PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Efficient Medical Reasoning with Minimal Fine-Tuning Data

Xinlin Zhuang, Feilong Tang, Haolin Yang, Xiwei Liu, Ming Hu, Huifa Li, Haochen Xue, Junjun He, Zongyuan Ge, Yichen Li, Ying Qian, Imran Razzak

arXiv:2508.01450v3cs.CL

TL;DR

Medical models still struggle to turn incomplete, ambiguous clinical information into reliable decisions, while medical SFT has explored scaling up CoT datasets. DIQ selects samples jointly by difficulty and model-dependent influence; across nine downstream tasks, fine-tuning with 1–10% of selected data matches or surpasses full-dataset training.

  • Problem

    Medical models still struggle to turn incomplete, ambiguous clinical information into reliable decisions, while medical SFT has explored scaling up CoT datasets.

    Source: §1 · p. 1

  • Method

    DIQ prioritizes medical-reasoning samples with jointly high difficulty and model-dependent influence, measured using a gradient inner product.

    Source: §1 · p. 2, §6 · p. 8

  • Results

    1–10% DIQ-selected data matches or surpasses full-dataset training across nine downstream tasks.

    Source: §6 · p. 9

  • Takeaways & Limitations

    DIQ-selected subsets preserve or improve accuracy while reducing the training-data footprint and computational cost.

    Source: §6 · p. 9

    DIQ has not yet been evaluated on very large LLMs (≥70B parameters) because of computational resource constraints.

    Source: §6 · p. 9

Abstract

from arXiv · show

Supervised Fine-Tuning (SFT) of the language backbone plays a pivotal role in adapting Vision-Language Models (VLMs) to specialized domains such as medical reasoning. However, existing SFT practices often rely on unfiltered textual datasets that contain redundant and low-quality samples, leading to substantial computational costs and suboptimal performance in complex clinical scenarios. Although existing methods attempt to alleviate this problem by selecting data based on sample difficulty, defined by knowledge and reasoning complexity, they overlook each sample's optimization utility reflected in its gradient. Interestingly, we find that gradient-based influence alone favors easy-to-optimize samples that cause large parameter shifts but lack deep reasoning chains, while difficulty alone selects noisy or overly complex textual cases that fail to guide stable optimization. Based on this observation, we propose a data selection strategy, Difficulty-Influence Quadrant (DIQ), which prioritizes samples in the "high-difficulty-high-influence" quadrant to balance complex clinical reasoning with substantial gradient influence. This enables efficient medical reasoning for VLMs with minimal fine-tuning data. Furthermore, Human and LLM-as-a-judge evaluations show that DIQ-selected subsets demonstrate higher data quality and generate clinical reasoning that is more aligned with expert practices in differential diagnosis, safety check, and evidence citation, as DIQ emphasizes samples that foster expert-like reasoning patterns. Extensive experiments on medical reasoning benchmarks demonstrate that DIQ enables VLM backbones fine-tuned on only 1% of selected data to match full-dataset performance, while using 10% consistently outperforms baseline methods, highlighting the superiority of principled data selection over brute-force scaling. The code is available at https://github.com/mihara-bot/DIQ.

1. Introduction

The introduction identifies a gap in medical SFT data selection: difficulty-based methods overlook optimization utility, while influence alone can favor samples with shallow reasoning. DIQ combines difficulty and influence, and the paper reports improved data efficiency and expert-aligned reasoning.

1. Introduction

  • Existing difficulty-based selection overlooks optimization utility, while unfiltered SFT data can contain redundant, low-value samples that raise computational costs.Earlier approaches used coarse difficulty heuristics, and later work separated knowledge and reasoning complexity but remained one-dimensional.Source: §1 · pp. 1–2
  • In the pilot, high-influence, low-difficulty Q2 data outperformed low-influence, high-difficulty Q3 data despite lower reasoning quality.Q2 samples had shallower reasoning chains, while Q3 samples provided weak gradient signals and were associated with unstable training and lower downstream performance.Source: §1 · p. 2
  • DIQ prioritizes samples in the high-difficulty, high-influence quadrant to balance complex clinical reasoning with training influence.Difficulty is estimated with a BiomedBERT-based classifier, and influence aggregates first-order gradient dot products across training and validation samples, tasks, and epochs.Source: §1 · p. 2
  • 1% of DIQ-selected data matches full-dataset performance, while 10% consistently outperforms baseline methods.Human and LLM-as-a-judge evaluations also find higher data quality and reasoning aligned with expert practices in differential diagnosis, safety checks, and evidence citation.Source: §1 · p. 2
  • Human and LLM-as-a-judge evaluations find that DIQ-selected subsets have higher data quality and reasoning aligned with expert practices.The evaluated practices include differential diagnosis, safety checks, and evidence citation.Source: §1 · p. 2

2. Related Work

Related work describes both large-scale medical reasoning dataset construction and efforts to elicit reasoning from fewer high-quality examples. It also traces medical benchmarks from factual recall toward complex clinical analysis.

Medical Reasoning Dataset Construction

  • Large medical reasoning datasets are built with synthetic-data generation and multi-agent verification, or with knowledge graphs and multistage training.Examples include UltraMedical [35] and ReasonMed [22] for synthetic-data pipelines, and MedReason [29], FineMedLM-o1 [34], and HuatuoGPT-o1 [2] for other approaches.Source: §2 · p. 2
  • A separate line of work argues that base models with extensive domain knowledge can elicit complex reasoning from few high-quality examples.This paper investigates selecting small subsets from existing medical SFT datasets as a data-efficient alternative to full-dataset fine-tuning.Source: §2 · p. 2

Medical Reasoning Benchmarking

  • Medical benchmarks have expanded from knowledge-recall questions to clinical case analysis and expert-level cross-domain reasoning.The paper's nine-benchmark evaluation suite spans fundamental knowledge recall through sophisticated clinical reasoning.Source: §2 · pp. 2–3

3. Method

DIQ combines intrinsic reasoning difficulty with expected validation utility to select a target-sized training subset. It estimates difficulty with one of three classifier-predicted dimensions and influence through training/validation-gradient alignment.

3.1. Overview

  • DIQ scores samples by model-agnostic difficulty and model-dependent influence, then selects by quadrant priority to meet the target retention ratio.The selected subset is intended to maximize downstream performance under SFT.Source: §3.1 · p. 3

3.2. Difficulty Estimation

  • Difficulty is one BiomedBERT-predicted, calibrated 5-point score for Knowledge, Reasoning, or Overall complexity, with a percentile threshold defining the high/low split.The method uses a single chosen dimension as its difficulty scalar.Source: §3.2 · p. 3

3.3. Dot-Product Influence

  • Dot measures alignment between a sample gradient and the mean validation gradient, approximating expected one-step validation-loss reduction up to a second-order remainder.The approximation is supported when the learning rate is small and local curvature is bounded; computation uses O(|Dval| + |D|) backward passes without Hessian products.Source: §3.3 · pp. 3–4

3.4. Quadrant-Based Data Selection

  • DIQ partitions samples at a difficulty percentile and the median Dot, then fills the subset in Q1-to-Q4 order, ranking within each quadrant by Dot and breaking ties by difficulty.This selection order is also described as prioritizing Q1 before Q2, Q3, and Q4 until the retention target is reached.Source: Figure 2 · p. 4, §3.4 · p. 4

4. Experiment

Experiments compare DIQ-selected subsets with full-data training and baseline selectors across medical datasets and benchmarks. DIQ performs strongly at low retention ratios, and its selected data and trained models score higher on clinical reasoning dimensions.

4.1. Experimental Setup

  • Training uses five medium-scale medical datasets and 410k-sample UltraMedical; evaluation spans three standard and six challenging benchmarks.The training datasets are Huatuo, Huatuo-DS, FineMed, MedReason, m1, and UltraMedical.Source: §4.1 · pp. 4–5

4.2. Main Results

  • 10% DIQ exceeds full-data AvgA on both datasets (Huatuo: 44.04 vs. 43.44; FineMed: 39.36 vs. 38.57).At 1%, FineMed AvgA is 41.46 vs. 38.57 for full data; Huatuo reaches 42.78 vs. 43.44.Source: Table 1 · p. 5, §4.2 · p. 6

4.1. Experimental Setup

  • Experiments train Qwen3-8B/14B/32B and Llama3.1-8B-Instruct, with general non-reasoning and reasoning models as references.The reference models include GPT-4.1, DeepSeek-V3-0324, Gemini-2.5-Flash, DeepSeek-R1-0528, QwQ-32B, o4-mini-medium, and Gemini-2.5-Pro.Source: §4.1 · p. 5
  • DIQ is compared with Random, PPL, Similarity, and LESS; fine-tuning uses LoRA with rank 8 for three epochs.The reported task accuracy uses a two-step answer check: exact option match, then presence of the ground-truth answer text.Source: §4.1 · pp. 5–6

4.2. Main Results

  • On MedReason-QA without CoT traces, DIQ beats random selection at 1% and 10% for Llama3.1-8B-Instruct and Qwen3-8B.The results average five runs; the passage reports that DIQ preserves coverage of rare, clinically important cases under QA-only supervision.Source: §4.2 · p. 6, Data DDx SC EC · p. 6

Data DDx SC EC

  • At 1%, DIQ scores higher on DDx, Safety Check, and Evidence Citation for both selected data and trained-model rationales.Data-level gains are +0.80, +0.35, and +0.46; model-level gains are +0.05, +0.16, and +0.15, respectively.Source: Data DDx SC EC · p. 6, Table 2 · p. 6

5. Analysis

The analysis examines DIQ’s computational cost, the value of combining selection signals, sensitivity to validation-set size, and generalization across models and preference learning. Across these settings, it reports efficient selection and competitive or improved downstream performance, with transfer qualified by model family and retention budget.

5.1. Efficiency Analysis

  • DIQ’s one-time selection cost on Huatuo is 9.05 normalized FLOPs, 1.85x below Llama3.1-8B-Instruct full-data fine-tuning (16.70) and 2.08x below Qwen3-8B (18.79).For larger models, the selection pass is reported as 9.80x cheaper than a run costing 88.71 and nearly 970x cheaper than one costing 8777.89; cached scores can be reused.Source: §5.1 · pp. 6–7, Figure 4 · p. 7

5.2. Ablation Study

  • At 1% and 10% retention, DIQ reaches 42.78% and 44.04% accuracy, exceeding the strongest single-criterion baseline by +0.89 and +0.88 points, respectively.Margins over the average of single-criterion baselines are +1.73 points at 1% and +1.59 points at 10%.Source: §5.2 · p. 7

5.3. Effect of Validation Set Size

  • Increasing the validation split from 90 to 450 examples raises AvgA from 42.04 to 43.89 (+1.85), with most gains realized by 360 examples.At 360, AvgA is 43.80 (+1.76 over 90); increasing to 450 adds only +0.09, while AvgS grows nearly monotonically and AvgC peaks at 360.Source: §5.3 · p. 7, Table 3 · p. 7

5.4. Generalization of DIQ

  • Qwen-sourced influence improves over random selection for Qwen3-14B and Qwen3-32B across all tested retention budgets, while Llama-sourced cross-family transfer is beneficial in 6/9 settings.Cross-family transfer can be neutral or negative at low budgets on larger targets; gains reach +2.11 for Qwen3-8B at 10% retention.Source: §5.4 · p. 8, Figure 5 · p. 8
  • After DPO, 1% DIQ reaches AvgA 56.52 (+1.00 over full-data SFT + DPO), and 10% DIQ reaches 55.70 (+0.18).At 50% retention, DIQ + DPO is 55.37 (−0.15); DIQ also beats random selection on AvgA and AvgC at matched budgets.Source: §5.4 · p. 8, Table 4 · p. 8

6. Conclusion

The conclusion presents DIQ as a compact-subset selection framework that combines sample difficulty with model-dependent influence. Across nine downstream tasks, subsets containing 1–10% of the data match or surpass full-dataset training, while evaluation on models of at least 70B parameters remains future work.

