arXiv subjects

Hexin Wang

Publications and source records attributed to Hexin Wang.

10 recordsLinked to original sources

arXiv2026

Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems

Artificial general intelligence (AGI) requires a method that can understand, generate, reason, and act across diverse domains and modalities. Rather than relying on a single model to master every capability, a promising path toward AGI now is to orchestrate a heterogeneous ecosystem of specialized models, tools, and policies (candidates) into a unified AI system. At the heart of this vision lies a fundamental question: how can an AI system dynamically select and coordinate the most capable candidate for each task while balancing performance and computational cost? In this work, we present Route Anything, a unified framework for studying model and policy routing across general-purpose AI systems. We develop a unified theoretical foundation that characterizes optimal routing under quality--cost trade-offs and establishes how the information available to a router determines its attainable performance. To systematically investigate routing across diverse capabilities, we first construct an omni-modal trajectory matrix spanning 25 benchmarks across language understanding, image, video, and speech generation, coding and workplace agents, and robotic manipulation, comprising 374,401 scored runs from 452 candidate configurations. Building on this framework, we train and test different routers to investigate the potential of omni-modal model routing, the capabilities of learned routers, and the conditions under which routing improves efficiency. Our results reveal substantial opportunities for capability orchestration: per-input oracle routing outperforms the best fixed candidate by 5.4 points, and learned routers outperform it by 0.3 points across 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robot manipulation while using 24% less policy time. More importantly, routing raises efficiency by 8.82x on average compared with the reference model.

cs.CV↗

arXiv2026

The Trinity: Capability-Driven Recursive Self-Improvement for Multimodal Agents

Multimodal intelligence can be viewed as human-guided co-evolution between worlds and agents through interaction. However, scaling this process faces two bottlenecks: constructing heterogeneous, verifiable environments is costly, and manual Harness iteration cannot keep pace with growing task diversity and complexity. A key insight is that tasks with different modalities and surface forms can share the same underlying capability requirements. We therefore introduce The Trinity, a recursive self-improvement framework coupling world generation with Harness evolution through three roles: Creator, Explorer, and Evolver. Drawing on verified execution trajectories, Evolver shifts the basis of task organization from surface categories to shared requirements, reorganizing the original task tree into a hierarchical capability tree. We introduce Capability-Driven Real2Sim2Real: for heterogeneous tasks at the same node, Creator abstracts their common requirements into simplified, verifiable simulated worlds for capability learning, while evaluation on original tasks measures transfer. Along this capability tree, Explorer progressively evolves reusable capability modules through selective inheritance, forming a Harness tree whose validated updates accumulate across iterations within a fixed topology. The resulting Harnesses specialize in different capabilities, motivating a Capability-Aware Harness Router that uses public task information to select compatible module configurations from the tree. Our evolved Harnesses improve mean performance by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks.

cs.AI↗

arXiv2026

REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control

Foundation models are becoming remarkably reliable at high-level reasoning, yet they remain too slow and costly to serve as fine-grained controllers. Conversely, fast policies like VLAs and whole-body controllers offer more efficient reflexes, but struggle to generalize. While pairing a deliberative brain (foundation models) with a reactive cerebellum (fast policies) seems natural for embodied systems, two obstacles remain: methods are built and scored on incompatible stacks, and their asynchronous coordination is difficult. For example, an agent that calls the VLA in the agentic loop can stall the fast execution, while a VLA that calls the agent on errors tends to re-trigger fallbacks repeatedly or completely miss silent failures. To address this, we propose REA (Routed Embodied Agent), a unified runtime where a deliberative System 2, an embodied router System 1 and reactive System 0 policies cooperate under a single golden rule - no running policy waits on the brain. Instead of querying the brain at each failure, REA employs a route memory that stores each policy's measured success ahead of time, a boundary-triggered judge gate that consults the brain on policy execution, and a shared-budget escalation ladder for bounded recovery. Evaluated on 3,000 test episodes across 60 LIBERO-Pro tasks, REA reaches 95.9% success and beats the best single-agent baseline in its own policy pool (94.7%) while reducing control steps by 25% and control time by 17%. In a separate 100-episode RoboTwin 2.0 study, REA achieves 84% success with VLA policies and 86% with VLA+code, compared with 80% for the strongest single VLA baseline MEDex. These results demonstrate learned and programmatic policies operating within the same runtime. REA also transfers to BEHAVIOR-1K and a real Unitree G1 humanoid, achieving the highest observed success among the compared methods.

cs.RO↗

arXiv2026

Defect-controlled twin activation in crystallographically equivalent magnesium micropillars

Tensile twinning plays a central role in accommodating -axis plasticity in Mg. In bulk Mg, it typically shows a relatively deterministic response with a low critical stress, whereas in confined volumes it exhibits broad yield-stress distributions that complicate the prediction of small-scale mechanical behavior. Here, site-specific compression tests are performed on 4 $μ$m-diameter pillars fabricated in a parent Mg crystal and an adjacent {10-12} twin. The two regions share the same [11-20] compression axis but experienced different prior deformation histories, allowing the influence of the residual microstructural state to be examined at fixed crystallographic orientation. Among 27 pillars, most parent-region pillars yield near 300 MPa, whereas pillars from the twin region span approximately 30 to 300 MPa. Interrupted tests combined with cross-sectional EBSD link individual load drops to discrete twin formation and further show that a pillar containing a pre-existing twin yields at approximately 80 MPa through the migration of the existing twin boundary. Molecular dynamics simulations of 30 nm-diameter pillars resolve possible atomistic pathways at the nanoscale. The simulations illustrate how contact geometry and pre-existing twin embryos alter event selection, and how an activated twin advances rapidly, while coherent twin boundary migration proceeds through disconnection motion accompanied by crystallographically required atomic shuffles. The results attribute the experimental scatter to the local availability of embryos and mobile interfaces, such that the first plastic event is governed by the twinning pathway accessible from the local microstructural state rather than by a single characteristic critical stress. Deformation history can therefore strongly modify the distribution of first plastic events even when the loading orientation is fixed.

cond-mat.mtrl-sci↗

arXiv2026

Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models

A common strategy for scaling world models is to train on more crawled video with more compute. We argue that this strategy is inefficient: scaling world models also requires a recursive data engine that offers grounded reward signals. The success of code agents illustrates why this matters. As code is executable, compilers and runtimes can provide high-quality rewards for Reinforcement Learning (RL) post-training of LLMs. By contrast, spatial generation still relies largely on fuzzy proxies such as CLIP scores. These signals are fuzzy and biased, making them hard to support RL post-training. Compared with these, game development provides a missing reward environment for spatial world models. A scene encoded by a game engine is an executable world specification: the engine can efficiently check collision, physics, navigability and bounded playability, while the developer provides the global verification signal by judging whether the scene should be accepted. Game development also provides real-world long-horizon trajectory data for RL post-training. We therefore propose Reinforcement Learning with Human-Engine Verification (RLHEV), a post-training paradigm that combines dense engine signals with implicit human acceptance feedback from the development process.

cs.AI↗

arXiv2026

Pretraining Large Language Models with NVFP4

Large Language Models (LLMs) today are powerful problem solvers across many domains, and they continue to get stronger as they scale in model size, training set size, and training set quality, as shown by extensive research and experimentation across the industry. Training a frontier model today requires on the order of tens to hundreds of yottaflops, which is a massive investment of time, compute, and energy. Improving pretraining efficiency is therefore essential to enable the next generation of even more capable LLMs. While 8-bit floating point (FP8) training is now widely adopted, transitioning to even narrower precision, such as 4-bit floating point (FP4), could unlock additional improvements in computational speed and resource utilization. However, quantization at this level poses challenges to training stability, convergence, and implementation, notably for large-scale models trained on long token horizons. In this study, we introduce a novel approach for stable and accurate training of large language models (LLMs) using the NVFP4 format. Our method integrates Random Hadamard transforms (RHT) to bound block-level outliers, employs a two-dimensional quantization scheme for consistent representations across both the forward and backward passes, utilizes stochastic rounding for unbiased gradient estimation, and incorporates selective high-precision layers. We validate our approach by training a 12-billion-parameter model on 10 trillion tokens -- the longest publicly documented training run in 4-bit precision to date. Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline. These findings highlight that NVFP4, when combined with our training approach, represents a major step forward in narrow-precision LLM training algorithms.

cs.CL↗

arXiv2025

NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model

We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.

cs.CL↗

arXiv2025

Solute Co-Segregation Mechanisms at Low-Angle Grain Boundaries in Magnesium: A Combined Atomic-Scale Experimental and Modeling Study

Solute segregation at low-angle grain boundaries (LAGBs) critically affects the microstructure and mechanical properties of magnesium (Mg) alloys. In modern alloys containing multiple substitutional elements, understanding solute-solute interactions at microstructural defects becomes essential for alloy design. This study investigates the co-segregation mechanisms of calcium (Ca), zinc (Zn), and aluminum (Al) at a LAGB in a dilute AZX010 Mg alloy by combining atomic-scale experimental and modeling techniques. Three-dimensional atom probe tomography (3D-APT) revealed significant segregation of Ca, Zn, and Al at the LAGB, with Ca forming linear segregation patterns along dislocation arrays characteristic of the LAGB. Clustering analysis showed increased Ca-Ca pairs at the boundary, indicating synergistic solute interactions. Atomistic simulations and elastic dipole calculations demonstrated that larger Ca atoms prefer tensile regions around dislocations, while smaller Zn and Al atoms favor compressive areas. These simulations also found that Ca-Ca co-segregation near dislocation cores is energetically more favorable than other solute pairings, explaining the enhanced Ca clustering observed experimentally. Thermodynamic modeling incorporating calculated segregation energies and solute-solute interactions accurately predicted solute concentrations at the LAGB, aligning with experimental data. The findings emphasize the importance of solute interactions at dislocation cores in Mg alloys, offering insights for improving mechanical performance through targeted alloying and grain boundary engineering.

cond-mat.mtrl-sci↗

arXiv2024

Grain boundary segregation spectrum in basal-textured Mg alloys: From solute decoration to structural transition

Mg alloys are promising lightweight structural materials due to their low density and excellent mechanical properties. However, their limited formability and ductility necessitate improvements in these properties, specifically through texture modification via grain boundary segregation. While significant efforts have been made, the segregation behavior in Mg polycrystals, particularly with basal texture, remains largely unexplored. In this study, we performed atomistic simulations to investigate grain boundary segregation in dilute and concentrated solid solution Mg-Al alloys. We computed the segregation energy spectrum of basal-textured Mg polycrystals, highlighting the contribution from specific grain boundary sites, such as junctions, and identified a newly discovered bimodal distribution which is distinct compared to the conventional skew-normal distribution found in randomly-oriented polycrystals. Using a hybrid molecular dynamics/Monte Carlo approach, we simulated segregation behavior at finite temperatures, identifying grain boundary structural transitions, particularly the varied fraction and morphology of topologically close-packed grain boundary phases when changing thermodynamic variables. The outcomes of this study offer crucial insights into basal-textured grain boundary segregation and phase formation, which can be extended to other relevant Mg alloys containing topologically close-packed intermetallics.

cond-mat.mtrl-sci↗

arXiv2023

Preliminary Design of CSNS-II Linac SRF LLRF

China Spallation Neutron Source(CSNS) target power will upgrade to 500 kW(CSNS-II) from 300kW, energy gain of H-Linac will up to 300 MeV from 80 MeV using about 50 superconductor cavities. LLRF is an important device for controlling the amplitude and phase of the SRF cavity field to be less than 0.6% and 0.6 deg. The parameters and requirements for CSNS-II Linac LLRF are presented here. The preliminary design work and algorithm verification progress and results at C-ADS Injector-I are introduced.

physics.acc-ph↗