United States

Clinical NLP systems are increasingly used for triage support, prediction, and decision assistance in EHR-based settings, where demographic fairness is a critical concern. A common evaluation approach is counterfactual demographic perturbation: modifying attributes such as age or sex while holding clinical evidence fixed and measuring output changes. 
 However, we show that such counterfactual audits can be misleading when interpreted in isolation. Across three clinical LLMs, we find that non-demographic control perturbations (e.g., paraphrases) often induce output variability comparable to or greater than demographic edits. 
 This can contribute to overestimation or misinterpretation of demographic bias.
 To address this, we propose a baseline-aware audit framework that explicitly compares demographic perturbations against control baselines. 
 Our analysis reveals that (i) label-level stability can mask substantial variation in generated rationales and recommendations, and (ii) age-based perturbations generally induce larger effects than sex-based ones in borderline cases. 
 Crucially, we identify a high intrinsic instability (&quot;noise floor&quot;; 0.46--0.71 Jaccard instability) in clinical LLM generations, while additional matched-metric analyses show that demographic perturbations are often comparable to non-demographic baseline variability.
 These findings highlight a key limitation of existing fairness evaluations: without establishing appropriate baselines, apparent demographic sensitivity may be over- or mis-attributed to bias rather than broader generative instability. We argue that baseline-aware counterfactual audits, which explicitly compare demographic effects against intrinsic model noise, provide a more reliable lens for evaluating clinical NLP systems in high-stakes settings.

ACL 2026

When Demographic Sensitivity Isn&#39;t What It Seems: Baseline-Aware Counterfactual Audits for Clinical NLP

generative instability

baseline-aware auditing

counterfactual fairness

large language models

clinical nlp

### Welcome to ACL 2023, the 61st Annual Meeting of the Association for Computational Linguistics! 
 The conference will be held in Toronto, Canada, July 9-14, 2023. 
Following the succession of the recent conferences in our field, ACL 2023 will adopt a hybrid format.
While the impact of Covid has considerably diminished in terms of traveling, obtaining visas to Canada
entails a very long process. Moreover, the global economic conditions pose challenges for many individuals to travel to conferences. Recognizing these circumstances, we know many participants may not be
able to attend the conference in person. Therefore, we are committed to providing a great virtual platform
so everyone has the opportunity to interact with other participants and enjoy the conference. Based on the
current registered participants, approxiately 30% have chosen to attend the conference virtually. Whether
you join us in person or virtually, we sincerely hope everyone has a remarkable conference experience. 
This General Chair’s message is where I express my gratitude to the many individuals who have made
enormous contributions to the conference over the past year.

Read [**ACL 2023 General Chair's message**](https://docs.google.com/document/d/1WobYM7norbG4dI48s75HfJoD89qgX5a_F-6U8AteLSA/edit?usp=sharing/) in full.

##### **[Conference Handbook](https://2023.aclweb.org/downloads/acl2023-handbook.pdf)**

ACL 2023

The Association for Computational Linguistics (ACL) is the premier international scientific and professional society for people working on computational problems involving human language, a field often referred to as either computational linguistics or natural language processing.

Clinical NLP systems are increasingly used for triage support, prediction, and decision assistance in EHR-based settings, where demographic fairness is a critical concern. A common evaluation approach is counterfactual demographic perturbation: modifying attributes such as age or sex while holding clinical evidence fixed and measuring output changes. 
 However, we show that such counterfactual audits can be misleading when interpreted in isolation. Across three clinical LLMs, we find that non-demographic control perturbations (e.g., paraphrases) often induce output variability comparable to or greater than demographic edits. 
 This can contribute to overestimation or misinterpretation of demographic bias.
 To address this, we propose a baseline-aware audit framework that explicitly compares demographic perturbations against control baselines. 
 Our analysis reveals that (i) label-level stability can mask substantial variation in generated rationales and recommendations, and (ii) age-based perturbations generally induce larger effects than sex-based ones in borderline cases. 
 Crucially, we identify a high intrinsic instability ("noise floor"; 0.46--0.71 Jaccard instability) in clinical LLM generations, while additional matched-metric analyses show that demographic perturbations are often comparable to non-demographic baseline variability.
 These findings highlight a key limitation of existing fairness evaluations: without establishing appropriate baselines, apparent demographic sensitivity may be over- or mis-attributed to bias rather than broader generative instability. We argue that baseline-aware counterfactual audits, which explicitly compare demographic effects against intrinsic model noise, provide a more reliable lens for evaluating clinical NLP systems in high-stakes settings.

