China

Safety evaluation of Large Language Models (LLMs) has made progress and attracted academic interest, but it remains challenging to keep pace with the rapid integration of LLMs across diverse applications. Different applications expose users to various harms, necessitating application-specific safety evaluations with tailored harms and policies. Another major gap is the lack of focus on the dynamic and conversational nature of LLM systems. Such potential oversights can lead to harms that go unnoticed in standard safety benchmarks. This paper identifies the above as key requirements for robust LLM safety evaluation and recognizing that current evaluation methodologies do not satisfy these, we introduce the textttSAGE (Safety AI Generic Evaluation) framework. textttSAGE is an automated modular framework designed for customized and dynamic harm evaluations. It utilizes adversarial user models that are system-aware and have unique personalities, enabling a holistic red-teaming evaluation. We demonstrate textttSAGE&#39;s effectiveness by evaluating seven state-of-the-art LLMs across three applications and harm policies. Our experiments with multi-turn conversational evaluations revealed a concerning finding that harm steadily increases with conversation length. Furthermore, we observe significant disparities in model behavior when exposed to different user personalities and scenarios. Our findings also reveal that some models minimize harmful outputs by employing severe refusal tactics that can hinder their usefulness. These insights highlight the necessity of adaptive and context-specific testing to ensure better safety alignment and safer deployment of LLMs in real-world scenarios.

EMNLP 2025

SAGE: A Generic Framework for LLM Safety Evaluation

responsible ai

ai safety

evaluation

### 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.

Safety evaluation of Large Language Models (LLMs) has made progress and attracted academic interest, but it remains challenging to keep pace with the rapid integration of LLMs across diverse applications. Different applications expose users to various harms, necessitating application-specific safety evaluations with tailored harms and policies. Another major gap is the lack of focus on the dynamic and conversational nature of LLM systems. Such potential oversights can lead to harms that go unnoticed in standard safety benchmarks. This paper identifies the above as key requirements for robust LLM safety evaluation and recognizing that current evaluation methodologies do not satisfy these, we introduce the textttSAGE (Safety AI Generic Evaluation) framework. textttSAGE is an automated modular framework designed for customized and dynamic harm evaluations. It utilizes adversarial user models that are system-aware and have unique personalities, enabling a holistic red-teaming evaluation. We demonstrate textttSAGE's effectiveness by evaluating seven state-of-the-art LLMs across three applications and harm policies. Our experiments with multi-turn conversational evaluations revealed a concerning finding that harm steadily increases with conversation length. Furthermore, we observe significant disparities in model behavior when exposed to different user personalities and scenarios. Our findings also reveal that some models minimize harmful outputs by employing severe refusal tactics that can hinder their usefulness. These insights highlight the necessity of adaptive and context-specific testing to ensure better safety alignment and safer deployment of LLMs in real-world scenarios.

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.

This paper presents an audio chatbot system designed to handle a wide range of audio-related queries by integrating multiple specialized audio processing models. The proposed system uses an intent classifier, trained on a diverse audio query dataset, to route queries about audio content to expert models such as Automatic Speech Recognition (ASR), Speaker Diarization, Music Identification, and Text-to-Audio generation. A novel audio intent classification dataset is developed for building the intent classifier. A 3.8 B LLM model then takes inputs from an Audio Context Detection (ACD) module extracting audio event information from the audio and post processes text domain outputs from the expert models to compute the final response to the user. We evaluated the system on custom audio tasks and MMAU sound set benchmarks. The custom datasets were motivated by target use cases not covered in industry benchmarks. We proposed ACD-timestamp-QA (Question Answering) as well as ACD-temporal-QA datasets to evaluate timestamp and temporal reasoning questions, respectively. First, we determined that a BERT based Intent Classifier outperforms LLM-fewshot intent classifier in routing queries. Experiments further show that our approach significantly improves accuracy on some custom tasks compared to state-of-the-art Large Audio Language Models and outperforms models in the 7B parameter size range on the sound testset of the MMAU benchmark, thereby offering an attractive option for on device deployment.

Audio Query Handling System with Integrated Expert Models and Contextual Understanding

Large language models (LLMs) remain unreliable for global enterprise applications due to substantial performance gaps between high-resource and mid/low-resource languages, driven by English-centric pretraining and internal reasoning biases. This inconsistency undermines customer experience and operational reliability in multilingual settings such as customer support, content moderation, and information retrieval. Even with advanced Retrieval-Augmented Generation (RAG) systems, we observe up to an 29\% accuracy drop in non-English languages compared to English.

