China

In this paper, we introduce a novel weighted co-training approach that is guided by Large Language Models (LLMs). Namely, in our co-training approach, we use LLM labels on unlabeled data as target labels and co-train two encoder-only based networks that train each other over multiple iterations: first, all samples are forwarded through each network and historical estimates of each network&#39;s confidence in the LLM label are recorded; second, a dynamic importance weight is derived for each sample according to each network&#39;s belief (or confidence) in the quality of the LLM label for that sample; finally, the two networks exchange importance weights with each other—each network back-propagates all samples weighted with the importance weights coming from its peer network and updates its own parameters. By strategically utilizing LLM-generated guidance, our approach significantly outperforms conventional SSL methods, particularly in settings with abundant unlabeled data. Empirical results show that it achieves state-of-the-art performance on 4 out of 5 benchmark datasets and ranks first among 14 compared methods according to the Friedman test. Our results highlight a new direction in semi-supervised learning—where LLMs serve as knowledge amplifiers, enabling backbone co-training models to achieve SOTA performance efficiently.

EMNLP 2025

LLM-Guided Co-Training for Text Classification

learning from noisy labels

semi supervised learning

large language models

text classification

### 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.

In this paper, we introduce a novel weighted co-training approach that is guided by Large Language Models (LLMs). Namely, in our co-training approach, we use LLM labels on unlabeled data as target labels and co-train two encoder-only based networks that train each other over multiple iterations: first, all samples are forwarded through each network and historical estimates of each network's confidence in the LLM label are recorded; second, a dynamic importance weight is derived for each sample according to each network's belief (or confidence) in the quality of the LLM label for that sample; finally, the two networks exchange importance weights with each other—each network back-propagates all samples weighted with the importance weights coming from its peer network and updates its own parameters. By strategically utilizing LLM-generated guidance, our approach significantly outperforms conventional SSL methods, particularly in settings with abundant unlabeled data. Empirical results show that it achieves state-of-the-art performance on 4 out of 5 benchmark datasets and ranks first among 14 compared methods according to the Friedman test. Our results highlight a new direction in semi-supervised learning—where LLMs serve as knowledge amplifiers, enabling backbone co-training models to achieve SOTA performance efficiently.

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.

Recently, researchers have turned to synthetic tasks for evaluation of large language models' long-context capabilities, as they offer more flexibility than realistic benchmarks in scaling both input length and dataset size. However, existing synthetic tasks typically target narrow skill sets such as retrieving information from massive input, limiting their ability to comprehensively assess model capabilities. Furthermore, existing benchmarks often pair each task with a different input context, creating confounding factors that prevent fair cross-task comparison. To address these limitations, we introduce SYNC, a new evaluation suite of synthetic tasks spanning domains including graph understanding and translation. Each domain includes three tasks designed to test a wide range of capabilities---from retrieval, to multi-hop tracking, and to global context understanding that that requires chain-of-thought (CoT) reasoning. Crucially, all tasks share the same context, enabling controlled comparisons of model performance. We evaluate 14 LLMs on SYNC and observe substantial performance drops on more challenging tasks, underscoring the benchmark’s difficulty. Additional experiments highlight the necessity of CoT reasoning and demonstrate that \benchmark poses a robust challenge for future models.

SYNC: A Synthetic Long-Context Understanding Benchmark for Controlled Comparisons of Model Capabilities

Large language model (LLM) personalization aims to align model outputs with individuals' unique preferences and opinions. While recent efforts have implemented various personalization methods, a unified theoretical framework that can systematically understand the drivers of effective personalization is still lacking. In this work, we integrate the well-established cognitive dual-memory model into LLM personalization, by mirroring episodic memory to historical user engagements and semantic memory to long-term, evolving user beliefs. Specifically, we systematically investigate memory instantiations and introduce a unified framework, PRIME, using episodic and semantic memory mechanisms. We further augment PRIME with a novel personalized thinking capability inspired by the slow thinking strategy. Moreover, recognizing the absence of suitable benchmarks, we introduce CMV dataset specifically designed to evaluate long-context personalization. Extensive experiments validate PRIME's effectiveness across both long- and short-context scenarios. Further analysis confirms that PRIME effectively captures dynamic personalization beyond mere popularity biases.

PRIME: Large Language Model Personalization with Cognitive Dual-Memory and Personalized Thought Process

In this work, we introduce the first benchmark for evaluating the capabilities of large language models (LLMs) in understanding and generating responses in Tunisian Arabic. To achieve this, we construct a dataset of Tunisian Arabic instructions and prompt ten widely-used LLMs that claim to support Arabic. We then assess the LLM responses through both human and LLM-based evaluations across four criteria: quality, correctness, relevance, and dialectal adherence. We analyze the agreement and correlation between these judgments and identify GPT-4o as our automated judge model based on its high correlation with human ratings, and generate a final leaderboard using this model. Our error analysis reveals that most LLMs struggle with recognizing and properly responding in Tunisian Arabic. To facilitate further research, we release our dataset and evaluation framework, allowing others to benchmark their own models.

TounsiBench: Benchmarking Large Language Models for Tunisian Arabic

While theories of discourse and cognitive science have long recognized the value of unhurried pacing, recent dialogue research tends to minimize friction in conversational systems. Yet, frictionless dialogue risks fostering uncritical reliance on AI outputs, which can obscure implicit assumptions and lead to unintended consequences. To meet this challenge, we propose integrating positive friction into conversational AI, which promotes user reflection on goals, critical thinking on system response, and subsequent re-conditioning of AI systems. We hypothesize systems can improve goal alignment, modeling of user mental states, and task success by deliberately slowing down conversations in strategic moments to ask questions, reveal assumptions, or pause. We present an ontology of positive friction and collect expert human annotations on multi-domain and embodied goal-oriented corpora. Experiments on these corpora, along with simulated interactions using state-of-the-art systems, suggest incorporating friction not only fosters accountable decision-making, but also enhances machine understanding of user beliefs and goals, and increases task success rates.

Better Slow than Sorry: Introducing Positive Friction for Reliable Dialogue Systems

This work demonstrates that LLM-based web navigation agents offer powerful automation capabilities but are vulnerable to Indirect Prompt Injection (IPI) attacks. We show that adversaries can embed universal adversarial triggers in webpage HTML to hijack agent behavior that utilizes the accessibility tree to parse HTML, causing unintended or malicious actions. Using the Greedy Coordinate Gradient (GCG) algorithm and a Browser Gym agent powered by Llama-3.1, our system demonstrates high success rates across real websites in both targeted and general attacks, including login credential exfiltration and forced ad clicks. Our empirical results highlight critical security risks and the need for stronger defenses as LLM-driven autonomous web agents become more widely adopted. The system software (https://github.com/sej2020/manipulating-web-agents) is released under the MIT License, with an accompanying publicly available demo website (http://lethaiq.github.io/attack-web-llm-agent).

Manipulating LLM Web Agents with Indirect Prompt Injection Attack via HTML Accessibility Tree

Scientific research often requires constructing high-quality datasets, yet the current workflows remain labor-intensive, and dependent on domain expertise. Existing approaches automate isolated steps such as retrieval or generation, but lack support for the full end-to-end data collection process. We present Quest2DataAgent, a general-purpose multi-agent framework for automating scientific data collection workflows. Given a natural language research question, it decomposes tasks into structured subtasks, retrieves relevant data using hybrid strategies, evaluates dataset quality, and generates visualizations through a conversational interface. We demonstrate its flexibility in two domains: EcoData for ecological research and PolyData for polymer materials. Both systems share the same core architecture but operate over distinct datasets and user needs. Human evaluations show that Quest2DataAgent significantly improves data relevance, usability, and time efficiency compared to manual collection and tool-assisted baselines. The framework is open-source and extensible to other domains.

Quest2DataAgent: Automating End-to-End Scientific Data Collection

Text Normalization (TN) is a key preprocessing step in Text-to-Speech (TTS) systems, converting written forms into their canonical spoken equivalents. Traditional TN systems can exhibit high accuracy, but involve substantial engineering effort, are difficult to scale, and pose challenges to language coverage, particularly in low-resource settings. We propose PolyNorm, a prompt-based approach to TN using Large Language Models (LLMs), aiming to reduce the reliance on manually crafted rules and enable broader linguistic applicability with minimal human intervention. Additionally, we present a language-agnostic pipeline for automatic data curation and evaluation, designed to facilitate scalable experimentation across diverse languages. Experiments across eight languages show consistent reductions in the word error rate (WER) compared to a production-grade-based system. To support further research, we release PolyNorm-Benchmark, a multilingual data set covering a diverse range of text normalization phenomena.

PolyNorm: Few-Shot LLM-Based Text Normalization for Text-to-Speech

Large Language Models (LLMs) deployed as autonomous agents with tool access present unique safety challenges that extend beyond standalone model vulnerabilities. Existing red-teaming frameworks like AgentHarm use static prompts and hardcoded toolsets, limiting their applicability to custom production systems.


We introduce a dual-component automated red-teaming framework: AgentHarm-Gen generates adversarial tasks and evaluation functions tailored to arbitrary toolsets, while Red-Agent-Reflect employs iterative prompt refinement with self-reflection to develop progressively more effective attacks.


Evaluating across 115 harmful tasks (71 generated, 44 from AgentHarm) spanning 8 risk categories, our method achieves substantial improvements: up to 162\% increase in attack success rate on o4-mini and 86\% success on Gemini 2.5 Pro. Successful attacks systematically decompose adversarial objectives into benign-appearing sub-tasks that circumvent safety alignment, highlighting the need for agent-specific guardrails.


We contribute our implementation to the AgentHarm repository, enabling systematic identification of safety vulnerabilities in custom agentic workflows before deployment.

Agent vs. Agent: Automated Data Generation and Red-Teaming for Custom Agentic Workflows

The Enterprise Intelligence Platform must integrate logs from numerous third-party vendors in order to perform various downstream tasks. However, vendor documentation is often unavailable at test time. It is either misplaced, mismatched, poorly formatted, or incomplete, which makes schema mapping challenging. We introduce a reinforcement learning agent that can self-improve without labeled examples or model weight updates. During inference, the agent: 1) Identifies ambiguous field-mapping attempts. 2) Generates targeted web-search queries to gather external evidence. 3) Applies a confidence-based reward to iteratively refine its mappings. To demonstrate this concept, we converted Microsoft Defender for Endpoint logs into a common schema. Our method increased mapping accuracy from 72.73\% to 93.94\% over 100 iterations using GPT-4o. At the same time, it reduced the number of low-confidence mappings requiring expert review by 85\%. This new approach provides an evidence-driven, transparent method for solving future industry problems, paving the way for more robust, accountable, scalable, efficient, flexible, adaptable, and collaborative solutions.

Mapping Smarter, Not Harder: A Test-Time Reinforcement Learning Agent That Improve Without Labels or Model Updates

Abstractive summarization is a core application in contact centers, where Large Language Models (LLMs) generate millions of summaries of call transcripts daily. Despite their apparent quality, it remains unclear whether LLMs systematically under- or over-attend to specific aspects of the transcript, potentially introducing biases in the generated summary. While prior work has examined social and positional biases, the specific forms of bias pertinent to contact center operations—which we term *Operational Bias*—have remained unexplored. To address this gap, we introduce ***BlindSpot***, a framework built upon a taxonomy of 15 operational bias dimensions (e.g., disfluency, speaker, topic) for the identification and quantification of these biases. ***BlindSpot*** leverages an LLM as a zero-shot classifier to derive categorical distributions for each bias dimension in a pair of transcript and its summary. The bias is then quantified using two metrics: *Fidelity Gap* (the JS Divergence between distributions) and *Coverage* (the percentage of source labels omitted). Using ***BlindSpot***, we conducted an empirical study with 2500 real call transcripts and their summaries generated by 20 LLMs of varying scales and families (e.g., GPT, Llama, Claude). Our analysis reveals that biases are systemic and present across all evaluated models, regardless of size or family.

Downloads

Next from EMNLP 2025

SYNC: A Synthetic Long-Context Understanding Benchmark for Controlled Comparisons of Model Capabilities

Similar lecture

Co-training for Low Resource Scientific Natural Language Inference

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES