Searcharxiv⌕ Search

arXiv subjects

Hao Sun

Publications and source records attributed to Hao Sun.

At least 127 records · Page 7Linked to original sources
arXiv2026

Training Large Reasoning Models Efficiently via Progressive Thought Encoding

Large reasoning models (LRMs) excel on complex problems but face a critical barrier to efficiency: reinforcement learning (RL) training requires long rollouts for outcome-based rewards, where autoregressive decoding dominates time and memory usage. While sliding-window cache strategies can bound memory, they disrupt long-context reasoning and degrade performance. We introduce Progressive Thought Encoding, a parameter-efficient fine-tuning method that enables LRMs to reason effectively under fixed-size caches. By progressively encoding intermediate reasoning into fixed-size vector representations, our approach eliminates the need to backpropagate through full-cache rollouts, thereby reducing memory usage, while maintaining constant memory during inference. Experiments on three models, including Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct, and DeepSeek-R1-Distill-Llama-8B, on six widely used challenging mathematical benchmarks show consistent gains: our method achieves +19.3% improvement over LoRA-based fine-tuning and +29.9% over LRMs without fine-tuning on average, with up to +23.4 accuracy improvement on AIME2024/2025 under the same tight cache budgets. These results demonstrate that Progressive Thought Encoding not only improves reasoning accuracy but also makes RL training of LRMs substantially more efficient and scalable under real-world memory constraints.

cs.LG↗
arXiv2026

Improving Interactive In-Context Learning from Natural Language Feedback

Adapting one's thought process based on corrective feedback is an essential ability in human learning, particularly in collaborative settings. In contrast, the current large language model training paradigm relies heavily on modeling vast, static corpora. While effective for knowledge acquisition, it overlooks the interactive feedback loops essential for models to adapt dynamically to their context. In this work, we propose a framework that treats this interactive in-context learning ability not as an emergent property, but as a distinct, trainable skill. We introduce a scalable method that transforms single-turn verifiable tasks into multi-turn didactic interactions driven by information asymmetry. We first show that current flagship models struggle to integrate corrective feedback on hard reasoning tasks. We then demonstrate that models trained with our approach dramatically improve the ability to interactively learn from language feedback. More specifically, the multi-turn performance of a smaller model nearly reaches that of a model an order of magnitude larger. We also observe robust out-of-distribution generalization: interactive training on math problems transfers to diverse domains like coding, puzzles and maze navigation. Our qualitative analysis suggests that this improvement is due to an enhanced in-context plasticity. Finally, we show that this paradigm offers a unified path to self-improvement. By training the model to predict the teacher's critiques, effectively modeling the feedback environment, we convert this external signal into an internal capability, allowing the model to self-correct even without a teacher.

cs.AI↗
arXiv2026

RPG: A Repository Planning Graph for Unified and Scalable Codebase Generation

Large language models excel at generating individual functions or single files of code, yet generating complete repositories from scratch remains a fundamental challenge. This capability is key to building coherent software systems from high-level specifications and realizing the full potential of automated code generation. The process requires planning at two levels: deciding what features and modules to build (proposal stage) and defining their implementation details (implementation stage). Current approaches rely on natural language planning, which often produces unclear specifications, misaligned components, and brittle designs due to its inherent ambiguity and lack of structure. To address these limitations, we introduce the Repository Planning Graph (RPG), a structured representation that encodes capabilities, file structures, data flows, and functions in a unified graph. By replacing free-form natural language with an explicit blueprint, RPG enables consistent long-horizon planning for repository generation. Building on RPG, we develop ZeroRepo, a graph-driven framework that operates in three stages: proposal-level planning, implementation-level construction, and graph-guided code generation with test validation. To evaluate, we construct RepoCraft, a benchmark of six real-world projects with 1,052 tasks. On RepoCraft, ZeroRepo produces nearly 36K Code Lines and 445K Code Tokens, on average 3.9$\times$ larger than the strongest baseline (Claude Code), and 68$\times$ larger than other baselines. It achieves 81.5% coverage and 69.7% test accuracy, improving over Claude Code by 27.3 and 35.8 points. Further analysis shows that RPG models complex dependencies, enables more sophisticated planning through near-linear scaling, and improves agent understanding of repositories, thus accelerating localization. Our data and code are available at https://github.com/microsoft/RPG-ZeroRepo.

