AgenticSum: An Agentic Inference-Time Framework for Faithful Clinical Text Summarization

Fahmida Liza Piya; Rahmatollah Beheshti

cs.CLFeb 23, 2026

AgenticSum: An Agentic Inference-Time Framework for Faithful Clinical Text Summarization

Authors: Fahmida Liza Piya, Rahmatollah Beheshti

Organizations: University of Delaware, Newark, DE, USA

Abstract

Large language models (LLMs) offer substantial promise for automating clinical text summarization, yet maintaining factual consistency remains challenging due to the length, noise, and heterogeneity of clinical documentation. We present AgenticSum, an inference-time, agentic framework that separates context selection, generation, verification, and targeted correction to reduce hallucinated content. The framework decomposes summarization into coordinated stages that compress task-relevant context, generate an initial draft, identify weakly supported spans using internal attention grounding signals, and selectively revise flagged content under supervisory control. We evaluate AgenticSum on two public datasets, using reference-based metrics, LLM-as-a-judge assessment, and human evaluation. Across various measures, AgenticSum demonstrates consistent improvements compared to vanilla LLMs and other strong baselines. Our results indicate that structured, agentic design with targeted correction offers an effective inference time solution to improve clinical note summarization using LLMs.

Figures & tables

Figure 1: Our agentic framework, AgenticSum , designed for hallucination-aware clinical text summarization.

Model	MIMIC-IV				SOAP Summary
Model	ROUGE-L	BLEU-1	BLEU-2	BERTScore	ROUGE-L	BLEU-1	BLEU-2	BERTScore
	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )
BioBART	$8.00\pm 3.20$	$6.88\pm 5.20$	$2.05\pm 2.10$	$78.10\pm 1.60$	$7.41\pm 4.90$	$5.27\pm 4.20$	$1.83\pm 2.80$	$77.32\pm 2.50$
T5-Large	$7.96\pm 3.30$	$4.95\pm 5.30$	$1.41\pm 2.20$	$79.84\pm 1.60$	$7.52\pm 5.20$	$3.86\pm 3.90$	$1.24\pm 2.40$	$78.46\pm 2.70$
Flan-T5	$7.82\pm 6.86$	$4.51\pm 6.91$	$2.20\pm 4.35$	$78.64\pm 4.38$	$8.91\pm 5.00$	$8.35\pm 5.10$	$2.17\pm 2.90$	$77.05\pm 2.80$
Gemma3-1B	$9.85\pm 3.40$	$7.89\pm 5.60$	$3.20\pm 2.40$	$79.78\pm 1.60$	$8.97\pm 5.30$	$7.25\pm 5.30$	$2.84\pm 3.20$	$79.25\pm 2.50$

Table 1: Summary quality comparison across medical datasets. We report lexical overlap (BLEU, ROUGE-L) and semantic similarity (BERTScore).

Model	MIMIC-IV				SOAP Summary
	Hallucination	Completeness	Coherence	Factual	Hallucination	Completeness	Coherence	Factual
	( $\downarrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\downarrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )
BioBART	$2.45\pm 0.72$	$2.95\pm 0.68$	$4.21\pm 0.89$	$3.68\pm 0.55$	$2.38\pm 0.65$	$2.87\pm 0.58$	$4.15\pm 0.74$	$3.72\pm 0.48$
T5-Large	$2.52\pm 0.78$	$2.88\pm 0.72$	$4.08\pm 0.95$	$3.61\pm 0.58$	$2.47\pm 0.69$	$2.82\pm 0.64$	$4.02\pm 0.82$	$3.65\pm 0.52$
Flan-T5	$2.37\pm 0.80$	$3.18\pm 0.80$	$4.60\pm 1.03$	$3.79\pm 0.62$	$2.41\pm 0.74$	$3.12\pm 0.76$	$4.52\pm 0.88$	$3.75\pm 0.59$
Gemma3-1B	$2.35\pm 0.68$	$3.08\pm 0.74$	$4.28\pm 0.82$	$3.71\pm 0.54$	$2.42\pm 0.71$	$2.98\pm 0.66$	$4.22\pm 0.78$	$3.68\pm 0.49$

Table 2: LLM-as-a-Judge Evaluation of Hallucination and Summary Quality Across Datasets. Scores are assigned by an instruction-tuned LLM acting as an evaluator using a 1–5 scale across four axes: hallucination frequency (lower is better), completeness, coherence, and factual consistency (higher is better). All metrics are rated on a 1–5 scale.

