ACL 2025 Papers

This paper presents CHARPEVAL, a challenging benchmark specifically designed to evaluate the ability of Large Language Models (LLMs) to perform contextualized reasoning in knowledge-grounded dialogue scenarios. The task involves selecting the correct response from 6 options, including 5 manually crafted distractors, given the conversation history. Extensive benchmarking experiments with a diverse set of state-of-the-art open-weight LLMs show poor performance on CHARPEVAL due to their inability to effectively reason over discontinuous chunks of text across the input. Our analysis reveals systematic error patterns across models with different properties, highlighting the need to improve LLMs beyond simply scaling-up data and compute. CHARPEVAL is publicly available at https://huggingface.co/datasets/huawei-noah/CHARP.

The adeptness of Large Language Models (LLMs) in comprehending and following natural language instructions is critical for their deployment in sophisticated real-world applications. Existing evaluations mainly focus on fragmented constraints or narrow scenarios, but they overlook the comprehensiveness and authenticity of constraints from the user’s perspective. To bridge this gap, we propose CFBench, a large-scale Chinese Comprehensive Constraints Following Benchmark for LLMs, featuring 1,000 curated samples that cover more than 200 real-life scenarios and over 50 NLP tasks. CFBench meticulously compiles constraints from real-world instructions and constructs an innovative systematic framework for constraint types, which includes 10 primary categories and over 25 subcategories, and ensures each constraint is seamlessly integrated within the instructions. To make certain that the evaluation of LLM outputs aligns with user perceptions, we propose an advanced methodology that integrates multi-dimensional assessment criteria with requirement prioritization, covering various perspectives of constraints, instructions, and requirement fulfillment. Evaluating current leading LLMs on CFBench reveals substantial room for improvement in constraints following, and we further investigate influencing factors and enhancement strategies. The data and code will be made available.

Ensuring the reliability of Large Language Models (LLMs) in complex reasoning tasks remains a formidable challenge, particularly in scenarios that demand precise mathematical calculations and knowledge-intensive open-domain generation. In this work, we introduce an uncertainty-aware framework designed to enhance the accuracy of LLM responses by systematically incorporating model confidence at critical decision points. We propose an approach that encourages multi-step reasoning in LLMs and quantify the confidence of intermediate answers such as numerical results in mathematical reasoning and proper nouns in open-domain generation. Then, the overall confidence of each reasoning chain is evaluated based on confidence of these critical intermediate steps. Finally, we aggregate the answer of generated response paths in a way that reflects the reliability of each generated content (as opposed to self-consistency in which each generated chain contributes equally to majority voting). We conducted extensive experiments in five datasets, three mathematical datasets and two open-domain datasets, using four LLMs. The results consistently validate the effectiveness of our novel confidence-aggregation method, leading to an accuracy improvement of up to 7.4% and 5.8% over baseline approaches in math and open-domain generation tasks, respectively. Code is publicly available at https://github.com/sharif-ml-lab/CER.

With the growing deployment of pre-trained models like Transformers on cloud platforms, privacy concerns about model parameters and inference data are intensifying. Existing Privacy-Preserving Transformer Inference (PPTI) frameworks face the “impossible trinity” of balancing privacy, efficiency, and performance: Secure Multi-Party Computation (SMPC)-based approaches ensure strong privacy but suffer from high computational overhead and performance losses; Conversely, permutation-based methods achieve near-plaintext efficiency and accuracy but compromise privacy by exposing sensitive model parameters and intermediate results. Bridging this gap with a single approach presents substantial challenges, motivating the introduction of CENTAUR, a groundbreaking PPTI framework that seamlessly integrates random permutations and SMPC to address the “impossible trinity”. By designing efficient PPTI algorithms tailored to the structural properties of Transformer models, CENTAUR achieves an unprecedented balance among privacy, efficiency, and performance. Our experiments demonstrate CENTAUR’s ability to resist diverse data reconstruction attacks, achieve plaintext-level inference accuracy, and boost inference speed by 5.0~30.4 times, unlocking new possibilities for secure and efficient AI deployment.

Most current automated essay quality assessment systems treat score prediction and feedback generation as separate tasks, overlooking the fact that scores provide a quantitative evaluation of quality, while feedback offers a qualitative assessment. Both aspects reflect essay quality from different perspectives, and they are inherently consistent and can reinforce each other. In this paper, we propose a novel bidirectional reinforcement learning framework that effectively utilizes this consistency constraint to jointly optimize score prediction and feedback generation, ensuring mutual reinforcement and alignment between them. In this way, our model is hope to obtain a simultaneous accurate ratings and consistent text feedback. We conducted extensive experiments on publicly available datasets. The results demonstrate that our approach surpasses the current state-of-the-art models, enhancing both scoring accuracy and feedback quality.