cs.CL↗
arXiv2026

HiST-VLA: A Hierarchical Spatio-Temporal Vision-Language-Action Model for End-to-End Autonomous Driving

Vision-Language-Action (VLA) models offer promising capabilities for autonomous driving through multimodal understanding. However, their utilization in safety-critical scenarios is constrained by inherent limitations, including imprecise numerical reasoning, weak 3D spatial awareness, and high sensitivity to context. To address these challenges, we propose HiST-VLA, a novel Hierarchical Spatio-Temporal VLA model designed for reliable trajectory generation. Our framework enhances 3D spatial and temporal reasoning by integrating geometric awareness with fine-grained driving commands and state history prompting. To ensure computational efficiency, we integrate dynamic token sparsification into the VLA architecture. This approach fuses redundant tokens rather than filtering them, effectively reducing redundancy without sacrificing model performance. Furthermore, we employ a hierarchical transformer-based planner to progressively refine coarse VLA waypoints into fine-grained trajectories. Crucially, the planner utilizes dynamic latent regularization to incorporate language commands, ensuring strict spatial grounding and temporal coherence. Extensive evaluation on the NAVSIM v2 benchmark demonstrates state-of-the-art performance on Navtest, achieving an EPDMS of 88.6, and EPDMS of 50.9 on pseudo closed-loop Navhard benchmark.

cs.CV↗
arXiv2026

Predictive Communications for Low-Altitude Networks

The emergence of dense, mission-driven aerial networks supporting the low-altitude economy presents unique communication challenges, including extreme channel dynamics and severe cross-tier interference. Traditional reactive communication paradigms are ill-suited to these environments, as they fail to leverage the network's inherent predictability. This paper introduces predictive communication, a novel paradigm transforming network management from reactive adaptation to proactive optimization. The approach is enabled by fusing predictable mission trajectories with stable, large-scale radio environment models (e.g., radio maps). Specifically, we present a hierarchical framework that decomposes the predictive cross-layer resource allocation problem into three layers: strategic (routing), tactical (timing), and operational (power). This structure aligns decision-making timescales with the accuracy levels and ranges of available predictive information. We demonstrate that this foresight-driven framework achieves an order-of-magnitude reduction in cross-tier interference, laying the groundwork for robust and scalable low-altitude communication systems.

eess.SP↗
arXiv2026

Geometric Image Editing via Effects-Sensitive In-Context Inpainting with Diffusion Transformers

Recent advances in diffusion models have significantly improved image editing. However, challenges persist in handling geometric transformations, such as translation, rotation, and scaling, particularly in complex scenes. Existing approaches suffer from two main limitations: (1) difficulty in achieving accurate geometric editing of object translation, rotation, and scaling; (2) inadequate modeling of intricate lighting and shadow effects, leading to unrealistic results. To address these issues, we propose GeoEdit, a framework that leverages in-context generation through a diffusion transformer module, which integrates geometric transformations for precise object edits. Moreover, we introduce Effects-Sensitive Attention, which enhances the modeling of intricate lighting and shadow effects for improved realism. To further support training, we construct RS-Objects, a large-scale geometric editing dataset containing over 120,000 high-quality image pairs, enabling the model to learn precise geometric editing while generating realistic lighting and shadows. Extensive experiments on public benchmarks demonstrate that GeoEdit consistently outperforms state-of-the-art methods in terms of visual quality, geometric accuracy, and realism.

cs.CV↗
arXiv2026

Physics-Aware Tensor Reconstruction for Radio Maps in Pixel-Based Fluid Antenna Systems

The deployment of pixel-based antennas and fluid antenna systems (FAS) is hindered by prohibitive channel state information (CSI) acquisition overhead. While radio maps enable proactive mode selection, reconstructing high-fidelity maps from sparse measurements is challenging. Existing physics-agnostic or data-driven methods often fail to recover fine-grained shadowing details under extreme sparsity. We propose a Physics-Regularized Low-Rank Tensor Completion (PR-LRTC) framework for radio map reconstruction. By modeling the signal field as a three-way tensor, we integrate environmental low-rankness with deterministic antenna physics. Specifically, we leverage Effective Aerial Degrees-of-Freedom (EADoF) theory to derive a differential gain topology map as a physical prior for regularization. The resulting optimization problem is solved via an efficient Alternating Direction Method of Multipliers (ADMM)-based algorithm. Simulations show that PR-LRTC achieves a 4 dB gain over baselines at a 10% sampling ratio. It effectively preserves sharp shadowing edges, providing a robust, physics-compliant solution for low-overhead beam management.

