arXiv subjects

Eunsol Choi

Publications and source records attributed to Eunsol Choi.

At least 37 records · Page 2Linked to original sources

arXiv2026

RARe: Retrieval Augmented Retrieval with In-Context Examples

While in-context learning is well-studied with decoder-only language models (LLMs), its utility for encoder-only models remains underexplored. We study in-context learning for encoder-only models for text retrieval tasks. Can incorporating in-context examples (query-document pairs) to the target query enhance retriever performance? Our approach, RARe, finetunes a pre-trained model with in-context examples whose query is semantically similar to the target query. This approach achieves performance gains of up to +2.72% nDCG across open-domain retrieval datasets (BeIR, RAR-b) compared to using the target query only as an input. In particular, we find RARe exhibits stronger out-of-domain generalization compared to models using queries without in-context examples, similar to what is seen for in-context learning in LLMs. We further provide analysis on the design choices of in-context example augmentation for retrievers and lay the foundation for future work.

cs.CL↗

arXiv2026

SAGE: Steerable Agentic Data Generation for Deep Search with Execution Feedback

Deep search agents, which aim to answer complex questions requiring reasoning across multiple documents, can significantly speed up the information-seeking process. Collecting human annotations for this application is prohibitively expensive due to long and complex exploration trajectories. We propose an agentic pipeline that automatically generates high quality, difficulty-controlled deep search question-answer pairs for a given corpus and a target difficulty level. Our pipeline, SAGE, consists of a data generator which proposes QA pairs and a search agent which attempts to solve the generated question and provide execution feedback for the data generator. The two components interact over multiple rounds to iteratively refine the question-answer pairs until they satisfy the target difficulty level. Our intrinsic evaluation shows SAGE generates questions that require diverse reasoning strategies, while significantly increases the correctness and difficulty of the generated data. Our extrinsic evaluation demonstrates up to 23% relative performance gain on popular deep search benchmarks by training deep search agents with our synthetic data. Additional experiments show that agents trained on our data can adapt from fixed-corpus retrieval to Google Search at inference time, without further training.

cs.AI↗

arXiv2025

Beyond Single Embeddings: Capturing Diverse Targets with Multi-Query Retrieval

Most text retrievers generate \emph{one} query vector to retrieve relevant documents. Yet, the conditional distribution of relevant documents for the query may be multimodal, e.g., representing different interpretations of the query. We first quantify the limitations of existing retrievers. All retrievers we evaluate struggle more as the distance between target document embeddings grows. To address this limitation, we develop a new retriever architecture, \emph{A}utoregressive \emph{M}ulti-\emph{E}mbedding \emph{R}etriever (AMER). Our model autoregressively generates multiple query vectors, and all the predicted query vectors are used to retrieve documents from the corpus. We show that on the synthetic vectorized data, the proposed method could capture multiple target distributions perfectly, showing 4x better performance than single embedding model. We also fine-tune our model on real-world multi-answer retrieval datasets and evaluate in-domain. AMER presents 4 and 21\% relative gains over single-embedding baselines on two datasets we evaluate on. Furthermore, we consistently observe larger gains on the subset of dataset where the embeddings of the target documents are less similar to each other. We demonstrate the potential of using a multi-query vector retriever and open up a new direction for future work.

cs.CL↗

arXiv2025

Learning to Reason Across Parallel Samples for LLM Reasoning

Scaling test-time compute brings substantial performance gains for large language models (LLMs). By sampling multiple answers and heuristically aggregate their answers (e.g., either through majority voting or using verifiers to rank the answers), one can achieve consistent performance gains in math domains. In this paper, we propose a new way to leverage such multiple sample set. We train a compact LLM, called Sample Set Aggregator (SSA), that takes a concatenated sequence of multiple samples and output the final answer, optimizing it for the answer accuracy with reinforcement learning. Experiments on five reasoning datasets demonstrate both the efficacy and efficiency of SSA. Notably, SSA improves over naive majority voting by 8% pass@5 on MATH. Furthermore, our 3B SSA surpasses model-based re-ranking with a much larger 72B process reward model. Our analysis also shows promising generalization ability of SSA, across sample set sizes, base model families and scales, and tasks. By separating LLMs to generate answers and LLMs to analyze and aggregate sampled answers, our approach can work with the outputs from premier black box models easily and efficiently.

cs.CL↗

arXiv2025

User Feedback in Human-LLM Dialogues: A Lens to Understand Users But Noisy as a Learning Signal

Once language models (LMs) are deployed, they can interact with users long-term, ideally evolving based on their feedback. Asking for direct user feedback can be disruptive; thus, we study harvesting implicit user feedback from user-LM interaction logs. We study two user-LM interaction datasets (WildChat and LMSYS). First, we analyze user feedback in the user-LLM conversation logs, providing insights into when and why such feedback occurs. Second, we study harvesting learning signals from such implicit user feedback. Specifically, we study whether incorporating the contents of user feedback (e.g., user wanted clarification), in addition to the polarity of the feedback, can improve the model performance. We observe mixed results, showing this helps in short human-designed questions (MTBench) but not on longer and more complex questions (WildBench). Together, we provide an in-depth study of implicit user feedback, showing its potential and limitations.

cs.CL↗

arXiv2025

Understanding Retrieval Augmentation for Long-Form Question Answering

How retrieved documents are used in language models (LMs) for long-form generation task is understudied. We present two controlled studies on retrieval-augmented LM for long-form question answering (LFQA): one fixing the LM and varying evidence documents and the other fixing evidence documents and varying the LMs. We study various attributes of generated answers (e.g., fluency, length, variance), with an emphasis on the attribution of generated answers to in-context evidence documents. We collect a dataset (SALAD) containing human annotations of sentence-level answer attribution in LFQA and evaluate existing methods for automatically judging attribution. We find that while LMs can leverage relevant in-context documents, the generated answer is only partially attributable towards the documents, especially for LMs trained without retrieval augmentation. Together, our analysis reveals how retrieval augmentation impacts long knowledge-rich text generation and provide directions for future work.

cs.CL↗

arXiv2025

Improving LLM-as-a-Judge Inference with the Judgment Distribution

Using language models to scalably approximate human preferences on text quality (LLM-as-a-judge) has become a standard practice applicable to many tasks. A judgment is often extracted from the judge's textual output alone, typically with greedy decoding. However, LLM judges naturally provide distributions over judgment tokens, inviting a breadth of inference methods for extracting fine-grained preferences. We find that taking the mean of the judgment distribution consistently outperforms taking the mode (i.e. greedy decoding) in all evaluation settings (i.e. pointwise, pairwise, and listwise). We further explore novel methods of deriving preferences from judgment distributions, and find that methods incorporating risk aversion often improve performance. Lastly, we analyze LLM-as-a-judge paired with chain-of-thought (CoT) prompting, showing that CoT can collapse the spread of the judgment distribution, often harming performance. Our findings show that leveraging distributional output improves LLM-as-a-judge, as opposed to using the text interface alone.

cs.CL↗

arXiv2025

Scaling Rich Style-Prompted Text-to-Speech Datasets

We introduce Paralinguistic Speech Captions (ParaSpeechCaps), a large-scale dataset that annotates speech utterances with rich style captions. While rich abstract tags (e.g. guttural, nasal, pained) have been explored in small-scale human-annotated datasets, existing large-scale datasets only cover basic tags (e.g. low-pitched, slow, loud). We combine off-the-shelf text and speech embedders, classifiers and an audio language model to automatically scale rich tag annotations for the first time. ParaSpeechCaps covers a total of 59 style tags, including both speaker-level intrinsic tags and utterance-level situational tags. It consists of 342 hours of human-labelled data (PSC-Base) and 2427 hours of automatically annotated data (PSC-Scaled). We finetune Parler-TTS, an open-source style-prompted TTS model, on ParaSpeechCaps, and achieve improved style consistency (+7.9% Consistency MOS) and speech quality (+15.5% Naturalness MOS) over the best performing baseline that combines existing rich style tag datasets. We ablate several of our dataset design choices to lay the foundation for future work in this space. Our dataset, models and code are released at https://github.com/ajd12342/paraspeechcaps .

eess.AS↗

arXiv2025

Rhapsody: A Dataset for Highlight Detection in Podcasts

Podcasts have become daily companions for half a billion users. Given the enormous amount of podcast content available, highlights provide a valuable signal that helps viewers get the gist of an episode and decide if they want to invest in listening to it in its entirety. However, identifying highlights automatically is challenging due to the unstructured and long-form nature of the content. We introduce Rhapsody, a dataset of 13K podcast episodes paired with segment-level highlight scores derived from YouTube's 'most replayed' feature. We frame the podcast highlight detection as a segment-level binary classification task. We explore various baseline approaches, including zero-shot prompting of language models and lightweight fine-tuned language models using segment-level classification heads. Our experimental results indicate that even state-of-the-art language models like GPT-4o and Gemini struggle with this task, while models fine-tuned with in-domain data significantly outperform their zero-shot performance. The fine-tuned model benefits from leveraging both speech signal features and transcripts. These findings highlight the challenges for fine-grained information access in long-form spoken media.

cs.CL↗

arXiv2025

Long-Form Answers to Visual Questions from Blind and Low Vision People

Vision language models can now generate long-form answers to questions about images - long-form visual question answers (LFVQA). We contribute VizWiz-LF, a dataset of long-form answers to visual questions posed by blind and low vision (BLV) users. VizWiz-LF contains 4.2k long-form answers to 600 visual questions, collected from human expert describers and six VQA models. We develop and annotate functional roles of sentences of LFVQA and demonstrate that long-form answers contain information beyond the question answer such as explanations and suggestions. We further conduct automatic and human evaluations with BLV and sighted people to evaluate long-form answers. BLV people perceive both human-written and generated long-form answers to be plausible, but generated answers often hallucinate incorrect visual details, especially for unanswerable visual questions (e.g., blurry or irrelevant images). To reduce hallucinations, we evaluate the ability of VQA models to abstain from answering unanswerable questions across multiple prompting strategies.

cs.CL↗

arXiv2025

CaLMQA: Exploring culturally specific long-form question answering across 23 languages

Despite rising global usage of large language models (LLMs), their ability to generate long-form answers to culturally specific questions remains unexplored in many languages. To fill this gap, we perform the first study of textual multilingual long-form QA by creating CaLMQA, a dataset of 51.7K culturally specific questions across 23 different languages. We define culturally specific questions as those that refer to concepts unique to one or a few cultures, or have different answers depending on the cultural or regional context. We obtain these questions by crawling naturally-occurring questions from community web forums in high-resource languages, and by hiring native speakers to write questions in under-resourced, rarely-studied languages such as Fijian and Kirundi. Our data collection methodologies are translation-free, enabling the collection of culturally unique questions like "Kuber iki umwami wa mbere w'uburundi yitwa Ntare?" (Kirundi; English translation: "Why was the first king of Burundi called Ntare (Lion)?"). We evaluate factuality, relevance and surface-level quality of LLM-generated long-form answers, finding that (1) for many languages, even the best models make critical surface-level errors (e.g., answering in the wrong language, repetition), especially for low-resource languages; and (2) answers to culturally specific questions contain more factual errors than answers to culturally agnostic questions -- questions that have consistent meaning and answer across many cultures. We release CaLMQA to facilitate future research in cultural and multilingual long-form QA.

cs.CL↗

arXiv2025

PropMEND: Hypernetworks for Knowledge Propagation in LLMs

Knowledge editing techniques for large language models (LLMs) can inject knowledge that is later reproducible verbatim, but they fall short on propagating that knowledge: models cannot answer questions that require reasoning with the injected knowledge. We present a hypernetwork-based approach for knowledge propagation, named PropMEND, where we meta-learn how to modify gradients of a language modeling loss to encourage injected information to propagate. Our approach extends the meta-objective of MEND [29] so that gradient updates on knowledge are transformed to enable answering multi-hop questions involving that knowledge. We show improved performance on the RippleEdit dataset, showing almost 2x accuracy on challenging multi-hop questions whose answers are not explicitly stated in the injected fact. We further introduce a new dataset, Controlled RippleEdit, to evaluate the generalization of our hypernetwork, testing knowledge propagation along relations and entities unseen during hypernetwork training. PropMEND still outperforms existing approaches in unseen entity-relation pairs, yet the performance gap decreases substantially, suggesting future work in propagating knowledge to a wide range of relations.

cs.CL↗

arXiv2025

From Distributional to Overton Pluralism: Investigating Large Language Model Alignment

The alignment process changes several properties of a large language model's (LLM's) output distribution. We analyze two aspects of post-alignment distributional shift of LLM responses. First, we re-examine previously reported reductions in response diversity post-alignment. Our analysis suggests that an apparent drop in the diversity of responses is largely explained by quality control and information aggregation. Alignment suppresses irrelevant and unhelpful content while shifting the output distribution toward longer responses that cover information spanning several responses from the base LLM, essentially presenting diverse information in a single response. Finding little evidence that alignment suppresses useful information, it is natural to ask the opposite question: do aligned models surface information that cannot be recovered from base models? Our second investigation shows this is not the case and the behavior of aligned models is recoverable from base models without fine-tuning. A combination of in-context examples and lower-resolution semantic hints about response content can elicit responses from base LLMs that are as similar to alignment-tuned LLM responses as alignment-tuned LLM responses are to each other. Taken together, these results indicate that current alignment techniques capture but do not extend the useful subset of assistant-like base LLM behavior, providing further evidence for the Superficial Alignment Hypothesis. They also show that in-context alignment can go surprisingly far as a strategy for imitating aligned LLMs without fine-tuning. Our code and data is available at https://github.com/thomlake/investigating-alignment.

cs.CL↗

arXiv2025

Open-World Evaluation for Retrieving Diverse Perspectives

We study retrieving a set of documents that covers various perspectives on a complex and contentious question (e.g., will ChatGPT do more harm than good?). We curate a Benchmark for Retrieval Diversity for Subjective questions (BERDS), where each example consists of a question and diverse perspectives associated with the question, sourced from survey questions and debate websites. On this data, retrievers paired with a corpus are evaluated to surface a document set that contains diverse perspectives. Our framing diverges from most retrieval tasks in that document relevancy cannot be decided by simple string matches to references. Instead, we build a language model-based automatic evaluator that decides whether each retrieved document contains a perspective. This allows us to evaluate the performance of three different types of corpus (Wikipedia, web snapshot, and corpus constructed on the fly with retrieved pages from the search engine) paired with retrievers. Retrieving diverse documents remains challenging, with the outputs from existing retrievers covering all perspectives on only 40% of the examples. We further study the effectiveness of query expansion and diversity-focused reranking approaches and analyze retriever sycophancy.

cs.CL↗

arXiv2025

On Language Models' Sensitivity to Suspicious Coincidences

Humans are sensitive to suspicious coincidences when generalizing inductively over data, as they make assumptions as to how the data was sampled. This results in smaller, more specific hypotheses being favored over more general ones. For instance, when provided the set {Austin, Dallas, Houston}, one is more likely to think that this is sampled from "Texas Cities" over "US Cities" even though both are compatible. Suspicious coincidence is strongly connected to pragmatic reasoning, and can serve as a testbed to analyze systems on their sensitivity towards the communicative goals of the task (i.e., figuring out the true category underlying the data). In this paper, we analyze whether suspicious coincidence effects are reflected in language models' (LMs) behavior. We do so in the context of two domains: 1) the number game, where humans made judgments of whether a number (e.g., 4) fits a list of given numbers (e.g., 16, 32, 2); and 2) by extending the number game setup to prominent cities. For both domains, the data is compatible with multiple hypotheses and we study which hypothesis is most consistent with the models' behavior. On analyzing five models, we do not find strong evidence for suspicious coincidences in LMs' zero-shot behavior. However, when provided access to the hypotheses space via chain-of-thought or explicit prompting, LMs start to show an effect resembling suspicious coincidences, sometimes even showing effects consistent with humans. Our study suggests that inductive reasoning behavior in LMs can be enhanced with explicit access to the hypothesis landscape.

cs.CL↗

arXiv2025

CodeUpdateArena: Benchmarking Knowledge Editing on API Updates

Large language models (LLMs) are increasingly being used to synthesize and reason about source code. However, the static nature of these models' knowledge does not reflect the fact that libraries and API functions they invoke are continuously evolving, with functionality being added or changing. While numerous benchmarks evaluate how LLMs can generate code, no prior work has studied how an LLMs' knowledge about code API functions can be updated. To fill this gap, we present CodeUpdateArena, a benchmark for knowledge editing in the code domain. An instance in our benchmark consists of a synthetic API function update paired with a program synthesis example that uses the updated functionality; our goal is to update an LLM to be able to solve this program synthesis example without providing documentation of the update at inference time. Compared to knowledge editing for facts encoded in text, success here is more challenging: a code LLM must correctly reason about the semantics of the modified function rather than just reproduce its syntax. Our dataset is constructed by first prompting GPT-4 to generate atomic and executable function updates. Then, for each update, we generate program synthesis examples whose code solutions are prone to use the update. Our benchmark covers updates of various types to 54 functions from seven diverse Python packages, with a total of 670 program synthesis examples. Our experiments show that prepending documentation of the update to open-source code LLMs (i.e., DeepSeek, CodeLlama) does not allow them to incorporate changes for problem solving, and existing knowledge editing techniques also have substantial room for improvement. We hope our benchmark will inspire new methods for knowledge updating in code LLMs.

cs.CL↗

arXiv2025

Modeling Future Conversation Turns to Teach LLMs to Ask Clarifying Questions

Large language models (LLMs) must often respond to highly ambiguous user requests. In such cases, the LLM's best response may be to ask a clarifying question to elicit more information. Existing LLMs often respond by presupposing a single interpretation of such ambiguous requests, frustrating users who intended a different interpretation. We speculate this is caused by current preference data labeling practice, where LLM responses are evaluated only on their prior contexts. To address this, we assign preference labels by simulating their expected outcomes in future turns. This allows LLMs to learn to ask clarifying questions when it can generate responses that are tailored to each user interpretation in future turns. On open-domain QA datasets with multiple annotations, we evaluate systems based on their ability to ask clarifying questions to recover each user's interpretation and expected answer. We compare systems trained using our proposed preference labeling methods against standard methods, which assign preferences based on only prior context. Our method achieves a 5% improvement in F1 measured against the answer set from different interpretations of each query, showing the value of modeling future conversation turns. We further demonstrate that our method can be used to train models to judiciously determine when to ask clarifying questions, directly answering the question when clarification is unnecessary. In our experiments, we find that our method achieves a 3% improvement in accuracy of such judgments over existing methods.

cs.CL↗

arXiv2025

RefreshKV: Updating Small KV Cache During Long-form Generation

Generating long sequences of tokens given a long-context input is a very compute-intensive inference scenario for large language models (LLMs). One prominent inference speed-up approach is to construct a smaller key-value (KV) cache, relieving LLMs from computing attention over a long sequence of tokens. While such methods work well to generate short sequences, their performance degrades rapidly for long-form generation. Most KV compression happens once, prematurely removing tokens that can be useful later in the generation. We propose a new inference method, RefreshKV, that flexibly alternates between full context attention and attention over a subset of input tokens during generation. After each full attention step, we update the smaller KV cache based on the attention pattern over the entire input. Applying our method to off-the-shelf LLMs achieves comparable speedup to eviction-based methods while improving performance for various long-form generation tasks. Lastly, we show that continued pretraining with our inference setting brings further gains in performance.

cs.CL↗