against repeated NAIT are about 26%, while transferred NAIT subsets are cheaper but less accurate. The evidence supports a useful tradeoff, with uncertainty about the practical magnitude of its benefit.
Questions for authors

Why were these six abilities chosen, and what justifies this representation over alternatives?

8 comments
AuthorsOpenReview ↗

Novelty and Rationale for the Six-Dimensional Ability Representation (W1/Q1, Part 1/2)

We thank the reviewer for recognizing MARS's reusable annotations, lightweight model calibration, and cross-model evaluation. Your comments have helped us clarify our contributions and motivated additional experiments comparing alternative representations and evaluating cross-task generalization.

1. Research Motivation and Key Contributions

Skill-It [1] dynamically adjusts data allocation based on skill dependencies and training losses, while STAT [2] diagnoses missing skills from student responses to guide data sampling or generation. These works provide important foundations for skill-guided data selection.

MARS investigates a distinct question: How can the same heterogeneous candidate pool serve different target tasks and base models while enabling both reusable candidate characterization and model-specific instance selection?

The key challenge is to decouple reusable data characteristics from the selection requirements of the current base model. To address this, MARS introduces three complementary designs:

  1. Reusable functional representations: We construct a continuous ability space independent of target tasks and base models, providing a unified characterization of the functional requirements of heterogeneous candidates and target references.
  2. Two-level model adaptation: Using only response NLLs on a small target reference set, we calibrate both ability-specific quotas (how many examples to select) and difficulty-weighted reference profiles (which examples to select), enabling instance-level selection.
  3. Cross-model reuse: The Candidate Ability Atlas is constructed once. Adapting to a new base model requires no recomputation of model-dependent features over the full candidate pool.

The central contribution of MARS is a mechanism that connects reusable candidate representations with lightweight model feedback, enabling model-specific selection while sharing candidate information across models. This research question and its associated mechanism distinguish MARS from prior approaches based on skill-level dynamic sampling or student skill diagnosis.

2. Why These Six Abilities?

Our goal is to construct a general-purpose functional ability space that supports a broad range of downstream tasks, without tying the representation to specific benchmarks or domains.

Although downstream tasks differ in their formats and objectives, their functional requirements can often be described through combinations of fundamental abilities, including instruction following, context utilization, knowledge retrieval, logical reasoning, quantitative reasoning, and procedural reasoning. For example, reading comprehension may involve both evidence integration and quantitative reasoning, while code generation may require constraint following and logical inference.

Motivated by prior research on these capabilities (Appendix A.1), we adopt six functional dimensions with cross-task applicability and use continuous distributions to represent the combination of abilities involved in each example. This allows new downstream tasks to be mapped into a shared candidate space using only small reference sets, without redefining the ability taxonomy or reannotating the candidate pool.

Continued in the next comment, where we provide controlled comparisons with alternative representations.

AuthorsOpenReview ↗

Novelty and Rationale for the Six-Dimensional Ability Representation (W1/Q1, Continued — Part 2/2)

Continuing our response to W1/Q1, we provide direct experimental evidence for the proposed functional representation.

3. Direct Comparison with Alternative Representations

To evaluate the contribution of the six-dimensional functional representation, we introduce two alternatives with the same dimensionality:

  • Semantic Embedding Soft Clustering (K=6): We encode candidate examples using a fixed pretrained embedding model, construct six semantic centroids through K-means, and derive continuous six-dimensional soft assignments based on distances to these centroids. Reference examples are mapped into the same space using the identical encoder and fixed centroids.
  • Automatically Induced Task-Function Taxonomy (K=6): Using the same LLM annotator as MARS, we automatically induce six task-function categories from candidate examples: information seeking and explanation, judgment and verification, content transformation and revision, content drafting and expression, problem solving and implementation, and advice and planning. After freezing these definitions, we annotate both candidates and references with six-dimensional scores and normalize them.

Both alternatives are integrated into the same MARS selection pipeline, with identical reference NLLs, quota computation, within-bucket profile matching, quality-based tie-breaking, and training configurations. Only the candidate and reference representations are changed.

Table R1: Comparison of alternative representations on Llama-3.1-8B with a 5% data budget.

RepresentationMMLUBBHTyDiQAGSM8KMBPPAvg.
Semantic Embedding Soft Clustering (K=6)61.5162.2269.4962.9346.8060.59
Automatically Induced Task-Function Taxonomy (K=6)62.1661.3970.5259.1447.6060.16
MARS Six-Dimensional Functional Representation64.8364.7273.2066.1149.6063.69

MARS outperforms both alternatives on all five tasks, with average gains of 3.10 and 3.53 points, respectively. Together with the original Table 13, where removing any ability dimension or merging the six dimensions into three consistently reduces performance, these results provide direct evidence for the proposed functional decomposition: the functional structure captured by the representation contributes meaningfully to reference calibration and instance selection.

