China

This paper explores the use of large language models (LLMs) for annotating document utility in training retrieval and retrieval-augmented generation (RAG) systems, aiming to reduce dependence on costly human annotations. We address the gap between retrieval relevance and generative utility by employing LLMs to annotate document utility. To effectively utilize multiple positive samples per query, we introduce a novel loss that maximizes their summed marginal likelihood. Using the Qwen-2.5-32B model, we annotate utility on the MS MARCO dataset and conduct retrieval experiments on MS MARCO and BEIR, as well as RAG experiments on MS MARCO QA, NQ, and HotpotQA. Our results show that LLM-generated annotations enhance out-of-domain retrieval performance and improve RAG outcomes compared to models trained solely on human annotations or downstream QA metrics. Furthermore, combining LLM annotations with just 20% of human labels achieves performance comparable to using full human annotations. Our study offers a comprehensive approach to utilizing LLM annotations for initializing QA systems on new corpora.

EMNLP 2025

Utility-Focused LLM Annotation for Retrieval and Retrieval-Augmented Generation

utility-focused llm annotation

retrieval-augmented generation

contrastive learning

curriculum learning

## 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."<br>

*Message from the General Chair, Dirk Hovy*





<br>

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

poster

Local citation recommendation (LCR) suggests a set of papers for a citation placeholder within a given context. This paper introduces CiteBART, citation-specific pre-training within an encoder-decoder architecture, where author-date citation tokens are masked to learn to reconstruct them to fulfill LCR. The global version (CiteBART-Global) extends the local context with the citing paper's title and abstract to enrich the learning signal. CiteBART-Global achieves state-of-the-art performance on LCR benchmarks except for the FullTextPeerRead dataset, which is quite small to see the advantage of generative pre-training. The effect is significant in the larger benchmarks, e.g., Refseer and ArXiv., with the Refseer pre-trained model emerging as the best-performing model. We perform comprehensive experiments, including an ablation study, a qualitative analysis, and a taxonomy of hallucinations with detailed statistics. Our analyses confirm that CiteBART-Global has a cross-dataset generalization capability; the macro hallucination rate (MaHR) at the top-3 predictions is 4%, and when the ground-truth is in the top-k prediction list, the hallucination tendency in the other predictions drops significantly. We publicly share our code to support reproducibility.

CiteBART: Learning to Generate Citations for Local Citation Recommendation

Vision Language Models (VLMs) have achieved remarkable success in a wide range of vision applications of increasing complexity and scales, yet choosing the right VLM model size involves a trade-off between response quality and cost. While smaller VLMs are cheaper to run, they typically produce responses only marginally better than random guessing on benchmarks such as MMMU. In this paper, we propose Cache of Thought (CoT), a master–apprentice framework for collaborative inference between large and small VLMs. CoT manages high-quality query results from large VLMs (master) in a cache, which are then selected via a novel multi-modal retrieval and in-context learning to aid the performance of small VLMs (apprentice). We extensively evaluate CoT on various widely-recognized and challenging general VQA benchmarks, and show that CoT increases overall VQA performance by up to 7.7% under the same budget, and specifically boosts the performance of apprentice VLMs by up to 36.6%.

Cache-of-Thought: Master-Apprentice Framework for Cost-Effective Vision Language Model Reasoning

Video large language models (VideoLLM) excel at video understanding, but face efficiency challenges due to the quadratic complexity of abundant visual tokens. Our systematic analysis of token compression methods for VideoLLMs reveals two critical issues: \textbf{(i)} overlooking distinctive visual signals across frames, leading to information loss; \textbf{(ii)} suffering from implementation constraints, causing incompatibility with modern architectures or efficient operators. To address these challenges, we distill three design principles for VideoLLM token compression and propose a plug-and-play inference acceleration framework ``\textbf{Vid}eo \textbf{Com}pression \textbf{Com}mander'' (\textbf{VidCom²}). By quantifying each frame’s uniqueness, VidCom² adaptively adjusts compression intensity across frames, effectively preserving essential information while reducing redundancy in video sequences. Extensive experiments across various VideoLLMs and benchmarks demonstrate the superior performance and efficiency of our VidCom². With only \textbf{25\%} visual tokens, VidCom² achieves \textbf{99.6\%} of the original performance on LLaVA-OV while reducing \textbf{70.8\%} of the LLM generation latency. Notably, our Frame Compression Adjustment strategy is compatible with other token compression methods to further improve their performance. \emph{Codes are available in the supplementary materials and will be released on GitHub. }

