论文检索

Zhonghao Yan, Muxi Diao, Yuxuan Yang, Ruoyan Jing, Jiayuan Xu, Kaizhou Zhang, Lele Yang, Yanxi Liu, Kongming Liang, Zhanyu Ma

Accurately grounding regions of interest (ROIs) is critical for diagnosis and treatment planning in medical imaging. While multimodal large language models (MLLMs) combine visual perception with natural language, current medical-grounding pipelines still rely on supervised fine-tuning with explicit spatial hints, making them ill-equipped to handle the implicit queries common in clinical practice. This work makes three core contributions. We first define Unified Medical Reasoning Grounding (UMRG), a novel vision–language task that demands clinical reasoning and pixel-level grounding. Second, we release U-MRG-14K, a dataset of 14K samples featuring pixel-level masks alongside implicit clinical queries and reasoning traces, spanning 10 modalities, 15 super-categories, and 108 specific categories. Finally, we introduce MedReasoner, a modular framework that distinctly separates reasoning from segmentation: an MLLM reasoner is optimized with reinforcement learning, while a frozen segmentation expert converts spatial prompts into masks, with alignment achieved through format and accuracy rewards. MedReasoner achieves state-of-the-art performance on U-MRG-14K and demonstrates strong generalization to unseen clinical queries, underscoring the significant promise of reinforcement learning for interpretable medical grounding.

Jierui Qu, Jianchun Zhao

Accurate whole-heart segmentation is a critical component in the precise diagnosis and interventional planning of cardiovascular diseases. Integrating complementary information from modalities such as computed tomography (CT) and magnetic resonance imaging (MRI) can significantly enhance segmentation accuracy and robustness. However, existing multi-modal segmentation methods face several limitations: severe spatial inconsistency between modalities hinders effective feature fusion; fusion strategies are often static and lack adaptability; and the processes of feature alignment and segmentation are decoupled and inefficient. To address these challenges, we propose a dual-branch U-Net architecture enhanced by reinforcement learning for feature alignment, termed RL-U2Net, designed for precise and efficient multi-modal 3D whole-heart segmentation. The model employs a dual-branch U-shaped network to process CT and MRI patches in parallel, and introduces a novel RL-XAlign module between the encoders. The module employs a cross‑modal attention mechanism to capture semantic correspondences between modalities and a reinforcement learning agent learns an optimal rotation strategy that consistently aligns anatomical pose and texture features. The aligned features are then reconstructed through their respective decoders. Finally, an ensemble‑learning–based decision module integrates the predictions from individual patches to produce the final segmentation result. Experimental results on the publicly available MM-WHS 2017 dataset demonstrate that the proposed RL-U2Net outperforms existing state-of-the-art methods, achieving Dice coefficients of 93.1% on CT and 87.0% on MRI, thereby validating the effectiveness and superiority of the proposed approach.

Miao Pan, Wangjie Gan, Jintao Chen, Wenqi Zhang, Sun Bing, Jianwei Yin, Xuhong Zhang

Multimodal large language models (MLLMs) have achieved significant results in various tasks, but their practical application is still severely constrained by hallucination issues, which are particularly prominent in reinforcement learning (RL) optimization processes. This paper systematically analyzes the causes of hallucinations in MLLM under RL training, identifying three key factors: (1) The model relies heavily on chained visual reasoning to guide decision-making during RL training. Thus, error and irrelevant information in visual reasoning can easily cause hallucinations, including inaccurate initial visual descriptions that anchor subsequent inferences to incorrect information, as well as redundant and broad inferential information; (2) Insufficient exploration diversity during the policy optimization phase, causing the model to output overly confident results; (3) The destructive conflict between different samples during optimization is a key factor that leads to false associations and unstable parameter updates. To address these issues, we propose a solution framework comprising three core modules. First, to improve the accuracy of visual localization, we add planning and caption stages before thinking and answer stages. To enhance initial visual descriptions ability, we allow LLMs to respond based solely on the caption and provide corresponding caption reward based on the quality of the response. Second, to enhance exploration capabilities, we classify samples based on the mean and variance of the reward distribution and select samples with high reward variance for training, thereby increasing the model's focus on diverse samples. Finally, to mitigate conflicts between training samples, we identify neural tangent kernel (NTK) similarity as the key factor. Rather than minimizing it uniformly, we regulate NTK similarity by grouping sample pairs based on a similarity threshold. An InfoNCE loss is then applied to pull dissimilar pairs closer and push overly similar ones apart, guiding interactions toward a balanced range. The experimental results demonstrate that the proposed method significantly reduces the hallucination rate and effectively improves the inference accuracy of MLLMs.

Yue Li, Meng Tian, Dechang Zhu, Jiangtong Zhu, Zhenyu Lin, Zhiwei Xiong, Xinhai Zhao

Large vision-language models (VLMs) for autonomous driving (AD) are evolving beyond perception and cognition tasks toward motion planning. However, we identify two critical challenges in this direction: (1) VLMs tend to learn shortcuts by relying heavily on history input information, achieving seemingly strong planning results without genuinely understanding the visual inputs; and (2) the chain-of-thought (COT) reasoning processes are always misaligned with the motion planning outcomes, and how to effectively leverage the complex reasoning capability to enhance planning remains largely underexplored. In this paper, we start from a small-scale domain-specific VLM and propose Drive-R1, designed to bridge the scenario reasoning and motion planning for AD. Drive-R1 first undergoes the supervised finetuning on an elaborate dataset containing both long and short COT data. Drive-R1 is encouraged to reason step-by-step from visual input to final planning decisions. Subsequently, Drive-R1 is trained within a reinforcement learning framework that incentivizes the discovery of reasoning paths that are more informative for planning, guided by rewards based on predicted trajectories and meta actions. Experimental evaluations on the nuScenes and DriveLM-nuScenes benchmarks demonstrate that Drive-R1 achieves superior performance compared to existing state-of-the-art VLMs. We believe that Drive-R1 presents a promising direction for bridging reasoning and planning in AD, offering methodological insights for future research and applications.

Yingxin Lai, Yufei Liu, Guoqing Yang, Jiaxing Chai, Zhiming Luo, Shaozi Li

Despite recent advancements in font generation, practitioners still grapple with a laborious trial-and-error workflow. To streamline this, we propose OneFont, an end-to-end framework that interprets user intents via free-form dialogue, seamlessly integrating both glyph synthesis and refinement modules. We introduce the Font with Thought (FwT) paradigm, reframing font design as a reasoning task where the model plans actions and articulates design rationales. OneFont’s core planner is trained via a two-stage regimen to master this paradigm. First, we instill reasoning abilities via Supervised Fine-Tuning (SFT) on a new, comprehensive benchmark of 1,500 font families we built. Second, we refine the model's policy with a novel reinforcement learning algorithm, Group Relative Policy Optimization (GRPO), guided by a hybrid reward that assesses visual fidelity, rationale coherence, and transformation correctness. Extensive experiments show OneFont significantly surpasses existing methods in design quality and stroke precision across diverse scripts, validated on our new benchmark. We will release our dataset, code, and models.

Boyu Chen, Zikang Wang, Zhengrong Yue, Kainan Yan, Chenyun Yu, Yi Huang, Zijun Liu, Yafei Wen, Xiaoxin Chen, Yang Liu 等

Most of the multi-agent video understanding frameworks adopt static and non-learnable tool invocation mechanisms, which limit the discovery of diverse clues essential for robust perception and reasoning regarding temporally or spatially complex videos. To address this challenge, we propose a novel Multi-agent system for video understanding, namely VideoChat-M1. Instead of using a single or fixed policy, we adopt a distinct Collaborative Policy Planning (CPP) paradigm with multiple policy agents, which comprises three key processes. (1)Policy Generation: Each agent generates its unique tool invocation policy tailored to the user's query. (2) Policy Execution: Each agent sequentially invokes relevant tools to execute its policy and explore the video content. (3) Policy Communication: During the intermediate stages of policy execution, agents interact with one another to update their respective policies. Through this collaborative framework, all agents work in tandem, dynamically refining their preferred policies based on contextual insights from peers. Moreover, we equip our CPP paradigm with Multi-Agent Reinforcement Learning (MARL). Consequently, policy agents can be jointly optimized to enhance the performance, guided by both the final answer reward and intermediate collaborative process feedback. Extensive experiments demonstrate that VideoChat-M1 achieves SOTA performance across eight benchmarks on four tasks. Notably, on LongVideoBench, our method outperforms Gemini 2.5 pro by 3.6% and GPT-4o by 15.6%.

Rujiao Long, Yang Li, Xingyao Zhang, Weixun Wang, Tianqianjin Lin, Xi Zhao, Yuchi Xu, Wenbo Su, Junchi Yan, Bo Zheng

Exploration capacity shapes both inference-time performance and reinforcement learning (RL) training for large (vision-) language models, as stochastic sampling often yields redundant reasoning paths with little high-level diversity. This paper proposes Reasoning Palette, a novel latent-modulation framework that endows the model with a stochastic latent variable for strategic contextualization, guiding its internal planning prior to token generation. This latent context is inferred from the mean-pooled embedding of a question-answer pair via a variational autoencoder (VAE), where each sampled latent potentially encodes a distinct reasoning context. During inference, a sampled latent is decoded into learnable token prefixes and prepended to the input prompt, modulating the model's internal reasoning trajectory. In this way, the model performs internal sampling over reasoning strategies prior to output generation, which shapes the style and structure of the entire response sequence. A brief supervised fine-tuning (SFT) warm-up phase allows the model to adapt to this latent conditioning. Within RL optimization, Reasoning Palette facilitates structured exploration by enabling on-demand injection for diverse reasoning modes, significantly enhancing exploration efficiency and sustained learning capability. Experiments across multiple reasoning benchmarks demonstrate that our method enables interpretable and controllable control over the (vision-) language model's strategic behavior, thereby achieving consistent performance gains over standard RL methods.

Xu Cao

Current video vision language models (VLMs) process information passively, lacking the ability to dynamically plan their analysis or perform joint reasoning across crucial modalities such as video and audio. To address this, we introduce Hippocampal Sensing (HippoSense), a learning paradigm inspired by hippocampus that shifts the focus from temporal predictive sensing to cross-modal predictive sensing. The core objective of HippoSense is to train the model to anticipate current status's audio-caption summarizations from video and vice versa. We present HippoVLM, a video VLM that operationalizes this paradigm. Instead of passively ingesting all data, HippoVLM actively reasons about its information needs using Chain-of-Thought (CoT). Our training process is twofold: we first finetune HippoVLM with HippoSense, and then apply a novel contrastive Reinforcement Learning (RL) algorithm, Video-Audio Negative-aware Optimization (VANAO), to optimize this selective co-reasoning process. This approach proves highly effective: despite their significantly smaller size, our HippoVLM model achieve competitive performance to massive MLLMs like GPT-4o and Gemini 1.5 Pro on multiple video VQA benchmarks.

Songping Wang, Rufan Qian, Yueming Lyu, Qinglong Liu, Linzhuang Zou, Jie Qin, Songhua Liu, Caifeng Shan

Image-to-Video (I2V) generation represents a frontier in content creation, where models synthesize dynamic visual sequences by jointly reasoning from both image and text prompts. This multimodal grounding enables diverse controllability over video attributes. However, it is precisely this capability that introduces a critical security blind spot: by exploiting the interplay between visual and textual cues, attackers can launch multimodal jailbreak attacks that severely compromise output security. Despite the increasing implementation of security mechanisms in real-world I2V systems, such cross-modal threats remain unexplored. Existing attack methods remain confined to single-modal settings, relying solely on isolated text or image perturbations, which severely limits their effectiveness. To bridge this gap, we propose Runaway Evil, the first multimodal jailbreaking framework for I2V models with dynamic evolutionary capability. Built on a Strategy-Tactic-Action paradigm, our framework exhibits self-amplifying attack through three core components: (1) a strategy-aware command unit that enables the attack to self-evolve its strategies through reinforcement learning-driven strategy customization and large language model (LLM)-based strategy exploration; (2) a multimodal tactical planning unit that generates synergistic text jailbreak instructions and image tampering guidelines based on the selected strategies; and (3) an tactical action Unit executes and evaluates the coordinated attacks. This self-evolving architecture allows the framework to continuously adapt and intensify its attack strategies without human intervention. Extensive experiments demonstrate that Runaway Evil achieves state-of-the-art attack success rates on commercial I2V models, such as Open-Sora 2.0 and CogVideoX. This work provides a critical tool for probing and mitigating multimodal vulnerabilities, laying a foundation for building more robust video generation systems. The code is available at: https://github.com/DeepSota/RunawayEvil.

Chuan Mao, Haoqi Yuan, Ziye Huang, Chaoyi Xu, Kai Ma, Zongqing Lu

Reinforcement learning (RL) has achieved great success in dexterous grasping, significantly improving grasp performance and generalization from simulation to the real world. However, fine-grained functional grasping, which is essential for downstream manipulation tasks, remains underexplored and faces several challenges: the complexity of specifying goals and reward functions for functional grasps across diverse objects, the difficulty of multi-task RL exploration, and the challenge of sim-to-real transfer. In this work, we propose DemoFunGrasp for universal dexterous functional grasping. We factorize functional grasping conditions into two complementary components -- grasping style and affordance -- and integrate them into an RL framework that can learn to grasp any object with any functional grasping condition. To address the multi-task optimization challenge, we leverage a single grasping demonstration and reformulate the RL problem as one-step demonstration editing, substantially enhancing sample efficiency and performance.Experimental results in both simulation and the real world show that DemoFunGrasp generalizes to unseen combinations of objects, affordances, and grasping styles, outperforming baselines in both success rate and functional grasping accuracy. In addition to strong sim-to-real capability, by incorporating a vision-language model (VLM) for planning, our system achieves autonomous instruction-following grasp execution.

Sirui Xu, Samuel Schulter, Morteza Ziyadi, Xialin He, Xiaohan Fei, Yu-Xiong Wang, Liang-Yan Gui

Humans rarely plan whole-body interactions with objects at the level of explicit whole-body movements. High-level intentions, such as affordance, define the goal, while coordinated balance, contact, and manipulation can emerge naturally from underlying physical and motor priors. Scaling such priors is key to enabling humanoids to compose and generalize loco-manipulation skills across diverse contexts while maintaining physically coherent whole-body coordination. To this end, we introduce InterPrior, a scalable framework that learns a unified generative controller through large-scale imitation pretraining and post-training by reinforcement learning. InterPrior first distills a full-reference imitation expert into a versatile, goal-conditioned variational policy that reconstructs motion from multimodal observations and high-level intent. While the distilled policy reconstructs training behaviors, it does not generalize reliably due to the vast configuration space of large-scale human-object interactions. To address this, we apply data augmentation with physical perturbations, and then perform reinforcement learning finetuning to improve competence on unseen goals and initializations. Together, these steps consolidate the reconstructed latent skills into a valid manifold, yielding a motion prior that generalizes beyond the training data, e.g., it can incorporate new behaviors such as interactions with unseen objects. We further demonstrate its effectiveness for user-interactive control and its potential for real robot deployment.

Runyu Jiao, Matteo Bortolon, Francesco Giuliari, Alice Fasoli, Sergio Povoli, Guofeng Mei, Yiming Wang, Fabio Poiesi

Successful robotic grasping in cluttered environments not only requires a model to visually ground a target object but also to reason about obstructions that must be cleared beforehand. While current vision-language embodied reasoning models show emergent spatial understanding, they remain limited in terms of obstruction reasoning and accessibility planning. To bridge this gap, we present UNOGrasp, a learning-based vision-language model capable of performing visually-grounded obstruction reasoning to infer the sequence of actions needed to unobstruct the path and grasp the target object. We devise a novel multi-step reasoning process based on obstruction paths originated by the target object. We anchor each reasoning step with obstruction-aware visual cues to incentivize reasoning capability. UNOGrasp combines supervised and reinforcement finetuning through verifiable reasoning rewards. Moreover, we construct UNOBench, a large-scale dataset for both training and benchmarking, based on MetaGraspNetV2, with over 100k obstruction paths annotated by humans with obstruction ratios, contact points, and natural-language instructions. Extensive experiments and real-robot evaluations show that UNOGrasp significantly improves obstruction reasoning and grasp success across both synthetic and real-world environments, outperforming generalist and proprietary alternatives.

Yuhan Wang, Zihan Li, Han Liu, Simon Arberet, Martin Kraus, Yuyin Zhou, Florin-Cristian Ghesu, Dorin Comaniciu, Ali Kamen, Riqiang Gao

Voxel-wise dose prediction is a critical yet challenging task in radiotherapy (RT) planning, as bespoke models trained from scratch often struggle to generalize across institutions, scanners, and planning protocols. Meanwhile, large generative backbones pretrained on billion-scale visual data have learned strong geometric and structural priors that may transfer beyond their source domain. We propose DiffKT3D, an Any2Any 3D diffusion framework that transfers prior knowledge from pretrained video diffusion models to clinically meaningful dose prediction. To support flexible conditioning over heterogeneous clinical inputs, including CT, target and organ masks, body mask, and beam descriptors, we introduce a modality- and role-aware Any2Any formulation that avoids cross-attention while retaining a unified conditioning interface. We further apply reinforcement-learning (RL) post-training with a clinically informed Scorecard reward to align generations with institutional treatment preferences. On the GDP--HMM challenge, DiffKT3D reduces voxel-wise MAE from 2.07 to 1.93 relative to the top challenge solution, while also improving image quality and clinical preference alignment. These results show that pretrained diffusion priors, combined with flexible multimodal conditioning and scorecard-aligned post-training, provide an effective approach to RT dose prediction.

Peiran Xu, Jiaqi Zheng, Yadong Mu

This paper focuses on embodied task planning, where an agent acquires visual observations from the environment and executes atomic actions to accomplish a given task. Although recent Vision-Language Models (VLMs) have achieved impressive results in multimodal understanding and reasoning, their performance remains limited when applied to embodied planning that involves multi-turn interaction, long-horizon reasoning, and extended context analysis. To bridge this gap, we propose RoboAgent, a capability-driven planning pipeline in which the model actively invokes different sub-capabilities. Each capability maintains its own context, and produces intermediate reasoning results or interacts with the environment according to the query given by a scheduler. This framework decomposes complex planning into a sequence of basic vision-language problems that VLMs can better address, enabling a more transparent and controllable reasoning process. The scheduler and all capabilities are implemented with a single VLM, without relying on external tools. To train this VLM, we adopt a multi-stage paradigm that consists of: (1) behavior cloning with expert plans, (2) DAgger training using trajectories collected by the model, and (3) reinforcement learning guided by an expert policy. Across these stages, we exploit the internal information of the environment simulator to construct high-quality supervision for each capability, and we further introduce augmented and synthetic data to enhance the model's performance in more diverse scenarios. Extensive experiments on widely used embodied task planning benchmarks validate the effectiveness of the proposed approach.

Huajie Tan, Peterson Co, Yijie Xu, Shanyu Rong, Yuheng Ji, Cheng Chi, Xiansheng Chen, Zhongxia Zhao, Pengwei Wang, Zhongyuan Wang 等

Long-horizon robotic manipulation is increasingly important for real-world deployment, requiring spatial disambiguation in complex layouts and temporal resilience under dynamic interaction. However, existing end-to-end and hierarchical Vision-Language-Action (VLA) policies often rely on text-only cues while keeping plan intent latent, which undermines referential grounding in cluttered or underspecified scenes, impedes effective task decomposition of long-horizon goals with close-loop interaction, and limits causal explanation by obscuring the rationale behind action choices. To address these issues, we first introduce Visual Sketch, an implausible visual intermediate that renders points, boxes, arrows, and typed relations in the robot's current views to externalize spatial intent, connect language to scene geometry, and provide a human-verifiable bridge between high-level reasoning and low-level control. Building on Visual Sketch, we present Action-Sketcher, a VLA framework that operates in a cyclic See -> Think -> Sketch -> Act workflow coordinated by adaptive token-gated strategy for reasoning triggers, sketch revision, and action issuance, thereby supporting reactive corrections and human interaction while preserving real-time action prediction. To enable scalable training and evaluation, we curate diverse corpus with interleaved images, text, Visual Sketch supervision, and action sequences, and train Action-Sketcher with a multi-stage curriculum recipe that combines interleaved sequence alignment for modality unification, language-to-sketch consistency for precise linguistic grounding, and imitation learning augmented with sketch-to-action reinforcement for robustness. Extensive experiments on cluttered scenes and multi-object tasks, in simulation and on real-world tasks, show improved long-horizon success, stronger robustness to dynamic scene changes, and enhanced interpretability via editable sketches and step-wise plans.

Yuechen Luo, Fang Li, Qimao Chen, Shaoqing Xu, Jiaxin Liu, Ziying Song, Zhi-xin Yang, Fuxi Wen

Vision-Language-Action (VLA) models for autonomous driving often hit a performance plateau during Reinforcement Learning (RL) optimization. This stagnation arises from exploration capabilities constrained by previous Supervised Fine-Tuning (SFT), leading to "persistent failures" in long-tail scenarios. In these critical situations, all explored actions yield a zero-value driving score. This information-sparse reward signals a failure, yet fails to identify its root cause--whether it is due to incorrect planning, flawed reasoning, or poor trajectory execution. To address this limitation, we propose **VLA** with **E**xplicit **L**earning from **F**ailures (**ELF-VLA**), a framework that augments RL with structured diagnostic feedback. Instead of relying on a vague scalar reward, our method produces detailed, interpretable reports that identify the specific failure mode. The VLA policy then leverages this explicit feedback to generate a **Feedback-Guided Refinement**. By injecting these corrected, high-reward samples back into the RL training batch, our approach provides a targeted gradient, which enables the policy to solve critical scenarios that unguided exploration cannot. Extensive experiments demonstrate that our method unlocks the latent capabilities of VLA models, achieving state-of-the-art (SOTA) performance on the public Navsim benchmark for overall PDMS, EPDMS score and high-level planning accuracy.

Zihan Wang, Seungjun Lee, Guangzhao Dai, Gim Hee Lee

Embodied agents face a critical dilemma that end-to-end models lack interpretability and explicit 3D reasoning, while modular systems ignore cross-component interdependencies and synergies. To bridge this gap, we propose the Dynamic 3D Vision-Language-Planning Model (D3D-VLP). Our model introduces two key innovations: 1) A Dynamic 3D Chain-of-Thought (3D CoT) that unifies planning, grounding, navigation, and question answering within a single 3D-VLM and CoT pipeline; 2) A Synergistic Learning from Fragmented Supervision (SLFS) strategy, which uses a masked autoregressive loss to learn from massive and partially-annotated hybrid data. This allows different CoT components to mutually reinforce and implicitly supervise each other. To this end, we construct a large-scale dataset with 10M hybrid samples from 5K real scans and 20K synthetic scenes that are compatible with online learning methods such as RL and DAgger. Our D3D-VLP achieves state-of-the-art results on multiple benchmarks, including Vision-and-Language Navigation (R2R-CE, REVERIE-CE, NavRAG-CE), Object-goal Navigation (HM3D-OVON), and Task-oriented Sequential Grounding and Navigation (SG3D). Real-world mobile manipulation experiments further validate the effectiveness.

Zhongyu Xia, Wenhao Chen, Yongtao Wang, Ming-Hsuan Yang

Visual-language reasoning, driving knowledge, and value alignment are essential for advanced autonomous driving systems. However, existing approaches largely rely on data-driven learning, making it difficult to capture the complex logic underlying decision-making through imitation or limited reinforcement rewards. To address this, we propose KnowVal, a new autonomous driving system that enables visual-language reasoning through the synergistic integration of open-world perception and knowledge retrieval. Specifically, we construct a comprehensive driving knowledge graph that encodes traffic laws, defensive driving principles, and ethical norms, complemented by an efficient LLM-based retrieval mechanism tailored for driving scenarios. Furthermore, we develop a human-preference dataset and train a Value Model to guide interpretable, value-aligned trajectory assessment. Experimental results show that our method substantially improves planning performance while remaining compatible with existing architectures. Notably, KnowVal achieves the lowest collision rate on nuScenes and state-of-the-art results on Bench2Drive and NAVSIM.

Ziyu Guo, Renrui Zhang, Hongyu Li, Manyuan Zhang, Xinyan Chen, Sifan Wang, Yan Feng, Peng Pei, Pheng-Ann Heng

Recent advances in visual generation have increasingly explored the integration of reasoning capabilities. They incorporate textual reasoning, i.e., think, either before (as pre-planning) or after (as post-refinement) the generation process, yet they lack on-the-fly multimodal interaction during the generation itself. In this preliminary study, we introduce Thinking-while-Generating (TwiG), the first interleaved framework that enables co-evolving textual reasoning throughout the visual generation process. As visual content is progressively generating, textual reasoning is interleaved to both guide upcoming local regions and reflect on previously synthesized ones. This dynamic interplay produces more context-aware and semantically rich visual outputs. To unveil the potential of this framework, we investigate three candidate strategies, zero-shot prompting, supervised fine-tuning (SFT) on our curated TwiG-50K dataset, and reinforcement learning (RL) via a customized TwiG-GRPO strategy, each offering unique insights into the dynamics of interleaved reasoning. We hope this work inspires further research into interleaving textual reasoning for enhanced visual generation. Code is released at: https://github.com/ZiyuGuo99/Thinking-while-Generating.

Zeyu Zhang, Danning Li, Ian Reid, Richard Hartley

Energy-based predictive world models provide a powerful approach for multi-step visual planning by reasoning over latent energy landscapes rather than generating pixels. However, existing approaches face two major challenges: (i) their latent representations are typically learned in Euclidean space, neglecting the underlying geometric and hierarchical structure among states, and (ii) they struggle with long-horizon prediction, which leads to rapid degradation across extended rollouts. To address these challenges, we introduce GeoWorld, a geometric world model that preserves geometric structure and hierarchical relations through a Hyperbolic JEPA, which maps latent representations from Euclidean space onto hyperbolic manifolds. We further introduce Geometric Reinforcement Learning for energy-based optimization, enabling stable multi-step planning in hyperbolic latent space. Extensive experiments on CrossTask and COIN demonstrate around 3% SR improvement in 3-step planning and 2% SR improvement in 4-step planning compared to the state-of-the-art V-JEPA-2.