eess.SP↗
arXiv2026

Benchmarking neural surrogates on realistic spatiotemporal multiphysics flows

Predicting multiphysics dynamics is computationally expensive and challenging due to the severe coupling of multi-scale, heterogeneous physical processes. While neural surrogates promise a paradigm shift, the field currently suffers from an "illusion of mastery", as repeatedly emphasized in top-tier commentaries: existing evaluations overly rely on simplified, low-dimensional proxies, which fail to expose the models' inherent fragility in realistic regimes. To bridge this critical gap, we present REALM (REalistic AI Learning for Multiphysics), a rigorous benchmarking framework designed to test neural surrogates on challenging, application-driven reactive flows. REALM features 11 high-fidelity datasets spanning from canonical multiphysics problems to complex propulsion and fire safety scenarios, alongside a standardized end-to-end training and evaluation protocol that incorporates multiphysics-aware preprocessing and a robust rollout strategy. Using this framework, we systematically benchmark over a dozen representative surrogate model families, including spectral operators, convolutional models, Transformers, pointwise operators, and graph/mesh networks, and identify three robust trends: (i) a scaling barrier governed jointly by dimensionality, stiffness, and mesh irregularity, leading to rapidly growing rollout errors; (ii) performance primarily controlled by architectural inductive biases rather than parameter count; and (iii) a persistent gap between nominal accuracy metrics and physically trustworthy behavior, where models with high correlations still miss key transient structures and integral quantities. Taken together, REALM exposes the limits of current neural surrogates on realistic multiphysics flows and offers a rigorous testbed to drive the development of next-generation physics-aware architectures.

cs.LG↗
arXiv2026

MARS: Co-evolving Dual-System Deep Research via Multi-Agent Reinforcement Learning

Large Reasoning Models (LRMs) face two fundamental limitations: excessive token consumption when overanalyzing simple information processing tasks, and inability to access up-to-date knowledge beyond their training data. We introduce MARS (Multi-Agent System for Deep ReSearch), a novel co-evolution framework that jointly optimizes dual cognitive systems through multi-agent reinforcement learning. Unlike prior approaches that employ fixed or independently-trained summarizers, MARS enables System 1 (fast, intuitive processing) and System 2 (deliberate reasoning) to co-adapt through shared trajectory rewards, developing complementary strategies where System 1 learns to distill information specifically useful for System 2's reasoning. We extend Group Relative Policy Optimization (GRPO) for multi-agent settings with three key innovations: (1) decoupled gradient computation ensuring proper credit assignment despite shared rewards, (2) bin-packing optimization for efficient parallel information processing, and (3) advantage-weighted balanced sampling preventing training imbalance. Extensive experiments demonstrate that MARS (8B), trained under a challenging Zero RL setting without any supervised fine-tuning, achieves 8.17% on HLE -- outperforming WebThinker (32B with SFT, 6.87%) and narrowing the gap with proprietary models like Claude 3.7 Sonnet (7.89%) -- while achieving an average gain of 8.9% across 7 knowledge-intensive tasks.

cs.AI↗
arXiv2026

An introduction to gauge theories and group theory in particle physics

In this review, the fundamental concepts of group theory and representation theory are introduced. Special emphasis is placed on the unitary irreducible representations of the $SU(N)$ Lie group, the Poincare group, Little Group, discrete group, and their applications in particle physics. Based on the principle of local gauge symmetry, the construction of gauge-invariant Lagrangians and their quantization procedure are discussed. To address gauge redundancy, the modern on-shell amplitude approach is applied to gauge theories, demonstrating both conceptual and computational advantages. From the perspective of symmetry, the Standard Model is presented through the identification of its gauge symmetry, its anomaly-free matter content, and its global symmetries, including flavor symmetry, custodial symmetry, and baryon and lepton number conservation, etc.