Video Compression Commander: Plug-and-Play Inference Acceleration for Video Large Language Models

Ensuring AI safety is crucial as large language models become increasingly integrated into real-world applications. A key challenge is jailbreak, where adversarial prompts bypass built-in safeguards to elicit harmful disallowed outputs. Inspired by psychological foot-in-the-door principles, we introduce FITD, a novel multi-turn jailbreak method that leverages the phenomenon where minor initial commitments lower resistance to more significant or more unethical transgressions. Our approach progressively escalates the malicious intent of user queries through intermediate bridge prompts and aligns the model's response by itself to induce toxic responses. Extensive experimental results on two jailbreak benchmarks demonstrate that FITD achieves an average attack success rate of 94% across seven widely used models, outperforming existing state-of-the-art methods. Additionally, we provide an in-depth analysis of LLM self-corruption, highlighting vulnerabilities in current alignment strategies and emphasizing the risks inherent in multi-turn interactions.

Foot-In-The-Door: A Multi-turn Jailbreak for LLMs

The efficacy of text embedding models in representing and retrieving information is crucial for many NLP applications, with performance significantly advanced by Large Language Models (LLMs). Despite this progress, existing benchmarks predominantly use general-purpose datasets, inadequately addressing the nuanced requirements of specialized domains like finance. To bridge this gap, we introduce the Finance Massive Text Embedding Benchmark (FinMTEB), a comprehensive evaluation suite specifically designed for the financial domain. FinMTEB encompasses 64 datasets across 7 task types, including classification, clustering, retrieval, pair classification, reranking, summarization, and semantic textual similarity (STS) in English and Chinese. Alongside this benchmark, we introduce Fin-E5, a state-of-the-art finance-adapted embedding model, ranking first on FinMTEB. Fin-E5 is developed by fine-tuning e5-Mistral-7B-Instruct on a novel persona-based synthetic dataset tailored for diverse financial embedding tasks. Evaluating 15 prominent embedding models on FinMTEB, we derive three key findings: (1) domain-specific models, including our Fin-E5, significantly outperform general-purpose models; (2) performance on general benchmarks is a poor predictor of success on financial tasks; and (3) surprisingly, traditional Bag-of-Words (BoW) models surpass dense embedding models on financial STS tasks. This work provides a robust benchmark for financial NLP and offers actionable insights for developing future domain-adapted embedding solutions. Both FinMTEB and Fin-E5 will be open-sourced for the research community.

FinMTEB: Finance Massive Text Embedding Benchmark

Recent advances in large language models have demonstrated remarkable performance on Contextual Question Answering (CQA). However, prior approaches typically employ elaborate reasoning strategies regardless of question complexity, leading to low adaptability. Recent efficient test-time scaling methods introduce budget constraints or early stop mechanisms to avoid overthinking for straightforward questions. But they add human bias to the reasoning process and fail to leverage models' inherent reasoning capabilities. To address these limitations, we present T²: Think-to-Think, a novel framework that dynamically adapts reasoning depth based on question complexity. T² leverages the insight that if an LLM can effectively solve similar questions using specific reasoning strategies, it can apply the same strategy to the original question. This insight enables to adoption of concise reasoning for straightforward questions while maintaining detailed analysis for complex problems. T² works through four key steps: decomposing questions into structural elements, generating similar examples with candidate reasoning strategies, evaluating these strategies against multiple criteria, and applying the most appropriate strategy to the original question. Experimental evaluation across seven diverse CQA benchmarks demonstrates that T² not only achieves higher accuracy than baseline methods but also reduces computational overhead by up to 25.2%.

