ACL 2025 Papers

It is crucial for large language models (LLMs) to follow instructions that involve multiple constraints. In real-world scenarios, user instructions often contain soft constraints, which are semantically related and cannot be rule-based verified, posing challenges for LLMs. To enhance the soft constraint following ability of LLMs, we initially design a pipeline to construct datasets with high-quality outputs for instructions containing soft constraints automatically. Additionally, to fully utilize the positive and negative samples generated during the data construction process, we choose Direct Preference Optimization (DPO) as the training method. Furthermore, taking into account the difficulty of soft constraints indicated by the number of constraints, we design a curriculum learning training paradigm based on the constraint quantity. We experimentally evaluate the effectiveness of our methods in improving LLMs’ soft constraint following ability and analyze the factors driving the improvements.

Steering methods for language models (LMs) have gained traction as lightweight alternatives to fine-tuning, enabling targeted modifications to model activations. However, prior studies primarily report results on a few models, leaving critical gaps in understanding the robustness of these methods. In this work, we systematically examine three prominent steering methods—DoLa, function vectors, and task vectors. In contrast to the original studies, which evaluated a handful of models, we test up to 36 models belonging to 14 families with sizes ranging from 1.5B to 70B parameters. Our experiments reveal substantial variability in the effectiveness of the steering approaches, with a large number of models showing no improvement and at times degradation in steering performance. Our analysis reveals fundamental flaws in the assumptions underlying these methods, challenging their reliability as scalable steering solutions.

Aligned representations across languages is a desired property in multilingual large language models (mLLMs), as alignment can improve performance in cross-lingual tasks. Typically alignment requires fine-tuning a model, which is computationally expensive, and sizable language data, which often may not be available. A data-efficient alternative to fine-tuning is model interventions — a method for manipulating model activations to steer generation into the desired direction. We analyze the effect of a popular intervention (finding experts) on the alignment of cross-lingual representations in mLLMs. We identify the neurons to manipulate for a given language and introspect the embedding space of mLLMs pre- and post-manipulation. We show that modifying the mLLM’s activations changes its embedding space such that cross-lingual alignment is enhanced. Further, we show that the changes to the embedding space translate into improved downstream performance on retrieval tasks, with up to 2x improvements in top-1 accuracy on cross-lingual retrieval.

Decentralized training has become a resource-efficient framework to democratize the training of large language models (LLMs). However, the privacy risks associated with this framework, particularly due to the potential inclusion of sensitive data in training datasets, remain unexplored. This paper identifies a novel and realistic attack surface: the privacy leakage from training data in decentralized training, and proposes activation inversion attack (AIA) for the first time. AIA first constructs a shadow dataset comprising text labels and corresponding activations using public datasets. Leveraging this dataset, an attack model can be trained to reconstruct the training data from activations in victim decentralized training. We conduct extensive experiments on various LLMs and publicly available datasets to demonstrate the susceptibility of decentralized training to AIA. These findings highlight the urgent need to enhance security measures in decentralized training to mitigate privacy risks in training LLMs.

Generative models are capable of producing human-expert level content across a variety of topics and domains. As the impact of generative models grows, it is necessary to develop statistical methods to understand collections of available models. These methods are particularly important in settings where the user may not have access to information related to a model’s pre-training data, weights, or other relevant model-level covariates. In this paper we extend recent results on representations of black-box generative models to model-level statistical inference tasks. We demonstrate that the model-level representations are effective for multiple inference tasks.

Tasks are central in machine learning, as they are the most natural objects to assess the capabilities of current models. The trend is to build general models able to address any task. Even though transfer learning and multitask learning try to leverage the underlying task space, no well-founded tools are available to study its structure. This study proposes a theoretically grounded setup to define the notion of task and to compute the inclusion between two tasks from a statistical deficiency point of view. We propose a tractable proxy as information sufficiency to estimate the degree of inclusion between tasks, show its soundness on synthetic data, and use it to reconstruct empirically the classic NLP pipeline.

Large Language Models (LLMs) excel in zero-shot and few-shot tasks, but achieving similar performance with encoder-only models like BERT and RoBERTa has been challenging due to their architecture. However, encoders offer advantages such as lower computational and memory costs. Recent work adapts them for zero-shot generalization using Statement Tuning, which reformulates tasks into finite templates. We extend this approach to multilingual NLP, exploring whether encoders can achieve zero-shot cross-lingual generalization and serve as efficient alternatives to memory-intensive LLMs for low-resource languages. Our results show that state-of-the-art encoder models generalize well across languages, rivaling multilingual LLMs while being more efficient. We also analyze multilingual Statement Tuning dataset design, efficiency gains, and language-specific generalization, contributing to more inclusive and resource-efficient NLP models. We release our code and models.

State Space Models (SSMs) have emerged as efficient alternatives to Transformers, mitigating their quadratic computational cost. However, the application of Parameter-Efficient Fine-Tuning (PEFT) methods to SSMs remains largely unexplored. In particular, prompt-based methods like Prompt Tuning and Prefix-Tuning, which are widely used in Transformers, do not perform well on SSMs. To address this, we propose **state-based methods** as a superior alternative to prompt-based methods. This new family of methods naturally stems from the architectural characteristics of SSMs. State-based methods adjust state-related features directly instead of depending on external prompts. Furthermore, we introduce a novel state-based PEFT method: **State-offset Tuning**. At every timestep, our method directly affects the state at the current step, leading to more effective adaptation. Through extensive experiments across diverse datasets, we demonstrate the effectiveness of our method. Code is available at https://github.com/furiosa-ai/ssm-state-tuning.

Research shows that two evaluation experiments reporting results for the same quality criterion name (e.g. Fluency) do not necessarily evaluate the same aspect of quality. Not knowing when two evaluations are comparable in this sense means we currently lack the ability to draw conclusions based on multiple independently conducted evaluations. It is hard to see how this issue can be fully addressed other than by the creation of a standard set of quality criterion names and definitions that the evaluations in use in NLP can be grounded in. Taking a descriptivist approach, the QCET Quality Criteria for Evaluation Taxonomy derives a standard set of 114 quality criterion names and definitions from three surveys of a combined total of 933 evaluation experiments in NLP, and structures them into a reference taxonomy. We present QCET and its uses in (i) establishing comparability of existing evaluations, (ii) guiding the design of new evaluations, and (iii) assessing regulation compliance.

The rapid advancement of large language models (LLMs) has spurred significant interest in tool learning, where LLMs are augmented with external tools to tackle complex tasks. However, existing tool environments face challenges in balancing stability, scale, and realism, particularly for benchmarking purposes. To address this, we propose MirrorAPI, a novel framework that trains specialized LLMs to accurately simulate real API responses, effectively acting as “mirrors” to tool environments. Using a comprehensive dataset of request-response pairs from 7,000+ APIs, we employ supervised fine-tuning and chain-of-thought reasoning to enhance simulation fidelity. MirrorAPI achieves superior accuracy and stability compared to state-of-the-art methods, as demonstrated by its performance on the newly constructed MirrorAPI-Bench and its integration into StableToolBench.

Emerging Large Language Model (LLM) applications require long input context in order to perform complex tasks like document analysis and code generation.For these long context length applications, the length of the input prompt poses a significant challenge in terms of inference efficiency since the inference costs increase linearly with sequence length.However, for many of these applications, much of the context in the prompt is fixed across different user inputs, thereby providing the opportunity to perform offline optimizations in order to process user inputs quickly, as they are received. We propose Squeezed Attention to accelerate LLM applications where a large portion of the input context is fixed.We first leverage K-means clustering offline to group the keys for the fixed context based on semantic similarity and represent each cluster with a single centroid value.During inference, we compare query tokens from the user input with the centroids to predict which keys from the fixed context are semantically relevant, and then compute exact attention using only the important keys, thereby reducing bandwidth and computational costs. We also present a hierarchical version of our algorithm which can reduce the complexity of attention from linear to logarithmic with respect to the fixed context length.We evaluate our method on various long-context benchmarks including LongBench, where it achieves a 3.1× reduction in KV budget with no noticeable accuracy loss and up to an 8× reduction with only a 0.5 point accuracy gap for the LLaMA-2-7B-32K, LWM-Text-Chat-1M, and Longchat-7B-v1.5-32K models.Futhermore, we implement kernels for centroid comparison and sparse FlashAttention with important keys, achieving more than 4× speedups during both the prefill and generation phases for long-context inference.Our code is available at https://github.com/SqueezeAILab/SqueezedAttention.

Social determinants of health (SDOH) extraction from clinical text is critical for downstream healthcare analytics. Although large language models (LLMs) have shown promise, they may rely on superficial cues leading to spurious predictions. Using the MIMIC portion of the SHAC (Social History Annotation Corpus) dataset and focusing on drug status extraction as a case study, we demonstrate that mentions of alcohol or smoking can falsely induce models to predict current/past drug use where none is present, while also uncovering concerning gender disparities in model performance. We further evaluate mitigation strategies—such as prompt engineering and chain-of-thought reasoning—to reduce these false positives, providing insights into enhancing LLM reliability in health domains.

Ensuring persona fidelity in large language models (LLMs) is essential for maintaining coherent and engaging human-AI interactions. However, LLMs often exhibit Out-of-Character (OOC) behavior, where generated responses deviate from an assigned persona, leading to inconsistencies that affect model reliability. Existing evaluation methods typically assign single scores to entire responses, struggling to capture subtle persona misalignment, particularly in long-form text generation. To address this limitation, we propose an atomic-level evaluation framework that quantifies persona fidelity at a finer granularity. Our three key metrics measure the degree of persona alignment and consistency within and across generations. Our approach enables a more precise and realistic assessment of persona fidelity by identifying subtle deviations that real users would encounter. Through our experiments, we demonstrate that our framework effectively detects persona inconsistencies that prior methods overlook. By analyzing persona fidelity across diverse tasks and personality types, we reveal how task structure and persona desirability influence model adaptability, highlighting challenges in maintaining consistent persona expression.

Common subword tokenization algorithms like BPE and UnigramLM assume that text can be split into meaningful units by concatenative measures alone. This is not true for languages such as Hebrew and Arabic, where morphology is encoded in root-template patterns, or Malay and Georgian, where split affixes are common. We present SPLINTER, a pre-processing step which rearranges text into a linear form that better represents such nonconcatenative morphologies, enabling meaningful contiguous segments to be found by the tokenizer. We demonstrate SPLINTER’s merit using both intrinsic measures evaluating token vocabularies in Hebrew, Arabic, and Malay; as well as on downstream tasks using BERT-architecture models trained for Hebrew.

Large Language Models (LLMs) have achieved impressive accomplishments in recent years. However, the increasing memory consumption of KV cache has possessed a significant challenge to the inference system. Eviction methods have revealed the inherent redundancy within the KV cache, demonstrating its potential for reduction, particularly in deeper layers. However, KV cache reduction for shallower layers has been found to be insufficient. Based on our observation that, the KV cache exhibits a high degree of similarity. Based on this observation, we proposed a novel KV cache reduction method, SpindleKV, which balances both shallow and deep layers. For deep layers, we employ an attention weight based eviction method, while for shallow layers, we apply a codebook based replacement approach which is learnt by similarity and merging policy. Moreover, SpindleKV addressed the Grouped-Query Attention (GQA) dilemma faced by other attention based eviction methods. Experiments on two common benchmarks with three different LLMs shown that SpindleKV obtained better KV cache reduction effect compared to baseline methods, while preserving similar or even better model performance.

Large language models (LLMs) have made significant strides in code acceleration (CA) tasks. Current works typically fine-tune LLMs using slow-fast code pairs mined from online programming platforms. Although these methods are widely recognized for their effectiveness, the training data often lack clear code acceleration patterns and offer only limited speed improvements. Moreover, existing training methods, such as direct instruction fine-tuning (IFT), tend to overlook the hierarchical relationships among acceleration patterns. In this work, we introduce BITE, a novel training paradigm designed to improve LLMs’ CA capabilities through two key innovations: (1) Bidirectional tree editing, which generates high-quality training data by incrementally transforming given code into both its most efficient and least efficient variants, and (2) Progressive code acceleration learning, which enables LLMs to internalize multi-level CA strategies by learning increasingly sophisticated acceleration patterns. Additionally, we introduce a new CA evaluation benchmark and metric for comprehensive assessment of model performance on CA tasks. Extensive experiments on both our benchmark and existing benchmarks demonstrate the effectiveness of our approach. Notably, BITE enables Qwen-1.5B to outperform prompt-enhanced GPT-4 and current training-based methods on average across five programming languages.

Large Language Models (LLMs) have been increasingly adopted for health-related tasks, yet their performance in depression detection remains limited when relying solely on text input. While Retrieval-Augmented Generation (RAG) typically enhances LLM capabilities, our experiments indicate that traditional text-based RAG systems struggle to significantly improve depression detection accuracy. This challenge stems partly from the rich depression-relevant information encoded in acoustic speech patterns — information that current text-only approaches fail to capture effectively. To address this limitation, we conduct a systematic analysis of temporal speech patterns, comparing healthy individuals with those experiencing depression. Based on our findings, we introduce Speech Timing-based Retrieval-Augmented Generation, SpeechT-RAG, a novel system that leverages speech timing features for both accurate depression detection and reliable confidence estimation. This integrated approach not only outperforms traditional text-based RAG systems in detection accuracy but also enhances uncertainty quantification through a confidence scoring mechanism that naturally extends from the same temporal features. Our unified framework achieves comparable results to fine-tuned LLMs without additional training while simultaneously addressing the fundamental requirements for both accuracy and trustworthiness in mental health assessment

We introduce Speech-based Intelligence Quotient (SIQ) as a new form of human cognition-inspired evaluation pipeline for voice understanding large language models (LLM_Voice), designed to assess their voice understanding ability. Moving beyond popular voice understanding metrics such as word error rate (WER), SIQ examines LLM_Voice across three cognitive levels motivated by Bloom’s Taxonomy: (1) Remembering (i.e., WER for verbatim accuracy); (2) Understanding (i.e., similarity of LLM’s interpretations); and (3) Application (i.e., QA accuracy for simulating downstream tasks). We demonstrate that SIQ not only quantifies voice understanding abilities but also provides unified comparisons between cascaded methods (e.g., ASR-LLM) and end-to-end models, identifies annotation errors in existing benchmarks, and detects hallucinations in LLM_Voice. Our framework represents a first-of-its-kind intelligence examination that bridges cognitive principles with voice-oriented benchmarks, while exposing overlooked challenges in multi-modal training. Our code and data will be open source to encourage future studies.

As speech generation technology advances, the risk of misuse through deepfake audio has become a pressing concern, which underscores the critical need for robust detection systems. However, many existing speech deepfake datasets are limited in scale and diversity, making it challenging to train models that can generalize well to unseen deepfakes. To address these gaps, we introduce SpeechFake, a large-scale dataset designed specifically for speech deepfake detection. SpeechFake includes over 3 million deepfake samples, totaling more than 3,000 hours of audio, generated using 40 different speech synthesis tools. The dataset encompasses a wide range of generation techniques, including text-to-speech, voice conversion, and neural vocoder, incorporating the latest cutting-edge methods. It also provides multilingual support, spanning 46 languages. In this paper, we offer a detailed overview of the dataset’s creation, composition, and statistics. We also present baseline results by training detection models on SpeechFake, demonstrating strong performance on both its own test sets and various unseen test sets. Additionally, we conduct experiments to rigorously explore how generation methods, language diversity, and speaker variation affect detection performance. We believe SpeechFake will be a valuable resource for advancing speech deepfake detection and developing more robust models for evolving generation techniques.