When Demographic Sensitivity Isn't What It Seems: Baseline-Aware Counterfactual Audits for Clinical NLP

workshop paper

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

MedCAT is an open-source framework for clinical named entity recognition and linking (NER+L) widely used in research and healthcare settings. We present MedCAT v2, a re-engineered version designed to improve modularity, extensibility, and maintainability while preserving the core functionality and performance of previous releases. The new architecture introduces a registry-based component system and a flexible pipeline that enables easy substitution of components, integration of alternative methods, and future expansion, including support for pre-trained components across the full NER+L and contextualisation workflow. This enables systematic exploration of clinical NER+L design trade-offs by evaluating different components in the pipeline. Evaluation across multiple public datasets shows equivalent or improved performance compared to earlier versions, with reduced integration overhead and improved runtime flexibility. The framework also supports optional extensions such as meta-annotation, relation extraction, providing a unified and reproducible environment for clinical NLP in real-world settings.

MedCAT v2: a modular, extensible architecture for clinical named entity recognition and linking under real-world privacy and compute constraints

The development of large language models (LLMs) has led to increased focus on their adaptation to specialized domains and languages, yet the effectiveness of domain adaptation strategies remains unclear. We present a study of medical domain adaptation using French medical question answering (QA) as a case study. We compare continual pretraining (CPT), supervised fine-tuning (SFT), and their combination across three model families, multiple sizes, and three initialization types, explicitly disentangling adaptation effects from base model choice.
 We evaluate both multiple-choice (MCQA) and open-ended QA (OEQA) under greedy and constrained decoding using automatic metrics and LLM-as-a-Judge evaluation. For MCQA, CPT+SFT most often achieves the best scores, but gains over SFT are small and frequently not statistically significant, making SFT a strong and cost-effective default. For OEQA, CPT consistently improves overlap-based metrics, while SFT often degrades generation quality; instruction tuning and CPT+SFT are preferred by LLM-based evaluation. Cross-lingual experiments further show effective transfer from French adaptation to English benchmarks. Overall, we provide practical guidelines for selecting adaptation strategies under computational constraints.

Trade-offs in Medical LLM Adaptation: An Empirical Study in French QA

Automatic report labeling facilitates the identification of clinical findings from unstructured text and enables large-scale annotation for medical imaging research.
 Existing rule-based labelers struggle with the diverse descriptions in clinical reports, while fine-tuning pre-trained language models (PLMs) requires large amounts of labeled data that are often unavailable in clinical settings.
 In this paper, we propose PromptRad, a knowledge-enhanced multi-label prompt-tuning approach for radiology report labeling under low-resource settings.
 PromptRad reformulates multi-label classification as masked language modeling and incorporates synonyms from the UMLS Metathesaurus into a multi-word verbalizer to enrich category representations.
 By fine-tuning the PLM without additional classification layers, PromptRad requires substantially less labeled data than conventional
 fine-tuning. Experiments on liver CT (computed tomography) reports show that PromptRad outperforms dictionary-based and fine-tuning baselines with only 32 labeled training examples, and achieves competitive performance with GPT-4 despite using a much smaller model.
 Further analysis demonstrates that PromptRad captures complex negation patterns more effectively than existing methods, making it a promising solution for report labeling in data-scarce clinical scenarios.
 Our code is available at https://github.com/ila-lab/PromptRad.

PromptRad: Knowledge-Enhanced Multi-Label Prompt-Tuning for Low-Resource Radiology Report Labeling

The global transition to the ICD-11 taxonomy demands robust automated medical coding, yet comprehensive benchmarks to evaluate Large Language Models (LLMs) on this task remain absent. We introduce MAX-EVAL-11, the first large-scale benchmark for full-spectrum ICD-11 medical coding. MAX-EVAL-11 comprises 10,000 MIMIC-III discharge summaries with mapped, expert-validated ICD-11 annotations spanning 99.87\% of the diagnostic taxonomy. To better reflect clinical utility, we propose a novel hierarchical evaluation framework that assigns partial credit based on ICD-11's 5-level structure, addressing the brittleness of traditional exact-match metrics.
 
 Our evaluation of state-of-the-art LLMs reveals significant performance gaps. The best-performing model (Claude 4 Sonnet) achieves a weighted score of 0.433, outperforming both general-purpose peers and specialized medical models (MedCoder). Crucially, all models exhibit near-zero exact match rates (0–4.8\%) and rely primarily on hierarchical credit, underscoring the extreme difficulty of precise ICD-11 code generation. Furthermore, the superiority of general-purpose LLMs over legacy ICD-10 medical models (with ICD-11 codelist) suggests that broad reasoning capabilities currently outweigh domain-specific training for complex taxonomy scaling.

MAX-EVAL-11: A Large Scale Benchmark for Evaluating Large Language Models on Full-Spectrum ICD-11 Medical Coding

Reliable extraction of structured information from radiology reports using Large Language Models (LLMs) remains a significant challenge, particularly for complex, non-English texts such as Hebrew. This study proposes an agent-based, uncertainty-aware framework to enhance the reliability and interpretability of LLM predictions in clinical contexts. A total of 9,683 Hebrew radiology reports from Crohn's disease patients (2010–2023) across three medical centers were analyzed. Of these, 512 reports were manually annotated for six gastrointestinal organs and 15 pathological findings, while the remainder were automatically labeled using HSMP-BERT. Structured data extraction was performed with Llama 3.1 (Llama 3-8b-instruct) employing Bayesian Prompt Ensembles (BayesPE), which utilized six semantically equivalent prompts to quantify uncertainty. An Agent-Based Decision Model aggregated prompt outputs into five calibrated confidence levels and was benchmarked against three entropy-based approaches. Model performance was assessed using accuracy, F1 score, precision, recall, and Cohen's Kappa before and after filtering high-uncertainty cases. The agent-based model outperformed all baselines, achieving an F1 score of 0.3967, recall of 0.6437, and Kappa of 0.3006; after excluding cases with uncertainty >= 0.5, the F1 score increased to 0.4787 and Kappa to 0.4258. The proposed framework improves uncertainty calibration and predictive reliability, advancing the safe deployment of LLMs in medical data extraction.

Trustworthy NLP for Low-Resource Languages: Agent-Based Uncertainty Modeling for Hebrew Radiology Report Structuring

NER requires token-level classification using both left and right context, which makes encoder-only models like BERT naturally well-suited for the task. Decoder-only LLMs, by contrast, use causal masking during training, so their token representations lack right-side context, limiting their effectiveness on structured prediction tasks like NER despite their strong general capabilities.
 To address this, the authors propose fine-tuning decoder-only LLMs with causal attention replaced by full attention, combined with label-supervised discriminative training. While similar ideas exist in prior work, those studies were limited in scope. This work evaluates seven LLMs across four model families (Gemma, Qwen2.5, Llama3.1, Llama3.2) and compares full fine-tuning against LoRA. Results show that the proposed approach with an appropriate LoRA configuration outperforms encoder baselines (BERT, RoBERTa, DeBERTa), and achieves strong NER performance without auxiliary data or architectural modifications, though it does not reach SOTA on BC5CDR and CoNLL2003.

Treating Decoder-Only LLMs as Encoders: A Simple and Effective Fine-tuning Approach for Named Entity Recognition

Nutrition misinformation on social media often arises from selective interpretation of scientific evidence rather than outright falsehoods, making it difficult to detect. We introduce a curated, expert-annotated Instagram dataset focused on seed oils and omega-6, two domains characterized by contested dietary claims. We evaluate feature-based, embedding-based, and transformer-based models under in-domain and cross-domain settings. Results show strong in-domain performance across all models, with Sentence-BERT achieving the highest AUPRC (up to 0.96). However, performance drops substantially under cross-domain transfer, indicating limited robustness to topic shift. Analysis suggests that while contextual embeddings capture strong in-domain semantic signals, linguistically and psychologically grounded features are more stable under distribution shift. These findings highlight the value of combining semantic and interpretable linguistic signals for robust misinformation detection.