Clinical Domain	Correct Guess $\uparrow$	Mean Hallucination Severity $\downarrow$	Effect Size	Dominance $p$
	Accuracy (%)	Vanilla / AgenticSum	( $r_{\mathrm{rb}}$ )
Psychiatric	91.3	3.74 / 1.91	0.905	$<0.001$
Cardiopulmonary	78.3	3.91 / 2.09	0.800	0.005
Oncology	87.0	4.13 / 1.96	0.909	$<0.001$

Table 3: Human evaluation of hallucination severity and annotator agreement across clinical domains . Correct Guess Accuracy (

\uparrow

) indicates the proportion of annotators identifying the same summary as more hallucinated. Hallucination severity is rated on a 1–5 ordinal scale, where lower values indicate fewer hallucinations (

\downarrow

). All Wilcoxon signed-rank tests were statistically significant (

p<0.001

for all domains). Rank-biserial correlation is reported as effect size. Dominance

p

-values are computed using a binomial test against chance agreement.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Model	MIMIC-IV				SOAP Summary
Model	BLEU-1	BLEU-2	BERT F1	ROUGE-L	BLEU-1	BLEU-2	BERT F1	ROUGE-L
	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )	( $\uparrow$ )
Vanilla LLM	$15.54\pm 6.55$	$7.17\pm 3.22$	$81.90\pm 1.60$	$10.65\pm 3.78$	$13.12\pm 9.35$	$8.10\pm 6.17$	$83.99\pm 2.34$	$12.53\pm 5.31$
DraftAgent (FOCUS+LLM)	$14.01\pm 8.01$	$5.61\pm 3.92$	$80.92\pm 1.70$	$11.30\pm 3.08$	$11.55\pm 5.50$	$6.09\pm 4.20$	$83.60\pm 2.20$	$10.70\pm 4.50$
FixAgent (W/out Supervisor)	$16.78\pm 5.80$	$7.85\pm 2.90$	$82.45\pm 1.50$	$11.42\pm 3.20$	$13.94\pm 7.20$	$8.92\pm 5.40$	$84.67\pm 2.10$	$11.21\pm 4.80$
AgenticSum (Final, w/ Clinical Supervisor)	$\mathbf{18.37\pm 5.80}$	$\mathbf{12.61\pm 2.90}$	$\mathbf{84.50\pm 3.50}$	$\mathbf{21.62\pm 3.20}$	$\mathbf{23.94\pm 7.20}$	$\mathbf{15.80\pm 5.40}$	$\mathbf{86.32\pm 2.10}$	$\mathbf{24.99\pm 4.80}$

Appendix

Table 4: Ablation study on MIMIC-IV and SOAP summaries. Module-wise contributions of FOCUS and FixAgent over the vanilla LLM.

Source Document (Excerpt)

Patient was in good health until 10 days ago, when he began experiencing gradually worsening headaches. Head CT revealed an intracranial mass lesion. An enhancing focus in the left parapharyngeal region likely represents an abnormally enlarged lymph node, though it is not fully visualized. The imaging raises the possibility of a benign neoplasm (question neurogenic tumor), and pathology is pending. Discharge medications include atenolol, phenytoin, docusate sodium, and dexamethasone.

Generated Summary

The patient is a 10-year-old male who was experiencing worsening headaches. A CT scan revealed an intracranial mass lesion, which was found to be an abnormally enlarged lymph node representing metastatic involvement . The pathology pending at the time of discharge is a benign neurogenic tumor . The patient was prescribed atenolol, phenytoin, docusate sodium, and dexamethasone.

