SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization

Agent skills, structured packages of procedural knowledge and executable resources that agents dynamically load at inference time, have become a reliable mechanism for augmenting LLM agents. Yet inference-time skill augmentation is fundamentally limited: retrieval noise introduces irrelevant guidance, injected skill content imposes substantial token overhead, and the model never truly acquires the knowledge it merely follows. We ask whether skills can instead be internalized into model parameters, enabling zero-shot autonomous behavior without any runtime skill retrieval. We introduce SKILL0, an in-context reinforcement learning framework designed for skill internalization. SKILL0 introduces a training-time curriculum that begins with full skill context and progressively withdraws it. Skills are grouped offline by category and rendered with interaction history into a compact visual context, teaching he model tool invocation and multi-turn task completion. A Dynamic Curriculum then evaluates each skill file's on-policy helpfulness, retaining only those from which the current policy still benefits within a linearly decaying budget, until the agent operates in a fully zero-shot setting. Extensive agentic experiments demonstrate that SKILL0 achieves substantial improvements over the standard RL baseline (+9.7\% for ALFWorld and +6.6\% for Search-QA), while maintaining a highly efficient context of fewer than 0.5k tokens per step. Our code is available at https://github.com/ZJU-REAL/SkillZero.

ReMemNav: A Rethinking and Memory-Augmented Framework for Zero-Shot Object Navigation

Zero-shot object navigation requires agents to locate unseen target objects in unfamiliar environments without prior maps or task-specific training which remains a significant challenge. Although recent advancements in vision-language models(VLMs) provide promising commonsense reasoning capabilities for this task, these models still suffer from spatial hallucinations, local exploration deadlocks, and a disconnect between high-level semantic intent and low-level control. In this regard, we propose a novel hierarchical navigation framework named ReMemNav, which seamlessly integrates panoramic semantic priors and episodic memory with VLMs. We introduce the Recognize Anything Model to anchor the spatial reasoning process of the VLM. We also design an adaptive dual-modal rethinking mechanism based on an episodic semantic buffer queue. The proposed mechanism actively verifies target visibility and corrects decisions using historical memory to prevent deadlocks. For low-level action execution, ReMemNav extracts a sequence of feasible actions using depth masks, allowing the VLM to select the optimal action for mapping into actual spatial movement. Extensive evaluations on HM3D and MP3D demonstrate that ReMemNav outperforms existing training-free zero-shot baselines in both success rate and exploration efficiency. Specifically, we achieve significant absolute performance improvements, with SR and SPL increasing by 1.7% and 7.0% on HM3D v0.1, 18.2% and 11.1% on HM3D v0.2, and 8.7% and 7.9% on MP3D.

Pose-Free Omnidirectional Gaussian Splatting for 360-Degree Videos with Consistent Depth Priors

Omnidirectional 3D Gaussian Splatting with panoramas is a key technique for 3D scene representation, and existing methods typically rely on slow SfM to provide camera poses and sparse points priors. In this work, we propose a pose-free omnidirectional 3DGS method, named PFGS360, that reconstructs 3D Gaussians from unposed omnidirectional videos. To achieve accurate camera pose estimation, we first construct a spherical consistency-aware pose estimation module, which recovers poses by establishing consistent 2D-3D correspondences between the reconstructed Gaussians and the unposed images using Gaussians' internal depth priors. Besides, to enhance the fidelity of novel view synthesis, we introduce a depth-inlier-aware densification module to extract depth inliers and Gaussian outliers with consistent monocular depth priors, enabling efficient Gaussian densification and achieving photorealistic novel view synthesis. The experiments show significant outperformance over existing pose-free and pose-aware 3DGS methods on both real-world and synthetic 360-degree videos. Code is available at https://github.com/zcq15/PFGS360.

CoVerRL: Breaking the Consensus Trap in Label-Free Reasoning via Generator-Verifier Co-Evolution

Label-free reinforcement learning enables large language models to improve reasoning capabilities without ground-truth supervision, typically by treating majority-voted answers as pseudo-labels. However, we identify a critical failure mode: as training maximizes self-consistency, output diversity collapses, causing the model to confidently reinforce systematic errors that evade detection. We term this the consensus trap. To escape it, we propose CoVerRL, a framework where a single model alternates between generator and verifier roles, with each capability bootstrapping the other. Majority voting provides noisy but informative supervision for training the verifier, while the improving verifier progressively filters self-consistent errors from pseudo-labels. This co-evolution creates a virtuous cycle that maintains high reward accuracy throughout training. Experiments across Qwen and Llama model families demonstrate that CoVerRL outperforms label-free baselines by 4.7-5.9\% on mathematical reasoning benchmarks. Moreover, self-verification accuracy improves from around 55\% to over 85\%, confirming that both capabilities genuinely co-evolve.

Code-A1: Adversarial Evolving of Code LLM and Test LLM via Reinforcement Learning

Reinforcement learning for code generation relies on verifiable rewards from unit test pass rates. Yet high-quality test suites are scarce, existing datasets offer limited coverage, and static rewards fail to adapt as models improve. Recent self-play methods unify code and test generation in a single model, but face a inherent dilemma: white-box access leads to self-collusion where the model produces trivial tests for easy rewards, yet black-box restriction yields generic tests that miss implementation-specific bugs. We introduce Code-A1, an adversarial co-evolution framework that jointly optimizes a Code LLM and a Test LLM with opposing objectives. The Code LLM is rewarded for passing more tests, while the Test LLM is rewarded for exposing more defects. This architectural separation eliminates self-collusion risks and safely enables white-box test generation, where the Test LLM can inspect candidate code to craft targeted adversarial tests. We further introduce a Mistake Book mechanism for experience replay and a composite reward balancing test validity with adversarial difficulty. Experiments on Qwen2.5-Coder models demonstrate that Code-A1 achieves code generation performance matching or exceeding models trained on human-annotated tests, while significantly improving test generation capability.

ReMix: Reinforcement routing for mixtures of LoRAs in LLM finetuning

Low-rank adapters (LoRAs) are a parameter-efficient finetuning technique that injects trainable low-rank matrices into pretrained models to adapt them to new tasks. Mixture-of-LoRAs models expand neural networks efficiently by routing each layer input to a small subset of specialized LoRAs of the layer. Existing Mixture-of-LoRAs routers assign a learned routing weight to each LoRA to enable end-to-end training of the router. Despite their empirical promise, we observe that the routing weights are typically extremely imbalanced across LoRAs in practice, where only one or two LoRAs often dominate the routing weights. This essentially limits the number of effective LoRAs and thus severely hinders the expressive power of existing Mixture-of-LoRAs models. In this work, we attribute this weakness to the nature of learnable routing weights and rethink the fundamental design of the router. To address this critical issue, we propose a new router designed that we call Reinforcement Routing for Mixture-of-LoRAs (ReMix). Our key idea is using non-learnable routing weights to ensure all active LoRAs to be equally effective, with no LoRA dominating the routing weights. However, our routers cannot be trained directly via gradient descent due to our non-learnable routing weights. Hence, we further propose an unbiased gradient estimator for the router by employing the reinforce leave-one-out (RLOO) technique, where we regard the supervision loss as the reward and the router as the policy in reinforcement learning. Our gradient estimator also enables to scale up training compute to boost the predictive performance of our ReMix. Extensive experiments demonstrate that our proposed ReMix significantly outperform state-of-the-art parameter-efficient finetuning methods under a comparable number of activated parameters.

Memory-Guided View Refinement for Dynamic Human-in-the-loop EQA

Embodied Question Answering (EQA) has traditionally been evaluated in temporally stable environments where visual evidence can be accumulated reliably. However, in dynamic, human-populated scenes, human activities and occlusions introduce significant perceptual non-stationarity: task-relevant cues are transient and view-dependent, while a store-then-retrieve strategy over-accumulates redundant evidence and increases inference cost. This setting exposes two practical challenges for EQA agents: resolving ambiguity caused by viewpoint-dependent occlusions, and maintaining compact yet up-to-date evidence for efficient inference. To enable systematic study of this setting, we introduce DynHiL-EQA, a human-in-the-loop EQA dataset with two subsets: a Dynamic subset featuring human activities and temporal changes, and a Static subset with temporally stable observations. To address the above challenges, we present DIVRR (Dynamic-Informed View Refinement and Relevance-guided Adaptive Memory Selection), a training-free framework that couples relevance-guided view refinement with selective memory admission. By verifying ambiguous observations before committing them and retaining only informative evidence, DIVRR improves robustness under occlusions while preserving fast inference with compact memory. Extensive experiments on DynHiL-EQA and the established HM-EQA dataset demonstrate that DIVRR consistently improves over existing baselines in both dynamic and static settings while maintaining high inference efficiency.

DREAM: Where Visual Understanding Meets Text-to-Image Generation

Unifying visual representation learning and text-to-image (T2I) generation within a single model remains a central challenge in multimodal learning. We introduce DREAM, a unified framework that jointly optimizes discriminative and generative objectives, while learning strong visual representations. DREAM is built on two key techniques: During training, Masking Warmup, a progressive masking schedule, begins with minimal masking to establish the contrastive alignment necessary for representation learning, then gradually transitions to full masking for stable generative training. At inference, DREAM employs Semantically Aligned Decoding to align partially masked image candidates with the target text and select the best one for further decoding, improving text-image fidelity (+6.3%) without external rerankers. Trained solely on CC12M, DREAM achieves 72.7% ImageNet linear-probing accuracy (+1.1% over CLIP) and an FID of 4.25 (+6.2% over FLUID), with consistent gains in few-shot classification, semantic segmentation, and depth estimation. These results demonstrate that discriminative and generative objectives can be synergistic, allowing unified multimodal models that excel at both visual understanding and generation.

CoMoL: Efficient Mixture of LoRA Experts via Dynamic Core Space Merging

Large language models (LLMs) achieve remarkable performance on diverse downstream and domain-specific tasks via parameter-efficient fine-tuning (PEFT). However, existing PEFT methods, particularly MoE-LoRA architectures, suffer from limited parameter efficiency and coarse-grained adaptation due to the proliferation of LoRA experts and instance-level routing. To address these issues, we propose Core Space Mixture of LoRA (\textbf{CoMoL}), a novel MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation. Specifically, CoMoL introduces two key components: core space experts and core space routing. Core space experts store each expert in a compact core matrix, preserving diversity while controlling parameter growth. Core space routing dynamically selects and activates the appropriate core experts for each token, enabling fine-grained, input-adaptive routing. Activated core experts are then merged via a soft-merging strategy into a single core expert, which is combined with a shared LoRA to form a specialized LoRA module. Besides, the routing network is projected into the same low-rank space as the LoRA matrices, further reducing parameter overhead without compromising expressiveness. Extensive experiments demonstrate that CoMoL retains the adaptability of MoE-LoRA architectures while achieving parameter efficiency comparable to standard LoRA, consistently outperforming existing methods across multiple tasks.

Path-Decoupled Hyperbolic Flow Matching for Few-Shot Adaptation

Recent advances in cross-modal few-shot adaptation treat visual-semantic alignment as a continuous feature transport problem via Flow Matching (FM). However, we argue that Euclidean-based FM overlooks fundamental limitations of flat geometry, where polynomial volume growth fails to accommodate diverse feature distributions, leading to severe path entanglement. To this end, we propose path-decoupled Hyperbolic Flow Matching (HFM), leveraging the Lorentz manifold's exponential expansion for trajectory decoupling. HFM structures the transport via two key designs: 1) Centripetal hyperbolic alignment: It constructs a centripetal hierarchy by anchoring textual roots, which pushes visual leaves to the boundary to initialize orderly flows. 2) Path-decoupled objective: It acts as a ``semantic guardrail'' rigidly confining trajectories within isolated class-specific geodesic corridors via step-wise supervision. Furthermore, we devise an adaptive diameter-based stopping to prevent over-transportation into the crowded origin based on the intrinsic semantic scale. Extensive ablations on 11 benchmarks have shown that HFM establishes a new state-of-the-art, consistently outperforming its Euclidean counterparts. Our codes and models will be released.