ACL 2025 Papers

While large language models have demonstrated exceptional performance across a wide range of tasks, they remain susceptible to hallucinations – generating plausible yet factually incorrect contents. Existing methods to mitigating such risk often rely on sampling multiple full-length generations, which introduces significant response latency and becomes ineffective when the model consistently produces hallucinated outputs with high confidence. To address these limitations, we introduce Monitoring Decoding (MD), a novel framework that dynamically monitors the generation process and selectively applies in-process interventions, focusing on revising crucial tokens responsible for hallucinations. Instead of waiting until completion of multiple full-length generations, we identify hallucination-prone tokens during generation using a monitor function, and further refine these tokens through a tree-based decoding strategy. This approach ensures an enhanced factual accuracy and coherence in the generated output while maintaining efficiency. Experimental results demonstrate that MD consistently outperforms self-consistency-based approaches in both effectiveness and efficiency, achieving higher factual accuracy while significantly reducing computational overhead.

Recent LLMs exhibit limited effectiveness on molecular property prediction task due to the semantic gap between molecular representations and natural language, as well as the lack of domain-specific knowledge. To address these challenges, we propose MolRAG, a Retrieval-Augmented Generation framework integrating Chain-of-Thought reasoning for molecular property prediction. MolRAG operates by retrieving structurally analogous molecules as contextual references to guide stepwise knowledge reasoning through chemical structure-property relationships. This dual mechanism synergizes molecular similarity analysis with structured inference, while generating human-interpretable rationales grounded in domain knowledge. Experimental results show MolRAG outperforms pre-trained LLMs on four datasets, and even matches supervised methods, achieving performance gains of 1.1%–45.7% over direct prediction approaches, demonstrating versatile effectiveness. Our code is available at https://github.com/AcaciaSin/MolRAG.

Multilingual sentence encoders (MSEs) are commonly obtained by training multilingual language models to map sentences from different languages into a shared semantic space. As such, they are subject to curse of multilinguality, a loss of monolingual representational accuracy due to parameter sharing. Another limitation of MSEs is the trade-off between different task performance: cross-lingual alignment training distorts the optimal monolingual structure of semantic spaces of individual languages, harming the utility of sentence embeddings in monolingual tasks; cross-lingual tasks, such as cross-lingual semantic similarity and zero-shot transfer for sentence classification, may also require conflicting cross-lingual alignment strategies. In this work, we address both issues by means of modular training of sentence encoders. We first train language-specific monolingual modules to mitigate negative interference between languages (i.e., the curse). We then align all non-English sentence embeddings to the English by training cross-lingual alignment adapters, preventing interference with monolingual specialization from the first step. We train the cross-lingual adapters with two different types of data to resolve the conflicting requirements of different cross-lingual tasks. Monolingual and cross-lingual results on semantic text similarity and relatedness, bitext mining and sentence classification show that our modular solution achieves better and more balanced performance across all the tasks compared to full-parameter training of monolithic multilingual sentence encoders, especially benefiting low-resource languages.

English as a Second Language (ESL) speakers often struggle to engage in group discussions due to language barriers. While moderators can facilitate participation, few studies assess conversational engagement and evaluate moderation effectiveness. To address this gap, we develop a dataset comprising 17 sessions from an online ESL conversation club, which includes both moderated and non-moderated discussions. We then introduce an approach that integrates automatic ESL dialogue assessment and a framework that categorizes moderation strategies. Our findings indicate that moderators help improve the flow of topics and start/end a conversation. Interestingly, we find active acknowledgement and encouragement to be the most effective moderation strategy, while excessive information and opinion sharing by moderators has a negative impact. Ultimately, our study paves the way for analyzing ESL group discussions and the role of moderators in non-native conversation settings. Code and data are available at https://github.com/RenaGao/L2Moderator.

We analyze the evolution of English noun compounds, which we represent as vectors of time-specific values. We implement a wide array of methods to create a rich set of features, using them to classify compounds for present-day compositionality and to assess the informativeness of the corresponding linguistic patterns. Our best results use BERT – reflecting the similarity of compounds and sentence contexts – and we further capture relevant and complementary information across approaches. Leveraging these feature differences, we find that the development of low-compositional meanings is reflected by a parallel drop in compositionality and sustained semantic change. The same distinction is echoed in transformer processing: compositionality estimates require far less contextualization than semantic change estimates.

Composed Image Retrieval (CIR) enables users to search for images using multimodal queries that combine text and reference images. While metric learning methods have shown promise, they rely on deterministic point embeddings that fail to capture the inherent uncertainty in the input data, in which user intentions may be imprecisely specified or open to multiple interpretations. We address this challenge by reformulating CIR through our proposed Composed Probabilistic Embedding (CoPE) framework, which represents both queries and targets as Gaussian distributions in latent space rather than fixed points. Through careful design of probabilistic distance metrics and hierarchical learning objectives, CoPE explicitly captures uncertainty at both instance and feature levels, enabling more flexible, nuanced, and robust matching that can handle polysemy and ambiguity in search intentions. Extensive experiments across multiple benchmarks demonstrate that CoPE effectively quantifies both quality and semantic uncertainties within Composed Image Retrieval, achieving state-of-the-art performance on recall rate. Code: https://github.com/tanghme0w/ACL25-CoPE.

Modeling relationships between concepts and entities is essential for many applications. While Large Language Models (LLMs) capture relational and commonsense knowledge effectively, they are computationally expensive and often underperform in tasks requiring efficient relational encoding, such as relation induction, extraction, and information retrieval. Despite advancements in learning relational embeddings, existing methods often fail to capture nuanced representations and the rich semantics needed for high-quality embeddings. In this work, we propose different relational encoders designed to capture diverse relational aspects and semantic properties of entity pairs. Although several datasets exist for training such encoders, they often rely on structured knowledge bases or predefined schemas, which primarily encode simple and static relations. To overcome this limitation, we also introduce a novel dataset generation method leveraging LLMs to create a diverse spectrum of relationships. Our experiments demonstrate the effectiveness of our proposed encoders and the benefits of our generated dataset.

Content moderation systems powered by large language models (LLMs) are increasingly deployed to detect hate speech; however, no systematic comparison exists between different systems. If different systems produce different outcomes for the same content, it undermines consistency and predictability, leading to moderation decisions that appear arbitrary or unfair. Analyzing seven leading models—dedicated Moderation Endpoints (OpenAI, Mistral), frontier LLMs (Claude 3.5 Sonnet, GPT-4o, Mistral Large, DeepSeek V3), and specialized content moderation APIs (Google Perspective API)—we demonstrate that moderation system choice fundamentally determines hate speech classification outcomes. Using a novel synthetic dataset of 1.3+ million sentences from a factorial design, we find identical content receives markedly different classification values across systems, with variations especially pronounced for specific demographic groups. Analysis across 125 distinct groups reveals these divergences reflect systematic differences in how models establish decision boundaries around harmful content, highlighting significant implications for automated content moderation.

Optimizing Large Language Model (LLM) performance requires well-crafted prompts, but manual prompt engineering is labor-intensive and often ineffective. Automated prompt optimization techniques address this challenge but the major of them rely on randomly selected evaluation subsets, which fail to represent the full dataset, leading to unreliable evaluations and suboptimal prompts. Existing coreset selection methods, designed for LLM benchmarking, are unsuitable for prompt optimization due to challenges in clustering similar samples, high data collection costs, and the unavailability of performance data for new or private datasets. To overcome these issues, we propose IPOMP, an Iterative evaluation data selection approach for effective Prompt Optimization using real time Model Performance. IPOMP is a two-stage approach that selects representative and diverse samples using semantic clustering and boundary analysis, followed by iterative refinement with real-time model performance data to replace redundant samples. Evaluations on two datasets BIG-bench and LIAR, and two models GPT-3.5 and GPT-4o-mini, show that IPOMP improves effectiveness by at least 1.6% to 3.1%, and stability by at least 50% to 55.5% compared with the best baseline across the studied datasets and models, with minimal computational overhead below 1%. Furthermore, the results demonstrate that our real-time performance-guided refinement approach can be universally applied to enhance existing coreset selection methods.

Given the high computational cost of preference alignment training of large language models (LLMs), exploring efficient methods to reduce the training overhead remains an important and compelling research problem. Motivated by the observation that alignment training typically involves only small parameter changes without injecting new knowledge into models, we propose a straightforward method called ExPO (model extrapolation) to expedite LLMs’ alignment with human preferences. Given a partially-trained model and its initial SFT checkpoint, ExPO improves the implicit optimization objective of alignment training by simply amplifying the parameter change based on a first-order approximation, without any additional training overhead. Through controlled experiments, we demonstrate that ExPO boosts a DPO model trained with only 20% steps to outperform the fully-trained one. Moreover, we show that ExPO notably improves existing open-source LLMs (ranging from 1.8B to 70B parameters) on the leading AlpacaEval 2.0 and MT-Bench benchmarks, which highlights ExPO’s broader utility in efficiently enhancing LLM alignment.

Generative models such as Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) trained on massive datasets can lead them to memorize and inadvertently reveal sensitive information, raising ethical and privacy concerns. While some prior works have explored this issue in the context of LLMs, it presents a unique challenge for MLLMs due to the entangled nature of knowledge across modalities, making comprehensive unlearning more difficult. To address this challenge, we propose Modality Aware Neuron Unlearning (MANU), a novel unlearning framework for MLLMs designed to selectively clip neurons based on their relative importance to the targeted forget data, curated for different modalities. Specifically, MANU consists of two stages: important neuron selection and selective pruning. The first stage identifies and collects the most influential neurons across modalities relative to the targeted forget knowledge, while the second stage is dedicated to pruning those selected neurons. MANU effectively isolates and removes the neurons that contribute most to the forget data within each modality, while preserving the integrity of retained knowledge. Our experiments conducted across various MLLM architectures illustrate that MANU can achieve a more balanced and comprehensive unlearning in each modality without largely affecting the overall model utility.

A modal dependency structure represents a web of connections between events and sources of information in a document that allows for tracing of who-said-what with what levels of certainty, thereby establishing factuality in an event-centric approach. Obtaining such graphs defines the task of modal dependency parsing, which involves event and source identification along with the modal relations between them. In this paper, we propose a simple yet effective solution based on biaffine attention that specifically optimizes against the domain-specific challenges of modal dependency parsing by integrating self-loop. We show that our approach, when coupled with data augmentation by leveraging the Large Language Models to translate annotations from one language to another, outperforms the previous state-of-the-art on English and Chinese datasets by 2% and 4% respectively.

In simultaneous interpreting, an interpreter renders the speech into another language with a very short lag, much sooner than sentences are finished. In order to understand and later reproduce this dynamic and complex task automatically, we need specialized datasets and tools for analysis, monitoring, and evaluation, such as parallel speech corpora, and tools for their automatic annotation. Existing parallel corpora of translated texts and associated alignment algorithms hardly fill this gap, as they fail to model long-range interactions between speech segments or specific types of divergences (e.g. shortening, simplification, functional generalization) between the original and interpreted speeches. In this work, we develop and explore MockConf, a student interpretation dataset that was collected from Mock Conferences run as part of the students’ curriculum. This dataset contains 7 hours of recordings in 5 European languages, transcribed and aligned at the level of spans and words. We further implement and release InterAlign, a modern web-based annotation tool for parallel word and span annotations on long inputs, suitable for aligning simultaneous interpreting. We propose metrics for the evaluation and a baseline for automatic alignment. Dataset and tools will be released to the community.

Deploying large language models (LLMs) with low-rank adaptation (LoRA) on mobile devices is promising due to their capability to complete diverse domain-specific tasks while ensuring privacy and accessibility. In this paper, we introduce MobiLoRA to accelerate LoRA-based LLM inference on mobile devices. MobiLoRA focuses on optimizing the key-value (KV) caches due to the limited computing and memory resources of mobile devices. The key insight of MobiLoRA lies in the utilization of two contexts for on-device LoRA serving: semantic-level contexts, such as prompts with shared prefixes, and system-level contexts, such as the application status (e.g., foreground or killed) of LLM requests. Specifically, for semantic-level contexts, MobiLoRA proposes similarity-aware delta encoding, which leverages token-wise similarity in KV caches across LoRA adapters for efficient storage and reuse. Furthermore, MobiLoRA advocates context-aware KV cache management to optimize cache retention and eviction considering the system-level contexts. We fully implement MobiLoRA and compare it with state-of-the-art LLM serving frameworks using real-world mobile device traces. Results show that MobiLoRA accelerates LoRA-based LLM inference by 57.6% on mobile devices.

With the rapid development of Large Language Models (LLMs), Parameter-Efficient Fine-Tuning (PEFT) methods have gained significant attention, which aims to achieve efficient fine-tuning of LLMs with fewer parameters. As a representative PEFT method, Low-Rank Adaptation (LoRA) introduces low-rank matrices to approximate the incremental tuning parameters and achieves impressive performance over multiple scenarios. After that, plenty of improvements have been proposed for further improvement. However, these methods either focus on single-task scenarios or separately train multiple LoRA modules for multi-task scenarios, limiting the efficiency and effectiveness of LoRA in multi-task scenarios. To better adapt to multi-task fine-tuning, in this paper, we propose a novel Mixture of Low-Rank Experts (MoRE) for multi-task PEFT. Specifically, instead of using an individual LoRA for each task, we align different ranks of LoRA module with different tasks, which we named low-rank experts. Moreover, we design a novel adaptive rank selector to select the appropriate expert for each task. By jointly training low-rank experts, MoRE can enhance the adaptability and efficiency of LoRA in multi-task scenarios. Finally, we conduct extensive experiments over multiple multi-task benchmarks along with different LLMs to verify model performance. Experimental results demonstrate that compared to traditional LoRA and its variants, MoRE significantly improves the performance of LLMs in multi-task scenarios and incurs no additional inference cost. We also release the model and code to facilitate the community.

