ChengXiang Zhai

Liliang Ren, Yang Liu, Shuohang Wang et al. · microsoft-research

Recent hybrid models combining Linear State Space Models (SSMs) with self-attention mechanisms have demonstrated impressive results across a range of sequence modeling tasks. However, current approaches apply attention modules statically and uniformly to all elements in the input sequences, leading to sub-optimal quality-efficiency trade-offs. To address this limitation, we introduce Sparse Modular Activation (SMA), a general mechanism enabling neural networks to sparsely and dynamically activate sub-modules for sequence elements in a differentiable manner. Through allowing each element to skip non-activated sub-modules, SMA reduces computation and memory consumption of neural networks at both training and inference stages. To validate the effectiveness of SMA on sequence modeling, we design a novel neural architecture, SeqBoat, which employs SMA to sparsely activate a Gated Attention Unit (GAU) based on the state representations learned from an SSM. By constraining the GAU to only conduct local attention on the activated inputs, SeqBoat can achieve linear inference complexity with theoretically infinite attention span, and provide substantially better quality-efficiency trade-off than the chunking-based models. With experiments on a wide range of tasks, including long sequence modeling, speech classification and language modeling, SeqBoat brings new state-of-the-art results among hybrid models with linear complexity, and reveals the amount of attention needed for each task through the learned sparse activation patterns. Our code is publicly available at https://github.com/renll/SeqBoat.

Liliang Ren, Mankeerat Sidhu, Qi Zeng et al. · microsoft-research

Existing reference-free turn-level evaluation metrics for chatbots inadequately capture the interaction between the user and the system. Consequently, they often correlate poorly with human evaluations. To address this issue, we propose a novel model-agnostic approach that leverages Conditional Pointwise Mutual Information (C-PMI) to measure the turn-level interaction between the system and the user based on a given evaluation dimension. Experimental results on the widely used FED dialogue evaluation dataset demonstrate that our approach significantly improves the correlation with human judgment compared with existing evaluation systems. By replacing the negative log-likelihood-based scorer with our proposed C-PMI scorer, we achieve a relative 62.6% higher Spearman correlation on average for the FED evaluation metric. Our code is publicly available at https://github.com/renll/C-PMI.

Liliang Ren, Zixuan Zhang, Han Wang et al. · microsoft-research

Modern large-scale Pre-trained Language Models (PLMs) have achieved tremendous success on a wide range of downstream tasks. However, most of the LM pre-training objectives only focus on text reconstruction, but have not sought to learn latent-level interpretable representations of sentences. In this paper, we manage to push the language models to obtain a deeper understanding of sentences by proposing a new pre-training objective, Sparse Latent Typing, which enables the model to sparsely extract sentence-level keywords with diverse latent types. Experimental results show that our model is able to learn interpretable latent type categories in a self-supervised manner without using any external knowledge. Besides, the language model pre-trained with such an objective also significantly improves Information Extraction related downstream tasks in both supervised and few-shot settings. Our code is publicly available at: https://github.com/renll/SparseLT.

Kung-Hsiang Huang, ChengXiang Zhai, Heng Ji

Fact-checking has gained increasing attention due to the widespread of falsified information. Most fact-checking approaches focus on claims made in English only due to the data scarcity issue in other languages. The lack of fact-checking datasets in low-resource languages calls for an effective cross-lingual transfer technique for fact-checking. Additionally, trustworthy information in different languages can be complementary and helpful in verifying facts. To this end, we present the first fact-checking framework augmented with cross-lingual retrieval that aggregates evidence retrieved from multiple languages through a cross-lingual retriever. Given the absence of cross-lingual information retrieval datasets with claim-like queries, we train the retriever with our proposed Cross-lingual Inverse Cloze Task (X-ICT), a self-supervised algorithm that creates training instances by translating the title of a passage. The goal for X-ICT is to learn cross-lingual retrieval in which the model learns to identify the passage corresponding to a given translated title. On the X-Fact dataset, our approach achieves 2.23% absolute F1 improvement in the zero-shot cross-lingual setup over prior systems. The source code and data are publicly available at https://github.com/khuangaf/CONCRETE.

Saket Reddy, Ke Yang, ChengXiang Zhai

Mitigating social bias in Large Language Models (LLMs) presents a distinct alignment challenge: unlike verifiable tasks, bias lacks a single ground truth, creating a high-variance, subjective reward landscape. Previous preference-based fine-tuning methods have major trade-offs: Direct Preference Optimization (DPO) is limited by the lack of exploration inherent in offline training, while Proximal Policy Optimization (PPO) can lead to training instability due to potentially unreliable critic estimates. In this paper, we propose BiasGRPO, a framework using Group Relative Policy Optimization (GRPO) to stabilize alignment by normalizing rewards across a group of sampled completions. By substituting the value function with a group-relative baseline, our approach reduces instability while maintaining the exploration benefits of online training. We find that BiasGRPO outperforms DPO and PPO across multiple benchmarks, indicating its effectiveness. To adapt GRPO, we synthetically extend a dataset spanning multiple domains and contexts. We also create and release a custom bias reward model that effectively guides generation while being highly compute-efficient and avoiding knowledge degradation, providing a valuable resource that can be seamlessly integrated into multi-objective RLHF pipelines.

Yu Zhang, Yunyi Zhang, Yucheng Jiang et al. · amazon-science

Given a few seed entities of a certain type (e.g., Software or Programming Language), entity set expansion aims to discover an extensive set of entities that share the same type as the seeds. Entity set expansion in software-related domains such as StackOverflow can benefit many downstream tasks (e.g., software knowledge graph construction) and facilitate better IT operations and service management. Meanwhile, existing approaches are less concerned with two problems: (1) How to deal with multiple types of seed entities simultaneously? (2) How to leverage the power of pre-trained language models (PLMs)? Being aware of these two problems, in this paper, we study the entity set co-expansion task in StackOverflow, which extracts Library, OS, Application, and Language entities from StackOverflow question-answer threads. During the co-expansion process, we use PLMs to derive embeddings of candidate entities for calculating similarities between entities. Experimental results show that our proposed SECoExpan framework outperforms previous approaches significantly.