Large Language Models (LLMs) have achieved significant advancements, but the increasing complexity of tasks and higher performance demands highlight the need for continuous improvement. Some approaches utilize synthetic data generated by advanced LLMs based on evaluation results to train models. However, conventional evaluation methods fail to provide detailed, fine-grained profiles of LLMs, limiting their guidance for data synthesis. In this paper, we introduce the **Cognitive Diagnostic Synthesis** (CDS) method, which incorporates a diagnostic process inspired by **Cognitive Diagnosis Theory** (CDT) to refine evaluation results and characterize model profiles at the knowledge component level. Based on these diagnostics, we propose two diagnosis-synthesis strategies for weakness-targeted data synthesis. Additionally, we present an enhanced data augmentation and selection pipeline to improve the quality and diversity of synthesized data. Our experiments with several open-source models show significant improvements across multiple benchmarks, achieving up to 6.00% improvement in code generation, 13.10% in mathematical reasoning, and 5.43% in academic exams. Code and data are available on GitHub https://anonymous.4open.science/r/cds-04D1.

Investigating hallucination issues in large language models (LLMs) within cross-lingual and cross-modal scenarios can greatly advance the large-scale deployment in real-world applications. Nevertheless, the current studies are limited to a single scenario, either cross-lingual or cross-modal, leaving a gap in the exploration of hallucinations in the joint cross-lingual and cross-modal scenarios. Motivated by this, we introduce a novel joint Cross-lingual and Cross-modal Hallucinations benchmark (CCHall) to fill this gap. Specifically, CCHall simultaneously incorporates both cross-lingual and cross-modal hallucination scenarios, which can be used to assess the cross-lingual and cross-modal capabilities of LLMs. Furthermore, we conduct a comprehensive evaluation on CCHall, exploring both mainstream open-source and closed-source LLMs. The experimental results highlight that current LLMs still struggle with CCHall. We hope CCHall can serve as a valuable resource to assess LLMs in joint cross-lingual and cross-modal scenarios.

Current large language models (LLMs) often exhibit imbalanced multilingual capabilities due to their English-centric training corpora. To address this, existing fine-tuning approaches operating at the data-level (e.g., through data augmentation or distillation) typically introduce implicit cross-lingual alignment, overlooking the potential for more profound, latent-level cross-lingual interactions. In this work, we propose CC-Tuning, a novel multilingual fine-tuning paradigm that explicitly establishes a cross-lingual connection mechanism at the latent level. During training, CC-Tuning fuses the feed forward activations from both English and non-English inputs, enabling the model to benefit from both linguistic resources. This process is facilitated with a trainable Decision Maker that identifies beneficial activations. Furthermore, during inference, a Transform Matrix is utilized to simulate the cross-lingual connection under monolingual setting through representation transformation. Our experiments on six benchmarks covering 22 languages show that CC-Tuning outperforms vanilla SFT and offers a strong latent-level alternative to data-level augmentation methods. Further analysis also highlights the practicality of CC-Tuning and the potential of latent-level cross-lingual interactions in advancing the multilingual performance of LLMs.

Security alignment enables the Large Language Model (LLM) to gain the protection against malicious queries, but various jailbreak attack methods reveal the vulnerability of this security mechanism. Previous studies have isolated LLM jailbreak attacks and defenses. We analyze the security protection mechanism of the LLM, and propose a framework that combines attack and defense. Our method is based on the linearly separable property of LLM intermediate layer embedding, as well as the essence of jailbreak attack, which aims to embed harmful problems and transfer them to the safe area. We utilize generative adversarial network (GAN) to learn the security judgment boundary inside the LLM to achieve efficient jailbreak attack and defense. The experimental results indicate that our method achieves an average jailbreak success rate of 88.85% across three popular LLMs, while the defense success rate on the state-of-the-art jailbreak dataset reaches an average of 84.17%. This not only validates the effectiveness of our approach but also sheds light on the internal security mechanisms of LLMs, offering new insights for enhancing model security.Warning: This paper contains some harmful text examples.

Cross-modal retrieval aims to search for instances, which are semantically related to the query through the interaction of different modal data. Traditional solutions utilize a single-tower or dual-tower framework to explicitly compute the score between queries and candidates, which is challenged by training cost and inference latency with large-scale data. Inspired by the remarkable performance and efficiency of generative models, we propose a generative cross-modal retrieval framework (CART) based on coarse-to-fine semantic modeling, which assigns identifiers to each candidate and treats the generating identifier as the retrieval target. Specifically, we explore an effective coarse-to-fine scheme, combining K-Means and RQ-VAE to discretize multimodal data into token sequences that support autoregressive generation. Further, considering the lack of explicit interaction between queries and candidates, we propose a feature fusion strategy to align their semantics. Extensive experiments demonstrate the effectiveness of the strategies in the CART, achieving excellent results in both retrieval performance and efficiency.

Reward modeling in large language models is known to be susceptible to reward hacking, causing models to latch onto superficial features such as the tendency to generate lists or unnecessarily long responses. In RLHF, and more generally during post-training, flawed reward signals often lead to outputs that optimize for these spurious correlates instead of genuine quality or correctness. We propose **Carmo (Context-Aware Reward Modeling)**, a novel approach that first generates dynamic, context-relevant criteria to ground the reward model prior to producing reward scores. Unlike prior methods that use static rubrics, Carmo leverages powerful LLMs to adaptively create evaluation criteria, e.g., logical consistency, clarity, and depth, tailored to the user query. Our theoretical analysis shows that such criteria generation can mitigate reward hacking. We further demonstrate how Carmo can be distilled into smaller models, thereby lowering the computational cost of alignment. We establish a new state-of-the-art performance on zero shot setting for generative models, with a 2.1% improvement on Reward Bench. Furthermore, alignment performed on the Carmo-curated preference dataset achieves **22.5% and 21.1% LC-WR (%) and WR (%) on Mistral-Base (7B)**. We release our datasets at [huggingface/CARMO](https://huggingface.co/datasets/Multi-preference-Optimization/CARMO-UltraFeedback).

In real-world applications, large language models (LLMs) often need to handle diverse and complex instructions. Specifically, when instructions are subject to multiple constraints, some of which are somewhat ambiguous, LLMs often fail to produce answers that satisfy all constraints, limiting their effectiveness in various tasks. To address this challenge, we examine the different constraints in the instructions and discover that the difficulty of determining whether an answer meets a constraint varies widely, from extremely straightforward to exceptionally perplexing. Correspondingly, we propose to assign constraints to different constraint levels. Furthermore, inspired by chain-of-thought (CoT) and self-taught reasoner (STaR), we propose a two-stage method named CARE-STaR (Constraint-AwaRE STaR). Our method distinguishes constraints within instructions by generating different CoTs and guides LLMs to autonomously learn optimal answers by setting the positive rewards to the CoTs that are beneficial to generating accurate answers and iteratively optimizing these answers. We have conducted extensive experiments on three instruction-following benchmarks, taking three existing LLMs as base LLMs, respectively. Experimental results indicate that our method substantially enhances the capability of these LLMs to handle complex instructions, outperforming supervised fine-tuning (SFT). Our code is available at https://github.com/lzl0124/carestar.

Conversational counselor agents have become essential tools for addressing the rising demand for scalable and accessible mental health support. This paper introduces CAMI, a novel automated counselor agent grounded in Motivational Interviewing (MI) – a client-centered counseling approach designed to address ambivalence and facilitate behavior change. CAMI employs a novel STAR framework, consisting of client’s state inference, motivation topic exploration, and response generation modules, leveraging large language models (LLMs). These components work together to evoke change talk, aligning with MI principles and improving counseling outcomes for diverse clients. We evaluate CAMI’s performance through both automated and expert evaluations, utilizing simulated clients to assess MI skill competency, client’s state inference accuracy, topic exploration proficiency, and overall counseling success. Results show that CAMI not only outperforms several state-of-the-art methods but also shows more realistic counselor-like behavior. Additionally, our ablation study underscores the critical roles of state inference and topic exploration in achieving this performance.

Computer-aided design (CAD) is crucial in prototyping 3D objects through geometric instructions (i.e., CAD programs). In practical design workflows, designers often engage in time-consuming reviews and refinements of these prototypes by comparing them with reference images. To bridge this gap, we introduce the CAD review task to automatically detect and correct potential errors, ensuring consistency between the constructed 3D objects and reference images. However, recent advanced multimodal large language models (MLLMs) struggle to recognize multiple geometric components and perform spatial geometric operations within the CAD program, leading to inaccurate reviews. In this paper, we propose the CAD program repairer (ReCAD) framework to effectively detect program errors and provide helpful feedback on error correction. Additionally, we create a dataset, CADReview, consisting of over 20K program-image pairs, with diverse errors for the CAD review task. Extensive experiments demonstrate that our ReCAD significantly outperforms existing MLLMs, which shows great potential in design applications.

Information retrieval has evolved from traditional sparse and dense retrieval methods to approaches driven by large language models (LLMs). Recent techniques, such as Generation-Augmented Retrieval (GAR) and Generative Document Retrieval (GDR), leverage LLMs to enhance retrieval but face key challenges: GAR’s generated content may not always align with the target document corpus, while GDR limits the generative capacity of LLMs by constraining outputs to predefined document identifiers. To address these issues, we propose Context-Aware Generation-Augmented Retrieval (CA-GAR), which enhances LLMs by integrating corpus information into their generation process. CA-GAR optimizes token selection by incorporating relevant document information and leverages a Distribution Alignment Strategy to extract corpus information using a lexicon-based approach. Experimental evaluations on seven tasks from the BEIR benchmark and four non-English languages from Mr.TyDi demonstrate that CA-GAR outperforms existing methods.

Recent advances in large language models (LLMs) have shown significant promise, yet their evaluation raises concerns, particularly regarding data contamination due to the lack of access to proprietary training data. To address this issue, we present C2LEVA, a comprehensive bilingual benchmark featuring systematic contamination prevention. C2LEVA firstly offers a holistic evaluation encompassing 22 tasks, each targeting a specific application or ability of LLMs, and secondly a trustworthy assessment due to our contamination-free tasks, ensured by a systematic contamination prevention strategy that fully automates test data renewal and enforces data protection during benchmark data release. Our large-scale evaluation of 15 open-source and proprietary models demonstrates the effectiveness of C2LEVA.

Sequential recommenders predict users’ next interactions based on historical behavior and are essential in modern recommendation systems. While Large Language Models (LLMs) show promise, their size and high inference costs limit deployment on resource-constrained devices. Small Language Models (SLMs) provide a more efficient alternative for edge devices, but bridging the recommendation performance gap between LLMs and SLMs remains challenging. Typical approaches like supervised fine-tuning or vanilla knowledge distillation (KD) often lead to suboptimal performance or even negative transfer. Our motivational experiments reveal key issues with vanilla KD methods: feature imitation suffers from redundancy and uneven recommendation ability across layers, while prediction mimicking faces conflicts caused by differing weight distributions of prediction heads. To address these challenges, we propose a simple yet effective framework, C2KD, to transfer task-relevant knowledge from two complementary dimensions. Specifically, our method incorporates: (1) cross-layer feature imitation, which uses a dynamic router to select the most relevant teacher layers and assimilate task-relevant knowledge from the teacher’s late layers, allowing the student to concentrate on the teacher’s specialized knowledge; and (2) cross-head logit distillation, which maps the intermediate features of the student to the teacher’s output head, thereby minimizing prediction discrepancies between the teacher and the student. Extensive experiments across diverse model families demonstrate that our approach enables 1B-parameter SLMs to achieve competitive performance compared to LLMs (e.g., Llama3-8B), offering a practical solution for real-world on-device sequential recommendations.

We introduce the Byte Latent Transformer (BLT), a new byte-level LLM architecture that, for the first time, matches tokenization-based LLM performance at scale with significant improvements in inference efficiency and robustness. BLT encodes bytes into dynamically sized patches, which serve as the primary units of computation. Patches are segmented based on the entropy of the next byte, allocating more compute and model capacity where increased data complexity demands it. We present the first FLOP controlled scaling study of byte-level models – up to 8B parameters and 4T training bytes – demonstrating the feasibility of scaling models trained on raw bytes without a fixed vocabulary. Both training and inference efficiency improve due to dynamically selecting long patches when data is predictable, along with qualitative improvements on reasoning and long tail generalization. For fixed inference costs, BLT shows significantly better scaling than tokenization-based models, by simultaneously growing both patch and model size.

Recent Large-Language Models (LLMs) pruning methods typically operate at the post-training phase without the expensive weight finetuning, however, their pruning criteria often rely on **heuristically hand-crafted metrics**, potentially leading to suboptimal performance. We instead propose a novel **optimization-based structural pruning** that learns the pruning masks in a probabilistic space directly by optimizing the loss of the pruned model. To preserve the efficiency, our method **eliminates the back-propagation** through the LLM *per se* during the optimization, requiring only **the forward pass of the LLM**. We achieve this by learning an underlying Bernoulli distribution to sample binary pruning masks, where we decouple the Bernoulli parameters from the LLM loss, thus facilitating an efficient optimization via *policy gradient estimator* without back-propagation. As a result, our method is able to 1) *support global and heterogeneous pruning* (*i.e.*, our method automatically determines different redundancy for different layers), and 2) *optionally initialize with a metric-based method* (for our Bernoulli distributions). Extensive experiments conducted on LLaMA, LLaMA-2, LLaMA-3, Vicuna, and Mistral models using the C4 and WikiText2 datasets demonstrate the promising performance of our method in efficiency and effectiveness.