We propose a practical, batch-wise alignment strategy for fine-tuning LLMs, leveraging semantically equivalent multilingual data in each training batch to directly align model outputs across languages. This approach improves non-English accuracy by up to 23.9\% without compromising English performance, model reasoning, or retrieval quality. Our method is simple to implement, scalable, and integrates seamlessly with existing LLM training \& deployment pipelines, enabling more robust and equitable multilingual AI solutions in industry.

Aligning LLMs for Multilingual Consistency in Enterprise Applications

Creating high-quality, large-scale datasets for large language models (LLMs) often relies on resource-intensive, GPU-accelerated models for quality filtering, making the process time-consuming and costly. This dependence on GPUs limits accessibility for organizations lacking significant computational infrastructure. To address this issue, we introduce the Lightweight, Purpose-driven (LP) Data Pipeline, a framework that operates entirely on CPUs to streamline the processes of dataset extraction, filtering, and curation. Based on our four core principles, the LP Data Pipeline significantly reduces preparation time and cost while maintaining high data quality. Importantly, our pipeline enables the creation of purpose-driven datasets tailored to specific domains and languages, enhancing the applicability of LLMs in specialized contexts. We anticipate that our pipeline will lower the barriers to LLM development, enabling a wide range of organizations to access LLMs more easily.

LP Data Pipeline: Lightweight, Purpose-driven Data Pipeline for Large Language Models

Talent search is a cornerstone of modern recruitment systems, yet existing approaches often struggle to capture nuanced job-specific preferences, model recruiter behavior at a fine-grained level, and mitigate noise from subjective human judgments. We present a novel framework that enhances talent search effectiveness and delivers substantial business value through two key innovations: (i) leveraging LLMs to extract fine-grained recruitment signals from job descriptions and historical hiring data, and (ii) employing a role-aware multi-gate MoE network to capture behavioral differences across recruiter roles. To further reduce noise, we introduce a multi-task learning module that jointly optimizes click-through rate (CTR), conversion rate (CVR), and resume matching relevance. Experiments on real-world recruitment data and online A/B testing show relative AUC gains of 1.70% (CTR) and 5.97% (CVR), and a 17.29% lift in click-through conversion rate. These improvements reduce dependence on external sourcing channels, enabling an estimated annual cost saving of millions of CNY.

Enhancing Talent Search Ranking with Role-Aware Expert Mixtures and LLM-based Fine-Grained Job Descriptions

The reliability of Multimodal Large Language Models (MLLMs) in real-world settings is often undermined by sensitivity to irrelevant or distracting visual context, an aspect not captured by existing evaluation metrics. We introduce the Patch Context Robustness Index (PCRI), the first systematic and interpretable score for quantifying MLLM robustness to variations in visual context granularity, measuring performance changes between localized image patches and full-image input.


Applying PCRI to 19 state-of-the-art MLLMs across 15 vision-language benchmarks, we find that most leading models remain brittle to background noise, with only a few, such as InternVL2-26B and Qwen2VL-72B, demonstrating consistent robustness across tasks. PCRI analysis also highlights how different model architectures handle and integrate visual context, offering actionable diagnostic insight for both researchers and practitioners.


PCRI enables rigorous comparison of context robustness, supporting principled model selection and guiding the development of future architectures and training strategies for robust, real-world deployment.

PCRI: Measuring Context Robustness in Multimodal Models for Enterprise Applications

We present a novel approach to conversational agent evaluation using Persona-driven User Simulations based on Large Language Models (LLMs). Our methodology first uses LLMs to generate diverse customer personas, which are then used to configure a single LLM-based user simulator. This simulator evaluates SalesBot 2.0, a proactive conversational sales agent. We introduce a dataset of these personas, along with corresponding goals and conversation scenarios, enabling comprehensive testing across different customer types with varying assertiveness levels and precision of needs. Our evaluation framework assesses both the simulator's adherence to persona instructions and the bot's performance across multiple dimensions, combining human annotation with LLM-as-a-judge assessments using commercial and open-source models. Results demonstrate that our LLM-based simulator effectively emulates nuanced customer roles, and that cross-selling strategies can be implemented with minimal impact on customer satisfaction, varying by customer type.

