PaperlayerView exact revision on arXiv

Source-linked AI summary

Recursive LLM Degradation in Biomedical Question Answering: A Cross-Generation Study

Bibek Bhandari, Kshitij Lingthep

arXiv:2609.34257v1cs.CLcs.LG

TL;DR

Recursive synthetic-data feedback raises questions about whether biomedical QA answers shift from the original data distribution, especially when fluent answers can still alter or omit medical entities. The paper compares recursive training with repeated training on human-authored data and evaluates biomedical and answer-alignment measures. Recursive training shows larger declines in several primary metrics, with longer answers but no similar increase in simple 3-gram repetition.

  • Problem

    Biomedical answers can remain fluent while omitting, altering, or introducing medically relevant entities, motivating domain-specific evaluation of recursive synthetic-data training.

    Source: §I · p. 1

  • Method

    The study compares recursive biomedical QA training with a Human-Control condition that repeatedly uses the original human-authored data, evaluating changes across biomedical, alignment, and behavioral metrics.

    Source: §II · p. 1

  • Results

    Recursive training showed larger declines than Human-Control in disease and chemical entity F1, context-supported rate, ROUGE-L, and cosine similarity, while answers tended to become longer without a similar increase in simple 3-gram repetition.

    Source: §VI · p. 5, §VIII · p. 6

  • Takeaways & Limitations

    The observed pattern is a domain-specific shift in biomedical QA answers, with longer responses that are less aligned with references and less supported by supplied context.

    Source: §VI · p. 5

    The metrics do not establish clinical hallucination or biomedical incorrectness without expert analysis, and comparisons rely on only two seeds, making findings primarily descriptive.

    Source: §VII · p. 6

Abstract

from arXiv · show

Repeatedly training language models on their own generated data may create a synthetic-data feedback loop in which errors and distributional biases are reintroduced into subsequent training datasets. This paper studies that process in biomedical question answering (QA) using PubMedQA and two Qwen2.5 model sizes, 0.5B and 3B parameters. The study compares a recursive synthetic-data condition, in which generation G(k+1) is trained on answers produced by G(k), against a Human-Control condition that repeatedly uses the original human training data. The study evaluates across four generations from G0-G3 with two random seeds (42 and 123) and a fixed evaluation set of 1,000 expert-labeled samples. The evaluation includes disease and chemical entity F1, context-supported rate, lexical and semantic similarity, answer length, repetition rate, and other evaluation metrics. The Recursive condition for both model sizes and both seeds showed larger declines than the Human-Control condition in disease entity F1, chemical entity F1, context-supported rate, ROUGE-L, and cosine similarity. Under the fixed no-repeat 3-gram decoding constraint, the main observed behavioral change was increased answer length, while the measured 3-gram repetition rate did not increase. The magnitude of the difference-in-change was larger for the 3B model than for the 0.5B model. This difference was particularly apparent in disease F1, context-supported rate, cosine similarity, and answer length. These results show domain-specific changes associated with using recursive synthetic-data training in biomedical QA, but do not establish clinical hallucination rates or universal model collapse.

I. INTRODUCTION

The study examines whether recursive synthetic-data training shifts biomedical QA answers, where fluent text can still omit, alter, or introduce medically relevant entities. It compares recursive training with a fixed-human-data control and reports larger declines in several biomedical and alignment metrics under recursion.

I. INTRODUCTION

  • Biomedical QA motivates domain-specific evaluation because fluent answers may omit, alter, or introduce medically relevant entities.The study therefore considers biomedical metrics alongside conventional lexical and semantic metrics.Source: §I · p. 1
  • The study compares generations trained on the previous generation’s answers with generations repeatedly trained on the original human training set.It uses the same evaluation set across generations and conditions.Source: §I · p. 1
  • Recursive training showed larger declines than Human-Control in disease and chemical entity F1, context-supported rate, and lexical and semantic alignment, while answers became longer.The pattern was generally larger in the tested 3B configuration; the study does not report uniform deterioration across all metrics.Source: §I · p. 1

II. CONTRIBUTIONS

  • The study contributes a controlled G0–G3 biomedical QA setup, cross-condition comparisons across seeds and model sizes, and reproducibility materials.Its evaluation spans biomedical entity, semantic, lexical, and behavioral metrics.Source: §II · p. 1

III. RELATED WORK

Prior work studies recursive synthetic-data feedback and iterative data enhancement, while biomedical QA adds domain-specific evaluation concerns. This paper applies those questions to entity preservation, context support, and reference-answer alignment.

