Large language models offer transformative potential for healthcare, yet their responsible and equitable development depends critically on a deeper understanding of how training data characteristics influence model behavior, including the potential for bias. Current practices in dataset curation and bias assessment often lack the necessary transparency, creating an urgent need for comprehensive evaluation frameworks to foster trust and guide improvements. In this study, we present an in-depth analysis of potential downstream biases in clinical language models, with a focus on differential opioid prescription tendencies across diverse demographic groups, such as ethnicity, gender, and age. As part of this investigation, we introduce HC4: Healthcare Comprehensive Commons Corpus, a novel and extensively curated pretraining dataset exceeding 89 billion tokens. Our evaluation leverages both established general benchmarks and a novel, healthcare-specific methodology, offering crucial insights to support fairness and safety in clinical AI applications.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
The rapid advancement of large language models (LLMs) has heightened concerns about benchmark data contamination (BDC), where models inadvertently memorize evaluation data during the training process, inflating performance metrics, and undermining genuine generalization assessment. This paper introduces the Data Contamination Risk (DCR) framework, a lightweight, interpretable pipeline designed to detect and quantify BDC risk across four granular levels: semantic, informational, data, and label. By synthesizing contamination scores via a fuzzy inference system, DCR produces a unified DCR Factor that adjusts raw accuracy to reflect contamination-aware performance. Validated on 9 LLMs (0.5B-72B) across sentiment analysis, fake news detection, and arithmetic reasoning tasks, the DCR framework reliably diagnoses contamination severity and with accuracy adjusted using the DCR Factor to within 4% average error across the three benchmarks compared to the uncontaminated baseline. Emphasizing computational efficiency and transparency, DCR provides a practical tool for integrating contamination assessment into routine evaluations, fostering fairer comparisons and enhancing the credibility of LLM benchmarking practices.
WildDoc: How Far Are We from Achieving Comprehensive and Robust Document Understanding in the Wild?
PDF ↗The rapid advancements in Multimodal Large Language Models (MLLMs) have significantly enhanced capabilities in Document Understanding. However, prevailing benchmarks like DocVQA and ChartQA predominantly comprise scanned or digital documents, inadequately reflecting the intricate challenges posed by diverse real-world scenarios such as variable illumination and physical distortions. This paper introduces WildDoc, the inaugural benchmark designed specifically for assessing document understanding in natural environments. WildDoc incorporates a diverse set of manually captured document images reflecting real-world conditions and leverages document sources from established benchmarks to facilitate comprehensive comparisons with digital or scanned documents. Further, to rigorously evaluate model robustness, each document is captured four times under different conditions. Evaluations of state-of-the-art MLLMs on WildDoc expose substantial performance declines and underscore the models’ inadequate robustness compared to traditional benchmarks, highlighting the unique challenges posed by real-world document understanding.
In an era where misinformation spreads freely, fact-checking (FC) plays a crucial role in verifying claims and promoting reliable information. While automated fact-checking (AFC) has advanced significantly, existing systems remain vulnerable to adversarial attacks that manipulate or generate claims, evidence, or claim-evidence pairs. These attacks can distort the truth, mislead decision-makers, and ultimately undermine the reliability of FC models. Despite growing research interest in adversarial attacks against AFC systems, a comprehensive, holistic overview of key challenges remains lacking. These challenges include understanding attack strategies, assessing the resilience of current models, and identifying ways to enhance robustness. This survey provides the first in-depth review of adversarial attacks targeting FC, categorizing existing attack methodologies and evaluating their impact on AFC systems. Additionally, we examine recent advancements in adversary-aware defenses and highlight open research questions that require further exploration. Our findings underscore the urgent need for resilient FC frameworks capable of withstanding adversarial manipulations in pursuit of preserving high verification accuracy.
Large Language Models (LLMs) are increasingly costly to fine-tune due to their size, with embedding layers alone accounting for up to 20% of model parameters. While Parameter-Efficient Fine-Tuning (PEFT) methods exist, they largely overlook the embedding layer. In this paper, we introduce TinyTE, a novel PEFT approach that steers model behavior via minimal translational transformations in the embedding space. TinyTE modifies input embeddings without altering hidden layers, achieving competitive performance while requiring approximately 0.0001% of the parameters needed for full fine-tuning. Experiments across architectures provide a new lens for understanding the relationship between input representations and model behavior—revealing them to be more flexible at their foundation than previously thought.
This paper presents R-BPE, a lightweight framework for adapting existing Byte-Pair Encoding (BPE) tokenizers to better support a specified target language. It reuses tokens from user-excluded languages and creates ID-based maps to resolve the new tokens of the chosen language. We evaluate R-BPE on Arabic as a target language. R-BPE reduced subword fertility by an average of 24.4% across the LLaMA 3.1 8B, Command R 35B, and Qwen 3 8B models. Applied to LLaMA 3.1 8B in continued pretraining mode, R-BPE yields a 7.33% reduction in training time. On the ArabicMMLU benchmark, the resulting model improved by 5.09 points on five in-domain topics and matched the original model’s overall performance. It also preserved performance on EnglishMMLU. R-BPE effectively leverages existing models’ tokenizers, embedding layers, and performance to better support target languages without incurring model size changes. We release an R-BPE implementation that is compatible with HuggingFace interfaces and thereby readily applicable to a wide range of existing models at https://acr.ps/1L9GPmL.
Maintaining mutual understanding is a key component in human-human conversation to avoid conversation breakdowns, in which repair, particularly Other-Initiated Repair (OIR, when one speaker signals trouble and prompts the other to resolve), plays a vital role. However, Conversational Agents (CAs) still fail to recognize user repair initiation, leading to breakdowns or disengagement. This work proposes a multimodal model to automatically detect repair initiation in Dutch dialogues by integrating linguistic and prosodic features grounded in Conversation Analysis. The results show that prosodic cues complement linguistic features and significantly improve the results of pretrained text and audio embeddings, offering insights into how different features interact. Future directions include incorporating visual cues, exploring multilingual and cross-context corpora to assess the robustness and generalizability.
Conventional retrieval-augmented generation(RAG) systems employ rigid retrieval strategies that create: (1) knowledge blind spots across domain boundaries, (2) reasoning fragmentation when processing interdependent concepts, and (3) contradictions from conflicting evidence sources. Motivated by these limitations, we introduce PathwiseRAG, which addresses these challenges through: intent-aware strategy selection to eliminate blind spots, dynamic reasoning networks that capture sub-problem interdependencies to overcome fragmentation, and parallel path exploration with adaptive refinement to resolve conflicts. The framework models query intent across semantic and reasoning dimensions, constructs a directed acyclic graph of interconnected sub-problems, and explores multiple reasoning trajectories while continuously adapting to emerging evidence. Evaluation across challenging benchmarks demonstrates significant improvements over state-of-the-art RAG systems, with average accuracy gains of 4.9% and up to 6.9% on complex queries, establishing a new paradigm for knowledge-intensive reasoning by transforming static retrieval into dynamic, multi-dimensional exploration.
Mixture of Weight-shared Heterogeneous Group Attention Experts for Dynamic Token-wise KV Optimization
PDF ↗Transformer models face scalability challenges in causal language modeling (CLM) due to inefficient memory allocation for growing key-value (KV) caches, which strains compute and storage resources. Existing methods like Grouped Query Attention (GQA) and token-level KV optimization improve efficiency but rely on rigid resource allocation, often discarding “low-priority” tokens or statically grouping them, failing to address the dynamic spectrum of token importance. We propose mixSGA, a novel mixture-of-expert (MoE) approach that dynamically optimizes token-wise computation and memory allocation. Unlike prior approaches, mixSGA retains all tokens while adaptively routing them to specialized experts with varying KV group sizes, balancing granularity and efficiency. Our key novelties include: (1) a token-wise expert-choice routing mechanism guided by learned importance scores, enabling proportional resource allocation without token discard; (2) weight-sharing across grouped attention projections to minimize parameter overhead; and (3) an auxiliary loss to ensure one-hot routing decisions for training-inference consistency in CLMs. Extensive evaluations across Llama3, TinyLlama, OPT, and Gemma2 model families show mixSGA’s superiority over static baselines. On instruction-following and continued pretraining tasks, mixSGA achieves higher ROUGE-L and lower perplexity under the same KV budgets.
A recent study showed that large language models (LLMs) can reconstruct surprisingly long texts — up to thousands of tokens — via autoregressive generation from just one trained input embedding. In this work, we explore whether autoregressive decoding is essential for such reconstruction. We show that frozen LLMs can generate hundreds of accurate tokens in just one token-parallel forward pass, when provided with only two learned embeddings. This reveals a surprising and underexplored multi-token generation capability of autoregressive LLMs. We examine these embeddings and characterize the information they encode. We also empirically show that, although these representations are not unique for a given text, they form connected and local regions in embedding space — suggesting the potential to train a practical encoder. The existence of such representations hints that multi-token generation may be natively accessible in off-the-shelf LLMs via a learned input encoder, eliminating heavy retraining and helping to overcome the fundamental bottleneck of autoregressive decoding while reusing already-trained models.
Machine Translation (MT) tools are widely used today, often in contexts where professional translators are not present. Despite progress in MT technology, a gap persists between system development and real-world usage, particularly for non-expert users who may struggle to assess translation reliability.This paper advocates for a human-centered approach to MT, emphasizing the alignment of system design with diverse communicative goals and contexts of use. We survey the literature in Translation Studies and Human-Computer Interaction to recontextualize MT evaluation and design to address the diverse real-world scenarios in which MT is used today.
Linguistic and Embedding-Based Profiling of Texts Generated by Humans and Large Language Models
PDF ↗The rapid advancements in large language models (LLMs) have significantly improved their ability to generate natural language, making texts generated by LLMs increasingly indistinguishable from human-written texts. While recent research has primarily focused on using LLMs to classify text as either human-written or machine-generated texts, our study focuses on characterizing these texts using a set of linguistic features across different linguistic levels such as morphology, syntax, and semantics. We select a dataset of human-written and machine-generated texts spanning 8 domains and produced by 11 different LLMs. We calculate different linguistic features such as dependency length and emotionality, and we use them for characterizing human-written and machine-generated texts along with different sampling strategies, repetition controls, and model release dates. Our statistical analysis reveals that human-written texts tend to exhibit simpler syntactic structures and more diverse semantic content. Furthermore, we calculate the variability of our set of features across models and domains. Both human- and machine-generated texts show stylistic diversity across domains, with human-written texts displaying greater variation in our features. Finally, we apply style embeddings to further test variability among human-written and machine-generated texts. Notably, newer models output text that is similarly variable, pointing to a homogenization of machine-generated texts.
Accurately measuring gender stereotypical bias in language models is a complex task with many hidden aspects. Current benchmarks have underestimated this multifaceted challenge and failed to capture the full extent of the problem. This paper examines the inconsistencies between intrinsic stereotype benchmarks. We propose that currently available benchmarks each capture only partial facets of gender stereotypes, and when considered in isolation, they provide just a fragmented view of the broader landscape of bias in language models. Using StereoSet and CrowS-Pairs as case studies, we investigated how data distribution affects benchmark results. By applying a framework from social psychology to balance the data of these benchmarks across various components of gender stereotypes, we demonstrated that even simple balancing techniques can significantly improve the correlation between different measurement approaches. Our findings underscore the complexity of gender stereotyping in language models and point to new directions for developing more refined techniques to detect and reduce bias.
Cross-lingual context retrieval (extracting contextual information in one language based on requests in another) is a fundamental aspect of cross-lingual alignment, but the performance and mechanism of it for large language models (LLMs) remains unclear. In this paper, we evaluate the cross-lingual context retrieval of over 40 LLMs across 12 languages, using cross-lingual machine reading comprehension (xMRC) as a representative scenario. Our results show that post-trained open LLMs show strong cross-lingual context retrieval ability, comparable to closed-source LLMs such as GPT-4o, and their estimated oracle performances greatly improve after post-training. Our mechanism analysis shows that the cross-lingual context retrieval process can be divided into two main phases: question encoding and answer retrieval, which are formed in pre-training and post-training respectively. The phasing stability correlates with xMRC performance, and the xMRC bottleneck lies at the last model layers in the second phase, where the effect of post-training can be evidently observed. Our results also indicate that larger-scale pretraining cannot improve the xMRC performance. Instead, larger LLMs need further multilingual post-training to fully unlock their cross-lingual context retrieval potential.
C3: A Bilingual Benchmark for Spoken Dialogue Models Exploring Challenges in Complex Conversations
PDF ↗Spoken Dialogue Models (SDMs) have recently attracted significant attention for their ability to generate voice responses directly to users’ spoken queries. Despite their increasing popularity, there exists a gap in research focused on comprehensively understanding their practical effectiveness in comprehending and emulating human conversations. This is especially true compared to text-based Large Language Models (LLMs), which benefit from extensive benchmarking. Human voice interactions are inherently more complex than text due to characteristics unique to spoken dialogue. Ambiguity poses one challenge, stemming from semantic factors like polysemy, as well as phonological aspects such as heterograph, heteronyms, and stress patterns. Additionally, context-dependency, like omission, coreference, and multi-turn interaction, adds further complexity to human conversational dynamics. To illuminate the current state of SDM development and to address these challenges, we present a benchmark dataset in this paper, which comprises 1,079 instances in English and Chinese. Accompanied by an LLM-based evaluation method that closely aligns with human judgment, this dataset facilitates a comprehensive exploration of the performance of SDMs in tackling these practical challenges.
Job Unfair: An Investigation of Gender and Occupational Bias in Free-Form Text Completions by LLMs
PDF ↗Disentangling how gender and occupations are encoded by LLMs is crucial to identify possible biases and prevent harms, especially given the widespread use of LLMs in sensitive domains such as human resources.In this work, we carry out an in-depth investigation of gender and occupational biases in English and Italian as expressed by 9 different LLMs (both base and instruction-tuned). Specifically, we focus on the analysis of sentence completions when LLMs are prompted with job-related sentences including different gender representations. We carry out a manual analysis of 4,500 generated texts over 4 dimensions that can reflect bias, we propose a novel embedding-based method to investigate biases in generated texts and, finally, we carry out a lexical analysis of the model completions. In our qualitative and quantitative evaluation we show that many facets of social bias remain unaccounted for even in aligned models, and LLMs in general still reflect existing gender biases in both languages. Finally, we find that models still struggle with gender-neutral expressions, especially beyond English.
Multimodal aspect-based sentiment classification (MABSC) requires fine-grained reasoning over both textual and visual content to infer sentiments toward specific aspects. However, existing methods often rely on superficial correlations—particularly between aspect terms and sentiment labels—leading to poor generalization and vulnerability to spurious cues. To address this limitation, we propose DPCI, a novel Dual-Path Counterfactual Integration framework that enhances model robustness by explicitly modeling counterfactual reasoning in multimodal contexts. Specifically, we design a dual counterfactual generation module that simulates two types of interventions: replacing aspect terms and rewriting descriptive content, thereby disentangling the spurious dependencies from causal sentiment cues. We further introduce a sample-aware counterfactual selection strategy to retain high-quality, diverse counterfactuals tailored to each generation path. Finally, a confidence-guided integration mechanism adaptively fuses counterfactual signals into the main prediction stream. Extensive experiments on standard MABSC benchmarks demonstrate that DPCI not only achieves state-of-the-art performance but also significantly improves model robustness.
LoRA has become one of the most widely used parameter-efficient fine-tuning methods due to its simplicity and effectiveness. However, numerous studies have shown that LoRA often introduces substantial parameter redundancy, which not only increases the number of trainable parameters but also hinders the effectiveness of fine-tuning. Since identifying redundant parameters in LoRA is inherently difficult, how to eliminate them efficiently and accurately remains a challenging problem. In this paper, we propose TASO, a redundancy reduction method that leverages importance information from the pretrained model’s weights to mitigate LoRA redundancy. Specifically, we estimate parameter importance on downstream tasks and identify task-specific core regions based on the distribution of importance scores. The location information of these core regions is then used to determine the sparse structure of LoRA modules, enabling redundancy removal before fine-tuning. Our approach significantly reduces the number of trainable parameters required for task adaptation, while providing a novel task-aligned perspective for LoRA redundancy reduction. Experimental results demonstrate that, with a parameter budget comparable to LoRA with rank r = 1, TASO consistently outperforms standard LoRA across multiple tasks, achieving strong fine-tuning performance while effectively eliminating redundant parameters.
This work presents a computational approach to analyze character development along the narrative timeline. The analysis characterizes changes in the protagonist’s views and behavior and the interplay between them. We consider transcripts of Holocaust survivor testimonies as a test case, each telling the story of an individual in first-person terms. We focus on the survivor’s religious trajectory, examining the evolution of their disposition toward religious belief and practice as it is reflected in the testimony. Clustering the resulting trajectories in the dataset, we identify common sequences in the data. Our findings highlight multiple common structures of religiosity across the narratives: in terms of belief, a constant disposition is common, while for practice, most present an oscillating structure, serving as valuable material for historical and sociological research. This work demonstrates the potential of natural language processing for analyzing character evolution through thematic trajectories in narratives.
Current benchmarks for evaluating Large Language Models (LLMs) often do not exhibit enough writing style diversity, with many adhering primarily to standardized conventions. Such benchmarks do not fully capture the rich variety of communication patterns exhibited by humans. Thus, it is possible that LLMs, which are optimized on these benchmarks, may demonstrate brittle performance when faced with “non-standard” input. In this work, we test this hypothesis by rewriting evaluation prompts using persona-based LLM prompting, a low-cost method to emulate diverse writing styles. Our results show that, even with identical semantic content, variations in writing style and prompt formatting significantly impact the estimated performance of the LLM under evaluation. Notably, we identify distinct writing styles that consistently trigger either low or high performance across a range of models and tasks, irrespective of model family, size, or recency. Our work offers a scalable approach to augment existing benchmarks, improving the external validity of the assessments they provide for LLM performance across linguistic variations.