Publications

Self-improving LLM agents refine answers, not the process that produces those answers. Systems that add a meta-level hold that level fixed, and those that edit themselves must leave part of their own editing machinery untouched to stay stable, capping the meta-depth they realize at roughly two. We present Meta^n, which keeps the meta-operation fixed and recurses on its input instead. That operation, Ω, is applied repeatedly to its own products, reading the traces of the solver stack below together with the code that produced them, then writing the next layer as a strategic pre-process and a library of callable helpers. Because Ω never changes, it cannot destabilize the system, and because its input strictly grows, each layer reasons from a higher vantage than the last. Depth is set by convergence rather than fixed in advance, and an evolutionary archive searches over layer chains. Across two backbones, Meta^n outperforms prior self-improving agents on all eight benchmark families. The sharpest case is ARC-AGI-2, built to resist skill memorization, where it alone scores above zero. Ablations indicate that most of the gain from recursion comes from the conditioning each layer passes to the next, and distinct layer roles emerge with depth although no prompt prescribes them. Code available at https://github.com/minnesotanlp/meta-n

Scientific discovery is an extended process of ideation–surveying prior work, forming hypotheses, and refining reasoning–yet existing approaches treat this phase as a brief preamble despite its central role in research. We introduce SCISENSE, a sensemaking-grounded framework that operationalizes ideation as a structured sequence of eight cognitive stages (Pirolli & Card, 2005). We construct SCISENSE-Traj, a 100K-scale dataset of citation-conditioned research trajectories in two modes: Target, where an LLM reconstructs the ideation path leading to a known paper from its cited works, and Infer, where the LLM proposes novel directions from the same citations. We distill these into SCISENSE-LM, a family of sensemaking LLMs spanning 3B to 70B parameters. Contrary to the assumption that looser supervision promotes greater exploration, Target-trained models achieve a 2.0% improvement in trajectory quality over Infer-trained models while also producing more novel and diverse outputs. This advantage propagates downstream: coding agents conditioned on Target trajectories produce research artifacts with higher executability and quality than those conditioned on Infer trajectories. This suggests that targeted ideation reduces cognitive burden on downstream agents, freeing them to explore more creatively. SCISENSE offers both a practical tool for augmenting LLM-driven research workflows and a principled testbed for studying how planning shapes scientific discovery.

Existing tool-use benchmarks for LLM agents are overwhelmingly linear: our analysis of six benchmarks shows 55 to 100% of instances are simple chains of 2 to 5 steps. We introduce The Amazing Agent Race (AAR), a benchmark featuring directed acyclic graph (DAG) puzzles (or "legs") with fork-merge tool chains. We release 1,400 instances across two variants: sequential (800 legs) and compositional (600 DAG legs). Agents must navigate Wikipedia, execute multi-step tool chains, and aggregate results into a verifiable answer. Legs are procedurally generated from Wikipedia seeds across four difficulty levels with live-API validation. Three complementary metrics (finish-line accuracy, pit-stop visit rate, and roadblock completion rate) separately diagnose navigation, tool-use, and arithmetic failures. Evaluating three agent frameworks on 1,400 legs, the best achieves only 37.2% accuracy. Navigation errors dominate (27 to 52% of trials) while tool-use errors remain below 17%, and agent architecture matters as much as model scale (Claude Code matches Codex CLI at 37% with 6x fewer tokens). The compositional structure of AAR reveals that agents fail not at calling tools but at navigating to the right pages, a blind spot invisible to linear benchmarks. The project page can be accessed at: https://minnesotanlp.github.io/the-amazing-agent-race

Generating long, coherent text remains a challenge for large language models (LLMs), as they lack hierarchical planning and structured organization in discourse generation. We introduce Structural Alignment, a novel method that aligns LLMs with human-like discourse structures to enhance long-form text generation. By integrating linguistically grounded discourse frameworks into reinforcement learning, our approach guides models to produce coherent and well-organized outputs. We employ a dense reward scheme within a Proximal Policy Optimization framework, assigning fine-grained, token-level rewards based on the discourse distinctiveness relative to human writing. Two complementary reward models are evaluated: the first improves readability by scoring surface-level textual features to provide explicit structuring, while the second reinforces deeper coherence and rhetorical sophistication by analyzing global discourse patterns through hierarchical discourse motifs, outperforming both standard and RLHF-enhanced models in tasks such as essay generation and long-document summarization. All training data and code will be publicly shared at https://github.com/minnesotanlp/struct_align.

A unified representation space in multi-modal learning is essential for effectively integrating diverse data sources, such as text, images, and audio, to enhance efficiency and performance across various downstream tasks. Recent binding methods, such as ImageBind, typically rely on a single, fixed anchor modality for aligning multi-modal data. We mathematically analyze these fixed anchor binding methods and uncover significant limitations: (1) over-reliance on the choice of the anchor modality, (2) inadequate capture of intra-modal information, and (3) failure to account for cross-modal correlation among non-anchored modalities. To address these issues, we propose the need for adaptive anchor binding methods, exemplified by our framework CentroBind. The proposed method uses adaptively adjustable centroid-based anchors generated from all available modalities, leading to a balanced and rich representation space. We theoretically demonstrate that our approach captures three critical properties of multi-modal learning – intra-modal learning, inter-modal learning, and multi-modal alignment – while constructing a unified representation that spans all modalities. Experiments on both synthetic and real-world datasets show that adaptive anchor methods such as CentroBind consistently outperform fixed anchor binding methods, verifying our analysis.