A. Recursive training and synthetic-data feedback loop

  • Earlier work described recursive synthetic-data training as model collapse after showing that repeated synthetic-data training can erode information from the original distribution.Related studies examine performance, data quality, stability, and the effects of generation and curation across generations –.Source: A. Recursive training and… · p. 2

B. Iterative synthetic-data enhancement

C. Biomedical question answering and medical language models

  • Biomedical and clinical language-model research includes GatorTron, PMC-LLaMA, BioMistral, and MEDITRON –[17], while PubMedQA uses expert-labeled biomedical data.This study uses a biomedical NER model as an auxiliary evaluation instrument rather than as its generation model.Source: C. Biomedical question answering and… · p. 2

D. Models, efficient adaptation, and domain-aware evaluation

The study combines Qwen2.5 with domain-aware entity evaluation and a fixed PubMedQA evaluation set. Its recursive condition trains each generation on answers from the previous generation, while Human-Control reuses the original human-authored training data.

D. Models, efficient adaptation, and domain-aware evaluation

  • Qwen2.5 is fine-tuned with QLoRA, and PubMedBERT underlies the NER model used to identify disease and chemical entities.The study evaluates reference-answer alignment and context support in addition to entity analysis.Source: D. Models, efficient adaptation, and… · p. 2

A. Dataset and fixed evaluation set

  • The study trains on 5,000 PQA-U samples and evaluates on the same 1,000 PQA-L samples across generations, conditions, seeds, and model sizes.Training and evaluation sets are checked for identifier overlap before training.Source: A. Dataset and fixed evaluation set · p. 2

B. Recursive and Human-Control conditions

  • In Recursive training, answers generated from training questions and contexts form the next training data, whereas Human-Control repeatedly uses the original human-authored set.The next Recursive model is fine-tuned using the previous generation’s adapter as initialization.Source: B. Recursive and Human-Control… · p. 2

C. Models and training

The training setup uses quantized, parameter-efficient adaptation with sequential adapter initialization. Answer generation uses fixed greedy-decoding settings across generations and both random seeds.

C. Models and training

  • Both Qwen2.5 models use LoRA/QLoRA adaptation with rank 16, alpha 32, dropout 0.05, and 4-bit NF4 quantization.The setup also specifies a per-device batch size of four with four gradient-accumulation steps.Source: C. Models and training · p. 2
  • Training runs for three epochs at learning rate 2 × 10−4, with maximum sequence length 1024 and gradient checkpointing.For generations after G0, each model is initialized with the preceding model’s adapter.Source: C. Models and training · p. 2
  • Greedy decoding uses a 256-token limit, repetition penalty 1.15, and no-repeat 3-gram constraint, fixed across generations and seeds 42 and 123.The same evaluation set and generation setup are used for both seeds.Source: C. Models and training · p. 2

D. Evaluation metrics

The evaluation combines biomedical entity, context-support, lexical, semantic, and behavioral measures to characterize generated answers.

D. Evaluation metrics

  • Disease and chemical entity F1 compare normalized extracted-entity sets with references, requiring matching entity type and identical normalized surface forms.Entity retention is reference-entity recall; F1 also accounts for additional predicted entities through precision.Source: D. Evaluation metrics · p. 2
  • Table I reports mean per-sample entity F1 for evaluation samples with at least one reference entity, whereas Figure 1 reports micro-F1 from pooled entity counts.Paired inferential analyses use per-sample F1 values.Source: D. Evaluation metrics · p. 3
  • Context-supported rate measures the fraction of predicted entities whose normalized surface forms occur in the biomedical context.It is averaged across evaluable answers; context-unmatched entities are not definitive evidence of clinical hallucination.Source: D. Evaluation metrics · p. 3
  • Lexical similarity uses BLEU-1, BLEU-4, ROUGE-L, METEOR, and chrF++; semantic similarity uses BERTScore F1 and embedding cosine similarity.Cosine similarity uses sentence-transformers/all-mpnet-base-v2.Source: D. Evaluation metrics · p. 3
  • Behavioral measures include answer length, 3-gram repetition rate, and answer-only conditional perplexity, which is descriptive rather than a direct measure of biomedical answer quality.Perplexity is computed over answer tokens conditioned on a fixed evaluation prompt.Source: D. Evaluation metrics · p. 3

