NAACL 2025 论文列表

Simultaneous machine translation (SMT) takes streaming input utterances and incrementally produces target text. Existing SMT methods only use the partial utterance that has already arrived at the input and the generated hypothesis. Motivated by human interpreters’ technique to forecast future words before hearing them, we propose Translation by Anticipating Future (TAF), a method to improve translation quality while retaining low latency. Its core idea is to use a large language model (LLM) to predict future source words and opportunistically translate without introducing too much risk. We evaluate our TAF and multiple baselines of SMT on four language directions. Experiments show that TAF achieves the best translation quality-latency trade-off and outperforms the baselines by up to 5 BLEU points at the same latency (three words).

The compositional generalization abilities of neural models have been sought after for human-like linguistic competence.The popular method to evaluate such abilities is to assess the models’ input-output behavior.However, that does not reveal the internal mechanisms, and the underlying competence of such models in compositional generalization remains unclear.To address this problem, we explore the inner workings of a Transformer model byfinding an existing subnetwork that contributes to the generalization performance and by performing causal analyses on how the model utilizes syntactic features.We find that the model depends on syntactic features to output the correct answer, but that the subnetwork with much better generalization performance than the whole model relies on a non-compositional algorithm in addition to the syntactic features.We also show that the subnetwork improves its generalization performance relatively slowly during the training compared to the in-distribution one, and the non-compositional solution is acquired in the early stages of the training.

Large language models (LLMs) have recently revolutionized natural language processing. These models, however, often suffer from instability or lack of coherence, that is the ability of the models to generate semantically equivalent outputs when receiving diverse yet semantically equivalent input variations. In this work, we analyze the behavior of multiple LLMs, including Mixtral-8x7B, Llama2-70b, Smaug-72b, and Phi-3, when dealing with multiple lexical variations of the same info-seeking questions. Our results suggest that various LLMs struggle to consistently answer diverse equivalent queries. To address this issue, we show how redundant information encoded as a prompt can increase the coherence of these models. In addition, we introduce a Retrieval-Augmented Generation (RAG) technique that supplements LLMs with the top-k most similar questions from a question retrieval engine. This knowledge-augmentation leads to 4-8 percentage point improvement in end-to-end performance in factual question answering tasks. These findings underscore the need to enhance LLM stability and coherence through semantic awareness.

The correlation between NLG automatic evaluation metrics and human evaluation is often regarded as a critical criterion for assessing the capability of an evaluation metric. However, different grouping methods and correlation coefficients result in various types of correlation measures used in meta-evaluation. In specific evaluation scenarios, prior work often directly follows conventional measure settings, but the characteristics and differences between these measures have not gotten sufficient attention. Therefore, this paper analyzes 12 common correlation measures using a large amount of real-world data from six widely-used NLG evaluation datasets and 32 evaluation metrics, revealing that different measures indeed impact the meta-evaluation results. Furthermore, we propose three perspectives that reflect the capability of meta-evaluation: discriminative power, ranking consistency, and sensitivity to score granularity. We find that the measure using global grouping and Pearson correlation coefficient exhibits the best performance in both discriminative power and ranking consistency. Besides, the measures using system-level grouping or Kendall correlation are the least sensitive to score granularity.

Large Language Models (LLMs) are prevalent in modern applications but often memorize training data, leading to privacy breaches and copyright issues. Existing research has mainly focused on post-hoc analyses—such as extracting memorized content or developing memorization metrics—without exploring the underlying architectural factors that contribute to memorization. In this work, we investigate memorization from an architectural lens by analyzing how attention modules at different layers impact its memorization and generalization performance. Using attribution techniques, we systematically intervene in the LLM’s architecture by bypassing attention modules at specific blocks while keeping other components like layer normalization and MLP transformations intact. We provide theorems analyzing our intervention mechanism from a mathematical view, bounding the difference in layer outputs with and without our attributions. Our theoretical and empirical analyses reveal that attention modules in deeper transformer blocks are primarily responsible for memorization, whereas earlier blocks are crucial for the model’s generalization and reasoning capabilities. We validate our findings through comprehensive experiments on different LLM families (Pythia and GPT-Neo) and five benchmark datasets. Our insights offer a practical approach to mitigate memorization in LLMs while preserving their performance, contributing to safer and more ethical deployment in real-world applications.

Autoencoders have been used for finding interpretable and disentangled features underlying neural network representations in both image and text domains. While the efficacy and pitfalls of such methods are well-studied in vision, there is a lack of corresponding results, both qualitative and quantitative, for the text domain. We aim to address this gap by training sparse autoencoders (SAEs) on a synthetic testbed of formal languages. Specifically, we train SAEs on the hidden representations of models trained on formal languages (Dyck-2, Expr, and English PCFG) under a wide variety of hyperparameter settings, finding interpretable latents often emerge in the features learned by our SAEs. However, similar to vision, we find performance turns out to be highly sensitive to inductive biases of the training pipeline. Moreover, we show latents correlating to certain features of the input do not always induce a causal impact on model’s computation. We thus argue that causality has to become a central target in SAE training: learning of causal features should be incentivized from the ground-up. Motivated by this, we propose and perform preliminary investigations for an approach that promotes learning of causally relevant features in our formal language setting.

African American English (AAE) presents unique challenges in natural language processing (NLP) This research systematically compares the performance of available NLP models—rule-based, transformer-based, and large language models (LLMs)—capable of identifying key grammatical features of AAE, namely Habitual Be and Multiple Negation. These features were selected for their distinct grammatical complexity and frequency of occurrence. The evaluation involved sentence-level binary classification tasks, using both zero-shot and few-shot strategies. The analysis reveals that while LLMs show promise compared to the baseline, they are influenced by biases such as recency and unrelated features in the text such as formality. This study highlights the necessity for improved model training and architectural adjustments to better accommodate AAE’s unique linguistic characteristics. Data and code are available.

Formulaic criteria for proportional analogies, which capture relational mappings between two ratios of terms, are mainly confined to the formal level. As analogy datasets grow more complex, especially in evaluating the cognitive abilities of Large Language Models (LLMs), assessing parallelism in them becomes increasingly challenging and often requires human annotation. In this work, we propose AnaScore, an automatic metric for evaluating the strength of semantic parallelism in sentence analogies. AnaScore systematically provides formalized explanations for shared relational patterns at the level of conceptual knowledge. We apply AnaScore to annotate several existing datasets, considering different directions of the relations, and uncover artifacts in data construction. Our experiments with various LLMs demonstrate the efficacy of the AnaScore metric in capturing the inherent quality of analogical relationships, showing a positive correlation between analogy quality and model performance. Thanks to this metric, we clearly demonstrate that formally explainable examples are more beneficial for analogical reasoning, while ambiguous analogies with no clear criterion tend to hinder inference.

Large language models (LLMs) can be leveraged to help write formulas in spreadsheets, but formula data resources are scarce, impacting both the base performance of pre-trained models and limiting the ability to fine-tune them. Given a corpus of formulas, we can use another model to generate synthetic natural language utterances for fine-tuning. However, it is important to validate whether the natural language (NL) generated by the LLM is accurate for it to be beneficial for fine-tuning. In this paper, we provide empirical results on the impact of validating these synthetic training examples with surrogate objectives that evaluate the accuracy of the synthetic annotations. We demonstrate that validation improves performance over raw data across four models (2 open and 2 closed weight). Interestingly, we show that although validation tends to prune more challenging examples, it increases the complexity of problems that models can solve after being fine-tuned on validated data.

Despite prior safety alignment efforts, LLMs can still generate harmful and unethical content when subjected to jailbreaking attacks. Existing jailbreaking methods fall into two main categories: template-based and optimization-based methods. The former requires significant manual effort and domain knowledge, while the latter, exemplified by GCG, which seeks to maximize the likelihood of harmful LLM outputs through token-level optimization, also encounters several limitations: requiring white-box access, necessitating pre-constructed affirmative phrase, and suffering from low efficiency. This paper introduces ECLIPSE, a novel and efficient black-box jailbreaking method with optimizable suffixes. We employ task prompts to translate jailbreaking objectives into natural language instructions, guiding LLMs to generate adversarial suffixes for malicious queries. A harmfulness scorer provides continuous feedback, enabling LLM self-reflection and iterative optimization to autonomously produce effective suffixes. Experimental results demonstrate that ECLIPSE achieves an average attack success rate (ASR) of 0.92 across three open-source LLMs and GPT-3.5-Turbo, significantly outperforming GCG by 2.4 times. Moreover, ECLIPSE matches template-based methods in ASR while substantially reducing average attack overhead by 83%, offering superior attack efficiency.

Generating insightful and actionable information from databases is critical in data analysis. This paper introduces a novel approach using Large Language Models (LLMs) to automatically generate textual insights. Given a multi-table database as input, our method leverages LLMs to produce concise, text-based insights that reflect interesting patterns in the tables. Our framework includes a Hypothesis Generator to formulate domain-relevant questions, a Query Agent to answer such questions by generating SQL queries against a database, and a Summarization module to verbalize the insights. The insights are evaluated for both correctness and subjective insightfulness using a hybrid model of human judgment and automated metrics. Experimental results on public and enterprise databases demonstrate that our approach generates more insightful insights than other approaches while maintaining correctness.

We analyse the cross-lingual transferability of a dialogue evaluation framework that assesses the relationships between micro-level linguistic features (e.g. backchannels) and macro-level interactivity labels (e.g. topic management), originally designed for English-as-a-second-language dialogues. To this end, we develop CNIMA (**C**hinese **N**on-Native **I**nteractivity **M**easurement and **A**utomation), a Chinese-as-a-second-language labelled dataset with 10K dialogues. We found the evaluation framework to be robust across languages, revealing language-specific and language-universal relationships between micro-level and macro-level features. Next, we propose an automated, interpretable approach with low data requirements that scores the overall quality of a second-language dialogue based on the framework. Our approach is interpretable in that it reveals the key linguistic and interactivity features that contributed to the overall quality score. As our approach does not require labelled data, it can also be adapted to other languages for second-language dialogue evaluation.

Incrementally fine-tuning foundational models on new tasks or domains is now the de facto approach in NLP. A known pitfall of this approach is the catastrophic forgetting of prior knowledge that happens during fine-tuning. A common approach to alleviate such forgetting is to rehearse samples from prior tasks during fine-tuning. Several existing works assume a fixed memory buffer to store prior task examples, while relying on inferences (forward passes) with the model at hand for choosing examples for rehearsal from the buffer. However, given the increasing computational cost of model inference, and decreasing cost of data storage, we focus on the setting to rehearse samples with a fixed computational budget instead of a fixed memory budget. We propose a sampling scheme, mix-cd, that prioritizes rehearsal of “collateral damage” samples, which are samples predicted correctly by the prior model but forgotten by the incrementally tuned one. The crux of our scheme is a procedure to efficiently estimate the density of collateral damage samples without incurring additional model inferences. Our approach is computationally efficient, easy to implement, and outperforms several leading continual learning methods in compute-constrained settings. All the code will be publicly available at https://github.com/jybai/mix-cd-rehearsal.

Gloss-free Sign Language Translation (SLT) converts sign videos into spoken language sentences without relying on glosses, which are the written representations of signs. Recently, Large Language Models (LLMs) have shown remarkable translation performance in gloss-free methods by harnessing their powerful natural language generation capabilities. However, these methods often rely on domain-specific fine-tuning of visual encoders to achieve optimal results. By contrast, we emphasize the importance of capturing the spatial configurations and motion dynamics in sign language. With this in mind, we introduce Spatial and Motion-based Sign Language Translation (SpaMo), a novel LLM-based SLT framework. The core idea of SpaMo is simple yet effective: instead of domain-specific tuning, we use off-the-shelf visual encoders to extract spatial and motion features, which are then input into an LLM along with a language prompt. Additionally, we employ a visual-text alignment process as a lightweight warm-up step before applying SLT supervision. Our experiments demonstrate that SpaMo achieves state-of-the-art performance on three popular datasets—PHOENIX14T, CSL-Daily, and How2Sign—without visual fine-tuning.

In human cognitive memory psychology, the context-dependent effect helps retrieve key memory cues essential for recalling relevant knowledge in problem-solving. Inspired by this, we introduce the context-dependent memory framework (CDMem), an efficient architecture miming human memory processes through multistage encoding, context-aware storage, and retrieval strategies for LLM-centric agents. We propose multistage memory encoding strategies for acquiring high-quality multilevel knowledge: expert encoding compresses raw trajectories from a domain-expert perspective, short-term encoding consolidates experiences from current tasks, and long-term encoding reflects insights from past tasks. For memory storage and retrieval, we design a graph-structured, context-dependent indexing mechanism that allows agents to efficiently and accurately recall the most relevant multilevel knowledge tailored to the current task and environmental context. Furthermore, the proposed CDMem framework is an online learning architecture, enabling agents to efficiently learn and update memory while adapting to novel environments and tasks in real-world applications. We conducted extensive experiments on two interactive decision-making benchmarks in the navigation and manipulation domain, ALFWorld and ScienceWorld. Using GPT-4o-mini, our method surpasses state-of-the-art online LLM-centric approaches, achieving success rates of 85.8% and 56.0%, respectively. We hope this work will serve as a valuable reference for the academic and industrial communities in advancing agent-based applications.

As premodern texts are passed down over centuries, errors inevitably accrue. These errors can be challenging to identify, as some have survived undetected for so long precisely because they are so elusive. While prior work has evaluated error detection methods on artificially-generated errors, we introduce the first dataset of real errors in premodern Greek, enabling the evaluation of error detection methods on errors that genuinely accumulated at some stage in the centuries-long copying process. To create this dataset, we use metrics derived from BERT conditionals to sample 1,000 words more likely to contain errors, which are then annotated and labeled by a domain expert as errors or not. We then propose and evaluate new error detection methods and find that our discriminator-based detector outperforms all other methods, improving the true positive rate for classifying real errors by 5%. We additionally observe that scribal errors are more difficult to detect than print or digitization errors. Our dataset enables the evaluation of error detection methods on real errors in premodern texts for the first time, providing a benchmark for developing more effective error detection algorithms to assist scholars in restoring premodern works.

Large Language Models (LLMs) inherently use autoregressive decoding, which lacks parallelism in inference and results in significantly slow inference speed. While methods such as Medusa constructs parallelized heads, they lack adequate information interaction across different prediction positions. To overcome this limitation, we introduce Amphista, an enhanced speculative decoding framework that builds upon Medusa. Specifically, Amphista models an *Auto-embedding Block* capable of parallel inference, incorporating bi-directional attention to enable interaction between different drafting heads. Additionally, Amphista integrates *Staged Adaptation Layers*, which ensure a seamless transition of semantic information from the target model’s autoregressive inference to the drafting heads’ non-autoregressive inference, effectively achieving paradigm shift and feature fusion. Experimental results on Vicuna models using MT-Bench and Spec-Bench demonstrate that Amphista achieves substantial acceleration while maintaining generation quality. On MT-Bench, Amphista delivers up to **2.75×** speedup over vanilla autoregressive decoding and **1.40×** over Medusa on Vicuna 33B in wall-clock time.

Current machine translation (MT) systems for low-resource languages have a particular failure mode: When translating words in a given domain, they tend to confuse words within that domain. So, for example, “lion” might be translated as “alligator”, and “orange” might be rendered as “purple.” We propose a recall-based metric for measuring this problem and show that the problem exists in 122 low-resource languages. We then show that this problem can be mitigated by using a large language model (LLM) to post-edit the MT output, specifically by including the entire GATITOS lexicon for the relevant language as a very long context prompt. We show gains in average ChrF score over the set of 122 languages, and we show that the recall score for relevant lexical items also improves. Finally, we demonstrate that a small dedicated MT system with a general-purpose LLM as a post-editor is outperforms a lexicon-based RAG-LLM translator, suggesting a new paradigm for LLM use.

Despite their impressive capabilities, large language models (LLMs) have been observed to generate responses that include inaccurate or fabricated information, a phenomenon commonly known as hallucination. In this work, we propose a simple Induce-then-Contrast Decoding (ICD) strategy to alleviate hallucinations. We first construct a factually weak LLM by inducing hallucinations from the original LLMs. Then, we penalize these induced hallucinations during decoding to enhance the factuality of the generated content. Concretely, we determine the final next-token predictions by amplifying the predictions from the original model and downplaying the induced untruthful predictions via contrastive decoding. Experimental results on both discrimination-based and generation-based hallucination evaluation benchmarks, such as TruthfulQA and FActScore, demonstrate that our proposed ICD methods can effectively enhance the factuality of LLMs across various task formats, model sizes, and model families. For example, when equipped with ICD, Llama2-7B-Chat and Mistral-7B-Instruct achieve performance comparable to ChatGPT and GPT4 on TruthfulQA, respectively, without compromising their generalization capabilities on other tasks.