Hayate Iso

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Bosung Kim, Hayate Iso, Nikita Bhutani et al. · nvidia

The task of triplet extraction aims to extract pairs of entities and their corresponding relations from unstructured text. Most existing methods train an extraction model on training data involving specific target relations, and are incapable of extracting new relations that were not observed at training time. Generalizing the model to unseen relations typically requires fine-tuning on synthetic training data which is often noisy and unreliable. We show that by reducing triplet extraction to a template infilling task over a pre-trained language model (LM), we can equip the extraction model with zero-shot learning capabilities and eliminate the need for additional training data. We propose a novel framework, ZETT (ZEro-shot Triplet extraction by Template infilling), that aligns the task objective to the pre-training objective of generative transformers to generalize to unseen relations. Experiments on FewRel and Wiki-ZSL datasets demonstrate that ZETT shows consistent and stable performance, outperforming previous state-of-the-art methods, even when using automatically generated templates. https://github.com/megagonlabs/zett/

Hayate Iso, Xiaolan Wang, Yoshi Suhara · nvidia

Opinion summarization research has primarily focused on generating summaries reflecting important opinions from customer reviews without paying much attention to the writing style. In this paper, we propose the stylized opinion summarization task, which aims to generate a summary of customer reviews in the desired (e.g., professional) writing style. To tackle the difficulty in collecting customer and professional review pairs, we develop a non-parallel training framework, Noisy Pairing and Partial Supervision (NAPA), which trains a stylized opinion summarization system from non-parallel customer and professional review sets. We create a benchmark ProSum by collecting customer and professional reviews from Yelp and Michelin. Experimental results on ProSum and FewSum demonstrate that our non-parallel training framework consistently improves both automatic and human evaluations, successfully building a stylized opinion summarization model that can generate professionally-written summaries from customer reviews. The code is available at https://github.com/megagonlabs/napa

Hayate Iso · nvidia

Lexically constrained text generation is one of the constrained text generation tasks, which aims to generate text that covers all the given constraint lexicons. While the existing approaches tackle this problem using a lexically constrained beam search algorithm or dedicated model using non-autoregressive decoding, there is a trade-off between the generated text quality and the hard constraint satisfaction. We introduce AutoTemplate, a simple yet effective lexically constrained text generation framework divided into template generation and lexicalization tasks. The template generation is to generate the text with the placeholders, and lexicalization replaces them into the constraint lexicons to perform lexically constrained text generation. We conducted the experiments on two tasks: keywords-to-sentence generations and entity-guided summarization. Experimental results show that the AutoTemplate outperforms the competitive baselines on both tasks while satisfying the hard lexical constraints. The code is available at https://github.com/megagonlabs/autotemplate

Haopeng Zhang, Hayate Iso, Sairam Gurajada et al. · nvidia

Text editing is a crucial task of modifying text to better align with user intents. However, existing text editing benchmark datasets contain only coarse-grained instructions and lack explainability, thus resulting in outputs that deviate from the intended changes outlined in the gold reference. To comprehensively investigate the text editing capabilities of large language models (LLMs), this paper introduces XATU, the first benchmark specifically designed for fine-grained instruction-based explainable text editing. XATU considers finer-grained text editing tasks of varying difficulty (simplification, grammar check, fact-check, etc.), incorporating lexical, syntactic, semantic, and knowledge-intensive edit aspects. To enhance interpretability, we combine LLM-based annotation and human annotation, resulting in a benchmark that includes fine-grained instructions and gold-standard edit explanations. By evaluating existing LLMs against our benchmark, we demonstrate the effectiveness of instruction tuning and the impact of underlying architecture across various editing tasks. Furthermore, extensive experimentation reveals the significant role of explanations in fine-tuning language models for text editing tasks. The benchmark will be open-sourced to support reproduction and facilitate future research at~\url{https://github.com/megagonlabs/xatu}.

Yunshu Wu, Hayate Iso, Pouya Pezeshkpour et al. · nvidia

Large Language Models (LLMs) have shown promising performance in summary evaluation tasks, yet they face challenges such as high computational costs and the Lost-in-the-Middle problem where important information in the middle of long documents is often overlooked. To address these issues, this paper introduces a novel approach, Extract-then-Evaluate, which involves extracting key sentences from a long source document and then evaluating the summary by prompting LLMs. The results reveal that the proposed method not only significantly reduces evaluation costs but also exhibits a higher correlation with human evaluations. Furthermore, we provide practical recommendations for optimal document length and sentence extraction methods, contributing to the development of cost-effective yet more accurate methods for LLM-based text generation evaluation.

Pouya Pezeshkpour, Hayate Iso, Thom Lake et al. · nvidia

Numerous HR applications are centered around resumes and job descriptions. While they can benefit from advancements in NLP, particularly large language models, their real-world adoption faces challenges due to absence of comprehensive benchmarks for various HR tasks, and lack of smaller models with competitive capabilities. In this paper, we aim to bridge this gap by introducing the Resume-Job Description Benchmark (RJDB). We meticulously craft this benchmark to cater to a wide array of HR tasks, including matching and explaining resumes to job descriptions, extracting skills and experiences from resumes, and editing resumes. To create this benchmark, we propose to distill domain-specific knowledge from a large language model (LLM). We rely on a curated skill-occupation graph to ensure diversity and provide context for LLMs generation. Our benchmark includes over 50 thousand triples of job descriptions, matched resumes and unmatched resumes. Using RJDB, we train multiple smaller student models. Our experiments reveal that the student models achieve near/better performance than the teacher model (GPT-4), affirming the effectiveness of the benchmark. Additionally, we explore the utility of RJDB on out-of-distribution data for skill extraction and resume-job description matching, in zero-shot and weak supervision manner. We release our datasets and code to foster further research and industry applications.

Catarina G. Belem, Pouya Pezeshkpour, Hayate Iso et al. · nvidia

Although many studies have investigated and reduced hallucinations in large language models (LLMs) for single-document tasks, research on hallucination in multi-document summarization (MDS) tasks remains largely unexplored. Specifically, it is unclear how the challenges arising from handling multiple documents (e.g., repetition and diversity of information) affect models outputs. In this work, we investigate how hallucinations manifest in LLMs when summarizing topic-specific information from multiple documents. Since no benchmarks exist for investigating hallucinations in MDS, we use existing news and conversation datasets, annotated with topic-specific insights, to create two novel multi-document benchmarks. When evaluating 5 LLMs on our benchmarks, we observe that on average, up to 75% of the content in LLM-generated summary is hallucinated, with hallucinations more likely to occur towards the end of the summaries. Moreover, when summarizing non-existent topic-related information, gpt-3.5-turbo and GPT-4o still generate summaries about 79.35% and 44% of the time, raising concerns about their tendency to fabricate content. To understand the characteristics of these hallucinations, we manually evaluate 700+ insights and find that most errors stem from either failing to follow instructions or producing overly generic insights. Motivated by these observations, we investigate the efficacy of simple post-hoc baselines in mitigating hallucinations but find them only moderately effective. Our results underscore the need for more effective approaches to systematically mitigate hallucinations in MDS. We release our dataset and code at github.com/megagonlabs/Hallucination_MDS.

Hayate Iso, Xiaolan Wang, Stefanos Angelidis et al.

Opinion summarization focuses on generating summaries that reflect popular subjective information expressed in multiple online reviews. While generated summaries offer general and concise information about a particular hotel or product, the information may be insufficient to help the user compare multiple different choices. Thus, the user may still struggle with the question "Which one should I pick?" In this paper, we propose the comparative opinion summarization task, which aims at generating two contrastive summaries and one common summary from two different candidate sets of reviews. We develop a comparative summarization framework CoCoSum, which consists of two base summarization models that jointly generate contrastive and common summaries. Experimental results on a newly created benchmark CoCoTrip show that CoCoSum can produce higher-quality contrastive and common summaries than state-of-the-art opinion summarization models. The dataset and code are available at https://github.com/megagonlabs/cocosum

Hayate Iso, Xiaolan Wang, Yoshihiko Suhara et al.

Recent advances in text autoencoders have significantly improved the quality of the latent space, which enables models to generate grammatical and consistent text from aggregated latent vectors. As a successful application of this property, unsupervised opinion summarization models generate a summary by decoding the aggregated latent vectors of inputs. More specifically, they perform the aggregation via simple average. However, little is known about how the vector aggregation step affects the generation quality. In this study, we revisit the commonly used simple average approach by examining the latent space and generated summaries. We found that text autoencoders tend to generate overly generic summaries from simply averaged latent vectors due to an unexpected $L_2$-norm shrinkage in the aggregated latent vectors, which we refer to as summary vector degeneration. To overcome this issue, we develop a framework Coop, which searches input combinations for the latent vector aggregation using input-output word overlap. Experimental results show that Coop successfully alleviates the summary vector degeneration issue and establishes new state-of-the-art performance on two opinion summarization benchmarks. Code is available at \url{https://github.com/megagonlabs/coop}.

Hayate Iso, Tiyasa Mitra, Sudipta Mondal et al.

RL post-training of frontier language models is increasingly bottlenecked by autoregressive rollout generation, making rollout acceleration a central systems challenge. Many existing efficiency methods improve throughput by changing the rollout or optimization regime, for example, through off-policy execution, replay, or lower-precision generation. We study speculative decoding as a lossless acceleration primitive for RL rollouts that preserves the target model's output distribution. We implement speculative decoding in NeMo-RL with a vLLM backend, supporting both synchronous and asynchronous pipelines and enabling speculation during RL rollouts. This benefit is realizable across speculation mechanisms, such as pretrained MTP heads, small external draft models or even techniques such as Eagle3, which are traditionally applied after RL phase. This yields a deployment path for state-of-the-art speculative decoding inside RL training. In a reasoning post-training workload at 8B scale under synchronous RL, speculative decoding improves rollout throughput by 1.8x. Using a high-fidelity performance simulator, we project that combining speculative decoding with asynchronous RL yields up to 2.5x end-to-end training speedup at 235B scale.

Seiji Maekawa, Hayate Iso, Sairam Gurajada et al. · nvidia

While large language models (LMs) demonstrate remarkable performance, they encounter challenges in providing accurate responses when queried for information beyond their pre-trained memorization. Although augmenting them with relevant external information can mitigate these issues, failure to consider the necessity of retrieval may adversely affect overall performance. Previous research has primarily focused on examining how entities influence retrieval models and knowledge recall in LMs, leaving other aspects relatively unexplored. In this work, our goal is to offer a more detailed, fact-centric analysis by exploring the effects of combinations of entities and relations. To facilitate this, we construct a new question answering (QA) dataset called WiTQA (Wikipedia Triple Question Answers). This dataset includes questions about entities and relations of various popularity levels, each accompanied by a supporting passage. Our extensive experiments with diverse LMs and retrievers reveal when retrieval does not consistently enhance LMs from the viewpoints of fact-centric popularity. Confirming earlier findings, we observe that larger LMs excel in recalling popular facts. However, they notably encounter difficulty with infrequent entity-relation pairs compared to retrievers. Interestingly, they can effectively retain popular relations of less common entities. We demonstrate the efficacy of our finer-grained metric and insights through an adaptive retrieval system that selectively employs retrieval and recall based on the frequencies of entities and relations in the question.

Ayana Niwa, Hayate Iso · nvidia

We introduce AmbigNLG, a novel task designed to tackle the challenge of task ambiguity in instructions for Natural Language Generation (NLG). Ambiguous instructions often impede the performance of Large Language Models (LLMs), especially in complex NLG tasks. To tackle this issue, we propose an ambiguity taxonomy that categorizes different types of instruction ambiguities and refines initial instructions with clearer specifications. Accompanying this task, we present AmbigSNI-NLG, a dataset comprising 2,500 instances annotated to facilitate research in AmbigNLG. Through comprehensive experiments with state-of-the-art LLMs, we demonstrate that our method significantly enhances the alignment of generated text with user expectations, achieving up to a 15.02-point increase in ROUGE scores. Our findings highlight the critical importance of addressing task ambiguity to fully harness the capabilities of LLMs in NLG tasks. Furthermore, we confirm the effectiveness of our method in practical settings involving interactive ambiguity mitigation with users, underscoring the benefits of leveraging LLMs for interactive clarification.

Seiji Maekawa, Hayate Iso, Nikita Bhutani · nvidia

The rapid increase in textual information means we need more efficient methods to sift through, organize, and understand it all. While retrieval-augmented generation (RAG) models excel in accessing information from large document collections, they struggle with complex tasks that require aggregation and reasoning over information spanning across multiple documents--what we call holistic reasoning. Long-context language models (LCLMs) have great potential for managing large-scale documents, but their holistic reasoning capabilities remain unclear. In this work, we introduce HoloBench, a novel framework that brings database reasoning operations into text-based contexts, making it easier to systematically evaluate how LCLMs handle holistic reasoning across large documents. Our approach adjusts key factors such as context length, information density, distribution of information, and query complexity to evaluate LCLMs comprehensively. Our experiments show that the amount of information in the context has a bigger influence on LCLM performance than the actual context length. Furthermore, the complexity of queries affects performance more than the amount of information, particularly for different types of queries. Interestingly, queries that involve finding maximum or minimum values are easier for LCLMs and are less affected by context length, even though they pose challenges for RAG systems. However, tasks requiring the aggregation of multiple pieces of information show a noticeable drop in accuracy as context length increases. Additionally, we find that while grouping relevant information generally improves performance, the optimal positioning varies across models. Our findings surface both the advancements and the ongoing challenges in achieving a holistic understanding of long contexts.

Hayate Iso, Pouya Pezeshkpour, Nikita Bhutani et al. · nvidia

Large Language Models (LLMs) offer the potential to automate hiring by matching job descriptions with candidate resumes, streamlining recruitment processes, and reducing operational costs. However, biases inherent in these models may lead to unfair hiring practices, reinforcing societal prejudices and undermining workplace diversity. This study examines the performance and fairness of LLMs in job-resume matching tasks within the English language and U.S. context. It evaluates how factors such as gender, race, and educational background influence model decisions, providing critical insights into the fairness and reliability of LLMs in HR applications. Our findings indicate that while recent models have reduced biases related to explicit attributes like gender and race, implicit biases concerning educational background remain significant. These results highlight the need for ongoing evaluation and the development of advanced bias mitigation strategies to ensure equitable hiring practices when using LLMs in industry settings.

Seiji Maekawa, Hayate Iso, Nikita Bhutani · nvidia

Effective decision-making often relies on identifying what makes each candidate distinctive. While existing benchmarks for LLMs emphasize retrieving or summarizing information relevant to a given query, they do not evaluate a model's ability to identify globally distinctive features across a set of documents. We introduce Distinctive Feature Mining (DFM), a new task that challenges models to analyze a small-to-medium collection (10-40 documents) and surface features that are rare in the global context (e.g., appearing in less than 10% of documents). This setting mirrors real-world scenarios such as candidate selection or product differentiation, where statistical reasoning, not retrieval, is key. To enable systematic evaluation of this capability, we present DiFBench, a configurable benchmark creation framework with controllable parameters such as document set size and distinctiveness thresholds. Using DiFBench, we perform a large-scale assessment of distinctive feature mining across ten state-of-the-art LLMs. Our findings reveal a significant performance gap between general-purpose and reasoning-enhanced models. All models, however, substantially degrade as the task complexity and document count increase. We also find that a common failure mode is misidentifying frequent features as distinctive. These insights reveal core limitations in contemporary LLMs' abilities to perform fine-grained, statistical reasoning and rarity detection.

Eser Kandogan, Sajjadur Rahman, Nikita Bhutani et al.

Large Language Models (LLMs) have showcased remarkable capabilities surpassing conventional NLP challenges, creating opportunities for use in production use cases. Towards this goal, there is a notable shift to building compound AI systems, wherein LLMs are integrated into an expansive software infrastructure with many components like models, retrievers, databases and tools. In this paper, we introduce a blueprint architecture for compound AI systems to operate in enterprise settings cost-effectively and feasibly. Our proposed architecture aims for seamless integration with existing compute and data infrastructure, with ``stream'' serving as the key orchestration concept to coordinate data and instructions among agents and other components. Task and data planners, respectively, break down, map, and optimize tasks and data to available agents and data sources defined in respective registries, given production constraints such as accuracy and latency.

Shogo Ujiie, Hayate Iso, Eiji Aramaki

We introduce BioCoM, a contrastive learning framework for biomedical entity linking that uses only two resources: a small-sized dictionary and a large number of raw biomedical articles. Specifically, we build the training instances from raw PubMed articles by dictionary matching and use them to train a context-aware entity linking model with contrastive learning. We predict the normalized biomedical entity at inference time through a nearest-neighbor search. Results found that BioCoM substantially outperforms state-of-the-art models, especially in low-resource settings, by effectively using the context of the entities.

Shogo Ujiie, Hayate Iso, Shuntaro Yada et al.

Disease name recognition and normalization, which is generally called biomedical entity linking, is a fundamental process in biomedical text mining. Recently, neural joint learning of both tasks has been proposed to utilize the mutual benefits. While this approach achieves high performance, disease concepts that do not appear in the training dataset cannot be accurately predicted. This study introduces a novel end-to-end approach that combines span representations with dictionary-matching features to address this problem. Our model handles unseen concepts by referring to a dictionary while maintaining the performance of neural network-based models, in an end-to-end fashion. Experiments using two major datasets demonstrate that our model achieved competitive results with strong baselines, especially for unseen concepts during training.

Hayate Iso, Chao Qiao, Hang Li

We propose a novel text editing task, referred to as \textit{fact-based text editing}, in which the goal is to revise a given document to better describe the facts in a knowledge base (e.g., several triples). The task is important in practice because reflecting the truth is a common requirement in text editing. First, we propose a method for automatically generating a dataset for research on fact-based text editing, where each instance consists of a draft text, a revised text, and several facts represented in triples. We apply the method into two public table-to-text datasets, obtaining two new datasets consisting of 233k and 37k instances, respectively. Next, we propose a new neural network architecture for fact-based text editing, called \textsc{FactEditor}, which edits a draft text by referring to given facts using a buffer, a stream, and a memory. A straightforward approach to address the problem would be to employ an encoder-decoder model. Our experimental results on the two datasets show that \textsc{FactEditor} outperforms the encoder-decoder approach in terms of fidelity and fluency. The results also show that \textsc{FactEditor} conducts inference faster than the encoder-decoder approach.

Hayate Iso, Yui Uehara, Tatsuya Ishigaki et al.

We propose a data-to-text generation model with two modules, one for tracking and the other for text generation. Our tracking module selects and keeps track of salient information and memorizes which record has been mentioned. Our generation module generates a summary conditioned on the state of tracking module. Our model is considered to simulate the human-like writing process that gradually selects the information by determining the intermediate variables while writing the summary. In addition, we also explore the effectiveness of the writer information for generation. Experimental results show that our model outperforms existing models in all evaluation metrics even without writer information. Incorporating writer information further improves the performance, contributing to content planning and surface realization.

Hayate Iso, Shoko Wakamiya, Eiji Aramaki

Nowadays, geographic information related to Twitter is crucially important for fine-grained applications. However, the amount of geographic information avail- able on Twitter is low, which makes the pursuit of many applications challenging. Under such circumstances, estimating the location of a tweet is an important goal of the study. Unlike most previous studies that estimate the pre-defined district as the classification task, this study employs a probability distribution to represent richer information of the tweet, not only the location but also its ambiguity. To realize this modeling, we propose the convolutional mixture density network (CMDN), which uses text data to estimate the mixture model parameters. Experimentally obtained results reveal that CMDN achieved the highest prediction performance among the method for predicting the exact coordinates. It also provides a quantitative representation of the location ambiguity for each tweet that properly works for extracting the reliable location estimations.