E. Statistical analysis

Recursive training shows larger declines than Human-Control on the reported biomedical and reference-answer alignment metrics. The analysis compares generation changes descriptively across conditions and seeds.

E. Statistical analysis

  • Within each condition and seed, paired G0–G3 changes use mean changes, bootstrap 95% confidence intervals, Cohen’s dz, and two-sided Wilcoxon signed-rank tests.The primary descriptive condition contrast is Recursive change minus Human-Control change, summarized across two seeds; comparisons across conditions and model sizes remain mainly descriptive.Source: E. Statistical analysis · p. 3

A. Overall recursive changes

  • Recursive training showed larger declines than Human-Control in both model sizes for disease F1, chemical F1, context-supported rate, ROUGE-L, and cosine similarity.The reported overall comparison uses mean G3 − G0 changes across seeds 42 and 123.Source: A. Overall recursive changes · p. 3
  • For Qwen2.5-0.5B, Recursive changes were disease F1 -0.1008, context-supported rate -0.1649, ROUGE-L -0.0462, and cosine similarity -0.0631.The corresponding Recursive-minus-Human-Control difference-in-change values were -0.0915, -0.2187, -0.0400, and -0.0416, respectively.Source: A. Overall recursive changes · p. 3
  • For Qwen2.5-3B, Recursive changes were disease F1 -0.1342, context-supported rate -0.2412, ROUGE-L -0.0555, and cosine similarity -0.0800.Human-Control disease F1 and ROUGE-L slightly increased over the same interval.Source: A. Overall recursive changes · p. 3

B. Seed consistency

The Recursive-versus-Human-Control difference-in-change has the same direction across the two seeds for the reported biomedical metrics. Within-condition tests and the limited seed count constrain interpretation.

B. Seed consistency

  • For 0.5B, disease F1 difference-in-change was -0.0925 and -0.0904, context-supported rate was -0.2481 and -0.1894, and ROUGE-L was -0.0431 and -0.0369 across Seeds 42 and 123.The direction of the condition contrast was reproduced across both seeds.Source: B. Seed consistency · p. 3
  • For 3B, disease F1 difference-in-change was -0.1280 and -0.1609, context-supported rate was -0.3670 and -0.2866, and ROUGE-L was -0.0630 and -0.0611 across Seeds 42 and 123.The seed-specific changes remain directionally consistent for the biomedical metrics within the tested configurations.Source: B. Seed consistency · p. 3

C. Within-condition inferential results

  • In 3B Recursive, disease F1 changes had Wilcoxon p-values of 0.00093 and 0.00051; ROUGE-L, BERTScore F1, and cosine similarity also had small p-values and large paired Cohen’s dz values in both seeds.In 0.5B Recursive, ROUGE-L, BERTScore F1, and cosine similarity likewise had small p-values in both seeds; entity-F1 uncertainty was wider because fewer examples had evaluable entities.Source: C. Within-condition inferential results · p. 3
  • The reported Wilcoxon p-values are within-condition tests, not direct tests of the Recursive-versus-Human-Control difference-in-change; two seeds make cross-condition and model-size comparisons mainly descriptive.The within-condition p-values are exploratory and unadjusted for multiple comparisons.Source: C. Within-condition inferential results · pp. 3–4
  • Figure 1 plots disease and chemical entity micro-F1 across G0–G3, with solid lines for Recursive, dashed lines for Human-Control, and colors for model sizes.Lines show means across Seeds 42 and 123.Source: Figure 1 · p. 4

D. Answer-length expansion without increased repetition

Recursive training expanded answers without increasing measured 3-gram repetition; the 3B model showed larger differences from Human-Control than the 0.5B model across several metrics.

