TREX: Automating LLM Fine-tuning via Agent-Driven Tree-based Exploration

While Large Language Models (LLMs) have empowered AI research agents to perform isolated scientific tasks, automating complex, real-world workflows, such as LLM training, remains a significant challenge. In this paper, we introduce TREX, a multi-agent system that automates the entire LLM training life-cycle. By orchestrating collaboration between two core modules-the Researcher and the Executor-the system seamlessly performs requirement analysis, open-domain literature and data research, formulation of training strategies, preparation of data recipes, and model training and evaluation. The multi-round experimental process is modeled as a search tree, enabling the system to efficiently plan exploration paths, reuse historical results, and distill high-level insights from iterative trials. To evaluate the capability of automated LLM training, we construct FT-Bench, a benchmark comprising 10 tasks derived from real-world scenarios, ranging from optimizing fundamental model capabilities to enhancing performance on domain-specific tasks. Experimental results demonstrate that the TREX agent consistently optimizes model performance on target tasks.

Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization

We present Kernel-Smith, a framework for high-performance GPU kernel and operator generation that combines a stable evaluation-driven evolutionary agent with an evolution-oriented post-training recipe. On the agent side, Kernel-Smith maintains a population of executable candidates and iteratively improves them using an archive of top-performing and diverse programs together with structured execution feedback on compilation, correctness, and speedup. To make this search reliable, we build backend-specific evaluation services for Triton on NVIDIA GPUs and Maca on MetaX GPUs. On the training side, we convert long-horizon evolution trajectories into step-centric supervision and reinforcement learning signals by retaining correctness-preserving, high-gain revisions, so that the model is optimized as a strong local improver inside the evolutionary loop rather than as a one-shot generator. Under a unified evolutionary protocol, Kernel-Smith-235B-RL achieves state-of-the-art overall performance on KernelBench with Nvidia Triton backend, attaining the best average speedup ratio and outperforming frontier proprietary models including Gemini-3.0-pro and Claude-4.6-opus. We further validate the framework on the MetaX MACA backend, where our Kernel-Smith-MACA-30B surpasses large-scale counterparts such as DeepSeek-V3.2-think and Qwen3-235B-2507-think, highlighting potential for seamless adaptation across heterogeneous platforms. Beyond benchmark results, the same workflow produces upstream contributions to production systems including SGLang and LMDeploy, demonstrating that LLM-driven kernel optimization can transfer from controlled evaluation to practical deployment.

Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale

We introduce Intern-S1-Pro, the first one-trillion-parameter scientific multimodal foundation model. Scaling to this unprecedented size, the model delivers a comprehensive enhancement across both general and scientific domains. Beyond stronger reasoning and image-text understanding capabilities, its intelligence is augmented with advanced agent capabilities. Simultaneously, its scientific expertise has been vastly expanded to master over 100 specialized tasks across critical science fields, including chemistry, materials, life sciences, and earth sciences. Achieving this massive scale is made possible by the robust infrastructure support of XTuner and LMDeploy, which facilitates highly efficient Reinforcement Learning (RL) training at the 1-trillion parameter level while ensuring strict precision consistency between training and inference. By seamlessly integrating these advancements, Intern-S1-Pro further fortifies the fusion of general and specialized intelligence, working as a Specializable Generalist, demonstrating its position in the top tier of open-source models for general capabilities, while outperforming proprietary models in the depth of specialized scientific tasks.

CoFL: Continuous Flow Fields for Language-Conditioned Navigation

Language-conditioned navigation pipelines often rely on brittle modular components or costly action-sequence generation. To address these limitations, we present CoFL, an end-to-end policy that directly maps a bird's-eye view (BEV) observation and a language instruction to a continuous flow field for navigation. Instead of predicting discrete action tokens or sampling action chunks via iterative denoising, CoFL outputs instantaneous velocities that can be queried at arbitrary 2D projected locations. Trajectories are obtained by numerical integration of the predicted field, producing smooth motion that remains reactive under closed-loop execution. To enable large-scale training, we build a dataset of over 500k BEV image-instruction pairs, each procedurally annotated with a flow field and a trajectory derived from BEV semantic maps built on Matterport3D and ScanNet. By training on a mixed distribution, CoFL significantly outperforms modular Vision-Language Model (VLM)-based planners and generative policy baselines on strictly unseen scenes. Finally, we deploy CoFL zero-shot in real-world experiments with overhead BEV observations across multiple layouts, maintaining reliable closed-loop control and a high success rate.

Design, Modeling and Direction Control of a Wire-Driven Robotic Fish Based on a 2-DoF Crank-Slider Mechanism

Robotic fish have attracted growing attention in recent years owing to their biomimetic design and potential applications in environmental monitoring and biological surveys. Among robotic fish employing the Body-Caudal Fin (BCF) locomotion pattern, motor-driven actuation is widely adopted. Some approaches utilize multiple servo motors to achieve precise body curvature control, while others employ a brushless motor to drive the tail via wire or rod, enabling higher oscillation and swimming speeds. However, the former approaches typically result in limited swimming speed, whereas the latter suffer from poor maneuverability, with few capable of smooth turning. To address this trade-off, we develop a wire-driven robotic fish equipped with a 2-degree-of-freedom (DoF) crank-slider mechanism that decouples propulsion from steering, enabling both high swimming speed and agile maneuvering. In this paper, we first present the design of the robotic fish, including the elastic skeleton, waterproof structure, and the actuation mechanism that realizes the decoupling. We then establish the actuation modeling and body dynamics to analyze the locomotion behavior. Furthermore, we propose a combined feedforward-feedback control strategy to achieve independent regulation of propulsion and steering. Finally, we validate the feasibility of the design, modeling, and control through a series of prototype experiments, demonstrating swimming, turning, and directional control.

Hierarchical Trajectory Planning of Floating-Base Multi-Link Robot for Maneuvering in Confined Environments

Floating-base multi-link robots can change their shape during flight, making them well-suited for applications in confined environments such as autonomous inspection and search and rescue. However, trajectory planning for such systems remains an open challenge because the problem lies in a high-dimensional, constraint-rich space where collision avoidance must be addressed together with kinematic limits and dynamic feasibility. This work introduces a hierarchical trajectory planning framework that integrates global guidance with configuration-aware local optimization. First, we exploit the dual nature of these robots - the root link as a rigid body for guidance and the articulated joints for flexibility - to generate global anchor states that decompose the planning problem into tractable segments. Second, we design a local trajectory planner that optimizes each segment in parallel with differentiable objectives and constraints, systematically enforcing kinematic feasibility and maintaining dynamic feasibility by avoiding control singularities. Third, we implement a complete system that directly processes point-cloud data, eliminating the need for handcrafted obstacle models. Extensive simulations and real-world experiments confirm that this framework enables an articulated aerial robot to exploit its morphology for maneuvering that rigid robots cannot achieve. To the best of our knowledge, this is the first planning framework for floating-base multi-link robots that has been demonstrated on a real robot to generate continuous, collision-free, and dynamically feasible trajectories directly from raw point-cloud inputs, without relying on handcrafted obstacle models.

Learning Agile and Robust Omnidirectional Aerial Motion on Overactuated Tiltable-Quadrotors

Tilt-rotor aerial robots enable omnidirectional maneuvering through thrust vectoring, but introduce significant control challenges due to the strong coupling between joint and rotor dynamics. While model-based controllers can achieve high motion accuracy under nominal conditions, their robustness and responsiveness often degrade in the presence of disturbances and modeling uncertainties. This work investigates reinforcement learning for omnidirectional aerial motion control on over-actuated tiltable quadrotors that prioritizes robustness and agility. We present a learning-based control framework that enables efficient acquisition of coordinated rotor-joint behaviors for reaching target poses in the $SE(3)$ space. To achieve reliable sim-to-real transfer while preserving motion accuracy, we integrate system identification with minimal and physically consistent domain randomization. Compared with a state-of-the-art NMPC controller, the proposed method achieves comparable six-degree-of-freedom pose tracking accuracy, while demonstrating superior robustness and generalization across diverse tasks, enabling zero-shot deployment on real hardware.

DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning

In the current landscape of Large Language Models (LLMs), the curation of large-scale, high-quality training data is a primary driver of model performance. A key lever is the \emph{data recipe}, which comprises a data processing pipeline to transform raw sources into training corpora. Despite the growing use of LLMs to automate individual data processing steps, such as data synthesis and filtering, the overall design of data recipes remains largely manual and labor-intensive, requiring substantial human expertise and iteration. To bridge this gap, we formulate \emph{end-to-end data recipe generation} for LLM adaptation. Given a target benchmark and a pool of available data sources, a model is required to output a complete data recipe that adapts a base LLM to the target task. We present DataChef-32B, which performs online reinforcement learning using a proxy reward that predicts downstream performance for candidate recipes. Across six held-out tasks, DataChef-32B produces practical recipes that reach comparable downstream performance to those curated by human experts. Notably, the recipe from DataChef-32B adapts Qwen3-1.7B-Base to the math domain, achieving 66.7 on AIME'25 and surpassing Qwen3-1.7B. This work sheds new light on automating LLM training and developing self-evolving AI systems.

From Failure to Mastery: Generating Hard Samples for Tool-use Agents

The advancement of LLM agents with tool-use capabilities requires diverse and complex training corpora. Existing data generation methods, which predominantly follow a paradigm of random sampling and shallow generation, often yield simple and homogeneous trajectories that fail to capture complex, implicit logical dependencies. To bridge this gap, we introduce HardGen, an automatic agentic pipeline designed to generate hard tool-use training samples with verifiable reasoning. Firstly, HardGen establishes a dynamic API Graph built upon agent failure cases, from which it samples to synthesize hard traces. Secondly, these traces serve as conditional priors to guide the instantiation of modular, abstract advanced tools, which are subsequently leveraged to formulate hard queries. Finally, the advanced tools and hard queries enable the generation of verifiable complex Chain-of-Thought (CoT), with a closed-loop evaluation feedback steering the continuous refinement of the process. Extensive evaluations demonstrate that a 4B parameter model trained with our curated dataset achieves superior performance compared to several leading open-source and closed-source competitors (e.g., GPT-5.2, Gemini-3-Pro and Claude-Opus-4.5). Our code, models, and dataset will be open-sourced to facilitate future research.

Exploring Superior Function Calls via Reinforcement Learning

Function calling capabilities are crucial for deploying Large Language Models in real-world applications, yet current training approaches fail to develop robust reasoning strategies. Supervised fine-tuning produces models that rely on superficial pattern matching, while standard reinforcement learning methods struggle with the complex action space of structured function calls. We present a novel reinforcement learning framework designed to enhance group relative policy optimization through strategic entropy based exploration specifically tailored for function calling tasks. Our approach addresses three critical challenges in function calling: insufficient exploration during policy learning, lack of structured reasoning in chain-of-thought generation, and inadequate verification of parameter extraction. Our two-stage data preparation pipeline ensures high-quality training samples through iterative LLM evaluation and abstract syntax tree validation. Extensive experiments on the Berkeley Function Calling Leaderboard demonstrate that this framework achieves state-of-the-art performance among open-source models with 86.02\% overall accuracy, outperforming standard GRPO by up to 6\% on complex multi-function scenarios. Notably, our method shows particularly strong improvements on code-pretrained models, suggesting that structured language generation capabilities provide an advantageous starting point for reinforcement learning in function calling tasks. We will release all the code, models and dataset to benefit the community.