T²: An Adaptive Test-Time Scaling Strategy for Contextual Question Answering

3D visual grounding (3DVG) is challenging due to the need to understand 3D spatial relations. While supervised approaches have achieved superior performance, they are constrained by the scarcity and high annotation costs of 3D vision-language datasets. Training-free approaches based on LLMs/VLMs eliminate the need for large-scale training data, but they either incur prohibitive grounding time and token costs or have unsatisfactory accuracy. To address the challenges, we introduce a novel method for training-free 3D visual grounding, namely **La**nguage-to-**S**pace **P**rogramming (LaSP). LaSP introduces LLM-generated codes to analyze 3D spatial relations among objects, along with a pipeline that evaluates and optimizes the codes automatically. Experimental results demonstrate that LaSP achieves 52.9% accuracy on the Nr3D benchmark, ranking among the best training-free methods. Moreover, it substantially reduces the grounding time and token costs, offering a balanced trade-off between performance and efficiency.

Language-to-Space Programming for Training-Free 3D Visual Grounding

Prompting-based conversational query reformulation has emerged as a powerful approach for conversational search, refining ambiguous user queries into standalone search queries. Best-of-N reformulation over the generated candidates via prompting shows impressive potential scaling capability. However, both the previous tuning methods (training time) and adaptation approaches (test time) can not fully unleash their benefits. In this paper, we propose AdaRewriter, a novel framework for query reformulation using an outcome-supervised reward model via test-time adaptation. By training a lightweight reward model with contrastive ranking loss, AdaRewriter selects the most promising reformulation during inference. Notably, it can operate effectively in black-box systems, including commercial LLM APIs. Experiments on five conversational search datasets show that AdaRewriter significantly outperforms the existing methods across most settings, demonstrating the potential of test-time adaptation for conversational query reformulation.

AdaRewriter: Unleashing the Power of Prompting-based Conversational Query Reformulation via Test-Time Adaptation

Multi-hop question answering (QA) remains challenging, as solutions must reliably integrate and reconcile evidence from multiple sources without succumbing to error propagation. While large language models (LLMs) have achieved substantial improvements via chain-of-thought (CoT) prompting and retrieval-augmented generation, these methods typically adopt a forward-only workflow—early mistakes persist throughout inference, and contradictions discovered later cannot systematically trigger re-evaluation. To address this limitation, we present ReAgent, a reversible multi-agent reasoning framework. Specifically, ReAgent enables agents to backtrack to earlier valid states when conflicts arise, thereby isolating and rectifying flawed assumptions before they undermine subsequent reasoning. Our approach combines explicit local and global rollback protocols with modular role specialization, resulting in a flexible and error-tolerant pipeline. Empirical evaluation on three multi-hop QA benchmarks demonstrates consistent performance gains of approximately 6% over forward-only baselines, in addition to enhanced interpretability. These findings highlight the value of non-monotonic, backtracking-driven inference in complex QA scenarios and point to broader implications for multi-agent collaboration in knowledge-intensive tasks.

ReAgent: Reversible Multi-Agent Reasoning for Knowledge-Enhanced Multi-Hop QA

Multimodal event extraction task aims to identify event types and arguments from visual and textual representations related to events. Due to the high cost of multimedia training data, previous methods mainly focused on weakly alignment of excellent unimodal encoders. However, they ignore the conflict between event understanding and image recognition, resulting in redundant feature perception affecting the understanding of multimodal events. In this paper, we propose a multimodal event extraction strategy with a multi-level redundant feature selection mechanism, which enhances the event understanding ability of multimodal large language models by leveraging knowledge editing techniques, and requires no additional parameter optimization work. Extensive experiments show that our method outperforms the state-of-the-art (SOTA) baselines on the M2E2 benchmark. Compared with the highest baseline, we achieve a 34% improvement of precision on event extraction and a 11% improvement of F1 on argument extraction.

Downloads

Next from EMNLP 2025

CiteBART: Learning to Generate Citations for Local Citation Recommendation

Similar lecture

MS-RAG: Simple and Effective Multi-Semantic Retrieval-Augmented Generation