Evaluating Conversational Agents with Persona-driven User Simulations based on Large Language Models: A Sales Bot Case Study

Online shoppers often initiate their journey with only a vague idea of what they need, forcing them to iterate over search results until they eventually discover a suitable product. We formulate this scenario as product demand clarification: starting from an ambiguous query, an agent must iteratively ask clarifying questions, progressively refine the user's intent, and retrieve increasingly relevant items. To tackle this challenge, we present **ProductAgent**, a fully autonomous conversational information-seeking agent that couples large language models with a set of domain-specific tools. ProductAgent maintains a structured memory of the dialogue, summarizes candidate products into concise feature statistics, generates strategic clarification questions, and performs retrieval over hybrid (symbolic + dense) indices in a closed decision loop. To measure real–world effectiveness, we further introduce **PROCLARE**, a PROduct CLArifying REtrieval benchmark that pairs ProductAgent with an LLM-driven user simulator, thereby enabling large-scale and reproducible evaluation without human annotation. On 2,000 automatically generated sessions, retrieval metrics improve monotonically with the number of turns, validating that ProductAgent captures and refines user intent through dialogue.

ProductAgent: Benchmarking Conversational Product Search Agent with Asking Clarification Questions

Chinese Spelling Correction (CSC) aims to detect and correct spelling errors in given sentences. 
Recently, multi-domain CSC has gradually attracted the attention of researchers because it is more practicable.
In this paper, we focus on the key flaw of the CSC model when adapting to multi-domain scenarios: the tendency to forget previously acquired knowledge upon learning new domain-specific knowledge (i.e., **catastrophic forgetting**).
To address this, we propose a novel model-agnostic **M**ulti-stage **K**nowledge **T**ransfer (**MKT**) framework with an evolving teacher model and dynamic distillation weights for knowledge transfer in each domain, rather than focusing solely on new domain knowledge.
It deserves to be mentioned that we are the first to apply continual learning methods to the multi-domain CSC task. 
Experiments. prove our method's effectiveness over traditional approaches, highlighting the importance of overcoming catastrophic forgetting to enhance model performance.

MKT: A Multi-Stage Knowledge Transfer Framework to Mitigate Catastrophic Forgetting in Multi-Domain Chinese Spelling Correction

This paper presents Ryt AI, an LLM-native agentic framework that powers Ryt Bank to enable customers to execute core financial transactions through natural language conversation. This represents the first global regulator-approved deployment worldwide where conversational AI functions as the primary banking interface, in contrast to prior assistants that have been limited to advisory or support roles. Built entirely in-house, Ryt AI is powered by ILMU, a closed-source LLM developed internally, and replaces rigid multi-screen workflows with a single dialogue orchestrated by four LLM-powered agents (Guardrails, Intent, Payment, and FAQ). Each agent attaches a task-specific LoRA adapter to ILMU, which is hosted within the bank's infrastructure to ensure consistent behavior with minimal overhead. Deterministic guardrails, human-in-the-loop confirmation, and a stateless audit architecture provide defense-in-depth for security and compliance. The result is Banking Done Right: demonstrating that regulator-approved natural-language interfaces can reliably support core financial operations under strict governance.

Banking Done Right: Redefining Retail Banking with Language-Centric AI

Fine-tuning large language models (LLMs) with backpropagation--even for a subset of parameters such as LoRA--can be much more memory-consuming than inference and is often deemed impractical for resource-constrained mobile devices. Alternative methods, such as zeroth-order optimization (ZO), can greatly reduce the memory footprint but come at the cost of significantly slower model convergence (10× to 100× more steps than backpropagation). We propose a memory-efficient implementation of backpropagation (MeBP) on mobile devices that allows flexible trade-offs between memory usage and compute time, while converging faster and achieving better performance than the ZO baseline. We verify the effectiveness of MeBP on an iPhone 15 Pro Max and show that various LLMs, ranging from 0.5B to 4B parameters, can be fine-tuned using less than 1GB of memory.

Downloads

Next from EMNLP 2025

Audio Query Handling System with Integrated Expert Models and Contextual Understanding

Similar lecture

From Representational Harms to Quality-of-Service Harms: A Case Study on Llama 2 Safety Safeguards

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 EMNLP 2025

Audio Query Handling System with Integrated Expert Models and Contextual Understanding

Similar lecture

From Representational Harms to Quality-of-Service Harms: A Case Study on Llama 2 Safety Safeguards

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads