acceptodds
Under review as a conference paper at ICLR 2027

Overcoming Prior Barriers: Supervised Fine-Tuning under Long-Tail Distribution

Abstract

Supervised fine-tuning (SFT) adapts pretrained large language models (LLMs) to downstream tasks, but the required concepts can receive substantially different levels of pretrained support. Frequent concepts are more likely to be well learned, whereas rare concepts may remain weakly represented. We introduce a novel notion named prior barrier to quantify how strongly the pretrained model supports competing concepts over the target concept. We observe that prior barriers follow a long-tail distribution, placing head and tail concepts at different starting points for SFT: head concepts face lower prior barriers, whereas tail concepts require additional instructions to overcome their higher prior barriers. Our theoretical analysis further derives a predictive risk bound for SFT under long-tail prior barriers, explicitly characterizing how the prior barrier and accumulated SFT evidence jointly determine predictive performance. Motivated by this prior barrier-dependent demand, we propose PASS, an adaptive SFT instruction selection method that constructs reference-derived concepts and estimates the distinguishing evidence provided by each instruction, and adaptively allocates the selection budget toward concepts that remain insufficiently covered under the current selection. In this way, PASS jointly considers which instructions can provide useful evidence and where additional supervision is needed under a limited budget. Experiments show that our method consistently outperforms seven state-of-the-art instruction selection methods on four backbone-budget settings. An ablation study further shows that PASS's adaptive allocation consistently improves over uniform allocation.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Instruction Selection, LLM, Long-Tail Distribution, Bayesian View of SFT

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AQvhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies instruction selection for supervised fine-tuning when target concepts receive unequal pretrained support. It introduces a prior barrier comparing a target concept's support with that of competing concepts, and derives posterior concentration and predictive risk bounds within a Bayesian concept-space model. The proposed PASS method constructs reference-derived concepts, estimates distinguishing evidence from desired and incorrect response representations, and selects instructions using barrier-aware weights, diminishing coverage rewards, and redundancy penalties. Its allocation weights satisfy Aθ∝eBθ/γθA_\theta \propto e^{B_\theta}/\gamma_\theta, combining the estimated barrier with existing pretrained support. Experiments compare two 7B backbones and two selection budgets across five downstream capabilities. PASS reports competitive aggregate performance and small improvements over equal per-capability allocation.

Critical strengths
  1. An explicit formulation of pretrained support asymmetry.

The paper separates the target concept's initial pretrained disadvantage from the distinguishing evidence provided by supervision. This provides a coherent motivation for considering both existing model support and candidate utility when allocating a limited fine-tuning budget.

  1. Promising empirical budget efficiency.

The evaluation includes two backbones, two budgets, and seven selection baselines under shared fine-tuning settings. On Mistral-7B, PASS with 5K examples achieves 51.5, exceeding the strongest reported 10K baseline score of 50.9. These results suggest potential practical value, subject to establishing their robustness.

Critical weaknesses
  1. The algorithmic increment over closely related adaptive selection methods is insufficiently isolated.

MIG, published in Findings of ACL 2025, already models diminishing information gains across semantic labels and explicitly evaluates the exponential saturation function 1−e−αx1-e^{-\alpha x}. BIDS, published in Findings of EMNLP 2025, normalizes candidate-target utilities and iteratively favors underrepresented targets. Its Appendix A.7 further demonstrates this mechanism using hidden-representation similarities through RDS-BIDS. Thus, adaptive coverage and representation-based balancing are established ingredients.

PASS's more distinctive components appear to be its desired/incorrect-response evidence and pretrained barrier/support weights. Although Section 6 acknowledges related approaches, the experiments do not establish the added value of these components over the shared coverage mechanism. In particular, Table 2 compares against equal allocation across five capabilities, rather than retaining the same concepts and objective while removing barrier-based weighting. This leaves the principal methodological contribution insufficiently demonstrated.

  1. The evidence supporting the barrier-based explanation and the reported improvements remains incomplete.

The theory concerns a Gibbs posterior with independent and identically distributed supervision, whereas the experiments use LoRA fine-tuning on adaptively selected heterogeneous instructions. The correspondence between the cosine-margin surrogates and theoretical prior ratios or likelihood evidence is not validated. Corollary 1 also establishes sufficient supervision conditions, rather than necessary head-tail sample requirements. Empirically, Table 1 reports only aggregate scores without repeated-run uncertainty. PASS's margins over the strongest baselines are approximately 0.9, 0.9, 0.2, and 0.0 points. PASS and RANDOM both report 65.6 for Qwen2.5-7B with a 10K budget, while Table 2's allocation gains are only 0.26–0.56 points. Consequently, neither the reliability of the improvements nor their attribution to prior-barrier modeling is sufficiently established.

Questions for authors
  1. Could the authors clarify the incremental contribution relative to MIG and BIDS, including RDS-BIDS? A controlled comparison using the same concept partition, evidence matrix, and coverage objective, with uniform versus barrier/support-based weights, would directly test the proposed contribution.

  2. Do the estimated barriers predict actual supervision requirements or subsequent fine-tuning gains beyond simpler concept-level signals such as reference error rate or response loss? Results stratified by barrier would help validate the proposed explanation. Please also clarify the intended scope of the connection between the Gibbs-posterior analysis and practical LoRA fine-tuning.

  3. Could the authors provide per-benchmark results and variability across repeated selection and fine-tuning runs? In particular, are the Table 2 improvements larger than run-to-run variation, and what is the unrounded PASS-RANDOM difference for Qwen2.5-7B with a 10K budget?

  4. Could the authors specify the calibration functions, aggregation rule, clustering criterion, and complete training and evaluation configuration? Please clarify how Equation 13 handles γθ=0\gamma_\theta=0 and how the redundancy term is defined when ∣S∣≤1|S|\leq 1 during greedy initialization.

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

This paper proposes PASS, an instruction selection method for SFT that accounts for unequal pretrained support across concepts. The authors introduce a prior barrier and derive theoretical bounds connecting it to the evidence provided by SFT data. PASS estimates this barrier and the usefulness of candidate instructions from response representations, then adaptively allocates a limited selection budget across concepts. Experiments on Mistral-7B and Qwen2.5-7B with two selection budgets show competitive average performance across five benchmarks, with additional comparisons supporting adaptive allocation over uniform allocation.

Critical strengths

(1) The prior-barrier perspective is interesting and well motivated. Within the Bayesian framework, the theoretical bounds make explicit how pretrained support and distinguishing evidence affect the supervision sufficient for learning a target concept. This provides a useful perspective on why different concepts may need different amounts of SFT data.

(2) PASS considers both which instructions are useful and which concepts need additional supervision. The data-efficiency result on Mistral-7B is encouraging: with 5K instructions, PASS obtains a higher average score than every evaluated baseline using 10K instructions.

Critical weaknesses

(1) My main concern is whether the empirical barrier actually predicts supervision demand. The theory uses a prior-density ratio, whereas Eq. 11 uses a margin based on representation similarity. I didn't find evidence that concepts with higher scores need more SFT examples to reach comparable performance. The comparison with uniform allocation in Table 2 doesn't resolve this, since the allocation weights depend on both the barrier and absolute pretrained support. It therefore remains unclear whether the empirical score captures the supervision needs that motivate the method.

(2) I find the claim in the abstract and conclusion that PASS outperforms seven baselines across all four settings too strong. In Table 1, PASS and Random both score 65.6 on Qwen2.5-7B at 10K, so there is no improvement at the reported precision. On the same backbone at 5K, the gain is only 0.2 points, and the allocation ablation shows gains of 0.26–0.56 points. Without knowing how much these results vary across runs, it is difficult to tell whether the small differences are reliable. The results are encouraging, but they do not yet establish the consistent advantage claimed in the paper.

Questions for authors

(1) Does BθB_\theta predict how much SFT data a concept needs to reach a given performance level? How do concepts with different barrier scores respond as the SFT budget increases?

(2) How many runs were used for Tables 1 and 2? Some of the gains over the best baseline and Uniform are quite small. Do they hold up across different seeds? Could you also give the scores for each benchmark, so I can see where the improvement in the average comes from?

Reviewer_8EupOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies SFT instruction selection when the concepts required by a known downstream task exhibit a long-tail distribution of prior barriers. Under a Bayesian concept-space view in which SFT yields a Gibbs posterior, the authors define the prior barrier and derive posterior concentration, a head–tail supervision gap, and a predictive risk bound. The proposed PASS clusters a benchmark-derived reference set into concepts and scores each candidate’s distinguishing evidence as a calibrated similarity margin. It then estimates each concept’s barrier and pretrained support to form allocation weights, and greedily optimizes an objective. Experiments use 7B models with LoRA against seven baselines.

Critical strengths
  • The paper links budget allocation to unequal pretrained support, unlike selectors based on relevance, quality, or distribution matching, and unlike skill-level allocation methods (e.g., Skill-It, BIDS).
  • Corollary 1 makes the barrier–evidence trade-off explicit: in the tail case. The proofs in Appendix A appear largely sound.
  • PASS keeps a concept-level evidence matrix rather than a single score per candidate.
Critical weaknesses
  1. The theory and the implementation are disconnected. Definition 1 defines B(r)B(r) as a log ratio of prior densities, whereas Eq. 11 computes BθB_\theta as a cosine-similarity gap of M0M_0's generations. Likewise, ϵiθ\epsilon_{i\theta} in Eq. 10 is a single-instance margin, whereas ϵ(r)\epsilon(r), defined via Eq. 1 and Definition 2, is an expected KL divergence, and no evidence is given that these surrogates track the theoretical quantities. Competing concepts are defined as Ar={θ:∥θ−θ⋆∥≥r}A_r=\{\theta:\|\theta-\theta^\star\|\ge r\}, but Sec. 4.1 implements them as wrong-response centroids nθkn_{\theta k} within the same cluster and never computes any inter-concept distance. The allocation weight in Eq. 13 is not derived either.

  2. The prior barrier measures the strength of competing concepts, not the rarity of the target concept. Definition 1 only asks whether a strongly supported competitor exists, yet the Abstract and Sec. 1 equate tail concepts with rare ones without verification. TruthfulQA is a counterexample.

  3. I am concerned that the barrier estimates may reflect pretrain-SFT format mismatch rather than a concept-level disadvantage. BθB_\theta and γθ\gamma_\theta in Eqs. 11–12, as well as part of nθkn_{\theta k}, rely on generations from the pretrained. Mistral-7B's BASE score of 26.9 in Table 1 suggests substantial format failures, so these quantities may partly capture instruction-following deficits, and their meaning likely differs across backbones given that Qwen2.5 already scores 52.8 without SFT. The generation setup, including prompting, few-shot examples, and decoding parameters, is not reported.

  4. The setting is benchmark-targeted selection, which limits the generality of the conclusions. The reference set is 20% of each evaluation benchmark and is therefore identically distributed with the test data. PASS also uses "available incorrect alternatives" described, presumably benchmark distractors, and thus exploits more benchmark information than baselines such as Random and DEITA. The general framing of the title and abstract does not match this setting, and no held-out benchmark is evaluated.

  5. The empirical evidence does not support the central claims. The gains in Table 1 are small and reported without variance. For ablations, the uniform variant imposes hard quotas per capability whereas PASS softly weights concepts; no other component such as γθ\gamma_\theta, Rθ−R^-_\theta, or D(S)D(S) is ablated, contrary to the claim that the ablation demonstrates "the necessity of each component." The sensitivity range in Tab 3 matches PASS's margin in the same setting, and the strong redundancy penalty beats the default, so the conclusion that PASS is "not sensitive" is unwarranted. Tab 3 also covers a single setting, and the data used for hyperparameter selection is unstated.

  6. The scale and efficiency evaluations are limited for the line of data selection research. Only two 7B backbones with LoRA are validated, leaving the effects of model scale and full fine-tuning untested.

Questions for authors
  1. How many concepts are produced, and how many per benchmark? How pure are the clusters with respect to task or subject labels?
  2. Implementation details are missing. E.g. How is the reference set sampled? Why are there only 200 GSM8K evaluation examples? Which prior work does the 20/80 split follow? Is the reference set excluded from the candidate pool and from SFT?
  3. Given that data selection is sensitive to randomness and settings, can the authors report mean ±\pm std over seeds, per-benchmark results, and results stratified by BθB_\theta?
  4. Does PASS outperform Random on held-out benchmarks not used for the reference set?
  5. On what data were hyper-parameters tuned?
  6. For Figure 4: which backbone and budget, and how many GPUs? What is the per-stage runtime, the complexity of the greedy step, and the acceleration used? How does PASS compare with Random at equal GPU-hours?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.