D. Answer-length expansion without increased repetition

E. Cross-Model difference-in-change between Recursive and Human-Control conditions

  • Difference-in-change was more negative for 3B than 0.5B in disease F1 (−0.1444 vs. −0.0915), context-supported rate (−0.3268 vs. −0.2187), ROUGE-L (−0.0620 vs. −0.0400), BERTScore F1 (−0.0175 vs. −0.0031), and cosine similarity (−0.0663 vs. −0.0416).Source: E. Cross-Model difference-in-change… · p. 4, Figure 2 · p. 5
  • Answer-length difference-in-change was +22.82 words for 3B versus +1.58 words for 0.5B, while both seeds showed lower context-supported rates under Recursive training than Human-Control.Source: E. Cross-Model difference-in-change… · p. 4
  • The two seed-specific changes were directionally consistent across reported metrics and both model sizes, so the cross-model contrasts are descriptive within tested configurations.Source: E. Cross-Model difference-in-change… · pp. 4–5

VI. DISCUSSION

In biomedical QA, recursive training produces longer, less context-supported and less reference-aligned answers; the contrast with Human-Control and the larger 3B differences are specific to tested configurations.

  • Recursive feedback shifts answers toward greater length but lower reference-answer similarity and support from biomedical context, rather than shorter or obviously repetitive outputs.Source: §VI · p. 5
  • Human-Control helps separate changes associated with recursive synthetic inputs from changes associated with repeated fine-tuning, though it does not remove all generation-to-generation variation.Source: §VI · p. 5
  • The tested 3B configuration shows larger separation than 0.5B on several biomedical and semantic metrics, alongside stronger answer-length expansion, without establishing a general scaling relationship.Source: §VI · p. 5
  • Entity F1, context support, lexical similarity, and semantic similarity capture different answer dimensions, so a single similarity metric may miss changes.Source: §VI · p. 5
  • These metrics are complementary indicators, not direct measures of biomedical correctness; unmatched entities or lexical and semantic differences do not necessarily imply incorrectness.Source: §VI · p. 5

VII. LIMITATIONS

The findings are constrained by the tested models, seeds, metrics, decoding setup, and low-resource fine-tuning regime.

  • Only two Qwen2.5 model sizes and two random seeds were tested, supporting directional consistency within these configurations rather than universal scaling behavior.Source: §VII · p. 5
  • Entity, context-match, and similarity measures are indirect, while repetition findings apply only to the no-repeat 3-gram decoding setting.Source: §VII · p. 6
  • Each generation used 5,000 training examples and three epochs, so observed changes may also reflect low-resource fine-tuning and could differ with more data or epochs.Source: §VII · p. 6
  • Condition and model-size comparisons rely on two seeds and are primarily descriptive; reported tests do not directly test the difference-in-change contrast.Source: §VII · p. 6

VIII. CONCLUSION

Across four generations, recursive training shows declines in biomedical entity preservation, context support, and reference alignment relative to Human-Control. The 3B configuration shows larger changes on several measures, but the findings do not establish universal model collapse.

VIII. CONCLUSION

  • Across seeds 42 and 123, Recursive training shows larger declines than Human-Control in disease and chemical entity F1, context-supported rate, ROUGE-L, and cosine similarity.Source: §VIII · p. 6
  • Changes were larger in tested 3B for disease F1, context-supported rate, cosine similarity, and answer length, without a similar increase in simple 3-gram repetition.Source: §VIII · p. 6
  • The findings describe recursive-training changes in biomedical QA, not proof of universal model collapse; further work should test more mixtures, model families, seeds, and expert factuality assessments.Source: §VIII · p. 6

IX. RESEARCH MATERIALS AND REPRODUCIBILITY

  • Research materials include code, configurations, notebooks, plotting scripts, derived datasets, prediction results, and adapter weights; the releases do not redistribute the full original PubMedQA dataset.Source: §IX · p. 6
Loading 2609.34257v1…