Austria

Large Language Models (LLMs) such as ChatGPT demonstrate significant potential in the medical domain and are often evaluated using multiple-choice questions (MCQs) modeled on exams like the USMLE. However, such benchmarks may overestimate true clinical understanding by rewarding pattern recognition and test-taking heuristics. To investigate this, we created a fictional medical benchmark centered on an imaginary organ, the Glianorex, allowing us to separate memorized knowledge from reasoning ability. We generated textbooks and MCQs in English and French using leading LLMs, then evaluated proprietary, open-source, and domain-specific models in a zero-shot setting. Despite the fictional content, models achieved an average score of 64%, while physicians scored only 27%. Fine-tuned medical models outperformed base models in English but not in French. Ablation and interpretability analyses revealed that models frequently relied on shallow cues, test-taking strategies, and hallucinated reasoning to identify the correct choice. These results suggest that standard MCQ-based evaluations may not effectively measure clinical reasoning and highlight the need for more robust, clinically meaningful assessment methods for LLMs.

ACL 2025

Pattern Recognition or Medical Knowledge? The Problem with Multiple-Choice Questions in Medicine

multiple choice questions

medicine

large language model

benchmark

evaluation

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s 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.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

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 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

poster

Transformers, as the fundamental deep learning architecture, have demonstrated great capability in reasoning. This paper studies the generalizable first-order logical reasoning ability of transformers with their *parameterized* knowledge and how to improve it. Transformers' capability of first-order reasoning is further captured by whether they can conduct first-order logical entailment, which is quantitatively measured by their performance in answering knowledge graph queries. We establish the connections between (1) two types of distribution shifts studied in out-of-distribution generalization and (2) unseen knowledge and query settings discussed in the task of knowledge graph query answering, which makes it possible to characterize the fine-grained generalizability. Results on our comprehensive dataset showed that transformers **outperform** previous methods designed particularly for this task and provided detailed empirical evidence about the impact of the input query syntax, token embedding, and transformer architectures on the reasoning capability of transformers. Interestingly, our results revealed the mismatch of positional encoding and other design choices of transformer architectures in previous practices. Motivated by this, we propose **TEGA**, a logic-aware architecture that significantly improves the performance in generalizable first-order logical entailment.

Enhancing Transformers for Generalizable First-Order Logical Entailment

As large language models (LLMs) become increasingly prevalent, concerns about their reliability, particularly due to hallucinations - factually inaccurate or irrelevant outputs - have grown. Our research investigates the relationship between the uncertainty in training dynamics and the emergence of hallucinations. Using models from the Pythia suite and several hallucination detection metrics, we analyze hallucination trends and identify significant variance during training. To address this, we propose Sensitivity Dropout (SenD), a novel training protocol designed to reduce hallucination variance during training by deterministically dropping embedding indices with significant variability. In addition, we develop an unsupervised hallucination detection metric, Efficient EigenScore (EES), which approximates the traditional EigenScore in 2x speed. This metric is integrated into our training protocol, allowing SenD to be both computationally scalable and effective at reducing hallucination variance. SenD improves test-time reliability of Pythia and Meta's Llama models by up to 17% and enhances factual accuracy in Wikipedia, Medical, Legal, and Coding domains without affecting downstream task performance.

Hallucination Detox: Sensitivity Dropout (SenD) for Large Language Model Training

Concept-based explanations work by mapping complex model computations to human-understandable concepts. Evaluating such explanations is very difficult, as it includes not only the quality of the induced space of possible concepts but also how effectively the chosen concepts are communicated to users. Existing evaluation metrics often focus solely on the former, neglecting the latter.

We introduce an evaluation framework for measuring concept explanations via automated simulatability: a simulator's ability to predict the explained model's outputs based on the provided explanations. This approach accounts for both the concept space and its interpretation in an end-to-end evaluation. Human studies for simulatability are notoriously difficult to enact, particularly at the scale of a wide, comprehensive empirical evaluation (which is the subject of this work). We propose using large language models (LLMs) as simulators to approximate the evaluation and report various analyses to make such approximations reliable. Our method allows for scalable and consistent evaluation across various models and datasets. We report a comprehensive empirical evaluation using this framework and show that LLMs provide consistent rankings of explanation methods. Code available at Anonymous GitHub.

ConSim: Measuring Concept-Based Explanations' Effectiveness with Automated Simulatability

Automated interpretability pipelines generate natural language descriptions for the concepts represented by features in large language models (LLMs), such as "plants" or "the first word in a sentence". These descriptions are derived using inputs that activate the feature, which may be a dimension or a direction in the model's representation space. However, identifying activating inputs is costly, and the mechanistic role of a feature in model behavior is determined both by how inputs cause a feature to activate and by how feature activation affects outputs. Using steering evaluations, we reveal that current pipelines provide descriptions that fail to capture the causal effect of the feature on outputs. To fix this, we propose efficient, output-centric methods for automatically generating feature descriptions. These methods use the tokens weighted higher after feature stimulation or the highest weight tokens after applying the vocabulary "unembedding" head directly to the feature. Our output-centric descriptions better capture the causal effect of a feature on model outputs than input-centric descriptions, but combining the two leads to the best performance on both input and output evaluations. Lastly, we show that output-centric descriptions can be used to find inputs that activate features previously thought to be "dead".

Enhancing Automated Interpretability with Output-Centric Feature Descriptions

Generative models such as Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) trained on massive datasets can lead them to memorize and inadvertently reveal sensitive information, raising ethical and privacy concerns. While some prior works have explored this issue in the context of LLMs, it presents a unique challenge for MLLMs due to the entangled nature of knowledge across modalities, making comprehensive unlearning more difficult. To address this challenge, we propose Modality Aware Neuron Unlearning (MANU), a novel unlearning framework for MLLMs designed to selectively clip neurons based on their relative importance to the targeted forget data, curated for different modalities. Specifically, MANU consists of two stages: important neuron selection and selective pruning. The first stage identifies and collects the most influential neurons across modalities relative to the targeted forget knowledge, while the second stage is dedicated to pruning those selected neurons. MANU effectively isolates and removes the neurons that contribute most to the forget data within each modality, while preserving the integrity of retained knowledge. Our experiments conducted across various MLLM architectures illustrate that MANU can achieve a more balanced and comprehensive unlearning in each modality without largely affecting the overall model utility.

Modality-Aware Neuron Pruning for Unlearning in Multimodal Large Language Models

Current unlearning methods for large language models usually rely on reverse optimization to reduce target token probabilities. However, this paradigm disrupts the subsequent tokens prediction, degrading model performance and linguistic coherence. Moreover, existing evaluation metrics overemphasize contextual forgetting while inadequately assessing response fluency and relevance. To address these challenges, we propose ReLearn, a data augmentation and fine-tuning pipeline for effective unlearning, along with a comprehensive evaluation framework. This framework introduces Knowledge Forgetting Ratio (KFR) and Knowledge Retention Ratio (KRR) to measure knowledge-level preservation, and Linguistic Score (LS) to evaluate generation quality. Our experiments show that ReLearn successfully achieves targeted forgetting while preserving high-quality outputs. Through mechanistic analysis, we further demonstrate how reverse optimization disrupts coherent text generation, while ReLearn preserves this essential capability.

ReLearn: Unlearning via Learning for Large Language Models

While Retrieval-Augmented Generation (RAG) has emerged as an effective approach for addressing the knowledge outdating problem in Large Language Models (LLMs), it still faces a critical challenge: the prevalence of outdated information in knowledge bases. Current research primarily focuses on incorporating up-to-date information, yet the impact of outdated information coexisting in retrieval sources remains inadequately addressed. To bridge this gap, we introduce HoH, the first benchmark specifically designed to evaluate the impact of outdated information on RAG. Our benchmark leverages token-level diff algorithms combined with LLM pipelines to efficiently create a large-scale QA dataset that accurately captures the evolution of temporal knowledge in real-world facts.
Through comprehensive experiments, we reveal that outdated information significantly degrades RAG performance in two critical ways: (1) it substantially reduces response accuracy by distracting models from correct information, and (2) it can mislead models into generating potentially harmful outputs, even when current information is available. Current RAG approaches struggle with both retrieval and generation aspects when handling outdated information. These findings highlight the urgent need for innovative solutions to address the temporal challenges in RAG.

HoH: A Dynamic Benchmark for Evaluating the Impact of Outdated Information on Retrieval-Augmented Generation

This paper studies how Transformer models with Rotary Position Embeddings (RoPE) develop emergent, wavelet-like properties that compensate for the positional encoding's theoretical limitations. Through an analysis spanning model scales, architectures, and training checkpoints, we show that attention heads evolve to implement multi-resolution processing analogous to wavelet transforms. We demonstrate that this scale-invariant behavior is unique to RoPE, emerges through distinct evolutionary phases during training, and statistically adheres to the fundamental uncertainty principle. Our findings suggest that the effectiveness of modern Transformers stems from their remarkable ability to spontaneously develop optimal, multi-resolution decompositions to address inherent architectural constraints.

Beyond Position: the emergence of wavelet-like properties in Transformers

Factual hallucinations are a major challenge for Large Language Models (LLMs). They undermine reliability and user trust by generating inaccurate or fabricated content. Recent studies suggest that when generating false statements, the internal states of LLMs encode information about truthfulness. However, these studies often rely on synthetic datasets that lack realism, which limits generalization when evaluating the factual accuracy of text generated by the model itself. In this paper, we challenge the findings of previous work by investigating truthfulness encoding capabilities, leading to the generation of a more realistic and challenging dataset. Specifically, we extend previous work by introducing: (1) a strategy for sampling plausible true-false factoid sentences from tabular data and (2) a procedure for generating realistic, LLM-dependent true-false datasets from Question Answering collections. Our analysis of two open-source LLMs reveals that while the findings from previous studies are partially validated, generalization to LLM-generated datasets remains challenging. This study lays the groundwork for future research on factuality in LLMs and offers practical guidelines for more effective evaluation.

Are the Hidden States Hiding Something? Testing the Limits of Factuality-Encoding Capabilities in LLMs

Large Language Models (LLMs) have rapidly become central to NLP, demonstrating their ability to adapt to various tasks through prompting techniques, including sentiment analysis. However, we still have a limited understanding of how these models capture sentiment-related information. This study probes the hidden layers of LLaMA models to pinpoint where sentiment features are most represented and to assess how this affects sentiment analysis.

Using probe classifiers, we analyze sentiment encoding across layers and scales, identifying the layers and pooling methods that best capture sentiment signals. Our results show that sentiment information is most concentrated in mid-layers for binary polarity tasks, with detection accuracy increasing up to 14% over prompting techniques. Additionally, we find that in decoder-only models, the last token is not consistently the most informative for sentiment encoding. Finally, this approach enables sentiment tasks to be performed with memory requirements reduced by an average of 57%.

These insights contribute to a broader understanding of sentiment in LLMs, suggesting layer-specific probing as an effective approach for sentiment tasks beyond prompting, with potential to enhance model utility and reduce memory requirements.

Downloads

Next from ACL 2025

Enhancing Transformers for Generalizable First-Order Logical Entailment

Similar lecture

MCQFormatBench: Robustness Tests for Multiple-Choice Questions