https://arxiv.org/api/+Lz/77DYzgfg0/STe33kT5nAkJs 2026-09-28T15:39:41Z 30 13 0 http://arxiv.org/abs/2609.03952v1 2026-09-03T14:53:53Z

Camera-conditioned world models generate interactive videos in which commanded actions should induce the expected scene changes while appearance, geometry, and temporal dynamics remain coherent. Existing rewards assess these requirements separately: geometry-based rewards estimate trajectory execution but cannot judge the visual quality of the executed motion, whereas image-based rewards measure frame quality without capturing action execution or temporal dynamics. We posit that a vision-language model (VLM) offers a shared reasoning space for relating actions to their visual outcomes. However, judging a complete long video against its full action sequence creates a lengthy, noisy context in which short-lived local action evidence can be missed or diluted. We present WorldReward, a VLM-based pairwise preference reward model that unifies action-consistency and visual-quality evaluation for camera-conditioned world models. WorldReward decomposes paired videos into action-aligned chunks, organizes each chunk into structured visual evidence, and aggregates chunk-level decisions by voting into separate video-level action and visual-quality preferences. To train it, we construct a large-scale reasoning-augmented preference dataset using structured judgments generated by a frontier VLM and refined through tool-based agent auditing and targeted human review. We further introduce WorldReward-Bench, a human-annotated benchmark measuring reward-model agreement with human preferences across action consistency, appearance quality, and motion quality. WorldReward achieves the highest agreement on all three dimensions, exceeding GPT-5.5 by 3.42, 1.45, and 3.56 percentage points, respectively. When used for RL post-training of HY-WorldPlay 1.5, it consistently improves both action execution and visual quality across short- to long-term horizons.

2026-09-03T14:53:53Z Website: https://codegoat24.github.io/WorldReward Yibin Wang Zehan Wang Junshu Tang Zhimin Li Yujie Zhou Jiazi Bu Pengyang Ling Feng Han Zhixiong Zhang Long Xing Shengyuan Ding Ziang Li Cheng Jin Yuhang Zang Jiaqi Wang Tianyu Pang http://arxiv.org/abs/2606.05922v3 2026-08-29T06:46:46Z

AI agents rely on a harness of skills, tools, and workflows to solve complex problems. Continually improving this harness is essential for adapting to new tasks. However, existing optimization methods typically require ground-truth validation sets, yet such labeled data is difficult to acquire in practical deployment settings. To address this problem, we introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes the agent harness using only past trajectories. Specifically, RHO selects a diverse coreset of challenging tasks from past trajectories and re-solves them in parallel. The agent analyzes these rollouts using self-validation and self-consistency, then generates candidate harness updates and selects the most effective one by its own pairwise self-preference. We evaluate RHO across three diverse domains, spanning software engineering, technical work, and knowledge work. Notably, a single optimization round improves the pass rate on SWE-Bench Pro from 59% to 78% without any external grading. Furthermore, our analysis demonstrates that RHO effectively targets prior failure modes. As a result, the optimized harness alters the agent's behavior patterns and sustains higher accuracy during long-horizon sessions.

2026-06-04T09:26:00Z Accepted to EMNLP 2026 (Findings). Code: https://github.com/wbopan/retro-harness ; Project website: https://paper-rho.wenbo.io Wenbo Pan Shujie Liu Chin-Yew Lin Jingying Zeng Xianfeng Tang Xiangyang Zhou Yan Lu Xiaohua Jia http://arxiv.org/abs/2605.11854v2 2026-05-18T06:27:28Z

