ContextGuard: Structured Self-Auditing for Context Learning in Language Models

Recent benchmarks reveal that despite strong reasoning capabilities, large language models (LLMs) still struggle to faithfully apply complex contextual knowledge. These failures are often not wholesale reasoning collapses: in context-rich tasks, models may follow the central reasoning path while missing peripheral, persistent, or format-sensitive requirements.

Context-CoT: Enhancing Context Learning via High-Quality Reasoning Synthesis

While LLMs excel at reasoning over prompts using static pretrained knowledge, they struggle significantly with context learning-the ability to dynamically extract, internalize, and apply new knowledge from complex, task-specific contexts. Recent evaluations on the CL-Bench reveal a critical capability gap: frontier models solve only 17.2% of context-dependent tasks on average.

VISD: Enhancing Video Reasoning via Structured Self-Distillation

Training VideoLLMs for complex reasoning remains challenging due to sparse sequence level rewards and the lack of fine grained credit assignment over long, temporally grounded reasoning trajectories. While reinforcement learning with verifiable rewards (RLVR) provides reliable supervision, it fails to capture token level contributions, leading to inefficient learning. Conversely, existing self distillation methods offer dense supervision but lack structure and diagnostic specificity, and often interact unstably with reinforcement learning. In this work, we propose VISD, a structured self distillation framework that introduces diagnostically meaningful privileged information for video reasoning. VISD employs a video aware judge model to decompose reasoning quality into multiple dimensions, including answer correctness, logical consistency, and spatio-temporal grounding, and uses this structured feedback to guide a teacher policy for token level supervision. To stably integrate dense supervision with RL, we introduce a direction magnitude decoupling mechanism, where rollout level advantages computed from rewards determine update direction, while structured privileged signals modulate token level update magnitudes. This design enables semantically aligned and fine grained credit assignment, improving both reasoning faithfulness and training efficiency. Additionally, VISD incorporates curriculum scheduling and EMA based teacher stabilization to support robust optimization over long video sequences. Experiments on diverse benchmarks show that VISD consistently outperforms strong baselines, improving answer accuracy and spatio temporal grounding quality. Notably, VISD reaches these gains with nearly 2x faster convergence in optimization steps, highlighting the effectiveness of structured self supervision in improving both performance and sample efficiency for VideoLLMs.

DGPO: Distribution Guided Policy Optimization for Fine Grained Credit Assignment

Reinforcement learning is crucial for aligning large language models to perform complex reasoning tasks. However, current algorithms such as Group Relative Policy Optimization suffer from coarse grained, sequence level credit assignment, which severely struggles to isolate pivotal reasoning steps within long Chain of Thought generations. Furthermore, the standard unbounded Kullback Leibler divergence penalty induces severe gradient instability and mode seeking conservatism, ultimately stifling the discovery of novel reasoning trajectories. To overcome these limitations, we introduce Distribution Guided Policy Optimization, a novel critic free reinforcement learning framework that reinterprets distribution deviation as a guiding signal rather than a rigid penalty.

HiMAC: Hierarchical Macro-Micro Learning for Long-Horizon LLM Agents

Large language model (LLM) agents have recently demonstrated strong capabilities in interactive decision-making, yet they remain fundamentally limited in long-horizon tasks that require structured planning and reliable execution. Existing approaches predominantly rely on flat autoregressive policies, where high-level reasoning and low-level actions are generated within a single token sequence, leading to inefficient exploration and severe error propagation over extended trajectories. In this work, we propose HiMAC, a hierarchical agentic RL framework that explicitly decomposes long-horizon decision-making into macro-level planning and micro-level execution. HiMAC models reasoning as a structured blueprint generation process followed by goal-conditioned action execution, enabling robust long-horizon planning within LLM-based agents. To train this hierarchy efficiently, we introduce a critic-free hierarchical policy optimization paradigm that extends group-based reinforcement learning to bi-level structures through hierarchical relative advantage estimation. Furthermore, we propose an iterative co-evolution training strategy that alternates between planner exploration and executor adaptation, mitigating the non-stationarity inherent in hierarchical learning. Extensive experiments on ALFWorld, WebShop, and Sokoban demonstrate that HiMAC consistently outperforms strong prompting and reinforcement learning baselines, achieving state-of-the-art performance and substantially improved sample efficiency across both text-based and visually grounded environments. Our results show that introducing structured hierarchy, rather than increasing model scale alone, is a key factor for enabling robust long-horizon agentic intelligence.

VideoCuRL: Video Curriculum Reinforcement Learning with Orthogonal Difficulty Decomposition

Reinforcement Learning (RL) is crucial for empowering VideoLLMs with complex spatiotemporal reasoning. However, current RL paradigms predominantly rely on random data shuffling or naive curriculum strategies based on scalar difficulty metrics. We argue that scalar metrics fail to disentangle two orthogonal challenges in video understanding: Visual Temporal Perception Load and Cognitive Reasoning Depth. To address this, we propose VideoCuRL, a novel framework that decomposes difficulty into these two axes. We employ efficient, training-free proxies, optical flow and keyframe entropy for visual complexity, Calibrated Surprisal for cognitive complexity, to map data onto a 2D curriculum grid. A competence aware Diagonal Wavefront strategy then schedules training from base alignment to complex reasoning. Furthermore, we introduce Dynamic Sparse KL and Structured Revisiting to stabilize training against reward collapse and catastrophic forgetting. Extensive experiments show that VideoCuRL surpasses strong RL baselines on reasoning (+2.5 on VSI-Bench) and perception (+2.9 on VideoMME) tasks. Notably, VideoCuRL eliminates the prohibitive inference overhead of generation-based curricula, offering a scalable solution for robust video post-training.

CoT-Vid: Dynamic Chain-of-Thought Routing with Self Verification for Training-Free Video Reasoning

System2 reasoning is developing rapidly these days with the emergence of Deep- Thinking Models and chain-of-thought technology, which has become a centralized discussion point in the AI community. However, there is a relative gap in the research on complex video reasoning at present. In this work, we propose CoT-Vid, a novel training-free paradigm for the video domain with a multistage complex reasoning design. Distinguishing from existing video LLMs, which rely heavily on perceptual abilities, it achieved surprising performance gain with explicit reasoning mechanism. The paradigm consists of three main components: dynamic inference path routing, problem decoupling strategy, and video self-consistency verification. In addition, we propose a new standard for categorization of video questions. CoT- Vid showed outstanding results on a wide range of benchmarks, and outperforms its base model by 9.3% on Egochema and 5.6% on VideoEspresso, rivalling or even surpassing larger and proprietary models, such as GPT-4V, GPT-4o and Gemini-1.5-flash. Our codebase will be publicly available soon.

TEOcc: Radar-camera Multi-modal Occupancy Prediction via Temporal Enhancement

As a novel 3D scene representation, semantic occupancy has gained much attention in autonomous driving. However, existing occupancy prediction methods mainly focus on designing better occupancy representations, such as tri-perspective view or neural radiance fields, while ignoring the advantages of using long-temporal information. In this paper, we propose a radar-camera multi-modal temporal enhanced occupancy prediction network, dubbed TEOcc. Our method is inspired by the success of utilizing temporal information in 3D object detection. Specifically, we introduce a temporal enhancement branch to learn temporal occupancy prediction. In this branch, we randomly discard the t-k input frame of the multi-view camera and predict its 3D occupancy by long-term and short-term temporal decoders separately with the information from other adjacent frames and multi-modal inputs. Besides, to reduce computational costs and incorporate multi-modal inputs, we specially designed 3D convolutional layers for long-term and short-term temporal decoders. Furthermore, since the lightweight occupancy prediction head is a dense classification head, we propose to use a shared occupancy prediction head for the temporal enhancement and main branches. It is worth noting that the temporal enhancement branch is only performed during training and is discarded during inference. Experiment results demonstrate that TEOcc achieves state-of-the-art occupancy prediction on nuScenes benchmarks. In addition, the proposed temporal enhancement branch is a plug-and-play module that can be easily integrated into existing occupancy prediction methods to improve the performance of occupancy prediction. The code and models will be released at https://github.com/VDIGPKU/TEOcc.