Austria

This work introduces a novel framework for evaluating LLMs&#39; capacity to balance instruction-following with critical reasoning when presented with multiple-choice questions containing no valid answers. Through systematic evaluation across arithmetic, domain-specific knowledge, and high-stakes medical decision tasks, we demonstrate that post-training aligned models often default to selecting invalid options, while base models exhibit improved refusal capabilities that scale with model size. Our analysis reveals that alignment techniques, though intended to enhance helpfulness, can inadvertently impair models&#39; reflective judgment--the ability to override default behaviors when faced with invalid options. We additionally conduct a parallel human study showing similar instruction-following biases, with implications for how these biases may propagate through human feedback datasets used in alignment. We provide extensive ablation studies examining the impact of model size, training techniques, and prompt engineering. Our findings highlight fundamental tensions between alignment optimization and preservation of critical reasoning capabilities, with important implications for developing more robust AI systems for real-world deployment.

ACL 2025

Wait, that&#39;s not an option: LLMs Robustness with Incorrect Multiple-Choice Options

reflections and critiques

model bias/fairness evaluation

transparency

### 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)

This work introduces a novel framework for evaluating LLMs' capacity to balance instruction-following with critical reasoning when presented with multiple-choice questions containing no valid answers. Through systematic evaluation across arithmetic, domain-specific knowledge, and high-stakes medical decision tasks, we demonstrate that post-training aligned models often default to selecting invalid options, while base models exhibit improved refusal capabilities that scale with model size. Our analysis reveals that alignment techniques, though intended to enhance helpfulness, can inadvertently impair models' reflective judgment--the ability to override default behaviors when faced with invalid options. We additionally conduct a parallel human study showing similar instruction-following biases, with implications for how these biases may propagate through human feedback datasets used in alignment. We provide extensive ablation studies examining the impact of model size, training techniques, and prompt engineering. Our findings highlight fundamental tensions between alignment optimization and preservation of critical reasoning capabilities, with important implications for developing more robust AI systems for real-world deployment.

Wait, that's not an option: LLMs Robustness with Incorrect Multiple-Choice Options

poster

Large Language Models for code often entail significant computational complexity, which grows significantly with the length of the input code sequence. We propose $LeanCode$ for code simplification to reduce training and prediction time, leveraging code contexts in utilizing attention scores to represent the tokens' importance. We advocate for the selective removal of tokens based on the average context-aware attention scores rather than average scores across all inputs. $LeanCode$ uses the attention scores of `CLS' tokens within the encoder for classification tasks, such as code search. It also employs the encoder-decoder attention scores to determine token significance for sequence-to-sequence tasks like code summarization. Our evaluation shows $LeanCode$'s superiority over the SOTAs $DietCode$ and $SlimCode$, with improvements of 60\% and 16\% for code search, and 29\% and 27\% for code summarization, respectively.

LEANCODE: Understanding Models Better for Code Simplification of Pre-trained Large Language Models

To enable Large Language Models (LLMs) to function as conscious agents with generalizable reasoning capabilities, it is crucial that they possess the ability to ***comprehend situational changes (transitions) in distribution*** triggered by environmental factors or actions from other agents. Despite its fundamental significance, this ability remains underexplored due to the complexity of modeling infinite possible changes in an event and their associated distributions, coupled with the lack of benchmark data with situational transitions. Addressing these gaps, we propose a novel formulation of ***reasoning with distributional changes as a three-step discriminative process***, termed as ***MetAphysical ReaSoning***. We then introduce the first-ever benchmark, **MARS**, comprising three tasks corresponding to each step. These tasks systematically assess LLMs’ capabilities in reasoning the plausibility of (i) changes in actions, (ii) states caused by changed actions, and (iii) situational transitions driven by changes in action. Extensive evaluations with 20 (L)LMs of varying sizes and methods indicate that all three tasks in this process pose significant challenges, even after fine-tuning. Further analyses reveal potential causes for the underperformance of LLMs and demonstrate that pre-training on large-scale conceptualization taxonomies can potentially enhance LMs’ metaphysical reasoning capabilities. Our data and models are publicly accessible at https://github.com/HKUST-KnowComp/MARS.

MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset

The rise of large language models (LLMs) offers new opportunities for automatic error detection in education, particularly for math word problems (MWPs). While prior studies demonstrate the promise of LLMs as error detectors, they overlook the presence of multiple valid solutions for a single MWP. Our preliminary analysis reveals a significant performance gap between conventional and alternative solutions in MWPs, a phenomenon we term conformity bias in this work. To mitigate this bias, we introduce the Ask-Before-Detect (AskBD) framework, which generates adaptive reference solutions using LLMs to enhance error detection. Experiments on 200 examples of GSM8K show that AskBD effectively mitigates bias and improves performance, especially when combined with reasoning-enhancing techniques like chain-of-thought prompting.

Ask-Before-Detection: Identifying and Mitigating Conformity Bias in LLM-Powered Error Detector for Math Word Problem Solutions

Semi-structured tables, with their varied layouts and formatting artifacts, remain a major obstacle for automated data processing and analytics. To address these challenges, we propose RelationalCoder, which uniformly converts semi-structured tables into relational data, enabling smooth integration with the rich ecosystem of data processing and analytics tools. By leveraging SQL code, RelationalCoder prevents schema errors and markedly improves normalization quality across multiple relational tables.

To address the challenge of large tables, we propose a new technique called Loop Reference Decoding (LRD): it identifies \emph{expandable groups}—repeating regions of similar structure and semantics—and replicates each group using a concise loop over its repetitive region by referencing cell addresses, rather than regenerating each individual cell. This design substantially reduces output length from $\mathcal{O}(N \times M)$—proportional to the table’s height ($N$) and width ($M$)—to approximately $\mathcal{O}(K)$, where $K$ is the total number of unique cell types within detected expandable groups. As a result, LRD is highly scalable: the larger the input table, the greater the compression ratio. It scales seamlessly to extremely large tables, achieving output reductions of up to $100{,}000\times$.

We further create the first human-labeled corpus for table transformation, created with a cost-efficient, actively supervised pipeline. Extensive experiments on HiTab and MultiHiertt show that RelationalCoder not only enables programmatic symbolic reasoning but also boosts QA accuracy—raising Llama-2 and Mistral models by more than 20\%, and GPT-4o by over 4\%. 
Project page: https://github.com/haoyudong/RelationalCoder.

RelationalCoder: Relational Representation of Complex Tables for Program-Based Processing and Reasoning

Videos are unique in their integration of temporal elements, including camera, scene, action, and attribute, along with their dynamic relationships over time. However, existing benchmarks for video understanding often treat these properties separately or narrowly focus on specific aspects, overlooking the holistic nature of video content. To address this, we introduce TUNA, a temporal-oriented benchmark for fine-grained understanding on dense dynamic videos, with two complementary tasks: captioning and QA. Our TUNA features diverse video scenarios and dynamics, assisted by interpretable and robust evaluation criteria. We evaluate several leading models on our benchmark, providing fine-grained performance assessments across various dimensions. This evaluation reveals key challenges in video temporal understanding, such as limited action description, inadequate multi-subject understanding, and insensitivity to camera motion, offering valuable insights for improving video understanding models.

TUNA: Comprehensive Fine-grained Temporal Understanding Evaluation on Dense Dynamic Videos

This paper introduces UnSeenTimeQA, a novel data contamination-free time-sensitive question-answering (TSQA) benchmark. It differs from existing TSQA benchmarks by avoiding web-searchable queries grounded in the real world. We present a series of time-sensitive event scenarios based on synthetically generated facts. It requires large language models (LLMs) to engage in genuine temporal reasoning without depending on the factual knowledge acquired during the pre-training phase. Our data generation framework enables on‐demand generation of new samples, mitigating the risk of data leakage. We designed three types of time-sensitive questions to test LLMs' temporal reasoning abilities over sequential and parallel event occurrences. Our evaluation of five LLMs on synthetic fact-based TSQA reveals mixed results: while they perform well on simpler subsets, their overall performance remains inferior as compared to real world fact-based TSQA. Error analysis indicates that LLMs face difficulties in reasoning over long-range event dependencies and parallel events.

UnSeenTimeQA: Time-Sensitive Question-Answering Beyond LLMs' Memorization

This paper addresses the challenge of aspect-based summarization in education by introducing Reflective ASPect-based summarization (ReflectASP), a novel dataset that summarizes student reflections on STEM lectures. Despite the promising performance of large language models in general summarization, their application to nuanced aspect-based summaries remains under-explored. 
ReflectASP eases the exploration of open-aspect-based summarization (OABS), overcoming the limitations of current datasets and comes with ample human annotations. We benchmarked different types of zero-shot summarization methods and proposed two refinement methods to improve summaries, supported by both automatic and human manual evaluations. Additionally, we analyzed suggestions and revisions made during the refinement process, offering a fine-grained study of the editing strategies employed by these methods. We make our models, dataset, and all human evaluation results available at https://github.com/cs329yangzhong/ReflectASP.

From Information to Insight: Leveraging LLMs for Open Aspect-Based Educational Summarization

People naturally vary in their annotations for subjective questions and some of this variation is thought to be due to the person's sociodemographic characteristics. LLMs have also been used to label data, but recent work has shown that models perform poorly when prompted with sociodemographic attributes, suggesting limited inherent sociodemographic knowledge. Here, we ask whether LLMs can be trained to be accurate sociodemographic models of annotator variation. Using a curated dataset of five tasks with standardized sociodemographics, we show that models do improve in sociodemographic prompting when trained but that this performance gain is largely due to models learning annotator-specific behaviour rather than sociodemographic behaviours. Across all tasks, our results suggest that models learn little meaningful connection between sociodemographics and annotation, raising doubts about the current use of LLMs for simulating sociodemographic variation and behaviour.

Beyond Demographics: Fine-tuning Large Language Models to Predict Individuals' Subjective Text Perceptions

Large language models (LLMs) inherit biases from their training data and alignment processes, influencing their responses in subtle ways. While many studies have examined these biases, little work has explored their robustness during interactions. In this paper, we introduce a novel approach where two instances of an LLM engage in self-debate, arguing opposing viewpoints to persuade a neutral version of the model. Through this, we evaluate how firmly biases hold and whether models are susceptible to reinforcing misinformation or shifting to harmful viewpoints. Our experiments span multiple LLMs of varying sizes, origins, and languages, providing deeper insights into bias persistence and flexibility across linguistic and cultural contexts.

Bias in the Mirror : Are LLMs opinions robust to their own adversarial attacks

Autonomous agents have become increasingly important for interacting with the real world. Android agents, in particular, have been a frequently-mentioned interaction method. However, existing studies for training and evaluating Android agents lack systematic research on both open-source and closed-source models. In this work, we propose AndroidLab as a systematic Android agent framework. It includes an operation environment with different modalities, action space, and a reproducible benchmark. It supports both large language models (LLMs) and multimodal models (LMMs) in the same action space. AndroidLab benchmark includes predefined Android virtual devices and 138 tasks across nine apps built on these devices. By using the AndroidLab environment, we develop an Android Instruction dataset and train six open-source LLMs and LMMs, lifting the average success rates from 4.59\% to 21.50\% for LLMs and from 1.93\% to 13.28\% for LMMs. AndroidLab is open-sourced and publicly available at https://github.com/THUDM/Android-Lab.

Downloads

Next from ACL 2025

LEANCODE: Understanding Models Better for Code Simplification of Pre-trained Large Language Models

Similar lecture

None of the Above, Less of the Right Parallel Patterns in Human and LLM Performance on Multi-Choice Questions Answering