Reward-based alignment methods for large language models (LLMs) face two key limitations: vulnerability to reward hacking, where models exploit flaws in the reward signal; and reliance on brittle, labor-intensive prompt engineering when LLMs are used as reward models. We introduce Meta Policy Optimization (MPO), a framework that addresses these challenges by integrating a meta-reward model that dynamically refines the reward model's prompt throughout training. In MPO, the meta-reward model monitors the evolving training context and continuously adjusts the reward model's prompt to maintain high alignment, providing an adaptive reward signal that resists exploitation by the policy. This meta-learning approach promotes a more stable policy optimization, and greatly reduces the need for manual reward prompt design. It yields performance on par with or better than models guided by extensively hand-crafted reward prompts. Furthermore, we show that MPO maintains its effectiveness across diverse tasks, from essay writing to mathematical reasoning, without requiring specialized reward designs. Beyond standard RLAIF, MPO's meta-learning formulation is readily extensible to higher-level alignment frameworks. Overall, this method addresses theoretical and practical challenges in reward-based RL alignment for LLMs, paving the way for more robust and adaptable alignment strategies. The code and data can be accessed at: https://github.com/minnesotanlp/mpo

With the advent of large language models (LLM), the line between human-crafted and machine-generated texts has become increasingly blurred. This paper delves into the inquiry of identifying discernible and unique linguistic properties in texts that were written by humans, particularly uncovering the underlying discourse structures of texts beyond their surface structures. Introducing a novel methodology, we leverage hierarchical parse trees and recursive hypergraphs to unveil distinctive discourse patterns in texts produced by both LLMs and humans. Empirical findings demonstrate that, although both LLMs and humans generate distinct discourse patterns influenced by specific domains, human-written texts exhibit more structural variability, reflecting the nuanced nature of human writing in different domains. Notably, incorporating hierarchical discourse features enhances binary classifiers' overall performance in distinguishing between human-written and machine-generated texts, even on out-of-distribution and paraphrased samples. This underscores the significance of incorporating hierarchical discourse features in the analysis of text patterns. The code and dataset are available at https://github.com/minnesotanlp/threads-of-subtlety.

Large Language Models are cognitively biased judges. Large Language Models (LLMs) have recently been shown to be effective as automatic evaluators with simple prompting and in-context learning. In this work, we assemble 15 LLMs of four different size ranges and evaluate their output responses by preference ranking from the other LLMs as evaluators, such as System Star is better than System Square. We then evaluate the quality of ranking outputs introducing the Cognitive Bias Benchmark for LLMs as Evaluators (CoBBLEr), a benchmark to measure six different cognitive biases in LLM evaluation outputs, such as the Egocentric bias where a model prefers to rank its own outputs highly in evaluation. We find that LLMs are biased text quality evaluators, exhibiting strong indications on our bias benchmark (average of 40% of comparisons across all models) within each of their evaluations that question their robustness as evaluators. Furthermore, we examine the correlation between human and machine preferences and calculate the average Rank-Biased Overlap (RBO) score to be 49.6%, indicating that machine preferences are misaligned with humans. According to our findings, LLMs may still be unable to be utilized for automatic annotation aligned with human preferences. Our project page is at: https://minnesotanlp.github.io/cobbler.

This work delves into the expanding role of large language models (LLMs) in generating artificial data. LLMs are increasingly employed to create a variety of outputs, including annotations, preferences, instruction prompts, simulated dialogues, and free text. As these forms of LLM-generated data often intersect in their application, they exert mutual influence on each other and raise significant concerns about the quality and diversity of the artificial data incorporated into training cycles, leading to an artificial data ecosystem. To the best of our knowledge, this is the first study to aggregate various types of LLM-generated text data, from more tightly constrained data like "task labels" to more lightly constrained "free-form text". We then stress test the quality and implications of LLM-generated artificial data, comparing it with human data across various existing benchmarks. Despite artificial data's capability to match human performance, this paper reveals significant hidden disparities, especially in complex tasks where LLMs often miss the nuanced understanding of intrinsic human-generated content. This study critically examines diverse LLM-generated data and emphasizes the need for ethical practices in data creation and when using LLMs. It highlights the LLMs' shortcomings in replicating human traits and behaviors, underscoring the importance of addressing biases and artifacts produced in LLM-generated content for future research and development. All data and code are available on our project page.

Capturing readers' engagement in fiction is a challenging but important aspect of narrative understanding. In this study, we collected 23 readers' reactions to 2 short stories through eye tracking, sentence-level annotations, and an overall engagement scale survey. We analyzed the significance of various qualities of the text in predicting how engaging a reader is likely to find it. As enjoyment of fiction is highly contextual, we also investigated individual differences in our data. Furthering our understanding of what captivates readers in fiction will help better inform models used in creative narrative generation and collaborative writing tools.

This survey paper provides a comprehensive review of the use of diffusion models in natural language processing (NLP). Diffusion models are a class of mathematical models that aim to capture the diffusion of information or signals across a network or manifold. In NLP, diffusion models have been used in a variety of applications, such as natural language generation, sentiment analysis, topic modeling, and machine translation. This paper discusses the different formulations of diffusion models used in NLP, their strengths and limitations, and their applications. We also perform a thorough comparison between diffusion models and alternative generative models, specifically highlighting the autoregressive (AR) models, while also examining how diverse architectures incorporate the Transformer in conjunction with diffusion models. Compared to AR models, diffusion models have significant advantages for parallel generation, text interpolation, token-level controls such as syntactic structures and semantic contents, and robustness. Exploring further permutations of integrating Transformers into diffusion models would be a valuable pursuit. Also, the development of multimodal diffusion models and large-scale diffusion language models with notable capabilities for few-shot learning would be important directions for the future advance of diffusion models in NLP.

Conversational implicatures are pragmatic inferences that require listeners to deduce the intended meaning conveyed by a speaker from their explicit utterances. Although such inferential reasoning is fundamental to human communication, recent research indicates that large language models struggle to comprehend these implicatures as effectively as the average human. This paper demonstrates that by incorporating Grice's Four Maxims into the model through chain-of-thought prompting, we can significantly enhance its performance, surpassing even the average human performance on this task.

Iterative text revision improves text quality by fixing grammatical errors, rephrasing for better readability or contextual appropriateness, or reorganizing sentence structures throughout a document. Most recent research has focused on understanding and classifying different types of edits in the iterative revision process from human-written text instead of building accurate and robust systems for iterative text revision. In this work, we aim to build an end-to-end text revision system that can iteratively generate helpful edits by explicitly detecting editable spans (where-to-edit) with their corresponding edit intents and then instructing a revision model to revise the detected edit spans. Leveraging datasets from other related text editing NLP tasks, combined with the specification of editable spans, leads our system to more accurately model the process of iterative text refinement, as evidenced by empirical results and human evaluations. Our system significantly outperforms previous baselines on our text revision tasks and other standard text revision tasks, including grammatical error correction, text simplification, sentence fusion, and style transfer. Through extensive qualitative and quantitative analysis, we make vital connections between edit intentions and writing quality, and better computational modeling of iterative text revisions.

Revision is an essential part of the human writing process. It tends to be strategic, adaptive, and, more importantly, iterative in nature. Despite the success of large language models on text revision tasks, they are limited to non-iterative, one-shot revisions. Examining and evaluating the capability of large language models for making continuous revisions and collaborating with human writers is a critical step towards building effective writing assistants. In this work, we present a human-in-the-loop iterative text revision system, Read, Revise, Repeat (R3), which aims at achieving high quality text revisions with minimal human efforts by reading model-generated revisions and user feedbacks, revising documents, and repeating human-machine interactions. In R3, a text revision model provides text editing suggestions for human writers, who can accept or reject the suggested edits. The accepted edits are then incorporated into the model for the next iteration of document revision. Writers can therefore revise documents iteratively by interacting with the system and simply accepting/rejecting its suggested edits until the text revision model stops making further revisions or reaches a predefined maximum number of revisions. Empirical experiments show that R3 can generate revisions with comparable acceptance rate to human writers at early revision depths, and the human-machine interaction can get higher quality revisions with fewer iterations and edits. The collected human-model interaction dataset and system code are available at https://github.com/vipulraheja/IteraTeR. Our system demonstration is available at https://youtu.be/lK08tIpEoaE.

Writing is, by nature, a strategic, adaptive, and more importantly, an iterative process. A crucial part of writing is editing and revising the text. Previous works on text revision have focused on defining edit intention taxonomies within a single domain or developing computational models with a single level of edit granularity, such as sentence-level edits, which differ from human's revision cycles. This work describes IteraTeR: the first large-scale, multi-domain, edit-intention annotated corpus of iteratively revised text. In particular, IteraTeR is collected based on a new framework to comprehensively model the iterative text revisions that generalize to various domains of formal writing, edit intentions, revision depths, and granularities. When we incorporate our annotated edit intentions, both generative and edit-based text revision models significantly improve automatic evaluations. Through our work, we better understand the text revision process, making vital connections between edit intentions and writing quality, enabling the creation of diverse corpora to support computational modeling of iterative text revisions.

This paper presents an interactive data dashboard that provides users with an overview of the preservation of discourse relations among 28 language pairs. We display a graph network depicting the cross-lingual discourse relations between a pair of languages for multilingual TED talks and provide a search function to look for sentences with specific keywords or relation types, facilitating ease of analysis on the cross-lingual discourse relations.

Recent studies on the analysis of the multilingual representations focus on identifying whether there is an emergence of language-independent representations, or whether a multilingual model partitions its weights among different languages. While most of such work has been conducted in a "black-box" manner, this paper aims to analyze individual components of a multilingual neural translation (NMT) model. In particular, we look at the encoder self-attention and encoder-decoder attention heads (in a many-to-one NMT model) that are more specific to the translation of a certain language pair than others by (1) employing metrics that quantify some aspects of the attention weights such as "variance" or "confidence", and (2) systematically ranking the importance of attention heads with respect to translation quality. Experimental results show that surprisingly, the set of most important attention heads are very similar across the language pairs and that it is possible to remove nearly one-third of the less important heads without hurting the translation quality greatly.

We release a multilingual neural machine translation model, which can be used to translate text in the biomedical domain. The model can translate from 5 languages (French, German, Italian, Korean and Spanish) into English. It is trained with large amounts of generic and biomedical data, using domain tags. Our benchmarks show that it performs near state-of-the-art both on news (generic domain) and biomedical test sets, and that it outperforms the existing publicly released models. We believe that this release will help the large-scale multilingual analysis of the digital content of the COVID-19 crisis and of its effects on society, economy, and healthcare policies. We also release a test set of biomedical text for Korean-English. It consists of 758 sentences from official guidelines and recent papers, all about COVID-19.

This paper describes the system submitted by Papago team for the quality estimation task at WMT 2020. It proposes two key strategies for quality estimation: (1) task-specific pretraining scheme, and (2) task-specific data augmentation. The former focuses on devising learning signals for pretraining that are closely related to the downstream task. We also present data augmentation techniques that simulate the varying levels of errors that the downstream dataset may contain. Thus, our PATQUEST models are exposed to erroneous translations in both stages of task-specific pretraining and finetuning, effectively enhancing their generalization capability. Our submitted models achieve significant improvement over the baselines for Task 1 (Sentence-Level Direct Assessment; EN-DE only), and Task 3 (Document-Level Score).

Disclosed are a terminal, a server and event suggesting methods thereof, the event suggesting method using word embedding, the method including: determining user preference information about a predetermined object by assigning values to a plurality of reference items with respect to user input data corresponding to the object; generating user intent information, which includes values based on the plurality of reference items, based on at least one piece of review data corresponding to a plurality of events for recommendation objects, and accessing a database built up by adding the generated user intent information to the plurality of events; and suggesting an event following the object among the plurality of events based on a similarity between the user preference information about the object and the user intent information of each event included in the database. Thus, the event based on to a user's preference is suggested, thereby providing the event-recommendation service to make a user's satisfaction higher.

Background: With the invention of fitness trackers, it has been possible to continuously monitor a user’s biometric data such as heart rates, number of footsteps taken, and amount of calories burned. This paper names the time series of these three types of biometric data, the user’s “activeness”, and investigates the feasibility in modeling and predicting the long-term activeness of the user.Methods: The dataset used in this study consisted of several months of biometric time-series data gathered by seven users independently. Four recurrent neural network (RNN) architectures–as well as a deep neural network and a simple regression model–were proposed to investigate the performance on predicting the activeness of the user under various length-related hyper-parameter settings. In addition, the learned model was tested to predict the time period when the user’s activeness falls below a certain threshold.Results: A preliminary experimental result shows that each type of activeness data exhibited a short-term autocorrelation; and among the three types of data, the consumed calories and the number of footsteps were positively correlated, while the heart rate data showed almost no correlation with neither of them. It is probably due to this characteristic of the dataset that although the RNN models produced the best results on modeling the user’s activeness, the difference was marginal; and other baseline models, especially the linear regression model, performed quite admirably as well. Further experimental results show that it is feasible to predict a user’s future activeness with precision, for example, a trained RNN model could predict–with the precision of 84%–when the user would be less active within the next hour given the latest 15 min of his activeness data.Conclusions: This paper defines and investigates the notion of a user’s “activeness”, and shows that forecasting the long-term activeness of the user is indeed possible. Such information can be utilized by a health-related application to proactively recommend suitable events or services to the user.

This paper introduces a vocabulary learning application called “Avocado” that aims to provide suitable learning materials to learners by modeling their language proficiency and topical interests. A learner's vocabulary level is estimated through aggregating words that he identifies as difficult in given text passages; and his topical interests are gathered by utilizing the social network (Facebook) profile. The application recommends a set of recent news articles that are 1) at an appropriate level to the learner, and 2) closely related to the topics that he finds interesting.

For the past few decades, action recognition has been attracting many researchers due to its wide use in a variety of applications. Especially with the increasing number of smartphone users, many studies have been conducted using sensors within a smartphone. However, a lot of these studies assume that the users carry the device in specific ways such as by hand, in a pocket, in a bag, etc. This paper investigates the impact of providing an action recognition system with the information of the possession-way of a smartphone, and vice versa. The experimental dataset consists of five possession-ways (hand, backpack, upper-pocket, lower-pocket, and shoulder-bag) and two actions (walking and running) gathered by seven users separately. Various machine learning models including recurrent neural network architectures are employed to explore the relationship between the action recognition and the possession-way recognition. The experimental results show that the assumption of possession-ways of smartphones do affect the performance of action recognition, and vice versa. The results also reveal that a good performance is achieved when both actions and possession-ways are recognized simultaneously.

This paper investigates the performance of Elman-type and Jordan-type recurrent neural networks (RNN) in extracting temporal information from textual data. The RNN architectures are applied to two tasks of TempEval-2 challenge: (1) extracting the extent of TIMEX3 tags and its TYPE, and (2) extracting the extent of EVENT tags and its CLASS attribute. For the first task, the performances of the RNN models are highly comparable to that of the wining entry for the challenge. For the second task, both models outperform the winning entry, attaining nearly full scores.

Previous research in spam detection, especially in email spam filtering, mainly focused on learning a set of discriminative features that are often present in the spam contents. Nowadays, these commercially oriented spams are well detected; the real challenge lies in filtering rather vague spams that do not exhibit distinctive spam keywords. We investigate two ways of detecting such spams: 1) By comparing the similarity between the publisher posts and user comments, and 2) by learning a single representative meta-feature such as user name or ID. The first measure relieves us from repetitively learning a set of domain-dependent spam features, and the second measure enables us to detect potential spam users even before the aggressive actions are performed. Prior to the language model comparison in the first method, we supplement the background information, normalize the text, perform co-reference resolution, and conduct word-to-word similarity measure in hope of enriching the language models to improve the classification accuracy. To evaluate the first measure, experiments on detecting blog-spam comments are conducted. As for the second measure, we employ SVM on the ID space of e-mail data collected by “Apache Spam Assassin”.

As documents tend to contain temporal information, extracting such information is attracting much research interests recently. In this paper, we propose a hybrid method that combines machine-learning models and hand-crafted rules for the task of extracting temporal information from unstructured Korean texts. We address Korean-specific research issues and propose a new probabilistic model to generate complementary features. The performance of our approach is demonstrated by experiments on the TempEval-2 dataset, and the Korean TimeBank dataset which we built for this study.

Generating long, coherent text remains a challenge for large language models (LLMs), as they lack hierarchical planning and structured organization in discourse generation. We introduce Structural Alignment, a novel method that aligns LLMs with human-like discourse structures to enhance long-form text generation. By integrating linguistically grounded discourse frameworks into reinforcement learning, our approach guides models to produce coherent and well-organized outputs. We employ a dense reward scheme within a Proximal Policy Optimization framework, assigning fine-grained, token-level rewards based on the discourse distinctiveness relative to human writing. Two complementary reward models are evaluated: the first improves readability by scoring surface-level textual features to provide explicit structuring, while the second reinforces deeper coherence and rhetorical sophistication by analyzing global discourse patterns through hierarchical discourse motifs, outperforming both standard and RLHF-enhanced models in tasks such as essay generation and long-document summarization. All training data and code will be publicly shared at https://github.com/minnesotanlp/struct_align.

With the advent of large language models (LLM), the line between human-crafted and machine-generated texts has become increasingly blurred. This paper delves into the inquiry of identifying discernible and unique linguistic properties in texts that were written by humans, particularly uncovering the underlying discourse structures of texts beyond their surface structures. Introducing a novel methodology, we leverage hierarchical parse trees and recursive hypergraphs to unveil distinctive discourse patterns in texts produced by both LLMs and humans. Empirical findings demonstrate that, although both LLMs and humans generate distinct discourse patterns influenced by specific domains, human-written texts exhibit more structural variability, reflecting the nuanced nature of human writing in different domains. Notably, incorporating hierarchical discourse features enhances binary classifiers' overall performance in distinguishing between human-written and machine-generated texts, even on out-of-distribution and paraphrased samples. This underscores the significance of incorporating hierarchical discourse features in the analysis of text patterns. The code and dataset are available at https://github.com/minnesotanlp/threads-of-subtlety.

Conversational implicatures are pragmatic inferences that require listeners to deduce the intended meaning conveyed by a speaker from their explicit utterances. Although such inferential reasoning is fundamental to human communication, recent research indicates that large language models struggle to comprehend these implicatures as effectively as the average human. This paper demonstrates that by incorporating Grice's Four Maxims into the model through chain-of-thought prompting, we can significantly enhance its performance, surpassing even the average human performance on this task.

This paper presents an interactive data dashboard that provides users with an overview of the preservation of discourse relations among 28 language pairs. We display a graph network depicting the cross-lingual discourse relations between a pair of languages for multilingual TED talks and provide a search function to look for sentences with specific keywords or relation types, facilitating ease of analysis on the cross-lingual discourse relations.

Self-improving LLM agents refine answers, not the process that produces those answers. Systems that add a meta-level hold that level fixed, and those that edit themselves must leave part of their own editing machinery untouched to stay stable, capping the meta-depth they realize at roughly two. We present Meta^n, which keeps the meta-operation fixed and recurses on its input instead. That operation, Ω, is applied repeatedly to its own products, reading the traces of the solver stack below together with the code that produced them, then writing the next layer as a strategic pre-process and a library of callable helpers. Because Ω never changes, it cannot destabilize the system, and because its input strictly grows, each layer reasons from a higher vantage than the last. Depth is set by convergence rather than fixed in advance, and an evolutionary archive searches over layer chains. Across two backbones, Meta^n outperforms prior self-improving agents on all eight benchmark families. The sharpest case is ARC-AGI-2, built to resist skill memorization, where it alone scores above zero. Ablations indicate that most of the gain from recursion comes from the conditioning each layer passes to the next, and distinct layer roles emerge with depth although no prompt prescribes them. Code available at https://github.com/minnesotanlp/meta-n

Reward-based alignment methods for large language models (LLMs) face two key limitations: vulnerability to reward hacking, where models exploit flaws in the reward signal; and reliance on brittle, labor-intensive prompt engineering when LLMs are used as reward models. We introduce Meta Policy Optimization (MPO), a framework that addresses these challenges by integrating a meta-reward model that dynamically refines the reward model's prompt throughout training. In MPO, the meta-reward model monitors the evolving training context and continuously adjusts the reward model's prompt to maintain high alignment, providing an adaptive reward signal that resists exploitation by the policy. This meta-learning approach promotes a more stable policy optimization, and greatly reduces the need for manual reward prompt design. It yields performance on par with or better than models guided by extensively hand-crafted reward prompts. Furthermore, we show that MPO maintains its effectiveness across diverse tasks, from essay writing to mathematical reasoning, without requiring specialized reward designs. Beyond standard RLAIF, MPO's meta-learning formulation is readily extensible to higher-level alignment frameworks. Overall, this method addresses theoretical and practical challenges in reward-based RL alignment for LLMs, paving the way for more robust and adaptable alignment strategies. The code and data can be accessed at: https://github.com/minnesotanlp/mpo

Large Language Models are cognitively biased judges. Large Language Models (LLMs) have recently been shown to be effective as automatic evaluators with simple prompting and in-context learning. In this work, we assemble 15 LLMs of four different size ranges and evaluate their output responses by preference ranking from the other LLMs as evaluators, such as System Star is better than System Square. We then evaluate the quality of ranking outputs introducing the Cognitive Bias Benchmark for LLMs as Evaluators (CoBBLEr), a benchmark to measure six different cognitive biases in LLM evaluation outputs, such as the Egocentric bias where a model prefers to rank its own outputs highly in evaluation. We find that LLMs are biased text quality evaluators, exhibiting strong indications on our bias benchmark (average of 40% of comparisons across all models) within each of their evaluations that question their robustness as evaluators. Furthermore, we examine the correlation between human and machine preferences and calculate the average Rank-Biased Overlap (RBO) score to be 49.6%, indicating that machine preferences are misaligned with humans. According to our findings, LLMs may still be unable to be utilized for automatic annotation aligned with human preferences. Our project page is at: https://minnesotanlp.github.io/cobbler.

Scientific discovery is an extended process of ideation–surveying prior work, forming hypotheses, and refining reasoning–yet existing approaches treat this phase as a brief preamble despite its central role in research. We introduce SCISENSE, a sensemaking-grounded framework that operationalizes ideation as a structured sequence of eight cognitive stages (Pirolli & Card, 2005). We construct SCISENSE-Traj, a 100K-scale dataset of citation-conditioned research trajectories in two modes: Target, where an LLM reconstructs the ideation path leading to a known paper from its cited works, and Infer, where the LLM proposes novel directions from the same citations. We distill these into SCISENSE-LM, a family of sensemaking LLMs spanning 3B to 70B parameters. Contrary to the assumption that looser supervision promotes greater exploration, Target-trained models achieve a 2.0% improvement in trajectory quality over Infer-trained models while also producing more novel and diverse outputs. This advantage propagates downstream: coding agents conditioned on Target trajectories produce research artifacts with higher executability and quality than those conditioned on Infer trajectories. This suggests that targeted ideation reduces cognitive burden on downstream agents, freeing them to explore more creatively. SCISENSE offers both a practical tool for augmenting LLM-driven research workflows and a principled testbed for studying how planning shapes scientific discovery.

Existing tool-use benchmarks for LLM agents are overwhelmingly linear: our analysis of six benchmarks shows 55 to 100% of instances are simple chains of 2 to 5 steps. We introduce The Amazing Agent Race (AAR), a benchmark featuring directed acyclic graph (DAG) puzzles (or "legs") with fork-merge tool chains. We release 1,400 instances across two variants: sequential (800 legs) and compositional (600 DAG legs). Agents must navigate Wikipedia, execute multi-step tool chains, and aggregate results into a verifiable answer. Legs are procedurally generated from Wikipedia seeds across four difficulty levels with live-API validation. Three complementary metrics (finish-line accuracy, pit-stop visit rate, and roadblock completion rate) separately diagnose navigation, tool-use, and arithmetic failures. Evaluating three agent frameworks on 1,400 legs, the best achieves only 37.2% accuracy. Navigation errors dominate (27 to 52% of trials) while tool-use errors remain below 17%, and agent architecture matters as much as model scale (Claude Code matches Codex CLI at 37% with 6x fewer tokens). The compositional structure of AAR reveals that agents fail not at calling tools but at navigating to the right pages, a blind spot invisible to linear benchmarks. The project page can be accessed at: https://minnesotanlp.github.io/the-amazing-agent-race

Iterative text revision improves text quality by fixing grammatical errors, rephrasing for better readability or contextual appropriateness, or reorganizing sentence structures throughout a document. Most recent research has focused on understanding and classifying different types of edits in the iterative revision process from human-written text instead of building accurate and robust systems for iterative text revision. In this work, we aim to build an end-to-end text revision system that can iteratively generate helpful edits by explicitly detecting editable spans (where-to-edit) with their corresponding edit intents and then instructing a revision model to revise the detected edit spans. Leveraging datasets from other related text editing NLP tasks, combined with the specification of editable spans, leads our system to more accurately model the process of iterative text refinement, as evidenced by empirical results and human evaluations. Our system significantly outperforms previous baselines on our text revision tasks and other standard text revision tasks, including grammatical error correction, text simplification, sentence fusion, and style transfer. Through extensive qualitative and quantitative analysis, we make vital connections between edit intentions and writing quality, and better computational modeling of iterative text revisions.

Revision is an essential part of the human writing process. It tends to be strategic, adaptive, and, more importantly, iterative in nature. Despite the success of large language models on text revision tasks, they are limited to non-iterative, one-shot revisions. Examining and evaluating the capability of large language models for making continuous revisions and collaborating with human writers is a critical step towards building effective writing assistants. In this work, we present a human-in-the-loop iterative text revision system, Read, Revise, Repeat (R3), which aims at achieving high quality text revisions with minimal human efforts by reading model-generated revisions and user feedbacks, revising documents, and repeating human-machine interactions. In R3, a text revision model provides text editing suggestions for human writers, who can accept or reject the suggested edits. The accepted edits are then incorporated into the model for the next iteration of document revision. Writers can therefore revise documents iteratively by interacting with the system and simply accepting/rejecting its suggested edits until the text revision model stops making further revisions or reaches a predefined maximum number of revisions. Empirical experiments show that R3 can generate revisions with comparable acceptance rate to human writers at early revision depths, and the human-machine interaction can get higher quality revisions with fewer iterations and edits. The collected human-model interaction dataset and system code are available at https://github.com/vipulraheja/IteraTeR. Our system demonstration is available at https://youtu.be/lK08tIpEoaE.

Writing is, by nature, a strategic, adaptive, and more importantly, an iterative process. A crucial part of writing is editing and revising the text. Previous works on text revision have focused on defining edit intention taxonomies within a single domain or developing computational models with a single level of edit granularity, such as sentence-level edits, which differ from human's revision cycles. This work describes IteraTeR: the first large-scale, multi-domain, edit-intention annotated corpus of iteratively revised text. In particular, IteraTeR is collected based on a new framework to comprehensively model the iterative text revisions that generalize to various domains of formal writing, edit intentions, revision depths, and granularities. When we incorporate our annotated edit intentions, both generative and edit-based text revision models significantly improve automatic evaluations. Through our work, we better understand the text revision process, making vital connections between edit intentions and writing quality, enabling the creation of diverse corpora to support computational modeling of iterative text revisions.

Recent studies on the analysis of the multilingual representations focus on identifying whether there is an emergence of language-independent representations, or whether a multilingual model partitions its weights among different languages. While most of such work has been conducted in a "black-box" manner, this paper aims to analyze individual components of a multilingual neural translation (NMT) model. In particular, we look at the encoder self-attention and encoder-decoder attention heads (in a many-to-one NMT model) that are more specific to the translation of a certain language pair than others by (1) employing metrics that quantify some aspects of the attention weights such as "variance" or "confidence", and (2) systematically ranking the importance of attention heads with respect to translation quality. Experimental results show that surprisingly, the set of most important attention heads are very similar across the language pairs and that it is possible to remove nearly one-third of the less important heads without hurting the translation quality greatly.

We release a multilingual neural machine translation model, which can be used to translate text in the biomedical domain. The model can translate from 5 languages (French, German, Italian, Korean and Spanish) into English. It is trained with large amounts of generic and biomedical data, using domain tags. Our benchmarks show that it performs near state-of-the-art both on news (generic domain) and biomedical test sets, and that it outperforms the existing publicly released models. We believe that this release will help the large-scale multilingual analysis of the digital content of the COVID-19 crisis and of its effects on society, economy, and healthcare policies. We also release a test set of biomedical text for Korean-English. It consists of 758 sentences from official guidelines and recent papers, all about COVID-19.

This paper describes the system submitted by Papago team for the quality estimation task at WMT 2020. It proposes two key strategies for quality estimation: (1) task-specific pretraining scheme, and (2) task-specific data augmentation. The former focuses on devising learning signals for pretraining that are closely related to the downstream task. We also present data augmentation techniques that simulate the varying levels of errors that the downstream dataset may contain. Thus, our PATQUEST models are exposed to erroneous translations in both stages of task-specific pretraining and finetuning, effectively enhancing their generalization capability. Our submitted models achieve significant improvement over the baselines for Task 1 (Sentence-Level Direct Assessment; EN-DE only), and Task 3 (Document-Level Score).

A unified representation space in multi-modal learning is essential for effectively integrating diverse data sources, such as text, images, and audio, to enhance efficiency and performance across various downstream tasks. Recent binding methods, such as ImageBind, typically rely on a single, fixed anchor modality for aligning multi-modal data. We mathematically analyze these fixed anchor binding methods and uncover significant limitations: (1) over-reliance on the choice of the anchor modality, (2) inadequate capture of intra-modal information, and (3) failure to account for cross-modal correlation among non-anchored modalities. To address these issues, we propose the need for adaptive anchor binding methods, exemplified by our framework CentroBind. The proposed method uses adaptively adjustable centroid-based anchors generated from all available modalities, leading to a balanced and rich representation space. We theoretically demonstrate that our approach captures three critical properties of multi-modal learning – intra-modal learning, inter-modal learning, and multi-modal alignment – while constructing a unified representation that spans all modalities. Experiments on both synthetic and real-world datasets show that adaptive anchor methods such as CentroBind consistently outperform fixed anchor binding methods, verifying our analysis.

This work delves into the expanding role of large language models (LLMs) in generating artificial data. LLMs are increasingly employed to create a variety of outputs, including annotations, preferences, instruction prompts, simulated dialogues, and free text. As these forms of LLM-generated data often intersect in their application, they exert mutual influence on each other and raise significant concerns about the quality and diversity of the artificial data incorporated into training cycles, leading to an artificial data ecosystem. To the best of our knowledge, this is the first study to aggregate various types of LLM-generated text data, from more tightly constrained data like "task labels" to more lightly constrained "free-form text". We then stress test the quality and implications of LLM-generated artificial data, comparing it with human data across various existing benchmarks. Despite artificial data's capability to match human performance, this paper reveals significant hidden disparities, especially in complex tasks where LLMs often miss the nuanced understanding of intrinsic human-generated content. This study critically examines diverse LLM-generated data and emphasizes the need for ethical practices in data creation and when using LLMs. It highlights the LLMs' shortcomings in replicating human traits and behaviors, underscoring the importance of addressing biases and artifacts produced in LLM-generated content for future research and development. All data and code are available on our project page.

Capturing readers' engagement in fiction is a challenging but important aspect of narrative understanding. In this study, we collected 23 readers' reactions to 2 short stories through eye tracking, sentence-level annotations, and an overall engagement scale survey. We analyzed the significance of various qualities of the text in predicting how engaging a reader is likely to find it. As enjoyment of fiction is highly contextual, we also investigated individual differences in our data. Furthering our understanding of what captivates readers in fiction will help better inform models used in creative narrative generation and collaborative writing tools.

This survey paper provides a comprehensive review of the use of diffusion models in natural language processing (NLP). Diffusion models are a class of mathematical models that aim to capture the diffusion of information or signals across a network or manifold. In NLP, diffusion models have been used in a variety of applications, such as natural language generation, sentiment analysis, topic modeling, and machine translation. This paper discusses the different formulations of diffusion models used in NLP, their strengths and limitations, and their applications. We also perform a thorough comparison between diffusion models and alternative generative models, specifically highlighting the autoregressive (AR) models, while also examining how diverse architectures incorporate the Transformer in conjunction with diffusion models. Compared to AR models, diffusion models have significant advantages for parallel generation, text interpolation, token-level controls such as syntactic structures and semantic contents, and robustness. Exploring further permutations of integrating Transformers into diffusion models would be a valuable pursuit. Also, the development of multimodal diffusion models and large-scale diffusion language models with notable capabilities for few-shot learning would be important directions for the future advance of diffusion models in NLP.

Disclosed are a terminal, a server and event suggesting methods thereof, the event suggesting method using word embedding, the method including: determining user preference information about a predetermined object by assigning values to a plurality of reference items with respect to user input data corresponding to the object; generating user intent information, which includes values based on the plurality of reference items, based on at least one piece of review data corresponding to a plurality of events for recommendation objects, and accessing a database built up by adding the generated user intent information to the plurality of events; and suggesting an event following the object among the plurality of events based on a similarity between the user preference information about the object and the user intent information of each event included in the database. Thus, the event based on to a user's preference is suggested, thereby providing the event-recommendation service to make a user's satisfaction higher.

Background: With the invention of fitness trackers, it has been possible to continuously monitor a user’s biometric data such as heart rates, number of footsteps taken, and amount of calories burned. This paper names the time series of these three types of biometric data, the user’s “activeness”, and investigates the feasibility in modeling and predicting the long-term activeness of the user.Methods: The dataset used in this study consisted of several months of biometric time-series data gathered by seven users independently. Four recurrent neural network (RNN) architectures–as well as a deep neural network and a simple regression model–were proposed to investigate the performance on predicting the activeness of the user under various length-related hyper-parameter settings. In addition, the learned model was tested to predict the time period when the user’s activeness falls below a certain threshold.Results: A preliminary experimental result shows that each type of activeness data exhibited a short-term autocorrelation; and among the three types of data, the consumed calories and the number of footsteps were positively correlated, while the heart rate data showed almost no correlation with neither of them. It is probably due to this characteristic of the dataset that although the RNN models produced the best results on modeling the user’s activeness, the difference was marginal; and other baseline models, especially the linear regression model, performed quite admirably as well. Further experimental results show that it is feasible to predict a user’s future activeness with precision, for example, a trained RNN model could predict–with the precision of 84%–when the user would be less active within the next hour given the latest 15 min of his activeness data.Conclusions: This paper defines and investigates the notion of a user’s “activeness”, and shows that forecasting the long-term activeness of the user is indeed possible. Such information can be utilized by a health-related application to proactively recommend suitable events or services to the user.

This paper introduces a vocabulary learning application called “Avocado” that aims to provide suitable learning materials to learners by modeling their language proficiency and topical interests. A learner's vocabulary level is estimated through aggregating words that he identifies as difficult in given text passages; and his topical interests are gathered by utilizing the social network (Facebook) profile. The application recommends a set of recent news articles that are 1) at an appropriate level to the learner, and 2) closely related to the topics that he finds interesting.

For the past few decades, action recognition has been attracting many researchers due to its wide use in a variety of applications. Especially with the increasing number of smartphone users, many studies have been conducted using sensors within a smartphone. However, a lot of these studies assume that the users carry the device in specific ways such as by hand, in a pocket, in a bag, etc. This paper investigates the impact of providing an action recognition system with the information of the possession-way of a smartphone, and vice versa. The experimental dataset consists of five possession-ways (hand, backpack, upper-pocket, lower-pocket, and shoulder-bag) and two actions (walking and running) gathered by seven users separately. Various machine learning models including recurrent neural network architectures are employed to explore the relationship between the action recognition and the possession-way recognition. The experimental results show that the assumption of possession-ways of smartphones do affect the performance of action recognition, and vice versa. The results also reveal that a good performance is achieved when both actions and possession-ways are recognized simultaneously.

This paper investigates the performance of Elman-type and Jordan-type recurrent neural networks (RNN) in extracting temporal information from textual data. The RNN architectures are applied to two tasks of TempEval-2 challenge: (1) extracting the extent of TIMEX3 tags and its TYPE, and (2) extracting the extent of EVENT tags and its CLASS attribute. For the first task, the performances of the RNN models are highly comparable to that of the wining entry for the challenge. For the second task, both models outperform the winning entry, attaining nearly full scores.

Previous research in spam detection, especially in email spam filtering, mainly focused on learning a set of discriminative features that are often present in the spam contents. Nowadays, these commercially oriented spams are well detected; the real challenge lies in filtering rather vague spams that do not exhibit distinctive spam keywords. We investigate two ways of detecting such spams: 1) By comparing the similarity between the publisher posts and user comments, and 2) by learning a single representative meta-feature such as user name or ID. The first measure relieves us from repetitively learning a set of domain-dependent spam features, and the second measure enables us to detect potential spam users even before the aggressive actions are performed. Prior to the language model comparison in the first method, we supplement the background information, normalize the text, perform co-reference resolution, and conduct word-to-word similarity measure in hope of enriching the language models to improve the classification accuracy. To evaluate the first measure, experiments on detecting blog-spam comments are conducted. As for the second measure, we employ SVM on the ID space of e-mail data collected by “Apache Spam Assassin”.

As documents tend to contain temporal information, extracting such information is attracting much research interests recently. In this paper, we propose a hybrid method that combines machine-learning models and hand-crafted rules for the task of extracting temporal information from unstructured Korean texts. We address Korean-specific research issues and propose a new probabilistic model to generate complementary features. The performance of our approach is demonstrated by experiments on the TempEval-2 dataset, and the Korean TimeBank dataset which we built for this study.