DRACULA: Hunting for the Actions Users Want Deep Research Agents to Execute

Scientific Deep Research (DR) agents answer user queries by synthesizing research papers into multi-section reports. User feedback can improve their utility, but existing protocols only score the final report, making it hard to study and learn which intermediate actions DR agents should take to improve reports. We collect DRACULA, the first dataset with user feedback on intermediate actions for DR. Over five weeks, nineteen expert CS researchers ask queries to a DR system that proposes actions (e.g., "Add a section on datasets"). Our users select actions they prefer, then judge whether an output report applied their selections successfully, yielding 8,103 action preferences and 5,230 execution judgments. After confirming a DR agent can execute DRACULA's actions, we study the predictability of user-preferred actions via simulation-how well LLMs predict the actions users select-a step toward learning to generate useful actions. We discover: (1) LLM judges initially struggle to predict action selections, but improve most when using a user's full selection history, rather than self-reported or extrapolated user context signals; (2) Users' selections for the same query differ based on unstated goals, bottlenecking simulation and motivating affordances that let users steer reports; and (3) Our simulation results inform an online intervention that generates new actions based on the user's past interactions, which users pick most often in follow-up studies. Overall, while work extensively studies execution, DRACULA reveals a key challenge is deciding which actions to execute in the first place. We open-source DRACULA's study design, user feedback, and simulation tasks to spur future work on action feedback for long-horizon agents.

Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models

Confirmation bias, the tendency to seek evidence that supports rather than challenges one's belief, hinders one's reasoning ability. We examine whether large language models (LLMs) exhibit confirmation bias by adapting the rule-discovery study from human psychology: given a sequence of three numbers (a "triple"), an agent engages in an interactive feedback loop where it (1) proposes a new triple, (2) receives feedback on whether it satisfies the hidden rule, and (3) guesses the rule. Across eleven LLMs of multiple families and scales, we find that LLMs exhibit confirmation bias, often proposing triples to confirm their hypothesis rather than trying to falsify it. This leads to slower and less frequent discovery of the hidden rule. We further explore intervention strategies (e.g., encouraging the agent to consider counter examples) developed for humans. We find prompting LLMs with such instruction consistently decreases confirmation bias in LLMs, improving rule discovery rates from 42% to 56% on average. Lastly, we mitigate confirmation bias by distilling intervention-induced behavior into LLMs, showing promising generalization to a new task, the Blicket test. Our work shows that confirmation bias is a limitation of LLMs in hypothesis exploration, and that it can be mitigated via injecting interventions designed for humans.

No Single Best Model for Diversity: Learning a Router for Sample Diversity

When posed with prompts that permit a large number of valid answers, comprehensively generating them is the first step towards satisfying a wide range of users. In this paper, we study methods to elicit a comprehensive set of valid responses. To evaluate this, we introduce \textbf{diversity coverage}, a metric that measures the total quality scores assigned to each \textbf{unique} answer in the predicted answer set relative to the best possible answer set with the same number of answers. Using this metric, we evaluate 18 LLMs, finding no single model dominates at generating diverse responses to a wide range of open-ended prompts. Yet, per each prompt, there exists a model that outperforms all other models significantly at generating a diverse answer set. Motivated by this finding, we introduce a router that predicts the best model for each query. On NB-Wildchat, our trained router outperforms the single best model baseline (26.3% vs $23.8%). We further show generalization to an out-of-domain dataset (NB-Curated) as well as different answer-generation prompting strategies. Our work lays foundation for studying generating comprehensive answers when we have access to a suite of models.

ParaSpeechCLAP: A Dual-Encoder Speech-Text Model for Rich Stylistic Language-Audio Pretraining

We introduce ParaSpeechCLAP, a dual-encoder contrastive model that maps speech and text style captions into a common embedding space, supporting a wide range of intrinsic (speaker-level) and situational (utterance-level) descriptors (such as pitch, texture and emotion) far beyond the narrow set handled by existing models. We train specialized ParaSpeechCLAP-Intrinsic and ParaSpeechCLAP-Situational models alongside a unified ParaSpeechCLAP-Combined model, finding that specialization yields stronger performance on individual style dimensions while the unified model excels on compositional evaluation. We further show that ParaSpeechCLAP-Intrinsic benefits from an additional classification loss and class-balanced training. We demonstrate our models' performance on style caption retrieval, speech attribute classification and as an inference-time reward model that improves style-prompted TTS without additional training. ParaSpeechCLAP outperforms baselines on most metrics across all three applications. Our models and code are released at https://github.com/ajd12342/paraspeechclap .

Language Models Don't Know What You Want: Evaluating Personalization in Deep Research Needs Real Users

Deep Research (DR) tools (e.g. OpenAI DR) help researchers cope with ballooning publishing counts. Such tools can synthesize scientific papers to answer researchers' queries, but lack understanding of their users. We change that in MyScholarQA (MySQA), a personalized DR tool that: 1) infers a profile of a user's research interests; 2) proposes personalized actions for a user's input query; and 3) writes a multi-section report for the query that follows user-approved actions. We first test MySQA with NLP's standard protocol: we design a benchmark of synthetic users and LLM judges, where MySQA beats baselines in citation metrics and personalized action-following. However, we suspect this process does not cover all aspects of personalized DR users value, so we interview users in an online version of MySQA to unmask them. We reveal nine nuanced errors of personalized DR undetectable by our LLM judges, and we study qualitative feedback to form lessons for future DR design. In all, we argue for a pillar of personalization that easy-to-use LLM judges can lead NLP to overlook: real progress in personalization is only possible with real users.

Are Dilemmas and Conflicts in LLM Alignment Solvable? A View from Priority Graph

As Large Language Models (LLMs) become more powerful and autonomous, they increasingly face conflicts and dilemmas in many scenarios. We first summarize and taxonomize these diverse conflicts. Then, we model the LLM's preferences to make different choices as a priority graph, where instructions and values are nodes, and the edges represent context-specific priorities determined by the model's output distribution. This graph reveals that a unified stable LLM alignment is very challenging, because the graph is neither static nor necessarily consistent in different contexts. Besides, it also reveals a potential vulnerability: priority hacking, where adversaries can craft deceptive contexts to manipulate the graph and bypass safety alignments. To counter this, we propose a runtime verification mechanism, enabling LLMs to query external sources to ground their context and resist manipulation. While this approach enhances robustness, we also acknowledge that many ethical and value dilemmas are philosophically irreducible, posing a long-term, open challenge for the future of AI alignment.

Bridging Latent Reasoning and Target-Language Generation via Retrieval-Transition Heads

Recent work has identified a subset of attention heads in Transformer as retrieval heads, which are responsible for retrieving information from the context. In this work, we first investigate retrieval heads in multilingual contexts. In multilingual language models, we find that retrieval heads are often shared across multiple languages. Expanding the study to cross-lingual setting, we identify Retrieval-Transition heads(RTH), which govern the transition to specific target-language output. Our experiments reveal that RTHs are distinct from retrieval heads and more vital for Chain-of-Thought reasoning in multilingual LLMs. Across four multilingual benchmarks (MMLU-ProX, MGSM, MLQA, and XQuaD) and two model families (Qwen-2.5 and Llama-3.1), we demonstrate that masking RTH induces bigger performance drop than masking Retrieval Heads (RH). Our work advances understanding of multilingual LMs by isolating the attention heads responsible for mapping to target languages.

BenchMarker: An Education-Inspired Toolkit for Highlighting Flaws in Multiple-Choice Benchmarks

Multiple-choice question answering (MCQA) is standard in NLP, but benchmarks lack rigorous quality control. We present BenchMarker, an education-inspired toolkit using LLM judges to flag three common MCQ flaws: 1) contamination - items appearing exactly online; 2) shortcuts - cues in the choices that enable guessing; and 3) writing errors - structural/grammatical issues based on a 19-rule education rubric. We validate BenchMarker with human annotations, then run the tool to audit 12 benchmarks, revealing: 2) contaminated MCQs tend to inflate accuracy, while writing errors tend to lower it and change rankings beyond random; and 3) prior benchmark repairs address their targeted issues (i.e., lowering accuracy with LLM-written distractors), but inadvertently add new flaws (i.e. implausible distractors, many correct answers). Overall, flaws in MCQs degrade NLP evaluation, but education research offers a path forward. We release BenchMarker to bridge the fields and improve MCQA benchmark design.

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.

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 multiple reasoning datasets show that SSA outperforms other test-time scaling methods such as reward model-based re-ranking. Our approach also shows a promising generalization ability, 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.