One of the primary challenges in optimizing large language models (LLMs) for long-context inference lies in the high memory consumption of the Key-Value (KV) cache. Existing approaches, such as quantization, have demonstrated promising results in reducing memory usage. However, current quantization methods cannot take both effectiveness and efficiency into account. In this paper, we propose MoQAE, a novel mixed-precision quantization method via mixture of quantization-aware experts. First, we view different quantization bit-width configurations as experts and use the traditional mixture of experts (MoE) method to select the optimal configuration. To avoid the inefficiency caused by inputting tokens one by one into the router in the traditional MoE method, we input the tokens into the router chunk by chunk. Second, we design a lightweight router-only fine-tuning process to train MoQAE with a comprehensive loss to learn the trade-off between model accuracy and memory usage. Finally, we introduce a routing freezing (RF) and a routing sharing (RS) mechanism to further reduce the inference overhead. Extensive experiments on multiple benchmark datasets demonstrate that our method outperforms state-of-the-art KV cache quantization approaches in both efficiency and effectiveness.

Retrieval-Augmented Generation (RAG), while serving as a viable complement to large language models (LLMs), often overlooks the crucial aspect of text chunking within its pipeline. This paper initially introduces a dual-metric evaluation method, comprising Boundary Clarity and Chunk Stickiness, to enable the direct quantification of chunking quality. Leveraging this assessment method, we highlight the inherent limitations of traditional and semantic chunking in handling complex contextual nuances, thereby substantiating the necessity of integrating LLMs into chunking process. To address the inherent trade-off between computational efficiency and chunking precision in LLM-based approaches, we devise the granularity-aware Mixture-of-Chunkers (MoC) framework, which consists of a three-stage processing mechanism. Notably, our objective is to guide the chunker towards generating a structured list of chunking regular expressions, which are subsequently employed to extract chunks from the original text. Extensive experiments demonstrate that both our proposed metrics and the MoC framework effectively settle challenges of the chunking task, revealing the chunking kernel while enhancing the performance of the RAG system.

The tendency of Large Language Models (LLMs) to generate hallucinations raises concerns regarding their reliability. Therefore, confidence estimations indicating the extent of trustworthiness of the generations become essential. However, current LLM confidence estimations in languages other than English remain underexplored. This paper addresses this gap by introducing a comprehensive investigation of Multilingual Confidence estimation (MlingConf) on LLMs, focusing on both language-agnostic (LA) and language-specific (LS) tasks to explore the performance and language dominance effects of multilingual confidence estimations on different tasks. The benchmark comprises four meticulously checked and human-evaluated high-quality multilingual datasets for LA tasks and one for the LS task tailored to specific social, cultural, and geographical contexts of a language. Our experiments reveal that on LA tasks English exhibits notable linguistic dominance in confidence estimations than other languages, while on LS tasks, using question-related language to prompt LLMs demonstrates better linguistic dominance in multilingual confidence estimations. The phenomena inspire a simple yet effective native-tone prompting strategy by employing language-specific prompts for LS tasks, effectively improving LLMs’ reliability and accuracy in LS scenarios.

In-context learning (ICL) adapts LLMs by providing demonstrations without fine-tuning the model parameters; however, it is very sensitive to the choice of in-context demonstrations, and processing many demonstrations can be computationally demanding. We propose Mixtures of In-Context Learners (MoICL), a novel approach that uses subsets of demonstrations to train a set of experts via ICL and learns a weighting function to merge their output distributions via gradient-based optimisation. In our experiments, we show performance improvements on 5 out of 7 classification datasets compared to a set of strong baselines (e.g., up to +13% compared to ICL and LENS). Moreover, we improve the Pareto frontier of ICL by reducing the inference time needed to achieve the same performance with fewer demonstrations. Finally, MoICL is more robust to out-of-domain (up to +11%), imbalanced (up to +49%) and perturbed demonstrations (up to +38%).