Krisztian Balog, ChengXiang Zhai

Information access systems, such as search engines, recommender systems, and conversational assistants, have become integral to our daily lives as they help us satisfy our information needs. However, evaluating the effectiveness of these systems presents a long-standing and complex scientific challenge. This challenge is rooted in the difficulty of assessing a system's overall effectiveness in assisting users to complete tasks through interactive support, and further exacerbated by the substantial variation in user behaviour and preferences. To address this challenge, user simulation emerges as a promising solution. This book focuses on providing a thorough understanding of user simulation techniques designed specifically for evaluation purposes. We begin with a background of information access system evaluation and explore the diverse applications of user simulation. Subsequently, we systematically review the major research progress in user simulation, covering both general frameworks for designing user simulators, utilizing user simulation for evaluation, and specific models and algorithms for simulating user interactions with search engines, recommender systems, and conversational assistants. Realizing that user simulation is an interdisciplinary research topic, whenever possible, we attempt to establish connections with related fields, including machine learning, dialogue systems, user modeling, and economics. We end the book with a detailed discussion of important future research directions, many of which extend beyond the evaluation of information access systems and are expected to have broader impact on how to evaluate interactive intelligent systems in general.

Kevin Pei, Ishan Jindal, Kevin Chen-Chuan Chang et al. · ibm-research

Open Information Extraction (OpenIE) has been used in the pipelines of various NLP tasks. Unfortunately, there is no clear consensus on which models to use in which tasks. Muddying things further is the lack of comparisons that take differing training sets into account. In this paper, we present an application-focused empirical survey of neural OpenIE models, training sets, and benchmarks in an effort to help users choose the most suitable OpenIE systems for their applications. We find that the different assumptions made by different models and datasets have a statistically significant effect on performance, making it important to choose the most appropriate model for one's applications. We demonstrate the applicability of our recommendations on a downstream Complex QA application.

Daniel Campos, ChengXiang Zhai, Alessandro Magnani

The success of contextual word representations and advances in neural information retrieval have made dense vector-based retrieval a standard approach for passage and document ranking. While effective and efficient, dual-encoders are brittle to variations in query distributions and noisy queries. Data augmentation can make models more robust but introduces overhead to training set generation and requires retraining and index regeneration. We present Contrastive Alignment POst Training (CAPOT), a highly efficient finetuning method that improves model robustness without requiring index regeneration, the training set optimization, or alteration. CAPOT enables robust retrieval by freezing the document encoder while the query encoder learns to align noisy queries with their unaltered root. We evaluate CAPOT noisy variants of MSMARCO, Natural Questions, and Trivia QA passage retrieval, finding CAPOT has a similar impact as data augmentation with none of its overhead.

Jiayu Liu, Zhenya Huang, Chengxiang Zhai et al.

Mathematical reasoning is one of the crucial abilities of general artificial intelligence, which requires machines to master mathematical logic and knowledge from solving problems. However, existing approaches are not transparent (thus not interpretable) in terms of what knowledge has been learned and applied in the reasoning process. In this paper, we propose a general Learning by Applying (LeAp) framework to enhance existing models (backbones) in a principled way by explicit knowledge learning. In LeAp, we perform knowledge learning in a novel problem-knowledge-expression paradigm, with a Knowledge Encoder to acquire knowledge from problem data and a Knowledge Decoder to apply knowledge for expression reasoning. The learned mathematical knowledge, including word-word relations and word-operator relations, forms an explicit knowledge graph, which bridges the knowledge "learning" and "applying" organically. Moreover, for problem solving, we design a semantics-enhanced module and a reasoning-enhanced module that apply knowledge to improve the problem comprehension and symbol reasoning abilities of any backbone, respectively. We theoretically prove the superiority of LeAp's autonomous learning mechanism. Experiments on three real-world datasets show that LeAp improves all backbones' performances, learns accurate knowledge, and achieves a more interpretable reasoning process.

Bhavya Bhavya, Jinjun Xiong, Chengxiang Zhai

We propose a novel application of prompting Pre-trained Language Models (PLMs) to generate analogies and study how to design effective prompts for two task settings: generating a source concept analogous to a given target concept (aka Analogous Concept Generation or ACG), and generating an explanation of the similarity between a given pair of target concept and source concept (aka Analogous Explanation Generation or AEG). We found that it is feasible to prompt InstructGPT to generate meaningful analogies and the best prompts tend to be precise imperative statements especially with a low temperature setting. We also systematically analyzed the sensitivity of the InstructGPT model to prompt design, temperature, and injected spelling errors, and found that the model is particularly sensitive to certain variations (e.g., questions vs. imperative statements). Further, we conducted human evaluation on 1.4k of the generated analogies and found that the quality of generations varies substantially by model size. The largest InstructGPT model can achieve human-level performance at generating meaningful analogies for a given target while there is still room for improvement on the AEG task.

Daniel Campos, ChengXiang Zhai

Vector-based retrieval systems have become a common staple for academic and industrial search applications because they provide a simple and scalable way of extending the search to leverage contextual representations for documents and queries. As these vector-based systems rely on contextual language models, their usage commonly requires GPUs, which can be expensive and difficult to manage. Given recent advances in introducing sparsity into language models for improved inference efficiency, in this paper, we study how sparse language models can be used for dense retrieval to improve inference efficiency. Using the popular retrieval library Tevatron and the MSMARCO, NQ, and TriviaQA datasets, we find that sparse language models can be used as direct replacements with little to no drop in accuracy and up to 4.3x improved inference speeds

Yuren Hao, Shuhaib Mehri, ChengXiang Zhai et al.

Large language models are increasingly used as personal assistants, yet most lack a persistent user model, forcing users to repeatedly restate preferences across sessions. We propose Vector-Adapted Retrieval Scoring (VARS), a pipeline-agnostic, frozen-backbone framework that represents each user with long-term and short-term vectors in a shared preference space and uses these vectors to bias retrieval scoring over structured preference memory. The vectors are updated online from weak scalar rewards from users' feedback, enabling personalization without per-user fine-tuning. We evaluate on \textsc{MultiSessionCollab}, an online multi-session collaboration benchmark with rich user preference profiles, across math and code tasks. Under frozen backbones, the main benefit of user-aware retrieval is improved interaction efficiency rather than large gains in raw task accuracy: our full VARS agent achieves the strongest overall performance, matches a strong Reflection baseline in task success, and reduces timeout rate and user effort. The learned long-term vectors also align with cross-user preference overlap, while short-term vectors capture session-specific adaptation, supporting the interpretability of the dual-vector design. Code, model, and data are available at https://github.com/YurenHao0426/VARS.

Pritom Saha Akash, Jie Huang, Kevin Chen-Chuan Chang et al.

We propose a probabilistic approach to select a subset of a \textit{target domain representative keywords} from a candidate set, contrasting with a context domain. Such a task is crucial for many downstream tasks in natural language processing. To contrast the target domain and the context domain, we adapt the \textit{two-component mixture model} concept to generate a distribution of candidate keywords. It provides more importance to the \textit{distinctive} keywords of the target domain than common keywords contrasting with the context domain. To support the \textit{representativeness} of the selected keywords towards the target domain, we introduce an \textit{optimization algorithm} for selecting the subset from the generated candidate distribution. We have shown that the optimization algorithm can be efficiently implemented with a near-optimal approximation guarantee. Finally, extensive experiments on multiple domains demonstrate the superiority of our approach over other baselines for the tasks of keyword summary generation and trending keywords selection.

Chenkai Sun, Jinning Li, Yi R. Fung et al.

Automatic response forecasting for news media plays a crucial role in enabling content producers to efficiently predict the impact of news releases and prevent unexpected negative outcomes such as social conflict and moral injury. To effectively forecast responses, it is essential to develop measures that leverage the social dynamics and contextual information surrounding individuals, especially in cases where explicit profiles or historical actions of the users are limited (referred to as lurkers). As shown in a previous study, 97% of all tweets are produced by only the most active 25% of users. However, existing approaches have limited exploration of how to best process and utilize these important features. To address this gap, we propose a novel framework, named SocialSense, that leverages a large language model to induce a belief-centered graph on top of an existent social network, along with graph-based propagation to capture social dynamics. We hypothesize that the induced graph that bridges the gap between distant users who share similar beliefs allows the model to effectively capture the response patterns. Our method surpasses existing state-of-the-art in experimental evaluations for both zero-shot and supervised settings, demonstrating its effectiveness in response forecasting. Moreover, the analysis reveals the framework's capability to effectively handle unseen user and lurker scenarios, further highlighting its robustness and practical applicability.

Ke Yang, Zixi Chen, Xuan He et al.

Long-term memory is essential for large language model (LLM) agents operating in complex environments, yet existing memory designs are either task-specific and non-transferable, or task-agnostic but less effective due to low task-relevance and context explosion from raw memory retrieval. We propose PlugMem, a task-agnostic plugin memory module that can be attached to arbitrary LLM agents without task-specific redesign. Motivated by the fact that decision-relevant information is concentrated as abstract knowledge rather than raw experience, we draw on cognitive science to structure episodic memories into a compact, extensible knowledge-centric memory graph that explicitly represents propositional and prescriptive knowledge. This representation enables efficient memory retrieval and reasoning over task-relevant knowledge, rather than verbose raw trajectories, and departs from other graph-based methods like GraphRAG by treating knowledge as the unit of memory access and organization instead of entities or text chunks. We evaluate PlugMem unchanged across three heterogeneous benchmarks (long-horizon conversational question answering, multi-hop knowledge retrieval, and web agent tasks). The results show that PlugMem consistently outperforms task-agnostic baselines and exceeds task-specific memory designs, while also achieving the highest information density under a unified information-theoretic analysis. Code and data are available at https://github.com/TIMAN-group/PlugMem.

Adam Davies, Jize Jiang, ChengXiang Zhai

Despite the recent successes of large, pretrained neural language models (LLMs), comparatively little is known about the representations of linguistic structure they learn during pretraining, which can lead to unexpected behaviors in response to prompt variation or distribution shift. To better understand these models and behaviors, we introduce a general model analysis framework to study LLMs with respect to their representation and use of human-interpretable linguistic properties. Our framework, CALM (Competence-based Analysis of Language Models), is designed to investigate LLM competence in the context of specific tasks by intervening on models' internal representations of different linguistic properties using causal probing, and measuring models' alignment under these interventions with a given ground-truth causal model of the task. We also develop a new approach for performing causal probing interventions using gradient-based adversarial attacks, which can target a broader range of properties and representations than prior techniques. Finally, we carry out a case study of CALM using these interventions to analyze and compare LLM competence across a variety of lexical inference tasks, showing that CALM can be used to explain behaviors across these tasks.

Daniel Campos, Alessandro Magnani, ChengXiang Zhai

In this paper, we consider the problem of improving the inference latency of language model-based dense retrieval systems by introducing structural compression and model size asymmetry between the context and query encoders. First, we investigate the impact of pre and post-training compression on the MSMARCO, Natural Questions, TriviaQA, SQUAD, and SCIFACT, finding that asymmetry in the dual encoders in dense retrieval can lead to improved inference efficiency. Knowing this, we introduce Kullback Leibler Alignment of Embeddings (KALE), an efficient and accurate method for increasing the inference efficiency of dense retrieval methods by pruning and aligning the query encoder after training. Specifically, KALE extends traditional Knowledge Distillation after bi-encoder training, allowing for effective query encoder compression without full retraining or index generation. Using KALE and asymmetric training, we can generate models which exceed the performance of DistilBERT despite having 3x faster inference.

Chenkai Sun, Tie Xu, ChengXiang Zhai et al.

In this paper, we present Tetris, a new task of Goal-Oriented Script Completion. Unlike previous work, it considers a more realistic and general setting, where the input includes not only the goal but also additional user context, including preferences and history. To address this problem, we propose a novel approach, which uses two techniques to improve performance: (1) concept prompting, and (2) script-oriented contrastive learning that addresses step repetition and hallucination problems. On our WikiHow-based dataset, we find that both methods improve performance. The dataset, repository, and models will be publicly available to facilitate further research on this new task.

Daniel Campos, Alexandre Marques, Mark Kurtz et al.

In this paper, we introduce the range of oBERTa language models, an easy-to-use set of language models which allows Natural Language Processing (NLP) practitioners to obtain between 3.8 and 24.3 times faster models without expertise in model compression. Specifically, oBERTa extends existing work on pruning, knowledge distillation, and quantization and leverages frozen embeddings improves distillation and model initialization to deliver higher accuracy on a broad range of transfer tasks. In generating oBERTa, we explore how the highly optimized RoBERTa differs from the BERT for pruning during pre-training and finetuning. We find it less amenable to compression during fine-tuning. We explore the use of oBERTa on seven representative NLP tasks and find that the improved compression techniques allow a pruned oBERTa model to match the performance of BERTbase and exceed the performance of Prune OFA Large on the SQUAD V1.1 Question Answering dataset, despite being 8x and 2x, respectively faster in inference. We release our code, training regimes, and associated model for broad usage to encourage usage and experimentation

Daniel Campos, ChengXiang Zhai

Sequence-to-sequence language models can be used to produce abstractive summaries which are coherent, relevant, and concise. Still, model sizes can make deployment in latency-sensitive or web-scale implementations difficult. This paper studies the relationship between model size, structured pruning, inference efficiency, and summarization accuracy on widely used summarization datasets. We show that model accuracy is tied to the encoder size while inference efficiency is connected to the decoder. Using asymmetric pruning can lead to nearly 3x improvement in inference latency with ~1 point loss in Rouge-2. Moreover, we find both the average degradation and the role of asymmetry to be consistent across model sizes and variations in datasets.

Daniel Campos, Alexandre Marques, Tuan Nguyen et al.

Large Language Models have become the core architecture upon which most modern natural language processing (NLP) systems build. These models can consistently deliver impressive accuracy and robustness across tasks and domains, but their high computational overhead can make inference difficult and expensive. To make using these models less costly, recent work has explored leveraging structured and unstructured pruning, quantization, and distillation to improve inference speed and decrease size. This paper studies how models pruned using Gradual Unstructured Magnitude Pruning can transfer between domains and tasks. Our experimentation shows that models that are pruned during pretraining using general domain masked language models can transfer to novel domains and tasks without extensive hyperparameter exploration or specialized approaches. We demonstrate that our general sparse model Sparse*BERT can become SparseBioBERT simply by pretraining the compressed architecture on unstructured biomedical text. Moreover, we show that SparseBioBERT can match the quality of BioBERT with only 10\% of the parameters.

Revanth Gangi Reddy, Hao Bai, Wentao Yao et al.

Open-domain dialog involves generating search queries that help obtain relevant knowledge for holding informative conversations. However, it can be challenging to determine what information to retrieve when the user is passive and does not express a clear need or request. To tackle this issue, we present a novel approach that focuses on generating internet search queries that are guided by social commonsense. Specifically, we leverage a commonsense dialog system to establish connections related to the conversation topic, which subsequently guides our query generation. Our proposed framework addresses passive user interactions by integrating topic tracking, commonsense response generation and instruction-driven query generation. Through extensive evaluations, we show that our approach overcomes limitations of existing query generation techniques that rely solely on explicit dialog information, and produces search queries that are more relevant, specific, and compelling, ultimately resulting in more engaging responses.

Ke Yang, Volodymyr Kindratenko, ChengXiang Zhai

Training language models (LMs) and their application agents is increasingly costly due to large datasets and models, making test failures difficult to bear. Simplified language environments serve as primordial training and testing grounds, retaining essential commonsense and communication skills but in a more digestible form, potentially enhancing the learning efficiency of LMs, and thus reducing the required model size and data volume for effective training and evaluation. In these simplified language environments, workable strategies for small models, datasets, and agents may be adaptable to larger models, datasets, and agents in complex language environments. To create such environments, we focus on two aspects: i) minimizing language dataset noise and complexity, and ii) preserving the essential text distribution characteristics. Unlike previous methods, we propose a pipeline to refine text data by eliminating noise, minimizing vocabulary, and maintaining genre-specific patterns (e.g., for books, conversation, code, etc.). Implementing this pipeline with large LMs, we have created a leaner suite of LM training and evaluation datasets: 71M Leaner-Pretrain, 7M Leaner-Instruct, Leaner-Glue for assessing linguistic proficiency, and Leaner-Eval for testing instruction-following ability. Our experiments show that leaner pre-training boosts LM learning efficiency. Tiny LMs trained on these datasets outperform those trained on original datasets in instruction-following across different language granularity levels. Moreover, the Leaner-Pretrain dataset's alignment with conventional large LM training sets enables resource-optimized analysis of how learning objectives, model architectures, and training techniques impact performance on language modeling and downstream tasks. Our code and datasets are available at https://github.com/EmpathYang/TinyHelen.git.

Yuren Hao, Xiang Wan, ChengXiang Zhai

In this paper, we introduce a systematic framework beyond conventional method to assess LLMs' mathematical-reasoning robustness by stress-testing them on advanced math problems that are mathematically equivalent but with linguistic and parametric variation. These transformations allow us to measure the sensitivity of LLMs to non-mathematical perturbations, thereby enabling a more accurate evaluation of their mathematical reasoning capabilities. Using this new evaluation methodology, we created PutnamGAP, a new benchmark dataset with multiple mathematically-equivalent variations of competition-level math problems. With the new dataset, we evaluate multiple families of representative LLMs and examine their robustness. Across 18 commercial and open-source models we observe sharp performance degradation on the variants. OpenAI's flagship reasoning model, O3, scores 51.5% on the originals but drops by 4.7 percentage points on surface-renaming variants, and by 12.9 percentage points on parametric variants, while smaller models fare far worse. Overall, the results show that the proposed new evaluation methodology is effective for deepening our understanding of the robustness of LLMs and generating new insights for further improving their mathematical reasoning capabilities.

Eric Modesitt, Ke Yang, Spencer Hulsey et al.

Recent advances in language modeling demonstrate the need for high-quality domain-specific training data, especially for tasks that require specialized knowledge. General-purpose models, while versatile, often lack the depth needed for expert-level tasks because of limited domain-specific information. Domain adaptation training can enhance these models, but it demands substantial, high-quality data. To address this, we propose ORBIT, a cost-efficient methodology for curating massive, high-quality domain-specific datasets from noisy web sources, tailored for training specialist large language models. Using astronomy as a primary case study, we refined the 1.3T-token FineWeb-Edu dataset into a high-quality, 10B-token subset focused on astronomy. Fine-tuning \textsc{LLaMA-3-8B} on a 1B-token astronomy subset improved performance on the MMLU astronomy benchmark from 69\% to 76\% and achieved top results on AstroBench, an astronomy-specific benchmark. Moreover, our model (Orbit-LLaMA) outperformed \textsc{LLaMA-3-8B-base}, with GPT-4o evaluations preferring it in 73\% of cases across 1000 astronomy-specific questions. Additionally, we validated ORBIT's generalizability by applying it to law and medicine, achieving a significant improvement of data quality compared to an unfiltered baseline. We open-source the ORBIT methodology, including the curated datasets, the codebase, and the resulting model at \href{https://github.com/ModeEric/ORBIT-Llama}{https://github.com/ModeEric/ORBIT-Llama}.

Xinyu Pi, Mingyuan Wu, Jize Jiang et al.

Smaller-scale Vision-Langauge Models (VLMs) often claim to perform on par with larger models in general-domain visual grounding and question-answering benchmarks while offering advantages in computational efficiency and storage. However, their ability to handle rare objects, which fall into the long tail of data distributions, is less understood. To rigorously evaluate this aspect, we introduce the "Uncontextualized Uncommon Objects" (UOUO) benchmark. This benchmark focuses on systematically testing VLMs with both large and small parameter counts on rare and specialized objects. Our comprehensive analysis reveals that while smaller VLMs maintain competitive performance on common datasets, they significantly underperform on tasks involving uncommon objects. We also propose an advanced, scalable pipeline for data collection and cleaning, ensuring the UOUO benchmark provides high-quality, challenging instances. These findings highlight the need to consider long-tail distributions when assessing the true capabilities of VLMs.

Dina Albassam, Kexin Quan, Mengke Wu et al.

YouTube is widely used for informal learning, where learners explore lectures and tutorials without a predefined curriculum. However, learning across videos remains fragmented: learners must decide what to watch, how videos relate, and how knowledge builds. Existing tools provide partial support but treat planning and learning as separate activities, lacking a persistent interaction structure that connects them. Grounded in self-regulated learning theory (SRLT), we introduce YT-Pilot, a pathway-aware learning system that operationalizes the learning pathway as a persistent, user-facing interaction structure spanning planning and learning. The pathway coordinates goal setting, planning, navigation, progress tracking, and cross-video assistance. Through a within-subjects study ($N=20$), we show that YT-Pilot significantly improves perceived goal clarity, pathway coherence, and progress tracking, while shifting interaction toward pathway-level reasoning across multiple resources.

Miri Liu, ChengXiang Zhai

The rigorous evaluation of the novelty of a scientific paper is, even for human scientists, a challenging task. With the increasing interest in AI scientists and AI involvement in scientific idea generation and paper writing, it also becomes increasingly important that this task be automatable and reliable, lest both human attention and compute tokens be wasted on ideas that have already been explored. Due to the challenge of quantifying ground-truth novelty, however, existing novelty metrics for scientific papers generally validate their results against noisy, confounded signals such as citation counts or peer review scores. These proxies can conflate novelty with impact, quality, or reviewer preference, which in turn makes it harder to assess how well a given metric actually evaluates novelty. We therefore propose an axiomatic benchmark for scientific novelty metrics. We first define a set of axioms that a well-behaved novelty metric should satisfy, grounded in human scientific norms and practice, then evaluate existing metrics across ten tasks spanning three domains of AI research. Our results reveal that no existing metric satisfies all axioms consistently, and that metrics fail on systematically different axioms, reflecting their underlying architectures. Additionally, we show that combining metrics of complementary architectures leads to consistent improvements on the benchmark, with per-axiom weighting achieving 90.1% versus 71.5% for the best individual metric, suggesting that developing architecturally diverse metrics is a promising direction for future work. We release the benchmark code as supplementary material to encourage the development of more robust scientific literature novelty metrics.

Duo Zhou, Yuji Zhang, Tianxin Wei et al.

Large language models (LLMs) can internalize private or harmful content, motivating unlearning that removes a forget set while preserving retaining knowledge. However, forgetting updates often cause collateral degradation on retaining knowledge, creating a persistent trade-off. Existing LLM unlearning methods are often heuristic, and other theoretical approaches rely on offline feature constructions that do not capture update-time forget-retain interaction in LLMs. To address this limitation, we aim to develop an LLM unlearning method that reduces the forget-retain trade-off with theoretical guarantees. We take a first-principles view by formalizing "no side effects" as local retain invariance under small parameter updates, and prove an equivalence under optimizer-induced geometry: the retain loss is locally invariant if and only if the update direction is orthogonal to the subspace spanned by retain gradients. Based on the insight, we propose Geometric-disentanglement Unlearning (GU), a lightweight and theoretically grounded projection that can be plug-and-play to existing gradient-based unlearning methods to mitigate forget-retain side effects. Experiments on TOFU, MUSE, and WMDP-cyber show that GU strengthens forgetting while reducing retain drift. When added to SimNPO, it achieves up to 62\% improved forgetting Extraction Strength (ES) and 31\% higher retain ES. We open-sourced our code in https://github.com/Lemutisme/Geometric-Unlearning.

Ke Yang, Jiateng Liu, John Wu et al.

The prominent large language models (LLMs) of today differ from past language models not only in size, but also in the fact that they are trained on a combination of natural language and formal language (code). As a medium between humans and computers, code translates high-level goals into executable steps, featuring standard syntax, logical consistency, abstraction, and modularity. In this survey, we present an overview of the various benefits of integrating code into LLMs' training data. Specifically, beyond enhancing LLMs in code generation, we observe that these unique properties of code help (i) unlock the reasoning ability of LLMs, enabling their applications to a range of more complex natural language tasks; (ii) steer LLMs to produce structured and precise intermediate steps, which can then be connected to external execution ends through function calls; and (iii) take advantage of code compilation and execution environment, which also provides diverse feedback for model improvement. In addition, we trace how these profound capabilities of LLMs, brought by code, have led to their emergence as intelligent agents (IAs) in situations where the ability to understand instructions, decompose goals, plan and execute actions, and refine from feedback are crucial to their success on downstream tasks. Finally, we present several key challenges and future directions of empowering LLMs with code.

Ke Yang, Kevin Ros, Shankar Kumar Senthil Kumar et al.

Just-in-time Information Recommendation (JIR) is a service designed to deliver the most relevant information precisely when users need it, , addressing their knowledge gaps with minimal effort and boosting decision-making and efficiency in daily life. Advances in device-efficient deployment of foundation models and the growing use of intelligent wearable devices have made always-on JIR assistants feasible. However, there has been no systematic effort to formally define JIR tasks or establish evaluation frameworks. To bridge this gap, we present the first mathematical definition of JIR tasks and associated evaluation metrics. Additionally, we introduce JIR-Arena, a multimodal benchmark dataset featuring diverse, information-request-intensive scenarios to evaluate JIR systems across critical dimensions: i) accurately inferring user information needs, ii) delivering timely and relevant recommendations, and iii) avoiding irrelevant content that may distract users. Developing a JIR benchmark dataset poses challenges due to subjectivity in estimating user information needs and uncontrollable system variables affecting reproducibility. To address these, JIR-Arena: i) combines input from multiple humans and large AI models to approximate information need distributions; ii) assesses JIR quality through information retrieval outcomes using static knowledge base snapshots; and iii) employs a multi-turn, multi-entity validation framework to improve objectivity and generality. Furthermore, we implement a baseline JIR system capable of processing real-time information streams aligned with user inputs. Our evaluation of this baseline system on JIR-Arena indicates that while foundation model-based JIR systems simulate user needs with reasonable precision, they face challenges in recall and effective content retrieval. To support future research in this new area, we fully release our code and data.

Mingyuan Wu, Jize Jiang, Haozhen Zheng et al.

Vision Language Models (VLMs) have achieved remarkable success in a wide range of vision applications of increasing complexity and scales, yet choosing the right VLM model size involves a trade-off between response quality and cost. While smaller VLMs are cheaper to run, they typically produce responses only marginally better than random guessing on benchmarks such as MMMU. In this paper, we propose Cache of Thought (CoT), a master apprentice framework for collaborative inference between large and small VLMs. CoT manages high quality query results from large VLMs (master) in a cache, which are then selected via a novel multi modal retrieval and in-context learning to aid the performance of small VLMs (apprentice). We extensively evaluate CoT on various widely recognized and challenging general reasoning benchmarks, and show that CoT increases overall reasoning performance by up to 7.7% under the same budget, and specifically boosts the performance of apprentice VLMs by up to 36.6%. Our code is available at https://github.com/UIUC-MONET/Cache-of-Thoughts

Himanshu Udupi, Xiaocong Yang, ChengXiang Zhai

Spiking Neural Networks (SNNs) have been proposed as biologically plausible and energy-efficient alternatives to conventional Artificial Neural Networks (ANNs). However, the training of SNN usually relies on surrogate gradients due to the non-differentiability of the spike function, introducing approximation errors that accumulate across layers. To address this challenge, we extend the work on convexification of parallel feedforward threshold networks to parallel recurrent threshold networks, which subsume parallel SNNs as a structured special case. Building on this theoretical framework, we propose a parameter reconstruction algorithm for SNN training that demonstrates consistent and significant advantages across various tasks, both as a standalone method and in combination with surrogate-gradient training. The ablations further demonstrate the data scalability and robustness to model configurations of our training algorithm, pointing toward its potential in large-scale SNN training.

Chenkai Sun, Ke Yang, Revanth Gangi Reddy et al.

The increasing demand for personalized interactions with large language models (LLMs) calls for methodologies capable of accurately and efficiently identifying user opinions and preferences. Retrieval augmentation emerges as an effective strategy, as it can accommodate a vast number of users without the costs from fine-tuning. Existing research, however, has largely focused on enhancing the retrieval stage and devoted limited exploration toward optimizing the representation of the database, a crucial aspect for tasks such as personalization. In this work, we examine the problem from a novel angle, focusing on how data can be better represented for more data-efficient retrieval in the context of LLM customization. To tackle this challenge, we introduce Persona-DB, a simple yet effective framework consisting of a hierarchical construction process to improve generalization across task contexts and collaborative refinement to effectively bridge knowledge gaps among users. In the evaluation of response prediction, Persona-DB demonstrates superior context efficiency in maintaining accuracy with a significantly reduced retrieval size, a critical advantage in scenarios with extensive histories or limited context windows. Our experiments also indicate a marked improvement of over 10% under cold-start scenarios, when users have extremely sparse data. Furthermore, our analysis reveals the increasing importance of collaborative knowledge as the retrieval capacity expands.

Yuji Zhang, Sha Li, Cheng Qian et al.

Hallucination is a persistent challenge in large language models (LLMs), where even with rigorous quality control, models often generate distorted facts. This paradox, in which error generation continues despite high-quality training data, calls for a deeper understanding of the underlying LLM mechanisms. To address it, we propose a novel concept: knowledge overshadowing, where model's dominant knowledge can obscure less prominent knowledge during text generation, causing the model to fabricate inaccurate details. Building on this idea, we introduce a novel framework to quantify factual hallucinations by modeling knowledge overshadowing. Central to our approach is the log-linear law, which predicts that the rate of factual hallucination increases linearly with the logarithmic scale of (1) Knowledge Popularity, (2) Knowledge Length, and (3) Model Size. The law provides a means to preemptively quantify hallucinations, offering foresight into their occurrence even before model training or inference. Built on overshadowing effect, we propose a new decoding strategy CoDa, to mitigate hallucinations, which notably enhance model factuality on Overshadow (27.9%), MemoTrap (13.1%) and NQ-Swap (18.3%). Our findings not only deepen understandings of the underlying mechanisms behind hallucinations but also provide actionable insights for developing more predictable and controllable language models.

Dean E. Alvarez, ChengXiang Zhai

One reason the Web is more useful than a simple collection of documents is that the structure created by hyperlinks enables flexible navigation from one web page to another. However, hyperlinks are typically created manually and cannot fully capture a corpus' implicit semantic structures. Is there a general way to make an arbitrary collection navigable? Recent work has formalized this problem generally as constructing a Hypergraph of Text (HoT), which provides a formal mathematical structure for supporting navigation and browsing. However, how to construct and evaluate a Hypergraph of Text remains a challenge. In this paper, we propose and study several methods for constructing a HoT. We also propose a novel quantitative metric, effort ratio, for evaluating the structural quality of a constructed HoT. Experimental results show that even simple TF-IDF baselines can match LLM-based methods on our proposed effort ratio metric.

Mingyuan Wu, Jingcheng Yang, Jize Jiang et al.

Reinforcement Learning Finetuning (RFT) has significantly advanced the reasoning capabilities of large language models (LLMs) by enabling long chains of thought, self-correction, and effective tool use. While recent works attempt to extend RFT to vision-language models (VLMs), these efforts largely produce text-only reasoning conditioned on static image inputs, falling short of true multimodal reasoning in the response. In contrast, test-time methods like Visual Sketchpad incorporate visual steps but lack training mechanisms. We introduce VTool-R1, the first framework that trains VLMs to generate multimodal chains of thought by interleaving text and intermediate visual reasoning steps. VTool-R1 integrates Python-based visual editing tools into the RFT process, enabling VLMs to learn when and how to generate visual reasoning steps that benefit final reasoning. Trained with outcome-based rewards tied to task accuracy, our approach elicits strategic visual tool use for reasoning without relying on process-based supervision. Experiments on structured visual question answering over charts and tables show that VTool-R1 enhances reasoning performance by teaching VLMs to "think with images" and generate multimodal chain of thoughts with tools.

Krisztian Balog, ChengXiang Zhai

User simulation is an emerging interdisciplinary topic with multiple critical applications in the era of Generative AI. It involves creating an intelligent agent that mimics the actions of a human user interacting with an AI system, enabling researchers to model and analyze user behaviour, generate synthetic data for training, and evaluate interactive AI systems in a controlled and reproducible manner. User simulation has profound implications for diverse fields and plays a vital role in the pursuit of Artificial General Intelligence. This paper provides an overview of user simulation, highlighting its key applications, connections to various disciplines, and outlining future research directions to advance this increasingly important technology.

Yiran Liu, Ke Yang, Zehan Qi et al.

We present a novel statistical framework for analyzing stereotypes in large language models (LLMs) by systematically estimating the bias and variation in their generation. Current alignment evaluation metrics often overlook stereotypes' randomness caused by LLMs' inconsistent generative behavior. For instance, LLMs may display contradictory stereotypes, such as those related to gender or race, for identical professions in different contexts. Ignoring this inconsistency risks misleading conclusions in alignment assessments and undermines efforts to evaluate the potential of LLMs to perpetuate or amplify social biases and unfairness. To address this, we propose the Bias-Volatility Framework (BVF), which estimates the probability distribution of stereotypes in LLM outputs. By capturing the variation in generative behavior, BVF assesses both the likelihood and degree to which LLM outputs negatively impact vulnerable groups, enabling a quantification of aggregated discrimination risk. Additionally, we introduce a mathematical framework to decompose this risk into bias risk (from the mean of the stereotype distribution) and volatility risk (from its variation). Applying BVF to 12 widely used LLMs, we find: i) Bias risk is the dominant contributor to discrimination; ii) Most LLMs exhibit substantial pro-male stereotypes across nearly all professions; iii) Reinforcement learning from human feedback reduces bias but increases volatility; iv) Discrimination risk correlates with socio-economic factors, such as professional salaries. Finally, we highlight BVF's broader applicability for assessing how generation inconsistencies in LLMs impact behavior beyond stereotypes.

Jiayu Liu, Zhenya Huang, Chaokun Wang et al.

Owing to the capability of in-context learning, large language models (LLMs) have shown impressive performance across diverse mathematical reasoning benchmarks. However, we find that few-shot demonstrations can sometimes bring negative performance and their effectiveness on LLMs' reasoning abilities remains unreliable. To this end, in this paper, we aim to theoretically analyze the impact of in-context demonstrations on LLMs' reasoning performance. We prove that the reasoning efficacy (measured by empirical prediction loss) can be bounded by a LLM-oriented semantic similarity and an inference stability of demonstrations, which is general for both one-shot and few-shot scenarios. Based on this finding, we propose a straightforward, generalizable, and low-complexity demonstration selection method named LMS3. It can adaptively facilitate to select the most pertinent samples for different LLMs and includes a novel demonstration rejection mechanism to automatically filter out samples that are unsuitable for few-shot learning. Through experiments on three representative benchmarks, two LLM backbones, and multiple few-shot settings, we verify that our LMS3 has superiority and achieves consistent improvements on all datasets, which existing methods have been unable to accomplish.

Cheng Qian, Hongyi Du, Hongru Wang et al.

Recent progress in large language models (LLMs) has enabled substantial advances in solving mathematical problems. However, existing benchmarks often fail to reflect the complexity of real-world problems, which demand open-ended, interdisciplinary reasoning and integration of computational tools. To address this gap, we introduce ModelingBench, a novel benchmark featuring real-world-inspired, open-ended problems from math modeling competitions across diverse domains, ranging from urban traffic optimization to ecosystem resource planning. These tasks require translating natural language into formal mathematical formulations, applying appropriate tools, and producing structured, defensible reports. ModelingBench also supports multiple valid solutions, capturing the ambiguity and creativity of practical modeling. We also present ModelingAgent, a multi-agent framework that coordinates tool use, supports structured workflows, and enables iterative self-refinement to generate well-grounded, creative solutions. To evaluate outputs, we further propose ModelingJudge, an expert-in-the-loop system leveraging LLMs as domain-specialized judges assessing solutions from multiple expert perspectives. Empirical results show that ModelingAgent substantially outperforms strong baselines and often produces solutions indistinguishable from those of human experts. Together, our work provides a comprehensive framework for evaluating and advancing real-world problem-solving in open-ended, interdisciplinary modeling challenges.

Yu Zhang, Yunyi Zhang, Yanzhen Shen et al. · amazon-science

Accurately typing entity mentions from text segments is a fundamental task for various natural language processing applications. Many previous approaches rely on massive human-annotated data to perform entity typing. Nevertheless, collecting such data in highly specialized science and engineering domains (e.g., software engineering and security) can be time-consuming and costly, without mentioning the domain gaps between training and inference data if the model needs to be applied to confidential datasets. In this paper, we study the task of seed-guided fine-grained entity typing in science and engineering domains, which takes the name and a few seed entities for each entity type as the only supervision and aims to classify new entity mentions into both seen and unseen types (i.e., those without seed entities). To solve this problem, we propose SEType which first enriches the weak supervision by finding more entities for each seen type from an unlabeled corpus using the contextualized representations of pre-trained language models. It then matches the enriched entities to unlabeled text to get pseudo-labeled samples and trains a textual entailment model that can make inferences for both seen and unseen types. Extensive experiments on two datasets covering four domains demonstrate the effectiveness of SEType in comparison with various baselines.

Yuji Zhang, Qingyun Wang, Cheng Qian et al.

Scientific texts often convey authority due to their technical language and complex data. However, this complexity can sometimes lead to the spread of misinformation. Non-experts are particularly susceptible to misleading claims based on scientific tables due to their high information density and perceived credibility. Existing table claim verification models, including state-of-the-art large language models (LLMs), often struggle with precise fine-grained reasoning, resulting in errors and a lack of precision in verifying scientific claims. Inspired by Cognitive Load Theory, we propose that enhancing a model's ability to interpret table-based claims involves reducing cognitive load by developing modular, reusable reasoning components (i.e., atomic skills). We introduce a skill-chaining schema that dynamically composes these skills to facilitate more accurate and generalizable reasoning with a reduced cognitive load. To evaluate this, we create SciAtomicBench, a cross-domain benchmark with fine-grained reasoning annotations. With only 350 fine-tuning examples, our model trained by atomic reasoning outperforms GPT-4o's chain-of-thought method, achieving state-of-the-art results with far less training data.

Jinrui Fang, Runhan Chen, Xu Yang et al.

Large language models (LLMs) achieve high accuracy in medical diagnosis when all clinical information is provided in a single turn, yet how they behave under multi-turn evidence accumulation closer to real clinical reasoning remains unexplored. We introduce MINT (Medical Incremental N-Turn Benchmark), a high-fidelity, multi-turn medical diagnosis benchmark comprising 1,035 cases with clinically labeled evidence shards, controlled turn granularity, and information-preserving decomposition. Through systematic evaluation of 11 LLMs on MINT, we uncover three persistent behavioral patterns that significantly impact diagnostic decisions: (1) intent to answer, models rush to answer before sufficient evidence has been observed, with over 55% of answers committed within the first two turns; (2) self-correction, incorrect-to-correct answer revisions occur at up to 10.6 times the rate of correct-to-incorrect flips, revealing a latent capacity for self-correction that premature commitment forecloses; and (3) strong lures, clinically salient information such as laboratory results trigger premature answering even when models are explicitly instructed to wait. We translate these findings into clinically actionable guidance: deferring the diagnostic question to later turns reduces premature answering and improves accuracy at the first point of commitment by up to 62.6%, while reserving salient clinical evidence for later turns prevents a catastrophic accuracy drop of up to 23.3% caused by premature commitment. Our work provides both a controlled evaluation framework and concrete recommendations for improving the reliability of LLMs in multi-turn medical diagnosis.

Chenkai Sun, Denghui Zhang, ChengXiang Zhai et al.

Given the growing influence of language model-based agents on high-stakes societal decisions, from public policy to healthcare, ensuring their beneficial impact requires understanding the far-reaching implications of their suggestions. We propose a proof-of-concept framework that projects how model-generated advice could propagate through societal systems on a macroscopic scale over time, enabling more robust alignment. To assess the long-term safety awareness of language models, we also introduce a dataset of 100 indirect harm scenarios, testing models' ability to foresee adverse, non-obvious outcomes from seemingly harmless user prompts. Our approach achieves not only over 20% improvement on the new dataset but also an average win rate exceeding 70% against strong baselines on existing safety benchmarks (AdvBench, SafeRLHF, WildGuardMix), suggesting a promising direction for safer agents.

Hanyang Chen, Mark Zhao, Rui Yang et al.

Recent advances in embodied AI highlight the potential of vision language models (VLMs) as agents capable of perception, reasoning, and interaction in complex environments. However, top-performing systems rely on large-scale models that are costly to deploy, while smaller VLMs lack the necessary knowledge and skills to succeed. To bridge this gap, we present \textit{Embodied Reasoning Agent (ERA)}, a two-stage framework that integrates prior knowledge learning and online reinforcement learning (RL). The first stage, \textit{Embodied Prior Learning}, distills foundational knowledge from three types of data: (1) Trajectory-Augmented Priors, which enrich existing trajectory data with structured reasoning generated by stronger models; (2) Environment-Anchored Priors, which provide in-environment knowledge and grounding supervision; and (3) External Knowledge Priors, which transfer general knowledge from out-of-environment datasets. In the second stage, we develop an online RL pipeline that builds on these priors to further enhance agent performance. To overcome the inherent challenges in agent RL, including long horizons, sparse rewards, and training instability, we introduce three key designs: self-summarization for context management, dense reward shaping, and turn-level policy optimization. Extensive experiments on both high-level planning (EB-ALFRED) and low-level control (EB-Manipulation) tasks demonstrate that ERA-3B surpasses both prompting-based large models and previous training-based baselines. Specifically, it achieves overall improvements of 8.4\% on EB-ALFRED and 19.4\% on EB-Manipulation over GPT-4o, and exhibits strong generalization to unseen tasks. Overall, ERA offers a practical path toward scalable embodied intelligence, providing methodological insights for future embodied AI systems.

Krisztian Balog, ChengXiang Zhai

Progress toward Artificial General Intelligence (AGI) faces significant bottlenecks, particularly in rigorously evaluating complex interactive systems and acquiring the vast interaction data needed for training adaptive agents. This paper posits that user simulation -- creating computational agents that mimic human interaction with AI systems -- is not merely a useful tool, but is a critical catalyst required to overcome these bottlenecks and accelerate AGI development. We argue that realistic simulators provide the necessary environments for scalable evaluation, data generation for interactive learning, and fostering the adaptive capabilities central to AGI. Therefore, research into user simulation technology and intelligent task agents are deeply synergistic and must advance hand-in-hand. This article elaborates on the critical role of user simulation for AGI, explores the interdisciplinary nature of building realistic simulators, identifies key challenges including those posed by large language models, and proposes a future research agenda.

Amogh Mannekote, Adam Davies, Guohao Li et al.

As LLMs are increasingly studied as role-playing agents to generate synthetic data for human behavioral research, ensuring that their outputs remain coherent with their assigned roles has become a critical concern. In this paper, we investigate how consistently LLM-based role-playing agents' stated beliefs about the behavior of the people they are asked to role-play ("what they say") correspond to their actual behavior during role-play ("how they act"). Specifically, we establish an evaluation framework to rigorously measure how well beliefs obtained by prompting the model can predict simulation outcomes in advance. Using an augmented version of the GenAgents persona bank and the Trust Game (a standard economic game used to quantify players' trust and reciprocity), we introduce a belief-behavior consistency metric to systematically investigate how it is affected by factors such as: (1) the types of beliefs we elicit from LLMs, like expected outcomes of simulations versus task-relevant attributes of individual characters LLMs are asked to simulate; (2) when and how we present LLMs with relevant information about Trust Game; and (3) how far into the future we ask the model to forecast its actions. We also explore how feasible it is to impose a researcher's own theoretical priors in the event that the originally elicited beliefs are misaligned with research objectives. Our results reveal systematic inconsistencies between LLMs' stated (or imposed) beliefs and the outcomes of their role-playing simulation, at both an individual- and population-level. Specifically, we find that, even when models appear to encode plausible beliefs, they may fail to apply them in a consistent way. These findings highlight the need to identify how and when LLMs' stated beliefs align with their simulated behavior, allowing researchers to use LLM-based agents appropriately in behavioral studies.

Dina Albassam, Adam Cross, Chengxiang Zhai

Electronic Health Records (EHRs) often lack explicit links between medications and diagnoses, making clinical decision-making and research more difficult. Even when links exist, diagnosis lists may be incomplete, especially during early patient visits. Discharge summaries tend to provide more complete information, which can help infer accurate diagnoses, especially with the help of large language models (LLMs). This study investigates whether LLMs can predict implicitly mentioned diagnoses from clinical notes and link them to corresponding medications. We address two research questions: (1) Does majority voting across diverse LLM configurations outperform the best single configuration in diagnosis prediction? (2) How sensitive is majority voting accuracy to LLM hyperparameters such as temperature, top-p, and summary length? To evaluate, we created a new dataset of 240 expert-annotated medication-diagnosis pairs from 20 MIMIC-IV notes. Using GPT-3.5 Turbo, we ran 18 prompting configurations across short and long summary lengths, generating 8568 test cases. Results show that majority voting achieved 75 percent accuracy, outperforming the best single configuration at 66 percent. No single hyperparameter setting dominated, but combining deterministic, balanced, and exploratory strategies improved performance. Shorter summaries generally led to higher accuracy.In conclusion, ensemble-style majority voting with diverse LLM configurations improves diagnosis prediction in EHRs and offers a promising method to link medications and diagnoses in clinical texts.