References

[1] Chen et al. Skill-It! A Data-Driven Skills Framework for Understanding and Training Language Models. NeurIPS 2023.

[2] He et al. STAT: Skill-Targeted Adaptive Training. ICLR 2026.

AuthorsOpenReview ↗

Novelty and Rationale for the Six-Dimensional Ability Representation (W1/Q1, Continued — Part 2/2)

4. Generalization to New Downstream Tasks

To further evaluate the generality of our six-dimensional representation, we extend the evaluation to five additional benchmarks, including tasks with substantially different objectives from the original five.

Importantly, we keep the six ability definitions, annotation protocol, Candidate Ability Atlas, and selection procedure unchanged. Each new task requires only a small target reference set for calibration, without redefining the ability taxonomy or reannotating the candidate pool.

(a) Generalization to new reasoning and QA benchmarks.

We first evaluate DROP, ARC-Challenge, and CommonsenseQA, which involve different combinations of evidence integration, quantitative reasoning, scientific knowledge, and commonsense reasoning. We also compare against the two alternative six-dimensional representations introduced in our previous comment.

Table R2: Generalization to additional QA and reasoning benchmarks (Llama-3.1-8B, 5% budget).

MethodDROP (F1)ARC-Challenge (Acc.)CommonsenseQA (Acc.)Avg.
NAIT16.9954.7872.9748.25
MoNA17.0055.2073.3848.53
Semantic Embedding Soft Clustering (K=6)16.7654.3572.6547.92
Automatically Induced Task-Function Taxonomy (K=6)16.9054.6172.8148.11
MARS18.5856.2374.6149.81

MARS achieves the best performance on all three benchmarks, exceeding the strongest selection baseline by 1.58, 1.03, and 1.23 points, respectively. It also outperforms semantic clustering and the automatically induced taxonomy by 1.89 and 1.70 points on average.

Together with Table R1, the proposed functional representation outperforms both alternatives across all eight benchmarks on which they were evaluated.

(b) Generalization to different task paradigms.

To more directly examine whether the representation is tied to the original benchmark taxonomy, we further evaluate machine translation (WMT14) and abstractive summarization (XSum). These tasks have different output objectives, and our six-dimensional taxonomy contains no translation- or summarization-specific dimensions.

Table R3: Generalization to translation and summarization (Llama-3.1-8B, 5% budget).

MethodWMT14 BLEUWMT14 chrFXSum ROUGE-1ROUGE-2ROUGE-L
NAIT16.4650.4824.758.4720.18
MoNA11.8447.5928.4210.6922.46
MARS17.5251.1429.0411.4723.14

MARS achieves the best results on all five reported metrics across these two tasks. Compared with the strongest baseline, it improves WMT14 BLEU by 1.06 points and XSum ROUGE-L by 0.68 points.

These results provide stronger evidence for the cross-task applicability of the proposed representation: the same six-dimensional ability space supports reasoning, QA, translation, and summarization without introducing task-specific dimensions or reconstructing candidate representations. This demonstrates the value of a reusable functional ability space for adapting data selection to diverse downstream objectives.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chairs and Reviewers,

We sincerely thank all reviewers for their thoughtful feedback and constructive suggestions. We appreciate their recognition of MARS's practical motivation, reusable candidate representations, lightweight model adaptation, and broad empirical evaluation. Their comments have helped us strengthen the experimental evidence and clarify the contributions of our work.

We have conducted additional experiments and analyses to address the main concerns, summarized below.

1. Novelty and ability representation (XbFt, FETu). We clarify the distinctions between MARS and existing capability-guided selection methods, and introduce comparisons with CDT, same-annotator scoring baselines, and two alternative six-dimensional representations. We also evaluate five additional benchmarks spanning reasoning, question answering, translation, and summarization.

2. Model adaptivity and statistical reliability (XbFt, xsBp, FETu). We provide controlled cross-model calibration experiments, extended component ablations, and three-seed evaluations across four base models. These results further support the contributions of model-specific calibration and the robustness of MARS.

3. Training and selection robustness (xsBp, FETu). We introduce matched-step and LoRA controls, evaluate selection budgets of 1%, 5%, and 10%, and analyze the effects of quality-based tie-breaking and ability allocation.

4. Efficiency and practical value (XbFt, FETu). We extend the cost analysis to additional baselines and direct subset transfer, explicitly account for one-time annotation costs, and clarify GPU-time measurements. These comparisons highlight the performance–cost trade-off of reusable candidate representations with lightweight model adaptation.

We hope that our responses and additional evidence have satisfactorily addressed the reviewers' concerns and clarified the contributions of our work.

Authors of Submission 12976

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.