A Multi-View Framework for Cross-Domain Nutrition Misinformation Detection in Social Media

Large language models (LLMs) can generate or synthesize clinical text for a wide range of applications, from improving clinical documentation to augmenting clinical text analytics. Yet evaluations typically focus on a narrow aspect -- such as similarity or utility comparisons -- even though these aspects are complementary and best viewed in parallel. In this study, we aim to conduct a systematic evaluation of LLM-generated clinical text, which includes intrinsic, extrinsic, and factuality evaluations of synthetic clinical notes rephrased from MIMIC databases at million-note scale. Our analysis demonstrates that synthetic notes preserve core clinical information and predictive utility for coarse-grained tasks despite substantial linguistic changes, but lose fine-grained details for task like ICD coding. We show this loss of detail can be substantially mitigated by rephrasing notes by chunks rather than by the whole note, but at the cost of reduced factual precision under incomplete context. Through fact-checking and error analysis, we further find that synthesis errors are dominated by misinterpretation of clinical context, alongside temporal confusion, measurement errors, and fabricated claims. Finally, we show that the synthetic notes -- despite their task-agnostic nature -- can effectively augment task-specific training for rare ICD codes.

Systematic Evaluation of the Quality of Synthetic Clinical Notes Rephrased by LLMs at Million-Note Scale

Clinical coding maps clinical documentation to standardized medical codes, an essential yet time-consuming administrative task that could benefit from automation. Current models on ICD coding are typically optimized for codes from a specific ICD version. However, in reality, ICD systems evolve continuously, and different versions are adopted across time periods and regions. Moreover, ICD coding suffers from the long-tail problem, and rare code performance can be a bottleneck for developing implementable models. We examine whether it is viable to train version-independent models by combining data annotated in different ICD versions, which may help address these challenges. We add ICD-9 data to the training of a modified label-wise attention model for ICD-10 prediction, and find that despite the version mismatch, adding ICD-9 yields a 27% increase in micro F1 for 18K rare ICD codes compared to training on ICD-10 alone. On 8K frequent ICD-10 codes, the multi-version training also substantially improves macro metrics, with far fewer model parameters.

Bridging the Version Gap: Multi-version Training Improves ICD Code Prediction, Especially for Rare Codes

The advent of single-cell RNA sequencing has enabled unprecedented resolution of cell fate decisions and regulatory mechanisms during peri-implantation human embryogenesis, in which accurate cell type annotation is a fundamental prerequisite and the first step for subsequent fate and mechanism inference. Large language models (LLMs) have demonstrated outstanding performance in various fields. However, current studies mostly rely on traditional methods and have not explored the application of LLMs in the field of human embryonic cell annotation. The main reason is the lack of instruction tuning datasets and evaluation benchmarks. In this paper, we proposed EmCellLLM, the first open sourced LLMs that are specialized for human embryonic cell type prediction task based on fine-tuning Qwen3-8B with EmCell4Instruction, the first embryonic cell type prediction instruction dataset. To support LLM instruction tuning, we also build EmCellBench, the first benchmark for evaluating human embryonic cell type prediction ability of LLMs. We compare our models with a variety of LLMs on EmCellBench, where our model outperforms all other open-sourced LLMs as well as DeepSeek.

Downloads

Next from ACL 2026

MedCAT v2: a modular, extensible architecture for clinical named entity recognition and linking under real-world privacy and compute constraints

Similar lecture

Enhancing textual counterfactual explanation intelligibility through Counterfactual Feature Importance

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

MedCAT v2: a modular, extensible architecture for clinical named entity recognition and linking under real-world privacy and compute constraints

Similar lecture

Enhancing textual counterfactual explanation intelligibility through Counterfactual Feature Importance

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads