The recent development of reasoning language models (RLMs) represents a novel evolution in large language models. In particular, the recent release of DeepSeek-R1 has generated widespread social impact and sparked enthusiasm in the research community for exploring the explicit reasoning paradigm of language models. However, the implementation details of the released models have not been fully open-sourced by DeepSeek, including DeepSeek-R1-Zero, DeepSeek-R1, and the distilled small models. As a result, many replication studies have emerged aiming to reproduce the strong performance achieved by DeepSeek-R1, reaching comparable performance through similar training procedures and fully open-source data resources. These works have investigated feasible strategies for supervised fine-tuning (SFT) and reinforcement learning from verifiable rewards (RLVR), focusing on data preparation and method design, yielding various valuable insights. In this report, we provide a summary of recent replication studies to inspire future research. We primarily focus on SFT and RLVR as two main directions, introducing the details for data construction, method design and training procedure of current replication studies. Moreover, we conclude key findings from the implementation details and experimental results reported by these studies, anticipating to inspire future research. We also discuss additional techniques of enhancing RLMs, highlighting the potential of expanding the application scope of these models, and discussing the challenges in development. By this survey, we aim to help researchers and developers of RLMs stay updated with the latest advancements, and seek to inspire new ideas to further enhance RLMs.

Long chain-of-thought (CoT) is an essential ingredient in effective usage of modern large language models, but our understanding of the reasoning strategies underlying these capabilities remains limited. While some prior works have attempted to categorize CoTs using predefined strategy types, such approaches are constrained by human intuition and fail to capture the full diversity of model behaviors. In this work, we introduce the CoT Encyclopedia, a bottom-up framework for analyzing and steering model reasoning. Our method automatically extracts diverse reasoning criteria from model-generated CoTs, embeds them into a semantic space, clusters them into representative categories, and derives contrastive rubrics to interpret reasoning behavior. Human evaluations show that this framework produces more interpretable and comprehensive analyses than existing methods. Moreover, we demonstrate that this understanding enables performance gains: we can predict which strategy a model is likely to use and guide it toward more effective alternatives. Finally, we provide practical insights, such as that training data format (e.g., free-form vs. multiple-choice) has a far greater impact on reasoning behavior than data domain, underscoring the importance of format-aware model design.

A central challenge in reinforcement learning (RL) is its dependence on extensive real-world interaction data to learn task-specific policies. While recent work demonstrates that large language models (LLMs) can mitigate this limitation by generating synthetic experience (noted as imaginary rollouts) for mastering novel tasks, progress in this emerging field is hindered due to the lack of a standard benchmark. To bridge this gap, we introduce ImagineBench, the first comprehensive benchmark for evaluating offline RL algorithms that leverage both real rollouts and LLM-imaginary rollouts. The key features of ImagineBench include: (1) datasets comprising environment-collected and LLM-imaginary rollouts; (2) diverse domains of environments covering locomotion, robotic manipulation, and navigation tasks; and (3) natural language task instructions with varying complexity levels to facilitate language-conditioned policy learning. Through systematic evaluation of state-of-the-art offline RL algorithms, we observe that simply applying existing offline RL algorithms leads to suboptimal performance on unseen tasks, achieving 35.44% success rate in hard tasks in contrast to 64.37% of method training on real rollouts for hard tasks. This result highlights the need for algorithm advancements to better leverage LLM-imaginary rollouts. Additionally, we identify key opportunities for future research: including better utilization of imaginary rollouts, fast online adaptation and continual learning, and extension to multi-modal tasks. Our code is publicly available at https://github.com/LAMDA-RL/ImagineBench.

Depth estimation and scene segmentation are two important tasks in intelligent transportation systems. A joint modeling of these two tasks will reduce the requirement for both the storage and training efforts. This work explores how the multi-task distillation could be used to improve such unified modeling. While existing solutions transfer multiple teachers' knowledge in a static way, we propose a self-adaptive distillation method that can dynamically adjust the knowledge amount from each teacher according to the student's current learning ability. Furthermore, as multiple teachers exist, the student's gradient update direction in the distillation is more prone to be erroneous where knowledge forgetting may occur. To avoid this, we propose a knowledge trajectory to record the most essential information that a model has learnt in the past, based on which a trajectory-based distillation loss is designed to guide the student to follow the learning curve similarly in a cost-effective way. We evaluate our method on multiple benchmarking datasets including Cityscapes and NYU-v2. Compared to the state-of-the-art solutions, our method achieves a clearly improvement. The code is provided in the supplementary materials.

Questioning has become increasingly crucial for both humans and artificial intelligence, yet there remains limited research comprehensively assessing question quality. In response, this study defines good questions and presents a systematic evaluation framework. We propose two key evaluation dimensions: appropriateness (sociolinguistic competence in context) and effectiveness (strategic competence in goal achievement). Based on these foundational dimensions, a rubric-based scoring system was developed. By incorporating dynamic contextual variables, our evaluation framework achieves structure and flexibility through semi-adaptive criteria. The methodology was validated using the CAUS and SQUARE datasets, demonstrating the ability of the framework to access both well-formed and problematic questions while adapting to varied contexts. As we establish a flexible and comprehensive framework for question evaluation, this study takes a significant step toward integrating questioning behavior with structured analytical methods grounded in the intrinsic nature of questioning.

This paper introduces an interactive continual learning paradigm where AI models dynamically learn new skills from real-time human feedback while retaining prior knowledge. This paradigm distinctively addresses two major limitations of traditional continual learning: (1) dynamic model updates using streaming, real-time human-annotated data, rather than static datasets with fixed labels, and (2) the assumption of clean labels, by explicitly handling the noisy feedback common in real-world interactions. To tackle these problems, we propose RiCL, a Reinforced interactive Continual Learning framework leveraging Large Language Models (LLMs) to learn new skills effectively from dynamic feedback. RiCL incorporates three key components: a temporal consistency-aware purifier to automatically discern clean from noisy samples in data streams; an interaction-aware direct preference optimization strategy to align model behavior with human intent by reconciling AI-generated and human-provided feedback; and a noise-resistant contrastive learning module that captures robust representations by exploiting inherent data relationships, thus avoiding reliance on potentially unreliable labels. Extensive experiments on two benchmark datasets (FewRel and TACRED), contaminated with realistic noise patterns, demonstrate that our RiCL approach substantially outperforms existing combinations of state-of-the-art online continual learning and noisy-label learning methods.

Modern robots face challenges shared by humans, where machines must learn multiple sensorimotor skills and express them adaptively. Equipping robots with a human-like memory of how it feels to do multiple stereotypical movements can make robots more aware of normal operational states and help develop self-preserving safer robots. Associative Skill Memories (ASMs) aim to address this by linking movement primitives to sensory feedback, but existing implementations rely on hard-coded libraries of individual skills. A key unresolved problem is how a single neural network can learn a repertoire of skills while enabling fault detection and context-aware execution. Here we introduce Neural Associative Skill Memories (ASMs), a framework that utilises self-supervised predictive coding for temporal prediction to unify skill learning and expression, using biologically plausible learning rules. Unlike traditional ASMs which require explicit skill selection, Neural ASMs implicitly recognize and express skills through contextual inference, enabling fault detection across learned behaviours without an explicit skill selection mechanism. Compared to recurrent neural networks trained via backpropagation through time, our model achieves comparable qualitative performance in skill memory expression while using local learning rules and predicts a biologically relevant speed-accuracy trade-off during skill memory expression. This work advances the field of neurorobotics by demonstrating how predictive coding principles can model adaptive robot control and human motor preparation. By unifying fault detection, reactive control, skill memorisation and expression into a single energy-based architecture, Neural ASMs contribute to safer robotics and provide a computational lens to study biological sensorimotor learning.

Vision language models (VLMs) exhibit vast knowledge of the physical world, including intuition of physical and spatial properties, affordances, and motion. With fine-tuning, VLMs can also natively produce robot trajectories. We demonstrate that eliciting wrenches, not trajectories, allows VLMs to explicitly reason about forces and leads to zero-shot generalization in a series of manipulation tasks without pretraining. We achieve this by overlaying a consistent visual representation of relevant coordinate frames on robot-attached camera images to augment our query. First, we show how this addition enables a versatile motion control framework evaluated across four tasks (opening and closing a lid, pushing a cup or chair) spanning prismatic and rotational motion, an order of force and position magnitude, different camera perspectives, annotation schemes, and two robot platforms over 220 experiments, resulting in 51% success across the four tasks. Then, we demonstrate that the proposed framework enables VLMs to continually reason about interaction feedback to recover from task failure or incompletion, with and without human supervision. Finally, we observe that prompting schemes with visual annotation and embodied reasoning can bypass VLM safeguards. We characterize prompt component contribution to harmful behavior elicitation and discuss its implications for developing embodied reasoning. Our code, videos, and data are available at: https://scalingforce.github.io/.

This study addresses the critical challenge of hallucination mitigation in Large Vision-Language Models (LVLMs) for Visual Question Answering (VQA) tasks through a Split Conformal Prediction (SCP) framework. While LVLMs excel in multi-modal reasoning, their outputs often exhibit hallucinated content with high confidence, posing risks in safety-critical applications. We propose a model-agnostic uncertainty quantification method that integrates dynamic threshold calibration and cross-modal consistency verification. By partitioning data into calibration and test sets, the framework computes nonconformity scores to construct prediction sets with statistical guarantees under user-defined risk levels ($α$). Key innovations include: (1) rigorous control of \textbf{marginal coverage} to ensure empirical error rates remain strictly below $α$; (2) dynamic adjustment of prediction set sizes inversely with $α$, filtering low-confidence outputs; (3) elimination of prior distribution assumptions and retraining requirements. Evaluations on benchmarks (ScienceQA, MMMU) with eight LVLMs demonstrate that SCP enforces theoretical guarantees across all $α$ values. The framework achieves stable performance across varying calibration-to-test split ratios, underscoring its robustness for real-world deployment in healthcare, autonomous systems, and other safety-sensitive domains. This work bridges the gap between theoretical reliability and practical applicability in multi-modal AI systems, offering a scalable solution for hallucination detection and uncertainty-aware decision-making.

Natural language image-caption datasets, widely used for training Large Multimodal Models, mainly focus on natural scenarios and overlook the intricate details of mathematical figures that are critical for problem-solving, hindering the advancement of current LMMs in multimodal mathematical reasoning. To this end, we propose leveraging code as supervision for cross-modal alignment, since code inherently encodes all information needed to generate corresponding figures, establishing a precise connection between the two modalities. Specifically, we co-develop our image-to-code model and dataset with model-in-the-loop approach, resulting in an image-to-code model, FigCodifier and ImgCode-8.6M dataset, the largest image-code dataset to date. Furthermore, we utilize FigCodifier to synthesize novel mathematical figures and then construct MM-MathInstruct-3M, a high-quality multimodal math instruction fine-tuning dataset. Finally, we present MathCoder-VL, trained with ImgCode-8.6M for cross-modal alignment and subsequently fine-tuned on MM-MathInstruct-3M for multimodal math problem solving. Our model achieves a new open-source SOTA across all six metrics. Notably, it surpasses GPT-4o and Claude 3.5 Sonnet in the geometry problem-solving subset of MathVista, achieving improvements of 8.9% and 9.2%. The dataset and models will be released at https://github.com/mathllm/MathCoder.