NAACL 2025 论文列表

Multi-label document classification (MLDC) aims to allocate more than one label to each document and attracts increasing attention in many practical applications. However, previous studies have failed to pay sufficient attention to the lack of semantic information on labels and the long-tail problem prevalent in the datasets. Additionally, most existing methods focus on optimizing document features, overlooking the potential of high-quality label features to enhance classification performance. In this paper, we propose a simple and effective paradigm for MLDC. Regarding the problem of insufficient label information and imbalance in the sample size of categories, we utilize large language models (LLMs) to semantically expand the label content and generate pseudo-samples for the tail categories. To optimize the features of both documents and labels, we design the contrastive learning boosted feature optimization module facilitated by the similarity matrices. Finally, we construct a label-guided feature selection module to incorporate the optimized label features into the input features to provide richer semantic information for the classifier. Extensive experiments have demonstrated that our proposed method significantly outperforms state-of-the-art baselines.

Language agents powered by large language models (LLMs) are increasingly valuable as decision-making tools in domains such as gaming and programming. However, these agents often face challenges in achieving high-level goals without detailed instructions and in adapting to environments where feedback is delayed. In this paper, we present SELFGOAL, a novel automatic approach designed to enhance agents’ capabilities to achieve high-level goals with limited human prior and environmental feedback. The core concept of SELFGOAL involves adaptively breaking down a high-level goal into a tree structure of more practical subgoals during the interaction with environments while identifying the most useful subgoals and progressively updating this structure. Experimental results demonstrate that SELFGOAL significantly enhances the performance of language agents across various tasks, including competitive, cooperative, and deferred feedback environments.

Large language models (LLMs) have achieved remarkable success in various natural language generation (NLG) tasks, but their performance in automatic text evaluation is not yet ready as human replacements. In this paper, we propose SEEval (Self-Explanation in Evaluation), a novel prompt-based text evaluator. Inspired by educational psychology, SEEval incorporates self-explanation, a metacognitive strategy, to enhance automatic text evaluation. Our experimental results show that SEEval, without probability normalization, is able to achieve competitive and often superior performance compared to the two state-of-the-art baselines – G-Eval and Analyze-Rate – across all evaluation dimensions and is 20 times more efficient in terms of run-time. The SEEval method is also generalizable as its results are consistent across three other selected LLMs – Claude 3.5 Sonnet, Command R+, and Mistral-Large 2.

Typical evaluations of Large Language Models (LLMs) report a single metric per dataset, often representing the model’s best-case performance under carefully selected settings. Unfortunately, this approach overlooks model robustness and reliability in real-world applications. For instance, simple paraphrasing of prompts on the MMLU-Pro dataset causes accuracy fluctuations of up to 10%, while reordering answer choices in the AGIEval dataset results in accuracy differences of up to 6.1%. While some studies discuss issues with LLM robustness, there is no unified or centralized framework for evaluating the robustness of language models. To address this gap and consolidate existing research on model robustness, we present SCORE (Systematic COnsistency and Robustness Evaluation), a comprehensive framework for non-adversarial evaluation of LLMs. The SCORE framework evaluates models by repeatedly testing them on the same benchmarks in various setups to give a realistic estimate of their accuracy and consistency. We will make the code publicly available to facilitate further development and research.

We investigate the mechanistic sources of uncertainty in large language models (LLMs), an area with important implications for language model reliability and trustworthiness. To do so, we conduct a series of experiments designed to identify whether the factuality of generated responses and a model’s uncertainty originate in separate or shared circuits in the model architecture. We approach this question by adapting the well-established mechanistic interpretability techniques of causal tracing and zero-ablation to study the effect of different circuits on LLM generations. Our experiments on eight different models and five datasets, representing tasks predominantly requiring factual recall, provide strong evidence that a model’s uncertainty is produced in the same parts of the network that are responsible for the factuality of generated responses.

