Source-linked AI summary
Abstracts or full text? Evaluating input scope and computational efficiency in biomedical lay summarization
Egecan Çelik Evgin, İlknur Karadeniz, Olcay Taner Yıldız
TL;DR
Biomedical lay summaries must make scientific information accessible and accurate for non-experts, while full-text summarization entails substantial computation. This study systematically evaluates abstract and full-text approaches using a consistent setup, reporting that abstract-based fine-tuning is 4.70 times faster than full-text fine-tuning.
Problem
Lay summaries must convey scientifically accurate information in language accessible to non-expert readers, while prior comparisons had not directly compared abstract-only and full-text fine-tuning under the same conditions.
Source: Introduction · p. 3, Related work · p. 5Method
The study systematically examines abstract and full-text summarization using the same models within a consistent experimental setup.
Source: Introduction · pp. 3–4, H2: Three-shot abstract prompting vs.… · p. 17Results
4.70 times faster than full-text LoRA fine-tuning was abstract-based LoRA fine-tuning, which achieved superior overall performance.
Source: Introduction · p. 4Takeaways & Limitations
Abstract-based approaches can achieve better overall lay summarization while full-text approaches require substantially more computation.
Source: H1: Abstract-based vs. full-text LoRA… · p. 16, Conclusion · p. 18The experiments were restricted to relatively small language models.
Source: Conclusion · p. 18
Abstract
from arXiv · showhide
Biomedical research articles are mainly written for experts and often contain terms that are difficult for the general public to understand. Automated lay summarization has become a solution to make biomedical knowledge easier for patients and the public to understand. However, generating lay summaries from full-text articles requires more computing power because the articles are long and costly to process. Fine-tuning large language models for this task also requires substantial resources. This study asks two questions: (1) whether abstracts contain enough information to produce lay summaries as good as those based on full-text articles, and (2) whether few-shot learning performs as well as task-specific fine-tuning. The study used two public biomedical lay summarization datasets and two recent models, evaluated for readability, factual consistency, semantic similarity, and runtime efficiency. Results show that abstract-based summarization performs better overall. Across both datasets and models, abstract fine-tuning is 4.70 times faster, while abstract zero- and few-shot prompting is 5.31 times faster than full-text methods. These results show that abstract-based prompting reduces runtime cost while maintaining lay summary quality.
H1: PLOS
On PLOS, abstract-based fine-tuning achieved higher overall scores than full-text fine-tuning for both models, with improved readability on three of four measures and shorter runtimes.
H1: PLOS
- Both models achieved higher overall scores with abstract-based fine-tuning than with full-text fine-tuning on PLOS.The comparison was significant for Qwen3.5-0.8B (Holm-adjusted p < .001, rrb = .68) and MobileLLM-Pro-1B (Holm-adjusted p < .001, rrb = .76).Source: Table 4 · p. 12, H1: PLOS · p. 12
- Abstract-based fine-tuning yielded lower scores on three of four readability metrics for both models, indicating improvement on those measures.Source: H1: PLOS · p. 12
- Fine-tuning was 4.90 times faster for Qwen3.5-0.8B and 4.56 times faster for MobileLLM-Pro-1B than full-text fine-tuning.Source: H1: PLOS · p. 12
H1: eLife
On eLife, abstract-based fine-tuning scored higher overall than full-text fine-tuning for both models and ran faster; the highest overall score among the reported fine-tuning conditions was achieved by MobileLLM-Pro-1B with abstracts.
H1: eLife
- Abstract-based fine-tuning with MobileLLM-Pro-1B achieved the highest overall score among the reported eLife fine-tuning conditions.Source: H1: eLife · p. 12
- Under full-text fine-tuning, Qwen3.5-0.8B achieved a higher overall score than MobileLLM-Pro-1B.Source: H1: eLife · p. 12
- On the eLife test set, both models had higher overall scores than with full-text fine-tuning: Qwen3.5-0.8B (Holm-adjusted p = .004, rrb = .62) and MobileLLM-Pro-1B (Holm-adjusted p < .001, rrb = .81).Source: Table 5 · p. 12, H1: eLife · p. 12
- 4.83 times faster for Qwen3.5-0.8B and 4.49 times faster for MobileLLM-Pro-1B was abstract-based fine-tuning than full-text fine-tuning.Source: H1: eLife · pp. 12–13
H2: PLOS
On PLOS, three-shot abstract prompting outperformed full-text fine-tuning overall for both models, was faster, and yielded easier-to-read summaries across all four readability metrics.
H2: PLOS
- Three-shot abstract prompting achieved higher overall scores than full-text fine-tuning for both models on PLOS.The comparison was p = .042, rrb = .517 for Qwen3.5-0.8B and p < .001, rrb = .69 for MobileLLM-Pro-1B.Source: Table 6 · p. 13, H2: PLOS · p. 13
- Three-shot abstract prompting yielded lower scores on all four readability metrics for both models, indicating easier-to-read summaries across those measures.Source: H2: PLOS · p. 13
- 7.63 times faster for Qwen3.5-0.8B and 7.41 times faster for MobileLLM-Pro-1B was three-shot abstract prompting than full-text fine-tuning.Source: H2: PLOS · p. 13
- Three-shot abstract prompting achieved a higher overall mean score than abstract-based fine-tuning for both models.Source: H2: PLOS · p. 13
H2: eLife
The supplied eLife results report higher overall scores for three-shot abstract prompting than full-text fine-tuning, alongside shorter runtimes; evaluation dimensions varied across models.
H2: eLife
- Results for relevance, readability, and factuality varied across the two models, so the overall advantage did not mean uniform improvement across evaluation dimensions.Source: H2: eLife · p. 13
- Three-shot abstract prompting achieved higher overall scores than full-text fine-tuning for both models.The comparisons were p < .001, rrb = .95 for Qwen3.5-0.8B and p < .001, rrb = .603 for MobileLLM-Pro-1B.Source: H2: eLife · p. 13, Table 7 · p. 14
- 10.14 times faster for Qwen3.5-0.8B and 9.14 times faster for MobileLLM-Pro-1B was three-shot abstract prompting than full-text fine-tuning.Source: H2: eLife · p. 14
- Abstract-based fine-tuning achieved a higher overall mean score for MobileLLM-Pro-1B.Source: H2: eLife · p. 14
H3: PLOS
The PLOS section compares abstract- and full-text prompting across zero-, one-, and three-shot settings for two models, including overall scores and runtime. It also calls for broader evaluation to determine whether the observed advantages of abstract-based approaches generalize to other models and practical biomedical communication scenarios.
H3: PLOS
- Three-shot abstract prompting achieved the highest overall mean score for both models.Source: H3: PLOS · p. 14
- Abstract prompting was 8.28 times faster for Qwen3.5-0.8B and 8.00 times faster for MobileLLM-Pro-1B than full-text prompting.The paper reports that these findings support H3 on PLOS for overall summary quality and runtime.Source: H3: PLOS · p. 14
- Abstract prompting scored higher overall than full-text prompting in all three settings for both models (Qwen3.5-0.8B: p = .002, rrb = .80; MobileLLM-Pro-1B: p < .001, rrb = .74).Table 8 presents the PLOS few-shot inference results.Source: H3: PLOS · p. 14, Table 8 · p. 15, H3: Abstract vs. full-text few-shot… · p. 17
- Evaluations of relevance, completeness, and readability are needed to test whether abstract-based efficiency and performance advantages extend to broader model settings and practical biomedical communication scenarios.Source: Conclusion · p. 18