Austria

Despite strong performance on vision-language tasks, Multimodal Large Language Models (MLLMs) struggle with mathematical problem-solving, with both open-source and state-of-the-art models falling short of human performance on visual-math benchmarks. To systematically examine visual-mathematical reasoning in MLLMs, we (1) evaluate their understanding of geometric primitives, (2) test multi-step reasoning, and (3) explore a potential solution to improve visual reasoning capabilities. Our findings reveal fundamental shortcomings in shape recognition, with top models achieving under 50% accuracy in identifying regular polygons. We analyze these failures through the lens of dual-process theory and show that MLLMs rely on System 1 (intuitive, memorized associations) rather than System 2 (deliberate reasoning). Consequently, MLLMs fail to count the sides of both familiar and novel shapes, suggesting they have neither learned the concept of ``sides&#39;&#39; nor effectively process visual inputs. Finally, we propose Visually Cued Chain-of-Thought (VC-CoT) prompting, which enhances multi-step mathematical reasoning by explicitly referencing visual annotations in diagrams, boosting GPT-4o’s accuracy on an irregular polygon side-counting task from 7% to 93%. Our findings suggest that System 2 reasoning in MLLMs remains an open problem, and visually-guided prompting is essential for successfully engaging visual reasoning.

ACL 2025

Forgotten Polygons: Multimodal Large Language Models are Shape-Blind

mathematical reasoning

multimodality

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)

Despite strong performance on vision-language tasks, Multimodal Large Language Models (MLLMs) struggle with mathematical problem-solving, with both open-source and state-of-the-art models falling short of human performance on visual-math benchmarks. To systematically examine visual-mathematical reasoning in MLLMs, we (1) evaluate their understanding of geometric primitives, (2) test multi-step reasoning, and (3) explore a potential solution to improve visual reasoning capabilities. Our findings reveal fundamental shortcomings in shape recognition, with top models achieving under 50% accuracy in identifying regular polygons. We analyze these failures through the lens of dual-process theory and show that MLLMs rely on System 1 (intuitive, memorized associations) rather than System 2 (deliberate reasoning). Consequently, MLLMs fail to count the sides of both familiar and novel shapes, suggesting they have neither learned the concept of ``sides'' nor effectively process visual inputs. Finally, we propose Visually Cued Chain-of-Thought (VC-CoT) prompting, which enhances multi-step mathematical reasoning by explicitly referencing visual annotations in diagrams, boosting GPT-4o’s accuracy on an irregular polygon side-counting task from 7% to 93%. Our findings suggest that System 2 reasoning in MLLMs remains an open problem, and visually-guided prompting is essential for successfully engaging visual reasoning.

poster

Large language models (LLMs) are increasingly used in decision-making contexts, yet their ability to reason over event structure—an important component in the situational awareness needed to make nuanced decisions—is not well understood. By operationalizing proto-role theory, which characterizes agents via properties such as *instigation* and *volition* and patients via properties such as *change of state*, we examine the ability of LLMs to answer questions that require complex, multi-step event reasoning. Specifically, we investigate the extent to which LLMs capture semantic roles such as "agent" and "patient" through zero-shot prompts, and whether incorporating semantic proto-role labeling (SPRL) context improves semantic role labeling (SRL) performance in a zero-shot setting. We find that, while SPRL context sometimes degrades SRL accuracy in high-performing models (e.g., GPT-4o), it also uncovers an internal consistency between SPRL and SRL predictions that mirrors linguistic theory, and provides evidence that LLMs implicitly encode consistent multi-dimensional event role knowledge. Furthermore, our experiments support prior work showing that LLMs underperform human annotators in complex semantic analysis.

On the Role of Semantic Proto-roles in Semantic Analysis: What do LLMs know about agency?

Retrieval-augmented Generation (RAG) relies on effective retrieval capabilities, yet traditional sparse and dense retrievers inherently struggle with multi-hop retrieval scenarios. In this paper, we introduce GeAR, a system that advances RAG performance through two key innovations: (i) an efficient graph expansion mechanism that augments any conventional base retriever, such as BM25, and (ii) an agent framework that incorporates the resulting graph-based retrieval into a multi-step retrieval framework. Our evaluation demonstrates GeAR’s superior retrieval capabilities across three multi-hop question answering datasets. Notably, our system achieves state-of-the-art results with improvements exceeding 10% on the challenging MuSiQue dataset, while consuming fewer tokens and requiring fewer iterations than existing multi-step retrieval systems. The project page is available at https://gear-rag.github.io.

GeAR: Graph-enhanced Agent for Retrieval-augmented Generation

Precise recognition of search intent in Retrieval-Augmented Generation (RAG) systems remains a challenging goal, especially under resource constraints and for complex queries with nested structures and dependencies. This paper presents QCompiler, a neuro-symbolic framework inspired by linguistic grammar rules and compiler design, to bridge this gap. It theoretically designs a minimal yet sufficient Backus-Naur Form (BNF) grammar G[q] to formalize complex queries. Unlike previous methods, this grammar maintains completeness while minimizing redundancy. Based on this, QCompiler includes a Query Expression Translator, a Lexical Syntax Parser, and a Recursive Descent Processor to compile queries into Abstract Syntax Trees (ASTs) for execution. The atomicity of the sub-queries in the leaf nodes ensures more precise document retrieval and response generation, significantly improving the RAG system's ability to address complex queries.

Neuro-Symbolic Query Compiler

Gradient Ascent (GA) has emerged as a promising approach for concept unlearning in Multimodal Generative Models (MGMs), such as Multimodal Large Language Models (MLLMs) and Stable Diffusion Models (SDMs). Despite its effectiveness in removing undesired knowledge, GA leads to severe utility degradation in MGMs. In this paper, we explore the mechanism behind this degradation by quantifying two distinct forms of knowledge in MGMs: (i) Conceptual Knowledge, which represents specific information about concepts; (ii) Natural Knowledge, which refers to the ability to produce coherent and logically structured outputs. Our analysis reveals that applying GA globally not only removes the targeted Conceptual Knowledge but also inadvertently diminishes Natural Knowledge, resulting in utility collapse. To address this issue, we propose Forget the Token and Pixel (FTTP), a novel approach that selectively applies GA to targeted Conceptual Knowledge while preserving Natural Knowledge through Gradient Descent (GD). FTTP eliminates the need for additional retain sets and a large number of training steps, thereby reducing computational resource costs. Extensive experiments demonstrate FTTP’s efficiency and superior utility-unlearning tradeoff for both text and image generation tasks. Our source code will be released in the near future\footnote{Our code is available in the supplementary material, along with a link to the anonymous GitHub repository provided in the Appendix.}.

Forget the Token and Pixel: Rethinking Gradient Ascent for Concept Unlearning in Multimodal Generative Models

We introduce Slam, a recipe for training high-quality Speech Language Models (SLMs) on a single academic GPU in 24 hours. We do so through empirical analysis of model initialisation and architecture, synthetic training data, preference optimisation with synthetic data and tweaking all other components. We empirically demonstrate that this training recipe also scales well with more compute getting results on par with leading SLMs in a fraction of the compute cost. We hope these insights will make SLM training and research more accessible. In the context of SLM scaling laws, our results far outperform predicted compute optimal performance, giving an optimistic view to SLM feasibility. We open source code, data, models, and provide samples at https://anonymous.4open.science/w/anon_demo/ .

Slamming: Training a Speech Language Model on One GPU in a Day

Discourse relations contribute to the structure of a text and can optionally be realized through explicit connectives such as "but" and "while". But when are these connectives necessary to avoid possible misunderstandings? We investigate this question by first building a corpus of 4,274 text revisions in each of which a connective was explicitly inserted. For a subset of 250 cases, we collect plausibility annotations on other connectives to check whether they would represent suitable alternative relations. The results of this annotation show that several relations are often perceived as plausible in our data. Furthermore, we analyze the extent to which large language models can identify instances with multiple plausible relations as a possible source of misunderstandings. We find that the models predict plausibility of individual connectives with up to 66% accuracy, but they are not reliable in estimating when multiple relations are plausible.

Clarifying Underspecified Discourse Relations in Instructional Texts

As large language models (LLM) become more and more capable in languages other than English, it is important to collect benchmark datasets in order to evaluate their multilingual performance, including on tasks like machine translation (MT). In this work, we extend the WMT24 dataset to cover 55 languages by collecting new human-written references and post-edits for 46 new languages/dialects in addition to post-edits of the references in 8 out of 9 languages in the original WMT24 dataset. We benchmark a variety of MT providers and LLMs on the collected dataset using automatic metrics and find that LLMs are the best-performing MT systems in all 55 languages. However, we caution against using our results to reach strong conclusions about MT quality without a human-based evaluation due to limitations of automatic evaluation metrics, which we leave for future work.

WMT24++: Expanding the Language Coverage of WMT24 to 55 Languages & Dialects

This study explores current limitations of learned image captioning evaluation metrics, specifically the lack of granular assessment for individual word misalignments within captions, and the reliance on single-point quality estimates without considering uncertainty. To address these limitations, we propose a simple yet effective strategy for generating and calibrating CLIPScore distributions. Leveraging a model-agnostic conformal risk control framework, we calibrate CLIPScore values for task-specific control variables, to tackle the aforementioned two limitations. Experimental results demonstrate that using conformal risk control, over the distributions produced with simple methods such as input masking, can achieve competitive performance compared to more complex approaches. Our method effectively detects misaligned words, while providing formal guarantees aligned with desired risk levels, and improving the correlation between uncertainty estimations and prediction errors, thus enhancing the overall reliability of caption evaluation metrics.

A Conformal Risk Control Framework for Granular Word Assessment and Uncertainty Calibration of CLIPScore Quality Estimates

Intelligent tutoring agents powered by large language models (LLMs) have been increasingly explored to deliver personalized knowledge in areas such as language learning and science education. However, their capabilities in guiding users to solve complex real-world tasks remain underexplored. To address this limitation, in this work, we focus on coding tutoring, a challenging problem that requires tutors to proactively guide students towards completing predefined coding tasks. We propose a novel agent workflow, Trace-and-Verify (TRAVER), which combines knowledge tracing to estimate a student's knowledge state and turn-by-turn verification to ensure effective guidance toward task completion. We introduce DICT, an automatic evaluation protocol that assesses tutor agents using controlled student simulation and code generation tests. Extensive experiments reveal the challenges of coding tutoring and demonstrate that TRAVER achieves a significantly higher success rate. Although we use code tutoring as an example in this paper, our approach can be extended beyond coding, providing valuable insights into advancing tutoring agents for human task learning.

Training Turn-by-Turn Verifiers for Dialogue Tutoring Agents: The Curious Case of LLMs as Your Coding Tutors

Ensuring factual consistency in summarization remains a challenge, especially for long-document evaluation. While automated, reference-free evaluation models are essential given the impracticality of large-scale human assessment for lengthy texts, challenges persist in evaluating different systems on how to handle different summary granularities and evolving model generations. In this work, we conduct a systematic study on diverse factual-consistency evaluation systems across four long-document datasets, encompassing summaries generated by models from non-LLMs to proprietary LLMs. Our analysis reveals that fine-grained continuous scores can provide more reliable assessments of different evaluation systems' capabilities than binary classification. We also examine the relationship between sentence-level and summary-level model performance, highlighting its dependency on dataset characteristics. Moreover, our study reveals that advanced systems can obtain higher recall on error detection in older summaries, yet struggle with false-positives and fine-grained error detection. Our study provides insights into designing robust factuality evaluation systems, which are becoming more in demand as generative models continue to advance rapidly.

Downloads

Next from ACL 2025

On the Role of Semantic Proto-roles in Semantic Analysis: What do LLMs know about agency?

Similar lecture

Mitigating Visual Forgetting via Take-along Visual Conditioning for Multi-modal Long CoT Reasoning