论文检索

Alexander Scarlatos, Nigel Fernandez, Christopher Ormerod, Susan Lottridge, Andrew Lan

Item (question) difficulties play a crucial role in educational assessments, enabling accurate and efficient assessment of student abilities and personalization to maximize learning outcomes. Traditionally, estimating item difficulties can be costly, requiring real students to respond to items, followed by fitting an item response theory (IRT) model to get difficulty estimates. This approach cannot be applied to the cold-start setting for previously unseen items either. In this work, we present SMART (Simulated Students Aligned with IRT), a novel method for aligning simulated students with instructed ability, which can then be used in simulations to predict the difficulty of open-ended items. We achieve this alignment using direct preference optimization (DPO), where we form preference pairs based on how likely responses are under a ground-truth IRT model. We perform a simulation by generating thousands of responses, evaluating them with a large language model (LLM)-based scoring model, and fit the resulting data to an IRT model to obtain item difficulty estimates. Through extensive experiments on two real-world student response datasets, we show that SMART outperforms other item difficulty prediction methods by leveraging its improved ability alignment.

Colten DiIanni, Daniel Deutsch

This paper introduces Pairwise Difference Pearson (PDP), a novel segment-level meta-evaluation metric for Machine Translation (MT) that addresses limitations in previous Pearson’s \rho-based and Kendall’s \tau-based meta-evaluation approaches. PDP is a correlation-based metric that utilizes pairwise differences rather than raw scores. It draws on information from all segments for a more robust understanding of score distributions and uses only pairwise differences to refine Global Pearson to intra-segment comparisons. Analysis on the WMT’24 shared task shows PDP properly ranks sentinel evaluation metrics and better aligns with human error weightings than acc_{eq}.

Yiwei Liu, Emma Jane Pretty, Jiahao Huang, Saku Sugawara

While recent studies explore Large Language Models’ (LLMs) performance on Theory of Mind (ToM) reasoning tasks, research on ToM abilities that require more nuanced social context is limited, such as white lies. We introduce TactfulToM, a novel English benchmark designed to evaluate LLMs’ ability to understand white lies within real-life conversations and reason about prosocial motivations behind them, particularly when they are used to spare others’ feelings and maintain social harmony. Our benchmark is generated through a multi-stage human-in-the-loop pipeline where LLMs expand manually designed seed stories into conversations to maintain the information asymmetry between participants necessary for authentic white lies. We show that TactfulToM is challenging for state-of-the-art models, which perform substantially below humans, revealing shortcomings in their ability to fully comprehend the ToM reasoning that enables true understanding of white lies.

Sasha Boguraev, Christopher Potts, Kyle Mahowald

Language Models (LMs) have emerged as powerful sources of evidence for linguists seeking to develop theories of syntax. In this paper, we argue that causal interpretability methods, applied to LMs, can greatly enhance the value of such evidence by helping us characterize the abstract mechanisms that LMs learn to use. Our empirical focus is a set of English filler–gap dependency constructions (e.g., questions, relative clauses). Linguistic theories largely agree that these constructions share many properties. Using experiments based in Distributed Interchange Interventions, we show that LMs converge on similar abstract analyses of these constructions. These analyses also reveal previously overlooked factors – relating to frequency, filler type, and surrounding context – that could motivate changes to standard linguistic theory. Overall, these results suggest that mechanistic, internal analyses of LMs can push linguistic theory forward.

Davis Brown, Prithvi Balehannina, Helen Jin, Shreya Havaldar, Hamed Hassani, Eric Wong

Language model evaluations often fail to characterize consequential failure modes, forcing experts to inspect outputs and build new benchmarks. We introduce task elicitation, a method that automatically builds new evaluations to profile model behavior. Task elicitation finds hundreds of natural-language tasks—an order of magnitude more than prior work—where frontier models exhibit systematic failures, in domains ranging from forecasting to online harassment. For example, we find that Sonnet 3.5 over-associates quantum computing and AGI and that o3-mini is prone to hallucination when fabrications are repeated in-context.

Sujoy Sarkar, Gourav Sarkar, Manoj Balaji Jagadeeshan, Jivnesh Sandhan, Amrith Krishna, Pawan Goyal

High lexical variation, ambiguous references, and long-range dependencies make entity resolution in literary texts particularly challenging. We present Mahānāma, the first large-scale dataset for end-to-end Entity Discovery and Linking (EDL) in Sanskrit, a morphologically rich and under-resourced language. Derived from the Mahābhārata , the world’s longest epic, the dataset comprises over 109K named entity mentions mapped to 5.5K unique entities, and is aligned with an English knowledge base to support cross-lingual linking. The complex narrative structure of Mahānāma, coupled with extensive name variation and ambiguity, poses significant challenges to resolution systems. Our evaluation reveals that current coreference and entity linking models struggle when evaluated on the global context of the test set. These results highlight the limitations of current approaches in resolving entities within such complex discourse. Mahānāma thus provides a unique benchmark for advancing entity resolution, especially in literary domains.

Hao Xu, Jiacheng Liu, Yejin Choi, Noah A. Smith, Hannaneh Hajishirzi

Language models are trained mainly on massive text data from the Internet, and it becomes increasingly important to understand this data source. Exact-match search engines enable searching in large text corpora – counting string appearances and retrieving the enclosing documents – yet the high storage overhead hinders their application on Internet-scale data. We present Infini-gram mini, an efficient and scalable system that can make petabyte-level text corpora searchable. Based on the FM-index data structure (Ferragina and Manzini, 2000), which simultaneously indexes and compresses text, our system creates indexes with size only 44% of the corpus. Infini-gram mini greatly improves upon the best existing implementation of FM-index in terms of indexing speed (18\times) and memory use during both indexing (3.2\times reduction) and querying (down to a negligible amount). We index 83TB of Internet text in 99 days with a single 128-core CPU node (or 19 hours if using 137 such nodes). We show one important use case of Infini-gram mini in a large-scale analysis of benchmark contamination. We find several core LM evaluation benchmarks to be heavily contaminated in Internet crawls (up to 74.2% in GSM8K), which could lead to overestimating the capabilities of language models if trained on such data. We host a benchmark contamination bulletin to share the contamination rate of many core and community-contributed benchmarks. We also release a web interface and an API endpoint to serve general search queries on Infini-gram mini indexes.

Kun Li, Yunxiang Li, Tianhua Zhang, Hongyin Luo, Xixin Wu, James R. Glass, Helen M. Meng

Robust evaluation is critical for deploying trustworthy retrieval-augmented generation (RAG) systems. However, current LLM-based evaluation frameworks predominantly rely on directly prompting resource-intensive models with complex multi-stage prompts, underutilizing models’ reasoning capabilities and introducing significant computational cost. In this paper, we present RAG-Zeval (RAG-Zero Evaluator), a novel end-to-end framework that formulates faithfulness and correctness evaluation of RAG systems as a rule-guided reasoning task. Our approach trains evaluators with reinforcement learning, facilitating compact models to generate comprehensive and sound assessments with detailed explanation in one-pass. We introduce a ranking-based outcome reward mechanism, using preference judgments rather than absolute scores, to address the challenge of obtaining precise pointwise reward signals. To this end, we synthesize the ranking references by generating quality-controlled responses with zero human annotation. Experiments demonstrate RAG-Zeval’s superior performance, achieving the strongest correlation with human judgments and outperforming baselines that rely on LLMs with 10-100\times more parameters. Our approach also exhibits superior interpretability in response evaluation.

Dingdong Wang, Junan Li, Mingyu Cui, Dongchao Yang, Xueyuan Chen, Helen M. Meng

With the rise of Speech Large Language Models (SpeechLLMs), two dominant approaches have emerged for speech processing: discrete tokens and continuous features. Each approach has demonstrated strong capabilities in audio-related processing tasks. However, the performance gap between these two paradigms has not been thoroughly explored. To address this gap, we present a fair comparison of self-supervised learning (SSL)-based discrete and continuous features under the same experimental settings. We evaluate their performance across six spoken language understanding-related tasks using both small and large-scale LLMs (Qwen1.5-0.5B and Llama3.1-8B). We further conduct in-depth analyses, including efficient comparison, SSL layer analysis, LLM layer analysis, and robustness comparison. Our findings reveal that continuous features generally outperform discrete tokens in various tasks. Each speech processing method exhibits distinct characteristics and patterns in how it learns and processes speech information. We hope our results will provide valuable insights to advance spoken language understanding in SpeechLLMs.

Hanxi Guo, Siyuan Cheng, Xiaolong Jin, Zhuo Zhang, Guangyu Shen, Kaiyuan Zhang, Shengwei An, Guanhong Tao, Xiangyu Zhang

With the increasing capabilities of Large Language Models (LLMs), the proliferation of AI-generated texts has become a serious concern. Given the diverse range of organizations providing LLMs, it is crucial for governments and third-party entities to identify the origin LLM of a given AI-generated text to enable accurate mitigation of potential misuse and infringement. However, existing detection methods, primarily designed to distinguish between human-generated and LLM-generated texts, often fail to accurately identify the origin LLM due to the high similarity of AI-generated texts from different LLMs. In this paper, we propose a novel black-box AI-generated text origin detection method, dubbed Profiler, which accurately predicts the origin of an input text by extracting distinct context inference patterns through calculating and analyzing novel context losses between the surrogate model’s output logits and the adjacent input context. Extensive experimental results show that Profiler outperforms 10 state-of-the-art baselines, achieving more than a 25% increase in AUC score on average across both natural language and code datasets when evaluated against five of the latest commercial LLMs under both in-distribution and out-of-distribution settings.

Leena Mathur, Marian Qian, Paul Pu Liang, Louis-Philippe Morency

Social reasoning abilities are crucial for AI systems to effectively interpret and respond to multimodal human communication and interaction within social contexts. We introduce Social Genome, the first benchmark for fine-grained, grounded social reasoning abilities of multimodal models. Social Genome contains 272 videos of interactions and 1,486 human-annotated reasoning traces related to inferences about these interactions. These traces contain 5,777 reasoning steps that reference evidence from visual cues, verbal cues, vocal cues, and external knowledge (contextual knowledge external to videos). Social Genome is also the first modeling challenge to study external knowledge in social reasoning. Social Genome computes metrics to holistically evaluate semantic and structural qualities of model-generated social reasoning traces. We demonstrate the utility of Social Genome through experiments with state-of-the-art models, identifying performance gaps and opportunities for future research to improve the grounded social reasoning abilities of multimodal models.

Benyamin Jamialahmadi, Parsa Kavehzadeh, Mehdi Rezagholizadeh, Parsa Farinneya, Hossein Rajabzadeh, Aref Jafari, Boxing Chen, Marzieh S. Tahaei

Deploying large language models (LLMs) in real-world applications is often hindered by strict computational and latency constraints. While dynamic inference offers the flexibility to adjust model behavior based on varying resource budgets, existing methods are frequently limited by hardware inefficiencies or performance degradation. In this paper, we introduce Balcony, a simple yet highly effective framework for depth-based dynamic inference. By freezing the pretrained LLM and inserting additional transformer layers at selected exit points, Balcony maintains the full model’s performance while enabling real-time adaptation to different computational budgets. These additional layers are trained using a straightforward self-distillation loss, aligning the sub-model outputs with those of the full model. This approach requires significantly fewer training tokens and tunable parameters, drastically reducing computational costs compared to prior methods. When applied to the LLaMA3-8B model, using only 0.2% of the original pretraining data, Balcony achieves minimal performance degradation while enabling significant speedups. Remarkably, we show that Balcony outperforms state-of-the-art methods such as Flextron and Layerskip on multiple models at various scales, as well as other leading compression techniques across a variety of benchmarks.

Michal Golovanevsky, William Rudman, Michael A. Lepori, Amir Bar, Ritambhara Singh, Carsten Eickhoff

Multimodal Large Language Models (MLLMs) perform well on tasks such as visual question answering, but it remains unclear whether their reasoning relies more on memorized world knowledge or on the visual information present in the input image. To investigate this, we introduce Visual CounterFact, a new dataset of visually-realistic counterfactuals that put world knowledge priors (e.g, red strawberry) into direct conflict with visual input (e.g, blue strawberry). Using Visual CounterFact, we show that model predictions initially reflect memorized priors, but shift toward visual evidence in mid-to-late layers. This dynamic reveals a competition between the two modalities, with visual input ultimately overriding priors during evaluation. To control this behavior, we propose Pixels Versus Priors (PvP) steering vectors, a mechanism for controlling model outputs toward either world knowledge or visual input through activation-level interventions. On average, PvP successfully shifts 99.3% of color and 80.8% of size predictions from priors to counterfactuals. Together, these findings offer new tools for interpreting and controlling factual behavior in multimodal models.

Sai Sundaresan, Harshita Chopra, Atanu R. Sinha, Koustava Goswami, Nagasai Saketh Naidu, Raghav Karan, N Anushka

A Large Language Model (LLM) offers versatility across domains and tasks, purportedly benefiting users with a wide variety of behaviors and preferences. We question this perception about an LLM when users have inherently subjective behaviors and preferences, as seen in their ubiquitous and idiosyncratic browsing of websites or apps. The sequential behavior logs of pages, thus generated, form something akin to each user’s self-constructed “language”, albeit without the structure and grammar imbued in natural languages. We ask: (i) Can a small LM represent the “language of browsing” better than a large LM? (ii) Can an LM with a single set of parameters (or, single LM) adequately capture myriad users’ heterogeneous, subjective behaviors and preferences? (iii) Can a single LM with high average performance, yield low variance in performance to make alignment good at user level? We introduce clusterwise LM training, HeTLM (Heterogeneity aware Training of Language Model), appropriate for subjective behaviors. We find that (i) a small LM trained using a page-level tokenizer outperforms large pretrained or finetuned LMs; (ii) HeTLM with heterogeneous cluster specific set of parameters outperforms a single LM of the same family, controlling for the number of parameters; and (iii) a higher mean and a lower variance in generation ensues, implying improved alignment.

Dongwon Jung, Wenxuan Zhou, Muhao Chen

Training large language models (LLMs) with chain-of-thought (CoT) supervision has proven effective for enhancing their reasoning abilities. However, obtaining reliable and accurate reasoning supervision remains a significant challenge. We propose a scalable method for generating a high-quality CoT supervision dataset by leveraging the determinism of program execution. Unlike existing reasoning dataset generation methods that rely on costly human annotations or error-prone LLM-generated CoT, our approach extracts verifiable, step-by-step reasoning traces from code execution and transforms them into a natural language CoT reasoning. Experiments on reasoning benchmarks across various domains show that our method effectively equips LLMs with transferable reasoning abilities across diverse tasks. Furthermore, the ablation studies validate that our method produces highly accurate reasoning data and reduces overall token length during inference by reducing meaningless repetition and overthinking.

Jiankun Zhang, Shenglai Zeng, Jie Ren, Tianqi Zheng, Hui Liu, Xianfeng Tang, Hui Liu, Yi Chang

Multimodal Retrieval-Augmented Generation (MRAG) systems enhance LMMs by integrating external multimodal databases, but introduce unexplored privacy vulnerabilities. While text-based RAG privacy risks have been studied, multimodal data presents unique challenges. We provide the first systematic analysis of MRAG privacy vulnerabilities across vision-language and speech-language modalities. Using a novel compositional structured prompt attack in a black-box setting, we demonstrate how attackers can extract private information by manipulating queries. Our experiments reveal that LMMs can both directly generate outputs resembling retrieved content and produce descriptions that indirectly expose sensitive information, highlighting the urgent need for robust privacy-preserving MRAG techniques.

Zhenyun Deng, Yulong Chen, Andreas Vlachos

Extracting individual sentences from a document as evidence or reasoning steps is commonly done in many NLP tasks. However, extracted sentences often lack context necessary to make them understood, e.g., coreference and background information. To this end, we propose a content selection and planning framework for zero-shot decontextualisation, which determines what content should be mentioned and in what order for a sentence to be understood out of context. Specifically, given a potentially ambiguous sentence and its context, we first segment it into basic semantically-independent units. We then identify potentially ambiguous units from the given sentence, and extract relevant units from the context based on their discourse relations. Finally, we generate a content plan to rewrite the sentence by enriching each ambiguous unit with its relevant units. Experimental results demonstrate that our approach is competitive for sentence decontextualisation, producing sentences that exhibit better semantic integrity and discourse coherence, outperforming existing methods.

Amruta Parulekar, Preethi Jyothi

Standard ASR evaluation metrics like Word Error Rate (WER) tend to unfairly penalize morphological and syntactic nuances that do not significantly alter sentence semantics. We introduce an LLM-based scoring rubric LASER that leverages state-of-the-art LLMs’ in-context learning abilities to learn from prompts with detailed examples. Hindi LASER scores using Gemini 2.5 Pro achieved a very high correlation score of 94% with human annotations. Hindi examples in the prompt were also effective in analyzing errors in other Indian languages such as Marathi, Kannada and Malayalam. We also demonstrate how a smaller LLM like Llama 3 can be finetuned on word-pair examples derived from reference and ASR predictions to predict what kind of penalty should be applied with close to 89% accuracy.

Rik Koncel-Kedziorski, Brihi Joshi, Tim Paek

As the adoption of language models advances, so does the need to better represent individual users to the model. Are there aspects of an individual’s belief system that a language model can utilize for improved alignment? Following prior research, we investigate this question in the domain of opinion prediction by developing PrimeX, a dataset of public opinion survey data from 858 US residents with two additional sources of belief information: written explanations from the respondents for why they hold specific opinions, and the Primal World Belief survey for assessing respondent worldview. We provide an extensive initial analysis of our data and show the value of belief explanations and worldview for personalizing language models. Our results demonstrate how the additional belief information in PrimeX can benefit both the NLP and psychological research communities, opening up avenues for further study.

Qianqi Yan, Hongquan Li, Shan Jiang, Yang Zhao, Xinze Guan, Ching-Chen Kuo, Xin Eric Wang

Multimodal large language models (MLLMs) are increasingly deployed in open-ended, real-world environments where inputs are messy, underspecified, and not always trustworthy. Unlike curated benchmarks, these settings frequently involve instructions that reference missing objects or contradictory facts, rely on ambiguous cues, or request infeasible actions. In such cases, success hinges not merely on task execution, but on the model’s ability to detect when something is silently wrong. This paper presents a systematic analysis of how current MLLMs handle such underspecified and misspecified scenarios: cases where flaws must be inferred from context rather than explicitly stated. Using a curated diagnostic suite spanning four categories of real-world failure modes, we evaluate nine MLLMs, including o3 and GPT-4o, and find that models often fail to surface hidden issues, even when they possess the necessary perceptual and reasoning skills. Explicit prompting reveals that the underlying capabilities exist but are frequently suppressed in favor of user compliance.We further show that simple inference-time interventions, such as cautious persona prompting and, in particular, requiring a clarifying question, can substantially recover performance. Our findings highlight a persistent gap between reasoning competence and behavioral compliance in current MLLMs, and suggest practical strategies for making these systems more trustworthy in underconstrained environments.