China

The rapid advancement of Chinese large language models (LLMs) underscores the need for domain-specific evaluations to ensure reliable applications. However, existing benchmarks often lack coverage in vertical domains and offer limited insights into the Chinese working context. Leveraging qualification exams as a unified framework for human expertise evaluation, we introduce QualBench, the first multi-domain Chinese QA benchmark dedicated to localized assessment of Chinese LLMs. The dataset includes over 17,000 questions across six vertical domains, with data selections grounded in 24 Chinese qualifications to closely align with national policies and working standards. Through comprehensive evaluations, the Qwen2.5 model outperformed the more advanced GPT-4o, with Chinese LLMs consistently surpassing non-Chinese models, highlighting the importance of localized domain knowledge in meeting qualification requirements. The best performance of 75.26% reveals the current gaps in domain coverage within model capabilities. Furthermore, we present the failure of LLM collaboration with crowdsourcing mechanisms and suggest the opportunities for multi-domain RAG knowledge enhancement and vertical domain LLM training with Federated Learning.

EMNLP 2025

QualBench: Benchmarking Chinese LLMs with Localized Professional Qualifications for Vertical Domain Evaluation

llm/ai agents

knowledge base qa

benchmarking

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

ACL 2024

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

poster

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Currently, Large Language Models (LLMs) have achieved remarkable results in machine translation. However, their performance in multi-domain translation (MDT) is less satisfactory, the meanings of words can vary across different domains, highlighting the significant ambiguity inherent in MDT. Therefore, evaluating the disambiguation ability of LLMs in MDT, remains an open problem. To this end, we present an evaluation and analysis of LLMs on disambiguation in multi-domain translation (DMDTEval), our systematic evaluation framework consisting of three critical aspects: (1) we construct a translation test set with multi-domain ambiguous word annotation, (2) we curate a diverse set of disambiguation prompt strategies, and (3) we design precise disambiguation metrics, and study the efficacy of various prompt strategies on multiple state-of-the-art LLMs. We conduct comprehensive experiments across 4 language pairs and 13 domains, our extensive experiments reveal a number of crucial findings that we believe will pave the way and also facilitate further research in the critical area of improving the disambiguation of LLMs. Our code and data will be released.

DMDTEval: An Evaluation and Analysis of LLMs on Disambiguation in Multi-domain Translation

Codes serve as the fundamental language for human to communicate with machines, and various Transformer-based models are trained to process codes in recent advancements. A unique symmetry of code is its semantic-preserving permutation, which allows certain lines to be rearranged without altering the overall meaning. To capture such symmetry, we propose a novel attention mechanism that incorporates semantic-preserving permutation equivariance, called the SPE attention. By leveraging the symmetry relationships within code, we introduce a directed layered graph to represent the code structure, and this graph is then summarized into a symmetry mask. The SPE attention integrates those symmetry masks, granting semantic-preserving permutations equivariance to the model. Experiments on various code related tasks, including code summarization and error detection, demonstrate the effectiveness of the proposed SPE attention.

SPE Attention: Making Attention Equivariant to Semantic-Preserving Permutation for Code Processing

Despite impressive progress in areas like mathematical reasoning, large language models still face challenges in consistently solving complex problems. Drawing inspiration from key human learning strategies, we propose two novel strategies to enhance the capability of large language models to solve these complex problems. First, Adaptive Difficulty Curriculum Learning (ADCL) is a novel curriculum learning strategy that tackles the Difficulty Shift phenomenon (i.e., a model's perception of problem difficulty dynamically changes during training) by periodically re-estimating difficulty within upcoming data batches to maintain alignment with the model's evolving capabilities. Second, Expert-Guided Self-Reformulation (EGSR) is a novel reinforcement learning strategy that bridges the gap between imitation learning and pure exploration by guiding models to reformulate expert solutions within their own conceptual framework, rather than relying on direct imitation, fostering deeper understanding and knowledge assimilation. Extensive experiments on challenging mathematical reasoning benchmarks, using Qwen2.5-7B as the base model, demonstrate that these human-inspired strategies synergistically and significantly enhance performance. Notably, their combined application improves performance over the standard Zero-RL baseline by 10\% on the AIME24 benchmark and 16.6\% on AIME25.

Learning Like Humans: Advancing LLM Reasoning Capabilities via Adaptive Difficulty Curriculum Learning and Expert-Guided Self-Reformulation

As Large Language Models (LLMs) are increasingly deployed in sensitive domains such as enterprise and government, ensuring that they adhere to **user-defined security policies** within context is critical-especially with respect to information non-disclosure. While prior LLM studies have focused on general safety and socially sensitive data, large-scale benchmarks for **contextual security** preservation against attacks remain lacking. To address this, we introduce a novel large-scale benchmark dataset, **CoPriva**, evaluating LLM adherence to contextual non-disclosure policies in question answering. Derived from realistic contexts, our dataset includes explicit policies and queries designed as direct and challenging indirect attacks seeking prohibited information. We evaluate 10 LLMs on our benchmark and reveal a significant vulnerability: many models violate user-defined policies and leak sensitive information. This failure is particularly severe against indirect attacks, highlighting a critical gap in current LLM safety alignment for sensitive applications. Our analysis reveals that while models can often identify the correct answer to a query, they struggle to incorporate policy constraints during generation. In contrast, they exhibit a partial ability to revise outputs when explicitly prompted. Our findings underscore the urgent need for more robust methods to guarantee contextual security.

Keep Security! Benchmarking Security Policy Preservation in Large Language Model Contexts Against Indirect Attacks in Question Answering

As Large Language Models (LLMs) expand across domains, LLM judges have become essential for systems evaluation. Current benchmarks typically compare system outputs against reference outputs from an encore model. This encore-mediated approach, though convenient, yields lower reliability than direct comparison between systems. We propose Arena-lite, which combines direct head-to-head comparison of outputs from competing systems with a tournament structure, eliminating the need for encore outputs, reducing the number of required comparisons, and achieving higher reliability in system rankings. We conducted two experiments: (1) controlled stochastic modeling and (2) empirical validation with a real LLM judge. Those experiments collectively demonstrate that Arena-lite consistently achieves higher reliability with fewer comparisons, even with smaller datasets or weaker judges. We release an easy-to-use web demonstration and code to foster adoption of Arena-lite, streamlining model selection across research and industry communities.

Arena-lite: Efficient and Reliable Large Language Model Evaluation via Tournament-Based Direct Comparisons

LLMs have shown impressive progress in natural language processing. However, they still face significant challenges in TableQA, where real-world complexities such as diverse table structures, multilingual data, and domain-specific reasoning are crucial. Existing TableQA benchmarks are often limited by their focus on simple flat tables and suffer from data leakage. Furthermore, most benchmarks are monolingual and fail to capture the cross-lingual and cross-domain variability in practical applications. To address these limitations, we introduce TableEval, a new benchmark designed to evaluate LLMs on realistic TableQA tasks. Specifically, TableEval includes tables with various structures (such as concise, hierarchical, and nested tables) collected from four domains (including government, finance, academia, and industry reports). Besides, TableEval features cross-lingual scenarios with tables in Simplified Chinese, Traditional Chinese, and English. To minimize the risk of data leakage, we collect all data from recent real-world documents. Considering that existing TableQA metrics fail to capture semantic accuracy, we further propose SEAT, a new evaluation framework that assesses the alignment between model responses and reference answers at the sub-question level. Experimental results have shown that SEAT achieves high agreement with human judgment. Extensive experiments on TableEval reveal critical gaps in the ability of state-of-the-art LLMs to handle these complex, real-world TableQA tasks, offering insights for future improvements.