Hallucination Detector agent Analysis

Factuality Verdict: UNSUPPORTED

Explanation: The patient’s age and gender are unspecified. The lymph node is described as potentially benign, not definitively metastatic. Pathology is still pending at discharge. The hallucinated spans introduce unwarranted specificity and certainty.

Appendix

Table 5: Example of hallucination detection using Hallucination Detector agent with span-level grounding analysis.

Source Document (Clinical Note)

<SEX> F

<SERVICE> SURGERY

<ALLERGIES> Iodine / contrast dye allergy.

The patient reports persistent abdominal pain that has worsened over the past week.

She states that despite taking her medications, the pain has not resolved.

She denies fever, chills, cough, or shortness of breath.

Appendix

Table 6: Example of sentence-level input filtering performed by FocusAgent , illustrating how clinically informative sentences are retained while low-information content is omitted.

Source Document (Clinical Note)

Male patient with major depressive disorder presents with chronic, intractable nausea that began in childhood and is associated with his depression. Last night, he experienced the worst nausea of his life, leading to suicidal ideation with a plan for carbon monoxide poisoning. He is currently undergoing ECT treatment (2 sessions completed). The patient reports anxiety exacerbates his nausea, and he describes himself as a ”worrier.” He no longer has active suicidal thoughts but worries about future episodes. Physical exam shows psychomotor retardation, depressed affect, slow speech with pauses, and fine tremor. Lab results are within normal limits. He is alert, oriented, and ambulatory. Current medications include: omeprazole, ranitidine, phenelzine, clonazepam, tamsulosin, and aripiprazole.

DraftAgent Summary (Initial)

The patient is a male who has been experiencing intractable nausea associated with depressive symptoms. He has a history of car accidents and has been prescribed ECT treatments. The patient reports that he is upset about his inability to throw up and is experiencing anxiety. He has a history of depression and anxiety, and has been taking medication for both conditions. The patient’s plan is to try the first round of ECT, and he is being treated by Dr. Stern . The patient’s physical examination reveals normal findings . The patient’s speech is slow and with many pauses, and his behavior is cooperative with psychomotor retardation. The patient’s thought process is linear and goal-directed. The patient is currently taking several medications, including omeprazole, ranitidine, phenelzine, clonazepam, tamsulosin, and aripiprazole. The patient is alert and interactive, and is ambulatory and independent.

FixAgent Summary (Corrected)

The patient is a male with major depressive disorder and persistent nausea. He has undergone two ECT sessions. The patient reports distress due to the intractable nausea, which is worsened by anxiety. He is undergoing ECT under medical supervision. The physical exam shows psychomotor retardation, slow speech, and cooperative behavior. The patient’s thought process is linear and goal-directed. He is currently taking omeprazole, ranitidine, phenelzine, clonazepam, tamsulosin, and aripiprazole. He is alert, interactive, and ambulatory.

FixAgent Analysis

Appendix

Table 7: End-to-end example of hallucination detection and correction using the AgenticSum pipeline.

Explore similar work

May 27, 2026

Hallucination Detection-Guided Preference Optimization for Clinical Summarization

Large language models (LLMs) have shown promise on summarization tasks, but they often produce hallucinations, which are unsupported or incorrect statements that limit their reliability in specialized healthcare applications. We introduce Hallucination Detection Guided Self-Refinement (HDSR), an inference-time method that leverages hallucination detectors to guide iterative summary revisions toward factual corrections. Building on this, we propose HDSR for Preference Learning (HDSR-PL), which converts detector-guided refinement trajectories into preference pairs for model finetuning. Extensive experiments show that our methods substantially reduce hallucinations for Llama and Gemma models in summarizing real-world clinical notes from MIMIC-IV-Note v2.2. For example, HDSR reduces 24% and HDSR-PL reduces 48% hallucinations in Llama-3.1-8B-Instruct. Importantly, both methods preserve summary fluency, coherence, and relevance according to human expert and LLM-Jury evaluations. Together, these results demonstrate that detection-informed refinement and preference learning offer an automated solution for improving factual faithfulness in clinical summarization.

