While large multimodal models (LMMs) have demonstrated strong performance across various Visual Question Answering (VQA) tasks, certain challenges require complex multi-step reasoning to reach accurate answers. One particularly challenging task is autonomous driving, which demands thorough cognitive processing before decisions can be made. In this domain, a sequential and interpretive understanding of visual cues is essential for effective perception, prediction, and planning. Nevertheless, common VQA benchmarks often focus on the accuracy of the final answer while overlooking the reasoning process that enables the generation of accurate responses. Moreover, existing methods lack a comprehensive framework for evaluating step-by-step reasoning in realistic driving scenarios. To address this gap, we propose DriveLMM-o1, a new dataset and benchmark specifically designed to advance step-wise visual reasoning for autonomous driving. Our benchmark features over 18k VQA examples in the training set and more than 4k in the test set, covering diverse questions on perception, prediction, and planning, each enriched with step-by-step reasoning to ensure logical inference in autonomous driving scenarios. We further introduce a large multimodal model that is fine-tuned on our reasoning dataset, demonstrating robust performance in complex driving scenarios. In addition, we benchmark various open-source and closed-source methods on our proposed dataset, systematically comparing their reasoning capabilities for autonomous driving tasks. Our model achieves a +7.49% gain in final answer accuracy, along with a 3.62% improvement in reasoning score over the previous best open-source model. Our framework, dataset, and model are available at https://github.com/ayesha-ishaq/DriveLMM-o1.

We introduce LUMOS, a language-conditioned multi-task imitation learning framework for robotics. LUMOS learns skills by practicing them over many long-horizon rollouts in the latent space of a learned world model and transfers these skills zero-shot to a real robot. By learning on-policy in the latent space of the learned world model, our algorithm mitigates policy-induced distribution shift which most offline imitation learning methods suffer from. LUMOS learns from unstructured play data with fewer than 1% hindsight language annotations but is steerable with language commands at test time. We achieve this coherent long-horizon performance by combining latent planning with both image- and language-based hindsight goal relabeling during training, and by optimizing an intrinsic reward defined in the latent space of the world model over multiple time steps, effectively reducing covariate shift. In experiments on the difficult long-horizon CALVIN benchmark, LUMOS outperforms prior learning-based methods with comparable approaches on chained multi-task evaluations. To the best of our knowledge, we are the first to learn a language-conditioned continuous visuomotor control for a real-world robot within an offline world model. Videos, dataset and code are available at http://lumos.cs.uni-freiburg.de.

Existing vision-language models (VLMs) often suffer from visual hallucination, where the generated responses contain inaccuracies that are not grounded in the visual input. Efforts to address this issue without model finetuning primarily mitigate hallucination by reducing biases contrastively or amplifying the weights of visual embedding during decoding. However, these approaches improve visual perception at the cost of impairing the language reasoning capability. In this work, we propose the Perception Magnifier (PM), a novel visual decoding method that iteratively isolates relevant visual tokens based on attention and magnifies the corresponding regions, spurring the model to concentrate on fine-grained visual details during decoding. Specifically, by magnifying critical regions while preserving the structural and contextual information at each decoding step, PM allows the VLM to enhance its scrutiny of the visual input, hence producing more accurate and faithful responses. Extensive experimental results demonstrate that PM not only achieves superior hallucination mitigation but also enhances language generation while preserving strong reasoning capabilities. Code is available at https://github.com/ShunqiM/PM .

This paper introduces the System 0/1/2/3 framework as an extension of dual-process theory, employing a quad-process model of cognition. Expanding upon System 1 (fast, intuitive thinking) and System 2 (slow, deliberative thinking), we incorporate System 0, which represents pre-cognitive embodied processes, and System 3, which encompasses collective intelligence and symbol emergence. We contextualize this model within Bergson's philosophy by adopting multi-scale time theory to unify the diverse temporal dynamics of cognition. System 0 emphasizes morphological computation and passive dynamics, illustrating how physical embodiment enables adaptive behavior without explicit neural processing. Systems 1 and 2 are explained from a constructive perspective, incorporating neurodynamical and AI viewpoints. In System 3, we introduce collective predictive coding to explain how societal-level adaptation and symbol emergence operate over extended timescales. This comprehensive framework ranges from rapid embodied reactions to slow-evolving collective intelligence, offering a unified perspective on cognition across multiple timescales, levels of abstraction, and forms of human intelligence. The System 0/1/2/3 model provides a novel theoretical foundation for understanding the interplay between adaptive and cognitive processes, thereby opening new avenues for research in cognitive science, AI, robotics, and collective intelligence.

Reasoning ability is crucial for solving challenging tasks. With the advancement of foundation models, such as the emergence of large language models (LLMs), a wide range of reasoning strategies has been proposed, including test-time enhancements, such as Chain-ofThought, and post-training optimizations, as used in DeepSeek-R1. While these reasoning strategies have demonstrated effectiveness across various challenging language or vision tasks, their applicability and impact on time-series forecasting (TSF), particularly the challenging zero-shot TSF, remain largely unexplored. In particular, it is unclear whether zero-shot TSF benefits from reasoning and, if so, what types of reasoning strategies are most effective. To bridge this gap, we propose ReC4TS, the first benchmark that systematically evaluates the effectiveness of popular reasoning strategies when applied to zero-shot TSF tasks. ReC4TS conducts comprehensive evaluations across datasets spanning eight domains, covering both unimodal and multimodal with short-term and longterm forecasting tasks. More importantly, ReC4TS provides key insights: (1) Self-consistency emerges as the most effective test-time reasoning strategy; (2) Group-relative policy optimization emerges as a more suitable approach for incentivizing reasoning ability during post-training; (3) Multimodal TSF benefits more from reasoning strategies compared to unimodal TSF. Beyond these insights, ReC4TS establishes two pioneering starting blocks to support future zero-shot TSF reasoning research: (1) A novel dataset, TimeThinking, containing forecasting samples annotated with reasoning trajectories from multiple advanced LLMs, and (2) A new and simple test-time scaling-law validated on foundational TSF models enabled by self-consistency reasoning strategy. All data and code are publicly accessible at: https://github.com/AdityaLab/OpenTimeR

In this note, we examine the aggregation of preferences achieved by the Group Policy Optimisation (GRPO) algorithm, a reinforcement learning method used to train advanced artificial intelligence models such as DeepSeek-R1-Zero and DeepSeekMath. The GRPO algorithm trains a policy using a reward preference model, which is computed by sampling a set of outputs for a given context, observing the corresponding rewards, and applying shift-and-scale normalisation to these reward values. Additionally, it incorporates a penalty function to discourage deviations from a reference policy. We present a framework that enables us to characterise the stationary policies of the GRPO algorithm. This analysis reveals that the aggregation of preferences differs fundamentally from standard logarithmic pooling, which is implemented by other approaches such as RLHF. The precise form of preference aggregation arises from the way the reward preference model is defined and from the penalty function, which we show to essentially correspond to the reverse Kullback-Leibler (KL) divergence between the aggregation policy and the reference policy. Interestingly, we demonstrate that for groups of size two, the reward preference model corresponds to pairwise comparison preferences, similar to those in other alignment methods based on pairwise comparison feedback. We provide explicit characterisations of the aggregate preference for binary questions, for groups of size two, and in the limit of large group size. This provides insights into the dependence of the aggregate preference on parameters such as the regularisation constant and the confidence margin of question answers. Finally, we discuss the aggregation of preferences obtained by modifying the GRPO algorithm to use direct KL divergence as the penalty or to use rewards without scale normalisation.

The most promising recent methods for AI reasoning require applying variants of reinforcement learning (RL) either on rolled out trajectories from the model, even for the step-wise rewards, or large quantities of human annotated trajectory data. The reliance on the rolled-out trajectory renders the compute cost and time prohibitively high. In particular, the correctness of a reasoning trajectory can typically only be judged at its completion, leading to sparse rewards in RL or requiring expensive synthetic data generation in expert iteration-like methods. In this work, we focus on the Automatic Theorem Proving (ATP) task and propose a novel verifier-in-the-loop design, which unlike existing approaches that leverage feedback on the entire reasoning trajectory, employs an automated verifier to give intermediate feedback at each step of the reasoning process. Using Lean as the verifier, we empirically show that the step-by-step local verification produces a global improvement in the model's reasoning accuracy and efficiency.

Post-training plays a crucial role in refining language models, ensuring they exhibit good reasoning capabilities, alignment with ethical and social values, and adaptability to user-specific preferences. Unlike pre-training, which demands extensive computational resources and large-scale datasets, post-training is a more efficient process that leverages targeted fine-tuning techniques such as reinforcement learning from human feedback (RLHF) [1, 2], instruction tuning [3], and Direct Preference Optimization (DPO) [4]. These approaches enable models to generalize better across various tasks and mitigate biases. Furthermore, post-training requires significantly low computational overhead compared to pre-training [5]. Reasoning has long been regarded as a cornerstone in the development of large language models (LLMs).

Large vision-language models (VLMs) face challenges in achieving robust, transferable reasoning abilities due to reliance on labor-intensive manual instruction datasets or computationally expensive self-supervised methods. To address these issues, we introduce MindGYM, a framework that enhances VLMs through synthetic self-challenging questions, consisting of three stages: (1) Seed Single-Hop Question Synthesis, generating cognitive questions across textual (e.g., logical deduction) and multimodal contexts (e.g., diagram-based queries) spanning eight semantic areas like ethical analysis; (2) Challenging Multi-Hop Question Synthesis, combining seed questions via diverse principles like bridging, visual-textual alignment, to create multi-step problems demanding deeper reasoning; and (3) Thinking-Induced Curriculum Fine-Tuning, a structured pipeline that progressively trains the model from scaffolded reasoning to standalone inference. By leveraging the model's self-synthesis capability, MindGYM achieves high data efficiency (e.g., +16% gains on MathVision-Mini with only 400 samples), computational efficiency (reducing both training and inference costs), and robust generalization across tasks. Extensive evaluations on seven benchmarks demonstrate superior performance over strong baselines, with notable improvements (+15.77% win rates) in reasoning depth and breadth validated via GPT-based scoring. MindGYM underscores the viability of self-challenging for refining VLM capabilities while minimizing human intervention and resource demands. Code and data are released to advance multimodal reasoning research.

A world model is essential for an agent to predict the future and plan in domains such as autonomous driving and robotics. To achieve this, recent advancements have focused on video generation, which has gained significant attention due to the impressive success of diffusion models. However, these models require substantial computational resources. To address these challenges, we propose a world model leveraging object-centric representation space using slot attention, guided by language instructions. Our model perceives the current state as an object-centric representation and predicts future states in this representation space conditioned on natural language instructions. This approach results in a more compact and computationally efficient model compared to diffusion-based generative alternatives. Furthermore, it flexibly predicts future states based on language instructions, and offers a significant advantage in manipulation tasks where object recognition is crucial. In this paper, we demonstrate that our latent predictive world model surpasses generative world models in visuo-linguo-motor control tasks, achieving superior sample and computation efficiency. We also investigate the generalization performance of the proposed method and explore various strategies for predicting actions using object-centric representations. A world model, or world simulator, enables an agent to perceive the current environment and predict future environmental states. With the remarkable success of diffusion models, there has been a growing interest in employing video-generation-based world models, particularly those that are conditioned on the current frame and language instructions, to perform planning and control tasks (Du et al., 2024a;b; Yang et al., 2024). However, the major drawback of language-guided video-generation models is the requirement of large-scale labeled language-video datasets and the corresponding high computational cost (Gu et al., 2024).