Conversational Recommender Systems (CRS) proactively engage users in interactive dialogues to elicit user preferences and provide personalized recommendations. Existing methods train Reinforcement Learning (RL)-based agent with greedy action selection or sampling strategy, and may suffer from suboptimal conversational planning. To address this, we present a novel Monte Carlo Tree Search (MCTS)-based CRS framework SAPIENT. SAPIENT consists of a conversational agent (S-agent) and a conversational planner (S-planner). S-planner builds a conversational search tree with MCTS based on the initial actions proposed by S-agent to find conversation plans. The best conversation plans from S-planner are used to guide the training of S-agent, creating a self-training loop where S-agent can iteratively improve its capability for conversational planning. Furthermore, we propose an efficient variant SAPIENT for trade-off between training efficiency and performance. Extensive experiments on four benchmark datasets validate the effectiveness of our approach, showing that SAPIENT outperforms the state-of-the-art baselines. Our code and data are accessible through https://github.com/ninglab/SAPIENT.

The rise of generative chat-based Large Language Models (LLMs) over the past two years has spurred a race to develop systems that promise near-human conversational and reasoning experiences. However, recent studies indicate that the language understanding offered by these models remains limited and far from human-like performance, particularly in grasping the contextual meanings of words—an essential aspect of reasoning. In this paper, we present a simple yet computationally efficient framework for multilingual Word Sense Disambiguation (WSD). Our approach reframes the WSD task as a cluster discrimination analysis over a semantic network refined from BabelNet using group algebra. We validate our methodology across multiple WSD benchmarks, achieving a new state of the art for all languages and tasks, as well as in individual assessments by part of speech. Notably, our model significantly surpasses the performance of current alternatives, even in low-resource languages, while reducing the parameter count by 72%.

In various natural language processing (NLP) tasks, fine-tuning Pre-trained Language Models (PLMs) often leads to the issue of spurious correlations, which negatively impacts performance, particularly when dealing with out-of-distribution data.To address this problem, we propose **SALAD** (**S**tructure **A**ware and **L**LM-driven **A**ugmented **D**ata), a novel approach designed to enhance model robustness and generalization by generating structure-aware and counterfactually augmented data for contrastive learning.Our method leverages a tagging-based approach to generate structure-aware positive samples and utilizes large language models (LLMs) to generate counterfactual negative samples with diverse sentence patterns. By applying contrastive learning, *SALAD* enables the model to focus on learning the structural relationships between key sentence components while minimizing reliance on spurious correlations.We validate our approach through experiments on three tasks: Sentiment Classification, Sexism Detection, and Natural Language Inference. The results demonstrate that *SALAD* not only improves model robustness and performance across different environments but also enhances generalization to out-of-distribution datasets and cross-domain scenarios.

Superposition refers to encoding representations of multiple features within a single neuron, which is common in deep neural networks. This property allows neurons to combine and represent multiple features, enabling the model to capture intricate information and handle complex tasks. Despite promising performance, the model’s interpretability has been diminished. This paper presents a novel approach to enhance model interpretability by regularizing feature superposition. We introduce SAFR, which simply applies regularizations to the loss function to promote monosemantic representations for important tokens while encouraging polysemanticity for correlated token pairs, where important tokens and correlated token pairs are identified via VMASK and attention weights respectively. We evaluate SAFR with a transformer model on two classification tasks. Experiments demonstrate the effectiveness of SAFR in improving model interpretability without compromising prediction performance. Besides, SAFR provides explanations by visualizing the neuron allocation within the intermediate layers.

Large language models (LLMs) have demonstrated remarkable capabilities across various natural language processing (NLP) scenarios, but they still face challenges when handling complex arithmetic and logical reasoning tasks. While Chain-Of-Thought (CoT) reasoning, self-consistency (SC) and self-correction strategies have attempted to guide models in sequential, multi-step reasoning, Multi-agent Debate (MAD) has emerged as a viable approach for enhancing the reasoning capabilities of LLMs. By increasing both the number of agents and the frequency of debates, the performance of LLMs improves significantly. However, this strategy results in a significant increase in token costs, presenting a barrier to scalability. To address this challenge, we introduce a novel sparsification strategy designed to reduce token costs within MAD. This approach minimizes ineffective exchanges of information and unproductive discussions among agents, thereby enhancing the overall efficiency of the debate process. We conduct comparative experiments on multiple datasets across various models, demonstrating that our approach significantly reduces the token costs in MAD to a considerable extent. Specifically, compared to MAD, our approach achieves an impressive reduction of up to 94.5% in token costs while maintaining performance degradation below 2.0%.

Automated construction of shopping cart frommedical prescriptions is a vital prerequisite forscaling up online pharmaceutical servicesin emerging markets due to the high prevalence of paper prescriptionsthat are challenging for customers to interpret.We present RxLens, a multi-step end-end Large Language Model (LLM)-based deployed solutionfor automated pharmacy cart construction comprisingmultiple steps: redaction of Personal Identifiable Information (PII),Optical Character Recognition (OCR), medication extraction, matching against the catalog, and bounding box detection for lineage. Our multi-step design leverages the synergy between retrieval and LLM-based generationto mitigate the vocabulary gaps in LLMs and fuzzy matching errors during retrieval.Empirical evaluation demonstrates that RxLens can yield up to 19% - 40% and 11% - 26% increase in Recall@3 relative to SOTA methods such as Medical Comprehend and vanilla retrieval augmentation of LLMs on handwritten and printed prescriptions respectively.We also explore LLM-based auto-evaluation as an alternative to costly manual annotations and observe a 76% - 100% match relative to human judgements on various tasks.

Text-to-image generation models have gained popularity among users around the world. However, many of these models exhibit a strong bias toward English-speaking cultures, ignoring or misrepresenting the unique characteristics of other language groups, countries, and nationalities. The lack of cultural awareness can reduce the generation quality and lead to undesirable consequences such as unintentional insult, and the spread of prejudice. In contrast to the field of natural language processing, cultural awareness in computer vision has not been explored as extensively. In this paper, we strive to reduce this gap. We propose a RusCode benchmark for evaluating the quality of text-to-image generation containing elements of the Russian cultural code. To do this, we form a list of 19 categories that best represent the features of Russian visual culture. Our final dataset consists of 1250 text prompts in Russian and their translations into English. The prompts cover a wide range of topics, including complex concepts from art, popular culture, folk traditions, famous people’s names, natural objects, scientific achievements, etc. We present the results of a human evaluation of the side-by-side comparison of Russian visual concepts representations using popular generative models.

Large language models (LLMs) still lack delicate controllability over their responses, which is critical to enhancing their performance and the user experience. However, curating supervised fine-tuning (SFT) datasets to improve LLM controllability usually relies on human experts or proprietary LLMs, which requires additional costs. To bridge this gap, we propose Rule-based Data Recycling (RuleR), a data augmentation method incorporating multiple constraints into the original data samples according to predefined rules, which creates new training tasks to consolidate the controllability of LLMs. Instead of creating new data from scratch, RuleR “recycles” existing data by simply applying rule-based edits to their responses and appending the rule-instructions in their original instructions. Experimental results demonstrate RuleR’s effectiveness in improving LLM controllability while maintaining general instruction-following capabilities.

Autoregressive (AR) Transformer-based sequence models are known to have difficulty generalizing to sequences longer than those seen during training. When applied to text-to-speech (TTS), these models tend to drop or repeat words or produce erratic output, especially for longer utterances. In this paper, we introduce enhancements aimed at AR Transformer-based encoder-decoder TTS systems that address these robustness and length generalization issues. Our approach uses an alignment mechanism to provide cross-attention operations with relative location information. The associated alignment position is learned as a latent property of the model via backpropagation and requires no external alignment information during training. While the approach is tailored to the monotonic nature of TTS input-output alignment, it is still able to benefit from the flexible modeling power of interleaved multi-head self- and cross-attention operations. A system incorporating these improvements, which we call Very Attentive Tacotron, matches the naturalness and expressiveness of a baseline T5-based TTS system, while eliminating problems with repeated or dropped words and enabling generalization to any practical utterance length.

There has been significant prior work using templates to study bias against demographic attributes in MLMs. However, these have limitations: they overlook random variability of templates and target concepts analyzed, assume equality amongst templates, and overlook bias quantification. Addressing these, we propose a systematic statistical approach to assess bias in MLMs, using mixed models to account for random effects, pseudo-perplexity weights for sentences derived from templates and quantify bias using statistical effect sizes. Replicating prior studies, we match on bias scores in magnitude and direction with small to medium effect sizes.Next, we explore the novel problem of gender bias in the context of *personality* and *character* traits, across seven MLMs (base and large). We find that MLMs vary; ALBERT is unbiased for binary gender but the most biased for non-binary *neo*, while RoBERTa-large is the most biased for binary gender but shows small to no bias for *neo*. There is some alignment of MLM bias and findings in psychology (human perspective) - in *agreeableness* with RoBERTa-large and *emotional stability* with BERT-large. There is general agreement for the remaining 3 personality dimensions: both sides observe at most small differences across gender. For character traits, human studies on gender bias are limited thus comparisons are not feasible.

While a large body of work inspects language models for biases concerning gender, race, occupation and religion, biases of geographical nature are relatively less explored. Some recent studies benchmark the degree to which large language models encode geospatial knowledge. However, the impact of the encoded geographical knowledge (or lack thereof) on real-world applications has not been documented. In this work, we examine large language models for two common scenarios that require geographical knowledge: (a) travel recommendations and (b) geo-anchored story generation. Specifically, we study five popular language models, and across about 100K travel requests, and 200K story generations, we observe that travel recommendations corresponding to poorer countries are less unique with fewer location references, and stories from these regions more often convey emotions of hardship and sadness compared to those from wealthier nations.

Reward models (RMs) are at the crux of successfully using RLHF to align pretrained models to human preferences, yet there has been relatively little study that focuses on evaluation of those models. Evaluating reward models presents an opportunity to understand the opaque technologies used for alignment of language models and which values are embedded in them. Resources for reward model training and understanding are sparse in the nascent open-source community around them. To enhance scientific understanding of reward models, we present RewardBench, a benchmark dataset and code-base for evaluation. The RewardBench dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety, to benchmark how reward models perform on challenging, structured and out-of-distribution queries. We create specific comparison datasets for RMs that have subtle, but verifiable reasons (e.g. bugs, incorrect facts) why one answer should be preferred to another. On the RewardBench leaderboard, we evaluate RMs trained with a variety of methods, such as the direct MLE training of classifiers and the implicit reward modeling of Direct Preference Optimization (DPO). We present many findings on propensity for refusals, reasoning limitations, and instruction following shortcomings of various reward models towards a better understanding of the RLHF process.

Inference-time computation methods enhance the performance of Large Language Models (LLMs) by leveraging additional computational resources to achieve superior results. Common techniques, such as Best-of-N sampling, Majority Voting, and variants of tree-search algorithm have proven to be effective in boosting the performance of LLMs. These approaches strategically trade increased computational resource for improved model responses. In this work, we proposed DARWIN, an inference-time alignment method that leverage the guidance of a reward model to achieve alignment through reward-guided tree search. Empirical evidences indicates that our method outperform other inference-time alignment methods such as Best-of-N and ARGS on two widely accepted alignment benchmarks AlpacaEval 2 and MT-Bench. Furthermore, we show that our inference-time approach achieves performance comparable to preference-tuned models on both benchmarks, highlighting the effectiveness of trading inference-time compute for enhanced performance during inference.

Online grooming is a severe social threat where sexual predators gradually entrap child victims with subtle and gradual manipulation. Therefore, timely intervention for online grooming is critical for proactive protection. However, previous methods fail to determine the optimal intervention points (i.e., jump to conclusions) as they rely on chat-level risk labels by causing weak supervision of risky utterances. For timely detection, we propose speed control reinforcement learning (SCoRL), incorporating a practical strategy derived from luring communication theory (LCT). To capture the predator’s turn-level entrapment, we use a turn-level risk label based on the LCT. Then, we design a novel speed control reward function that balances the trade-off between speed and accuracy based on turn-level risk label; thus, SCoRL can identify the optimal intervention moment. In addition, we introduce a turn-level metric for precise evaluation, identifying limitations in previously used chat-level metrics. Experimental results show that SCoRL effectively preempted online grooming, offering a more proactive and timely solution. Further analysis reveals that our method enhances performance while intuitively identifying optimal early intervention points.