Shamanth Kuthpadi Seethakantha, Dung Ngoc Thai, Vara Prasad Gudi +6

1Manning College of Information and Computer Sciences · 2Ensemble HP · 3Columbia University +1

LLM Self-Refinement LLM Hallucination Mitigation Healthcare Hallucination in Language Models Text Summarization Preference Optimization

Ask PDF

Jun 8, 2026

CARE: A Conformal Safety Layer for Medical Summarization

Large language models (LLMs) are increasingly used for medical summarization, but their outputs can omit medically important information and introduce unsupported claims. Existing error-detection methods produce heuristic or uncalibrated scores, providing no formal control over missed errors and no principled way to trade off safety against clinician review burden. We introduce Conformal Assessment for Risk Evaluation (CARE), a post-hoc, model-agnostic safety layer that uses conformal risk control to overlay calibrated omission and hallucination flags onto summaries from any LLM without retraining. CARE provides finite-sample, distribution-free guarantees through two controllers: a hallucination controller that bounds the probability of a document containing any unflagged hallucinated sentence, and an omission controller that bounds the expected fraction of important omissions not surfaced for review. Unlike hallucination detection, omissions depend jointly on whether a source sentence is important and whether it is covered by the summary. We show that calibrating only one dimension can violate the target risk bound, while marginal decompositions remain valid but overly conservative. By jointly calibrating over the full

(τ,γ)

threshold space, CARE preserves formal guarantees while surfacing up to 5

\times

fewer sentences than alternative calibrated baselines. Across five medical summarization tasks, CARE satisfies the target risk bound at

α= 0.15

with 95% confidence across 100 calibration/test resplits, using only ~100 labeled documents per domain. In a preliminary clinician study (75 document reviews), calibrated flags improved omission detection by 28.6 percentage points on average. These results show that sentence-level safety guarantees are feasible for LLM-assisted medical summarization and offer a tunable mechanism for balancing residual risk and review effort.

Suhana Bedi, Bridget Lin, Anson Y. Zhou +5

Department of Biomedical Data Science, Stanford University · Department of Anesthesiology, Perioperative and Pain Medicine, Stanford University · Department of Computer Science, Stanford University +2

Text Summarization Conformal Risk Control LLM Hallucination Detection Medical LLMs

Ask PDF

Jul 10, 2026

Faithful by Design: Evaluating and Improving LLM-Generated Clinical Trial Summaries for Multi-Stakeholder Audiences

Large language models are increasingly used to summarize clinical trial results for healthcare providers, patients, and payers, but their tendency to hallucinate poses significant risks in this high-stakes context. This study introduces a benchmark evaluation framework for measuring the faithfulness of LLM-generated clinical trial summaries across three stakeholder audiences. The framework consists of 200 stratified trials drawn from the Aggregate Analysis of ClinicalTrials.gov database, evaluated using audience-specific prompt templates and a six-dimension faithfulness annotation schema. Baseline measurements were established for GPT-4o, Claude Sonnet 4.6, and Gemini 2.5 Flash across 1,800 generated summaries scored using a cross-encoder natural language inference (NLI) model. Unsupported Claims was identified as the dominant failure mode across all three models, with a mean annotation score of 1.55 out of three. A knowledge-graph-augmented retrieval system was developed and evaluated against the baseline, producing statistically significant improvements in NLI-based faithfulness scores (entailment +0.0125, faithfulness +0.0130, p < 0.0001). Improvement pathways were model-dependent, with GPT-4o improving primarily through contradiction reduction while Claude Sonnet 4.6 and Gemini 2.5 Flash improved through increased entailment.

Robert Williams

University of Texas at Austin Computer & Data Science Online Austin, Texas, USA

Hallucination in Language Models Text Summarization RAG Evaluation Clinical Language Model Evaluation

Ask PDF