Yapei Chang

Yapei Chang, Kyle Lo, Tanya Goyal et al.

Summarizing book-length documents (>100K tokens) that exceed the context window size of large language models (LLMs) requires first breaking the input document into smaller chunks and then prompting an LLM to merge, update, and compress chunk-level summaries. Despite the complexity and importance of this task, it has yet to be meaningfully studied due to the challenges of evaluation: existing book-length summarization datasets (e.g., BookSum) are in the pretraining data of most public LLMs, and existing evaluation methods struggle to capture errors made by modern LLM summarizers. In this paper, we present the first study of the coherence of LLM-based book-length summarizers implemented via two prompting workflows: (1) hierarchically merging chunk-level summaries, and (2) incrementally updating a running summary. We obtain 1193 fine-grained human annotations on GPT-4 generated summaries of 100 recently-published books and identify eight common types of coherence errors made by LLMs. Because human evaluation is expensive and time-consuming, we develop an automatic metric, BooookScore, that measures the proportion of sentences in a summary that do not contain any of the identified error types. BooookScore has high agreement with human annotations and allows us to systematically evaluate the impact of many other critical parameters (e.g., chunk size, base LLM) while saving $15K USD and 500 hours in human evaluation costs. We find that closed-source LLMs such as GPT-4 and Claude 2 produce summaries with higher BooookScore than those generated by open-source models. While LLaMA 2 falls behind other models, Mixtral achieves performance on par with GPT-3.5-Turbo. Incremental updating yields lower BooookScore but higher level of detail than hierarchical merging, a trade-off sometimes preferred by annotators.

Kalpesh Krishna, Yapei Chang, John Wieting et al. · deepmind

Given an input sequence (or prefix), modern language models often assign high probabilities to output sequences that are repetitive, incoherent, or irrelevant to the prefix; as such, model-generated text also contains such artifacts. To address these issues we present RankGen, a 1.2B parameter encoder model for English that scores model generations given a prefix. RankGen can be flexibly incorporated as a scoring function in beam search and used to decode from any pretrained language model. We train RankGen using large-scale contrastive learning to map a prefix close to the ground-truth sequence that follows it and far away from two types of negatives: (1) random sequences from the same document as the prefix, and (2) sequences generated from a large language model conditioned on the prefix. Experiments across four different language models (345M-11B parameters) and two domains show that RankGen significantly outperforms decoding algorithms like nucleus, top-k, and typical sampling, as well as contrastive decoding and search, on both automatic metrics (85.0 vs 77.3 MAUVE over nucleus) as well as human evaluations with English writers (74.5% human preference over nucleus sampling). Analysis reveals that RankGen outputs are more relevant to the prefix and improve continuity and coherence compared to baselines. We release our model checkpoints, code, and human preference data with explanations to facilitate future research.

Katherine Thai, Yapei Chang, Kalpesh Krishna et al. · deepmind

Humanities scholars commonly provide evidence for claims that they make about a work of literature (e.g., a novel) in the form of quotations from the work. We collect a large-scale dataset (RELiC) of 78K literary quotations and surrounding critical analysis and use it to formulate the novel task of literary evidence retrieval, in which models are given an excerpt of literary analysis surrounding a masked quotation and asked to retrieve the quoted passage from the set of all passages in the work. Solving this retrieval task requires a deep understanding of complex literary and linguistic phenomena, which proves challenging to methods that overwhelmingly rely on lexical and semantic similarity matching. We implement a RoBERTa-based dense passage retriever for this task that outperforms existing pretrained information retrieval baselines; however, experiments and analysis by human domain experts indicate that there is substantial room for improvement over our dense retriever.

Vinay Samuel, Yapei Chang, Mohit Iyyer

Many open-ended instructions have multiple valid answers that users can benefit from seeing, but post-training often narrows an LLM's output space toward a small set of canonical responses. We introduce REDIPO, an offline DPO data-construction pipeline for recovering distinct valid answer modes while preserving the alignment benefits of the instruct model. For each prompt, REDIPO samples responses from both base and instruct models, rewrites base-model responses with the instruct model, filters candidates for safety and instruction-following quality, and builds preference pairs that favor marginally diverse responses among candidates with similar instruction-following reward. Across Qwen3-4B, OLMo-3-7B, and LLaMA-3.1-8B, REDIPO improves NoveltyBench distinct_k by 134%, 33%, and 44% relative to the instruct checkpoints, while DivPO changes diversity by 0%, -6%, and -4% on the same models. These gains largely maintain MTBench, IFEval, and Arena-Hard performance, and reduce direct-category HarmBench attack success rate. Ablations show that marginal-diversity pair selection and base-response rewriting drive the diversity gains, while filtering and quality-bounded pairing help maintain alignment. Overall, our results show that diverse valid answers from base-model generations can be reintroduced through carefully constructed preference data while retaining the alignment benefits of post-training. We release our code and data at https://github.com/vsamuel2003/RiDiPO.

Yekyung Kim, Yapei Chang, Chau Minh Pham et al.

As LLMs are increasingly used to draft public-facing arguments, they may flatten public debate by repeatedly introducing the same polished, plausible arguments. We study argument collapse, the tendency of essays generated by different LLMs to converge to a smaller set of main arguments, sub-arguments, and paragraph-level structures. We compare 1,039 human responses from 195 New York Times (NYT) debates, 448 human responses from 61 longer-form Boston Review (BR) forums, and 23,384 LLM-generated essays. In the NYT corpus, 65.3% of human main arguments are unique within a debate, compared to 3.4% of LLM main arguments. Asking LLMs to generate diverse answers adds variation, but a typical model recovers only about half of the distinct human main arguments, with much of the added variation falling outside the observed human argument space. Collapse also appears in sub-arguments, where among essays with the same main argument, 41.0% of human sub-arguments are unique versus 9.1% from LLM responses. Qualitatively, LLMs often reuse generalized and hedged sub-arguments, while humans prefer more concrete and topic-specific ones. Structure-wise, LLM-generated essays tend to follow a more fixed arc, often opening with a direct claim and moving quickly toward proposals. The same patterns hold in longer BR essays, suggesting that argument collapse extends beyond short-form responses.

Yapei Chang, Kyle Lo, Mohit Iyyer et al.

Generating step-by-step "how-to" procedures is a key LLM capability: how-to advice is commonly requested in chatbots, and step-by-step planning is critical for reasoning over complex tasks. Yet, measuring and improving procedural validity at scale on real-world tasks remains challenging and understudied. To address this, we introduce How2Everything, a scalable framework to evaluate and improve goal-conditioned procedure generation. Our framework includes How2Mine, which mines 351K procedures from 980K web pages across 14 topics and readily scales to larger corpora. From this pool we build How2Bench, a 7K-example evaluation set balanced across topics. To reliably score model outputs, we develop How2Score, an evaluation protocol that uses an LLM judge to detect whether a generation contains any critical failure that would prevent achieving the goal. For low-cost, reproducible evaluation, we distill a frontier model into an open 8B model, achieving 80.5% agreement with human annotators. How2Bench reveals clear scaling trends across model sizes and training stages, providing signal early in pretraining. Finally, RL using How2Score as a reward improves performance on How2Bench by >10 points across three models without systematic regressions on standard benchmarks, with gains robust to superficial source-document memorization or format compliance. Taken together, How2Everything shows how pretraining web data can support a closed loop of capability evaluation and improvement at scale.

Yekyung Kim, Yapei Chang, Marzena Karpinska et al.

While long-context large language models (LLMs) can technically summarize book-length documents (>100K tokens), the length and complexity of the documents have so far prohibited evaluations of input-dependent aspects like faithfulness. In this paper, we conduct the first large-scale human evaluation of faithfulness and content selection on LLM-generated summaries of fictional books. Our study mitigates the issue of data contamination by focusing on summaries of books published in 2023 or 2024, and we hire annotators who have fully read each book prior to the annotation task to minimize cost and cognitive burden. We collect FABLES, a dataset of annotations on 3,158 claims made in LLM-generated summaries of 26 books, at a cost of $5.2K USD, which allows us to rank LLM summarizers based on faithfulness: Claude-3-Opus significantly outperforms all closed-source LLMs, while the open-source Mixtral is on par with GPT-3.5-Turbo. An analysis of the annotations reveals that most unfaithful claims relate to events and character states, and they generally require indirect reasoning over the narrative to invalidate. While LLM-based auto-raters have proven reliable for factuality and coherence in other settings, we implement several LLM raters of faithfulness and find that none correlates strongly with human annotations, especially with regard to detecting unfaithful claims. Our experiments suggest that detecting unfaithful claims is an important future direction not only for summarization evaluation but also as a testbed for long-context understanding. Finally, we move beyond faithfulness by exploring content selection errors in book-length summarization: we develop a typology of omission errors related to crucial narrative elements and also identify a systematic over-emphasis on events occurring towards the end of the book.

Team Olmo, Allyson Ettinger, Amanda Bertsch et al. · uw

We introduce Olmo 3, a family of state-of-the-art, fully-open language models at the 7B and 32B parameter scales. Olmo 3 model construction targets long-context reasoning, function calling, coding, instruction following, general chat, and knowledge recall. This release includes the entire model flow, i.e., the full lifecycle of the family of models, including every stage, checkpoint, data point, and dependency used to build it. Our flagship model, Olmo 3 Think 32B, is the strongest fully-open thinking model released to-date.

Zongxia Li, Yapei Chang, Yuhang Zhou et al.

Evaluating open-ended long-form generation is challenging because it is hard to define what clearly separates good from bad outputs. Existing methods often miss key aspects like coherence, style, or relevance, or are biased by pretraining data, making open-ended long-form evaluation an underexplored problem. To address this gap, we propose PrefBERT, a scoring model for evaluating open-ended long-form generation in GRPO and guiding its training with distinct rewards for good and bad outputs. Trained on two response evaluation datasets with diverse long-form styles and Likert-rated quality, PrefBERT effectively supports GRPO by offering better semantic reward feedback than traditional metrics ROUGE-L and BERTScore do. Through comprehensive evaluations, including LLM-as-a-judge, human ratings, and qualitative analysis, we show that PrefBERT, trained on multi-sentence and paragraph-length responses, remains reliable across varied long passages and aligns well with the verifiable rewards GRPO needs. Human evaluations confirm that using PrefBERT as the reward signal to train policy models yields responses better aligned with human preferences than those trained with traditional metrics. Our code is available at https://github.com/zli12321/long_form_rl.

Yapei Chang, Kalpesh Krishna, Amir Houmansadr et al.

The most effective techniques to detect LLM-generated text rely on inserting a detectable signature -- or watermark -- during the model's decoding process. Most existing watermarking methods require access to the underlying LLM's logits, which LLM API providers are loath to share due to fears of model distillation. As such, these watermarks must be implemented independently by each LLM provider. In this paper, we develop PostMark, a modular post-hoc watermarking procedure in which an input-dependent set of words (determined via a semantic embedding) is inserted into the text after the decoding process has completed. Critically, PostMark does not require logit access, which means it can be implemented by a third party. We also show that PostMark is more robust to paraphrasing attacks than existing watermarking methods: our experiments cover eight baseline algorithms, five base LLMs, and three datasets. Finally, we evaluate the impact of PostMark on text quality using both automated and human assessments, highlighting the trade-off between quality and robustness to paraphrasing. We release our code, outputs, and annotations at https://github.com/lilakk/PostMark.

Yixiao Song, Katherine Thai, Chau Minh Pham et al.

Modern web agents possess computer use abilities that allow them to interact with webpages by sending commands to a virtual keyboard and mouse. While such agents have considerable potential to assist human users with complex tasks, evaluating their capabilities in real-world settings poses a major challenge. To this end, we introduce BEARCUBS, a "smallbut mighty" benchmark of 111 information-seeking questions designed to evaluate a web agent's ability to search, browse, and identify factual information from the web. Unlike prior web agent benchmarks, solving BEARCUBS requires (1) accessing live web content rather than synthetic or simulated pages, which captures the unpredictability of real-world web interactions; and (2) performing a broad range of multimodal interactions (e.g., video understanding, 3D navigation) that cannot be bypassed via text-based workarounds. Each question in BEARCUBS has a corresponding short, unambiguous answer and a human-validated browsing trajectory, allowing for transparent evaluation of agent performance and strategies. A human study confirms that BEARCUBS questions are solvable but non-trivial (84.7% human accuracy), revealing domain knowledge gaps and overlooked details as common failure points. We find that ChatGPT Agent significantly outperforms other computer-using agents with an overall accuracy of 65.8% (compared to e.g., Operator's 23.4%), showcasing substantial progress in tasks involving real computer use, such as playing web games and navigating 3D environments. Nevertheless, closing the gap to human performance requires improvements in areas like fine control, complex data filtering, and execution speed. To facilitate future research, BEARCUBS will be updated periodically to replace invalid or contaminated questions, keeping the benchmark fresh for future generations of web agents.

Chau Minh Pham, Yapei Chang, Mohit Iyyer

LLM developers are increasingly reliant on synthetic data, but generating high-quality data for complex long-context reasoning tasks remains challenging. We introduce CLIPPER, a compression-based approach for generating synthetic data tailored to narrative claim verification - a task that requires reasoning over a book to verify a given claim. Instead of generating claims directly from the raw text of the book, which results in artifact-riddled claims, CLIPPER first compresses the book into chapter outlines and book summaries and then uses these intermediate representations to generate complex claims and corresponding chain-of-thoughts. Compared to naive approaches, CLIPPER produces claims that are more valid, grounded, and complex. Using CLIPPER, we construct a dataset of 19K synthetic book claims paired with their source texts and chain-of-thought reasoning, and use it to fine-tune three open-weight models. Our best model achieves breakthrough results on narrative claim verification (from 28% to 76% accuracy on our test set) and sets a new state-of-the-art for sub-10B models on the NoCha leaderboard. Further analysis shows that our models generate more detailed and grounded chain-of-thought reasoning while also improving performance on other narrative understanding tasks (e.g., NarrativeQA).

Roshita Bhonsle, Rishav Dutta, Sneha Vavilapalli et al.

The increasing adoption of foundation models as agents across diverse domains necessitates a robust evaluation framework. Current methods, such as LLM-as-a-Judge, focus only on final outputs, overlooking the step-by-step reasoning that drives agentic decision-making. Meanwhile, existing Agent-as-a-Judge systems, where one agent evaluates another's task completion, are typically designed for narrow, domain-specific settings. To address this gap, we propose a generalizable, modular framework for evaluating agent task completion independent of the task domain. The framework emulates human-like evaluation by decomposing tasks into sub-tasks and validating each step using available information, such as the agent's output and reasoning. Each module contributes to a specific aspect of the evaluation process, and their outputs are aggregated to produce a final verdict on task completion. We validate our framework by evaluating the Magentic-One Actor Agent on two benchmarks, GAIA and BigCodeBench. Our Judge Agent predicts task success with closer agreement to human evaluations, achieving 4.76% and 10.52% higher alignment accuracy, respectively, compared to the GPT-4o based LLM-as-a-Judge baseline. This demonstrates the potential of our proposed general-purpose evaluation framework.