TableEval: A Real-World Benchmark for Complex, Multilingual, and Multi-Structured Table Question Answering

We introduce the Phonetics-Semantics Evaluation Testbed (PSET), a new English-based testbed to evaluate phonetic embeddings. Our testbed is built on the assumption that phonetic embeddings should always prioritize phonetics over semantics, and it therefore leverages homophones and synonyms. We use PSET to test three phonetic embedding models: articulatory embeddings, Phoneme2Vec, and XPhoneBERT. The phonetic-based embeddings solve the task with varying degrees of success, with Phoneme2Vec performing the best. We also test five recent LLMs, GPT-4o, Gemini 2.5 Flash, Llama 3.1-8B, OLMo-7B and OLMo 2-7B. Gemini 2.5 Flash performs better than the other models. With this testbed, we hope to advance the development and evaluation of phonetic embedding models.

PSET: a Phonetics-Semantics Evaluation Testbed

Large Language Models (LLMs) often generate errors in their reasoning chains that can propagate and complicate checking the correctness of intermediate claims. Current LLM-based error detection methods usually take in the full reasoning chain as the context and output a score for each step. However, the model can be misled when there are incorrect steps in the context, and these errors are propagated to later steps. To address this problem, we leverage how humans typically check the soundness of claims in a reasoning chain, and introduce Reasoning Entailment Stability (RES), a novel probabilistic framework that inductively judges each step in a reasoning chain based solely on the previously validated claims. RES achieves 72.1% F1 (+8.2 points) across four benchmarks and 90.3% F1 (+27.6 points) on our controllable dataset with long reasoning chains.

Probabilistic Soundness Guarantees in LLM Reasoning Chains

Recent work has demonstrated the remarkable potential of Large Language Models (LLMs) in test-time scaling. By making the models think before answering, they are able to achieve much higher accuracy with extra inference computation. However, in many real-world scenarios, models are used under time constraints, where an answer should be given to the user within a certain output length. It is unclear whether and how the reasoning abilities of LLMs remain effective under such constraints. We take a first look at this problem by conducting an in-depth empirical study. Specifically, we test more than 25 LLMs on common reasoning datasets under a wide range of output length budgets, and we analyze the correlation between the inference accuracy and various properties including model type, model size, prompt style, etc. We also consider the mappings between the token budgets and the actual on-device latency budgets. The results have demonstrated several interesting findings regarding the budget-aware LLM reasoning that differ from the unconstrained situation, e.g. the optimal choices of model sizes and prompts change under different budgets. These findings offer practical guidance for users to deploy LLMs under real-world latency constraints.

An Empirical Study of LLM Reasoning Ability Under Strict Output Length Constraint

Vision-Language Models (VLMs) now generate discourse-level, multi-sentence visual descriptions, challenging text scene graph parsers originally designed for single-sentence caption-to-graph mapping. Current approaches typically merge sentence-level parsing outputs for discourse input, often missing phenomena like cross-sentence coreference, resulting in fragmented graphs and degraded downstream VLM task performance. To address this, we introduce textbfDiscourse-level text textbfScene textbfGraph parsing (DiscoSG), supported by our dataset DiscoSG-DS, which comprises 400 expert-annotated and 8,430 synthesized multi-sentence caption–graph pairs for images. Each graph contains sim15times more objects and relations per caption than prior datasets. While fine-tuning large PLMs (i.e., GPT-4o) on DiscoSG-DS improves SPICE by sim 48% over the best sentence-merging baseline, high inference cost and restrictive licensing hinder its open-source use, and smaller fine-tuned PLMs struggle with complex graphs. We propose DiscoSG-Refiner, which drafts a base graph using one small PLM, then employs a second PLM to iteratively propose graph edits, reducing full-graph generation overhead. Using two Flan-T5-BASE models, DiscoSG-Refiner still improves SPICE by sim 30% over the best baseline while achieving 86times faster inference than GPT-4o. It also consistently improves downstream VLM tasks like discourse-level caption evaluation and hallucination detection. Code and data are available at~\url{https://anonymous.4open.science/r/DiscoSG/}.

Downloads

Next from EMNLP 2025

DMDTEval: An Evaluation and Analysis of LLMs on Disambiguation in Multi-domain Translation

Similar lecture

CMMLU: Measuring massive multitask language understanding in Chinese

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES