Zae Myung Kim

Wanyu Du, Zae Myung Kim, Vipul Raheja et al. · deepmind

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 \url{https://github.com/vipulraheja/IteraTeR}. Our system demonstration is available at \url{https://youtu.be/lK08tIpEoaE}.

Ryan Koo, Minhwa Lee, Vipul Raheja et al. · deepmind

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.

Wanyu Du, Vipul Raheja, Dhruv Kumar et al. · deepmind

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.

Zae Myung Kim, Wanyu Du, Vipul Raheja et al. · deepmind

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.

Zae Myung Kim, Dongseok Lee, Jaehyung Kim et al. · deepmind

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

James Mooney, Zae Myung Kim, Young-Jun Lee et al.

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.

Rose Neis, Karin de Langis, Zae Myung Kim et al.

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.

Zae Myung Kim, Kwang Hee Lee, Preston Zhu et al. · deepmind

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.

Zae Myung Kim, Anand Ramachandran, Farideh Tavazoee et al.

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.

Zae Myung Kim, Chanwoo Park, Vipul Raheja et al. · deepmind

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

Minhwa Lee, Zae Myung Kim, Vivek Khetan et al.

Large Language Models (LLMs) have assisted humans in several writing tasks, including text revision and story generation. However, their effectiveness in supporting domain-specific writing, particularly in business contexts, is relatively less explored. Our formative study with industry professionals revealed the limitations in current LLMs' understanding of the nuances in such domain-specific writing. To address this gap, we propose an approach of human-AI collaborative taxonomy development to perform as a guideline for domain-specific writing assistants. This method integrates iterative feedback from domain experts and multiple interactions between these experts and LLMs to refine the taxonomy. Through larger-scale experiments, we aim to validate this methodology and thus improve LLM-powered writing assistance, tailoring it to meet the unique requirements of different stakeholder needs.

Debarati Das, Karin De Langis, Anna Martin-Boyle et al.

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.

Hao Zou, Zae Myung Kim, Dongyeop Kang

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.

Zae Myung Kim, David E. Taylor, Dongyeop Kang

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.

Zae Myung Kim, Laurent Besacier, Vassilina Nikoulina et al.

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.

Alexandre Bérard, Zae Myung Kim, Vassilina Nikoulina et al.

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.