6. Conclusion

  • DIQ selects compact medical-reasoning training subsets by jointly considering sample difficulty and model-dependent influence, prioritizing the high-difficulty/influence region.The framework describes this selection as a principled, model-aware curriculum.Source: §6 · pp. 8–9
  • Across nine downstream tasks, fine-tuning on 1–10% of DIQ-selected data matches or surpasses full-dataset training while reducing the training data footprint and computational cost.The conclusion reports preserved or improved accuracy across these tasks.Source: §6 · p. 9
  • DIQ has not yet been evaluated on very large LLMs, including models with at least 70B parameters, because of computational resource constraints.The authors identify scaling to such models as future work.Source: §6 · p. 9

A. Details of Pilot Experiment

The pilot experiment divides FineMed samples into four difficulty–influence quadrants and fine-tunes a separate Qwen3-8B model on a 1% subset from each. It assesses reasoning quality and downstream accuracy using a defined evaluation procedure.

A. Details of Pilot Experiment

  • The pilot partitions samples using a difficulty threshold of 3 and the median influence score, defining four high/low difficulty and influence quadrants.Separate Qwen3-8B instances are fine-tuned on 1% subsets drawn exclusively from each quadrant.Source: A. Details of Pilot Experiment · p. 11
  • Each quadrant is evaluated on reasoning quality with Gemini-2.5-pro’s five-level Likert scale and on accuracy across nine downstream datasets.The prompt asks for a holistic assessment of information completeness, synthesis, logical progression, differential diagnosis, medical knowledge, guidelines, and conclusion justification.Source: A. Details of Pilot Experiment · p. 11
  • The rubric assigns one overall score from 1 to 5, ranging from fundamentally flawed reasoning to exemplary, expert-like reasoning.Intermediate levels distinguish adequate reasoning with noticeable shortcomings from strong reasoning with minor presentation omissions.Source: A. Details of Pilot Experiment · pp. 11–12

B. Case Study

The case study traces a DIQ-1% FineMed model’s clinical reasoning from patient evidence to diagnosis and treatment. The supplied difficulty prompt rates knowledge and reasoning separately, then synthesizes them into an overall score.

B. Case Study

  • The model links severe immunosuppression and cerebrospinal-fluid findings, including a positive India ink stain, to cryptococcal meningitis and standard induction therapy.It distinguishes Amphotericin B and flucytosine for severe disease from treatments used for other pathogens or less severe presentations.Source: B. Case Study · p. 12

C.1. Prompt for Difficulty Score Annotation

Dimension 2: Reasoning Complexity (1-5 Levels)

  • Reasoning Complexity ranges from direct recall to expert-level problem solving involving complex, ambiguous data and multistep clinical judgment.The rubric considers information integration, reasoning-step complexity, ambiguity, and clinical judgment.Source: Dimension 2: Reasoning Complexity (1-5… · p. 13

Dimension 3: Overall Difficulty (1-5 Levels)

  • Overall Difficulty combines knowledge and reasoning demands, rating exceptionally high demands in both as Very Hard.The rubric describes high knowledge with low reasoning, or low knowledge with high reasoning, as moderate challenge.Source: Dimension 3: Overall Difficulty (1-5… · pp. 13–14

C.2. Difficulty Classifier Training

The appendix describes difficulty and influence scoring, their distributions, and a clinician-grounded rubric for assessing clinical reasoning. BiomedBERT is selected as the difficulty classifier; the supplied passages do not report numeric classifier results.

C.2. Difficulty Classifier Training

  • BiomedBERT outperformed ClinicalBERT and ModernBERT on difficulty prediction and was selected as the difficulty classifier.Table 6 is identified as reporting test-set F1 scores, but the supplied text gives no values.Source: §C.2 · p. 14, Table 6 · p. 14

C.3. Difficulty Score Distribution

D.1. Influence Score Computation

  • Influence scores are computed directly for five medium-scale datasets, while a trained rater predicts scores for large-scale UltraMedical to conserve compute.The rater is ModernBERT, selected after evaluation on 45,000 training and 5,000 test instances sampled from medium-scale datasets.Source: §D.1 · p. 14

D.2. Influence Score Distribution

  • Influence-score distributions are provided for FineMed, Huatuo, Huatuo-DS, m1, and MedReason.Figures 13–17 show these distributions.Source: §D.2 · p. 14

E. Details of Clinical Value Assessment

  • Three experienced clinicians identified differential diagnosis, safety check, and evidence citation as components of clinical value, which form the automated evaluation prompt.The prompt asks evaluators to score each category from 1 to 5 and provide specific rationales.Source: E. Details of Clinical Value Assessment · pp. 14, 16

1. Differential Diagnosis (DDx)

  • The DDx rubric scores whether responses generate relevant alternatives and systematically compare and narrow them using logical reasoning.Its levels range from comprehensive, clinically astute comparison to no meaningful differential diagnosis.Source: §1 · pp. 16–17

2. Safety Check

  • The safety rubric assesses identification, prioritization, and mitigation of patient risks, from proactive risk-benefit analysis to dangerous recommendations.The supplied criteria distinguish proactive foresight from merely avoiding obvious risks.Source: §2 · pp. 17–18

3. Evidence Citation

  • The evidence-citation rubric assesses whether responses connect patient-specific findings and medical knowledge to their conclusions.Its lowest level includes unsupported claims, misinterpreted evidence, or irrelevant justification.Source: §3 · p. 18

F. Details of Efficiency Analysis

The appendix estimates training and inference FLOPs with transformer-model formulas and adapts the training estimate for LoRA. It also identifies tables containing full downstream results for the main, QA, and ablation experiments.

F. Details of Efficiency Analysis

  • Training FLOPs are approximated as 6 × L × H^2 × T × |Dtrain| × E, using layer count, hidden size, tokens per sample, training samples, and epochs.For LoRA, the hidden-size-squared term is replaced by a term proportional to rank.Source: F. Details of Efficiency Analysis · pp. 18–19

G.1. Main Results

  • Table 8 provides full downstream task results for general reasoning and non-reasoning models; the cited text also says results for Llama-3.1-8B-Instruct trained on different selection-setting data are provided.Source: §G.1 · p. 20, Table 8 · p. 21

G.2. QA Experiment

  • Full downstream results for the QA experiment are reported in Table 10.The supplied passages give no QA result values.Source: §G.2 · p. 20

G.3. Ablation Study Experiment

  • Full downstream results for the DIQ ablation study are reported in Table 11.The supplied passages give no ablation result values.Source: §G.3 · p. 21

G.4. Generalization Experiment

The supplied appendix passages identify tables containing cross-scale, cross-model, and preference-learning results, but do not provide the results themselves.

G.4. Generalization Experiment

  • Tables 12 and 13 contain full results for cross-scale and cross-model experiments and for preference learning.Table 12 covers Qwen-series models trained with different influence-score settings; Table 13 covers DIQ-selected or random data and DPO.Source: §G.4 · p. 21, Table 12 · p. 25, Table 13 · p. 25
  • Table 9 reports full downstream results for Llama-3.1-8B-Instruct trained on full data, random subsets, and DIQ-selected data at 1%, 10%, and 50% keep ratios.The supplied caption does not include the result values.Source: Table 9 · p. 22

Related papers

Loading 2508.01450v3…