Diffusion Language Models (DLMs) have recently emerged as a promising alternative to autoregressive language models, offering stronger global awareness and highly parallel generation. However, post-training DLMs with standard Negative Evidence Lower Bound (NELBO)-based supervised fine-tuning remains inefficient: training reconstructs randomly masked tokens in a single step, whereas inference follows a confidence-guided, multi-step easy-to-hard denoising trajectory. Recent trajectory-based self-distillation methods exploit such inference trajectories mainly for sampling-step compression and acceleration, often improving decoding efficiency without substantially enhancing the model's underlying capability, and may even degrade performance under full diffusion decoding. In this work, we ask whether self-distilled trajectories can be used not merely for faster inference, but for genuine knowledge acquisition. Although these trajectories lie on the pretrained DLM's own distributional manifold and thus offer a potentially lower optimization barrier, we find that naively fine-tuning on them with standard NELBO objectives yields only marginal gains. To address this limitation, we propose \textbf{T}rajectory-\textbf{A}ligned optimization via \textbf{Bo}ltzmann \textbf{M}odeling (\textbf{TABOM}), a self-distilled trajectory-based post-training framework that aligns training with the easy-to-hard structure of inference. TABOM models the inference unmasking preference as a Boltzmann distribution over predictive entropies and derives a tractable pairwise ranking objective to align the model's certainty ordering with the observed decoding trajectory. Empirically, TABOM achieves substantial gains in new domains, expands the effective knowledge boundary of DLMs, and significantly mitigates catastrophic forgetting compared with standard SFT.

2026-05-12T09:39:06Z Project website: https://tonyckc.github.io/TABOM-web/ Kecheng Chen Ziru Liu Xijia Tao Hui Liu Yibing Liu Xinyu Fu Shi Wu Suiyun Zhang Dandan Tu Lingpeng Kong Rui Liu Haoliang Li http://arxiv.org/abs/2604.09876v2 2026-09-14T04:29:50Z

Generative user interfaces (GenUIs) create new opportunities to adapt interfaces to individual users on demand. Yet personalization is difficult because it is not possible to provide settings for screens that have not yet been generated, making it necessary to learn preferences from users' feedback on generated interfaces. Such feedback is sparse, subjective, and often difficult to articulate. We study this problem through a new dataset in which 20 participants each judge the same 600 pairs of GenUIs. Their judgments differ substantially (Krippendorff's alpha = 0.25), and written rationales show that even when participants consider similar UI properties, such as color tone or hierarchy, they often make different choices and prioritize them differently. We develop a sample-efficient personalization method that uses a few pairwise judgments to weight prior users' preferences rather than relying on a fixed rubric of UI attributes. Our method outperforms a pretrained UI evaluator and a larger multimodal model offline. In an online study with 12 new users, participants preferred interfaces personalized by our method over all baselines, including shared guidelines and users' own written preferences.

2026-04-10T20:05:51Z Yi-Hao Peng Jeffrey P. Bigham Jason Wu http://arxiv.org/abs/2602.16321v1 2026-02-18T10:01:21Z

Multi-objective optimisation using BRIGHT has proven insightful and effective in prostate cancer brachytherapy treatment planning. BRachytherapy via artificially Intelligent GOMEA-Heuristic based Treatment planning (BRIGHT) generates multiple treatment plans, each with a different trade-off between tumour coverage and organs-at-risk sparing. BRIGHT was recently extended to cervical cancer brachytherapy. In this study, we present a novel, custom-developed graphical user interface (GUI) that enables plan navigation, pairwise comparisons, dose distribution visualisation, and possibility for adjustments - essential for efficient clinical use of BRIGHT. End-user validation of BRIGHT with the dedicated GUI was conducted for cervical cancer brachytherapy by emulating clinical practice in ten previously treated patients. A multidisciplinary brachytherapy team used BRIGHT to create new treatment plans. GUI usability was assessed using the System Usability Scale (SUS). BRIGHT plan quality was compared to clinical practice via blinded one-on-one comparisons. The GUI offered helpful features for plan navigation and evaluation, giving users quick insight into whether planning aims are achievable and what treatment options are available. The overall SUS score was 83.3, indicating an 'excellent' system. BRIGHT outperformed clinical practice in five out of ten patients regarding the coverage-sparing trade-off and performed equally well in the remaining five. The BRIGHT plan was preferred over the clinical plan in eight out of ten patients, four of which showed clinically relevant differences. The clinical plan was preferred in two patients, neither with clinically relevant differences. In conclusion, BRIGHT, with its dedicated GUI, is a clinically viable and user-friendly tool for treatment planning in cervical cancer brachytherapy.

2026-02-18T10:01:21Z Leah R. M. Dickhoff Ellen M. Kerkhof Heloisa H. Deuzeman Laura A. Velema Stephanie M. de Boer Lavinia A. L. Verhagen Danique L. J. Barten Bradley R. Pieters Lukas J. A. Stalpers Renzo J. Scholman Pedro M. Matos Anton Bouter Carien L. Creutzberg Peter A. N. Bosman Tanja Alderliesten http://arxiv.org/abs/2511.20513v1 2025-11-25T17:19:10Z

Generative models, such as large language models and text-to-image diffusion models, are increasingly used to create visual designs like user interfaces (UIs) and presentation slides. Finetuning and benchmarking these generative models have often relied on datasets of human-annotated design preferences. Yet, due to the subjective and highly personalized nature of visual design, preference varies widely among individuals. In this paper, we study this problem by introducing DesignPref, a dataset of 12k pairwise comparisons of UI design generation annotated by 20 professional designers with multi-level preference ratings. We found that among trained designers, substantial levels of disagreement exist (Krippendorff's alpha = 0.25 for binary preferences). Natural language rationales provided by these designers indicate that disagreements stem from differing perceptions of various design aspect importance and individual preferences. With DesignPref, we demonstrate that traditional majority-voting methods for training aggregated judge models often do not accurately reflect individual preferences. To address this challenge, we investigate multiple personalization strategies, particularly fine-tuning or incorporating designer-specific annotations into RAG pipelines. Our results show that personalized models consistently outperform aggregated baseline models in predicting individual designers' preferences, even when using 20 times fewer examples. Our work provides the first dataset to study personalized visual design evaluation and support future research into modeling individual design taste.

2025-11-25T17:19:10Z Yi-Hao Peng Jeffrey P. Bigham Jason Wu http://arxiv.org/abs/2506.18123v2 2025-11-29T22:53:38Z

Comprehensive, unbiased, and comparable evaluation of modern generalist policies is uniquely challenging: existing approaches for robot benchmarking typically rely on heavy standardization, either by specifying fixed evaluation tasks and environments, or by hosting centralized ''robot challenges'', and do not readily scale to evaluating generalist policies across a broad range of tasks and environments. In this work, we propose RoboArena, a new approach for scalable evaluation of generalist robot policies in the real world. Instead of standardizing evaluations around fixed tasks, environments, or locations, we propose to crowd-source evaluations across a distributed network of evaluators. Importantly, evaluators can freely choose the tasks and environments they evaluate on, enabling easy scaling of diversity, but they are required to perform double-blind evaluations over pairs of policies. Then, by aggregating preference feedback from pairwise comparisons across diverse tasks and environments, we can derive a ranking of policies. We instantiate our approach across a network of evaluators at seven academic institutions using the DROID robot platform. Through more than 600 pairwise real-robot evaluation episodes across seven generalist policies, we demonstrate that our crowd-sourced approach can more accurately rank the performance of existing generalist policies than conventional, centralized evaluation approaches, while being more scalable, resilient, and trustworthy. We open our evaluation network to the community and hope that it can enable more accessible comparisons of generalist robot policies.

2025-06-22T18:13:31Z Website: https://robo-arena.github.io/ Pranav Atreya Karl Pertsch Tony Lee Moo Jin Kim Arhan Jain Artur Kuramshin Clemens Eppner Cyrus Neary Edward Hu Fabio Ramos Jonathan Tremblay Kanav Arora Kirsty Ellis Luca Macesanu Marcel Torne Villasevil Matthew Leonard Meedeum Cho Ozgur Aslan Shivin Dass Jie Wang William Reger Xingfang Yuan Xuning Yang Abhishek Gupta Dinesh Jayaraman Glen Berseth Kostas Daniilidis Roberto Martin-Martin Youngwoon Lee Percy Liang Chelsea Finn Sergey Levine http://arxiv.org/abs/2410.12851v7 2025-04-19T23:14:54Z

Large language models (LLMs) often exhibit subtle yet distinctive characteristics in their outputs that users intuitively recognize, but struggle to quantify. These "vibes" -- such as tone, formatting, or writing style -- influence user preferences, yet traditional evaluations focus primarily on the singular axis of correctness. We introduce VibeCheck, a system for automatically comparing a pair of LLMs by discovering identifying traits of a model (vibes) that are well-defined, differentiating, and user-aligned. VibeCheck iteratively discovers vibes from model outputs and then utilizes a panel of LLM judges to quantitatively measure the utility of each vibe. We validate that the vibes generated by VibeCheck align with those found in human discovery and run VibeCheck on pairwise preference data from real-world user conversations with Llama-3-70b vs GPT-4. VibeCheck reveals that Llama has a friendly, funny, and somewhat controversial vibe. These vibes predict model identity with 80% accuracy and human preference with 61% accuracy. Lastly, we run VibeCheck on a variety of models and tasks including summarization, math, and captioning to provide insight into differences in model behavior. VibeCheck discovers vibes like Command X prefers to add concrete intros and conclusions when summarizing in comparison to TNGL, Llama-405b often overexplains its thought process on math problems compared to GPT-4o, and GPT-4 prefers to focus on the mood and emotions of the scene when captioning compared to Gemini-1.5-Flash. Code and vibe visualizer found at https://bench-mark.org/

2024-10-10T17:59:17Z unironic use of the word 'vibe', added more analysis and cooler graphs. added website link Lisa Dunlap Krishna Mandal Trevor Darrell Jacob Steinhardt Joseph E Gonzalez http://arxiv.org/abs/2410.05582v1 2024-10-08T00:45:49Z

Autonomous driving necessitates the ability to reason about future interactions between traffic agents and to make informed evaluations for planning. This paper introduces the \textit{Gen-Drive} framework, which shifts from the traditional prediction and deterministic planning framework to a generation-then-evaluation planning paradigm. The framework employs a behavior diffusion model as a scene generator to produce diverse possible future scenarios, thereby enhancing the capability for joint interaction reasoning. To facilitate decision-making, we propose a scene evaluator (reward) model, trained with pairwise preference data collected through VLM assistance, thereby reducing human workload and enhancing scalability. Furthermore, we utilize an RL fine-tuning framework to improve the generation quality of the diffusion model, rendering it more effective for planning tasks. We conduct training and closed-loop planning tests on the nuPlan dataset, and the results demonstrate that employing such a generation-then-evaluation strategy outperforms other learning-based approaches. Additionally, the fine-tuned generative driving policy shows significant enhancements in planning performance. We further demonstrate that utilizing our learned reward model for evaluation or RL fine-tuning leads to better planning performance compared to relying on human-designed rewards. Project website: https://mczhi.github.io/GenDrive.

2024-10-08T00:45:49Z Zhiyu Huang Xinshuo Weng Maximilian Igl Yuxiao Chen Yulong Cao Boris Ivanovic Marco Pavone Chen Lv http://arxiv.org/abs/2409.13940v1 2024-09-20T23:04:08Z

This paper presents a novel technique for incorporating user input when learning and inferring user preferences. When trying to provide users of black-box machine learning models with actionable recourse, we often wish to incorporate their personal preferences about the ease of modifying each individual feature. These recourse finding algorithms usually require an exhaustive set of tuples associating each feature to its cost of modification. Since it is hard to obtain such costs by directly surveying humans, in this paper, we propose the use of the Bradley-Terry model to automatically infer feature-wise costs using non-exhaustive human comparison surveys. We propose that users only provide inputs comparing entire recourses, with all candidate feature modifications, determining which recourses are easier to implement relative to others, without explicit quantification of their costs. We demonstrate the efficient learning of individual feature costs using MAP estimates, and show that these non-exhaustive human surveys, which do not necessarily contain data for each feature pair comparison, are sufficient to learn an exhaustive set of feature costs, where each feature is associated with a modification cost.

2024-09-20T23:04:08Z "Recourse for Humans", paper 49 from the Participatory Approaches to Machine Learning workshop at the International Conference on Machine Learning (ICML) 2020. For workshop website, see https://participatoryml.github.io/#49 Kaivalya Rawal Himabindu Lakkaraju http://arxiv.org/abs/2111.12182v2 2021-11-30T15:00:02Z

Terms and conditions (T&Cs) are pervasive on the web and often contain important information for consumers, but are rarely read. Previous research has explored methods to surface alarming privacy policies using manual labelers, natural language processing, and deep learning techniques. However, this prior work used pre-determined categories for annotations, and did not investigate what consumers really deem as important from their perspective. In this paper, we instead combine crowdsourcing with an open definition of "what is important" in T&Cs. We present a workflow consisting of pairwise comparisons, agreement validation, and Bradley-Terry rank modeling, to effectively establish rankings of T&C statements from non-expert crowdworkers on this open definition, and further analyzed consumers' preferences. We applied this workflow to 1,551 T&C statements from 27 e-commerce websites, contributed by 3,462 unique crowd workers doing 203,068 pairwise comparisons, and conducted thematic and readability analysis on the statements considered as important/unimportant. We found that consumers especially cared about policies related to after-sales and money, and tended to regard harder-to-understand statements as more important. We also present machine learning models to identify T&C clauses that consumers considered important, achieving at best a 92.7% balanced accuracy, 91.6% recall, and 89.2% precision. We foresee using our workflow and model to efficiently and reliably highlight important T&Cs on websites at a large scale, improving consumers' awareness

2021-11-23T22:39:52Z Xingyu Liu Annabel Sun Jason I. Hong http://arxiv.org/abs/2010.06848v2 2020-11-05T04:42:58Z

Group buying, as an emerging form of purchase in social e-commerce websites, such as Pinduoduo, has recently achieved great success. In this new business model, users, initiator, can launch a group and share products to their social networks, and when there are enough friends, participants, join it, the deal is clinched. Group-buying recommendation for social e-commerce, which recommends an item list when users want to launch a group, plays an important role in the group success ratio and sales. However, designing a personalized recommendation model for group buying is an entirely new problem that is seldom explored. In this work, we take the first step to approach the problem of group-buying recommendation for social e-commerce and develop a GBGCN method (short for Group-Buying Graph Convolutional Network). Considering there are multiple types of behaviors (launch and join) and structured social network data, we first propose to construct directed heterogeneous graphs to represent behavioral data and social networks. We then develop a graph convolutional network model with multi-view embedding propagation, which can extract the complicated high-order graph structure to learn the embeddings. Last, since a failed group-buying implies rich preferences of the initiator and participants, we design a double-pairwise loss function to distill such preference signals. We collect a real-world dataset of group-buying and conduct experiments to evaluate the performance. Empirical results demonstrate that our proposed GBGCN can significantly outperform baseline methods by 2.69%-7.36%. The codes and the dataset are released at https://github.com/Sweetnow/group-buying-recommendation.

2020-10-14T07:20:21Z IEEE International Conference on Data Engineering (ICDE), 2021 Jun Zhang Chen Gao Depeng Jin Yong Li http://arxiv.org/abs/1605.07722v3 2017-04-30T17:43:02Z

Nutrient-based meal recommendations have the potential to help individuals prevent or manage conditions such as diabetes and obesity. However, learning people's food preferences and making recommendations that simultaneously appeal to their palate and satisfy nutritional expectations are challenging. Existing approaches either only learn high-level preferences or require a prolonged learning period. We propose Yum-me, a personalized nutrient-based meal recommender system designed to meet individuals' nutritional expectations, dietary restrictions, and fine-grained food preferences. Yum-me enables a simple and accurate food preference profiling procedure via a visual quiz-based user interface, and projects the learned profile into the domain of nutritionally appropriate food options to find ones that will appeal to the user. We present the design and implementation of Yum-me, and further describe and evaluate two innovative contributions. The first contriution is an open source state-of-the-art food image analysis model, named FoodDist. We demonstrate FoodDist's superior performance through careful benchmarking and discuss its applicability across a wide array of dietary applications. The second contribution is a novel online learning framework that learns food preference from item-wise and pairwise image comparisons. We evaluate the framework in a field study of 227 anonymous users and demonstrate that it outperforms other baselines by a significant margin. We further conducted an end-to-end validation of the feasibility and effectiveness of Yum-me through a 60-person user study, in which Yum-me improves the recommendation acceptance rate by 42.63%.

2016-05-25T04:13:49Z Longqi Yang Cheng-Kang Hsieh Hongjian Yang Nicola Dell Serge Belongie Curtis Cole Deborah Estrin