hep-ph↗
arXiv2026

Complete Operator Basis for the modular invariant SMEFT

We implement modular flavor symmetries within the Standard Model Effective Field Theory (SMEFT) framework, using the flavor group $A_4^{(q)} \times A_4^{(e)}$ with distinct moduli $τ_q$ and $τ_e$, and assigning different modular weights to right-handed quarks using simplest weight assignment. By treating the moduli as non-dynamical spurions, adopting the MFV-like assumption, and neglecting effects associated with $\mathrm{Im}\,τ$, we systematically construct a finite set of independent modular-invariant higher-dimensional operators via the Hilbert-series techniques. In the holomorphic $A_4$ scenario, where all modular forms derive from the weight-2 triplet $Y^{(2)}_{\mathbf{3}}$, we present two equivalent Hilbert-series bases. This establishes that higher-dimensional operators can be formally organized as $[Y_{\mathbf{r}}^{(k_Y)},{Y_{\mathbf{r}'}^{(k_Y')}}^{*},\mathcal{O}]_{\mathbf{1}}$ singlets. We subsequently enumerate all independent operators up to dimension 7 under this assumption and provide explicit constructions for all dimension-5 operators as well as baryon- and lepton-number conserving dimension-6 operators. Relaxing holomorphicity to the non-holomorphic case of polyharmonic Maas forms, considering that non-holomorphic modular forms are not closed under multiplication, adopting the holomorphic organizing idea would generically lead to an infinite proliferation of modular-invariant structures. To retain a finite and complete operator basis, we therefore impose the same minimal formal organizing principle, which reproduces the benchmark Weinberg operator and the corresponding dimension-$6$ operators.

hep-ph↗
arXiv2026

C3Box: A CLIP-based Class-Incremental Learning Toolbox

Traditional machine learning systems are typically designed for static data distributions, which suffer from catastrophic forgetting when learning from evolving data streams. Class-Incremental Learning (CIL) addresses this challenge by enabling learning systems to continuously learn new classes while preserving prior knowledge. With the rise of pre-trained models (PTMs) such as CLIP, leveraging their strong generalization and semantic alignment capabilities has become a promising direction in CIL. However, existing CLIP-based CIL methods are often scattered across disparate codebases, rely on inconsistent configurations, hindering fair comparisons, reproducibility, and practical adoption. Therefore, we propose C3Box (CLIP-based Class-inCremental learning toolBOX), a modular and comprehensive Python toolbox. C3Box integrates representative traditional CIL methods, ViT-based CIL methods, and state-of-the-art CLIP-based CIL methods into a unified CLIP-based framework. By inheriting the streamlined design of PyCIL, C3Box provides a JSON-based configuration and standardized execution pipeline. This design enables reproducible experimentation with low engineering overhead and makes C3Box a reliable benchmark platform for continual learning research. Designed to be user-friendly, C3Box relies only on widely used open-source libraries and supports major operating systems. The code is available at https://github.com/LAMDA-CL/C3Box.

cs.LG↗
arXiv2026

Copositive criteria for a two-component dark matter model

We consider a two-component scalar dark matter model in this work, where the scalars are stabilized by extra $Z_2 \times Z'_2$ symmetry. To guarantee the stability of the vacuum, we consider the copositive criteria and different choices of the signs of the couplings will contribute to different viable parameter spaces. Based on the copositive criteria, we systematically analyze the possible conditions and pick up 17 different cases with fixing some parameters to be negative. We randomly scan the parameter space under dark matter relic density constraint and direct detection constraint and focus on the quartic couplings with $λ_{13}$,$λ_{23}$,$λ_{14}$ and $λ_{24}$. The shapes of the viable parameter space for $|λ_{14}|-|λ_{24}|$ are almost similar among these cases, while for $|λ_{13}|$ and $|λ_{23}|$, the larger value are excluded as long as $λ_{13}\leqslant 0$ and $λ_{23}\leqslant 0$ due to the large interference effect of the $2 \to 2$ processes.

hep-ph↗
arXiv2026

Nowcast3D: Reliable precipitation nowcasting via gray-box learning

Reliable nowcasting of extreme precipitation remains difficult because convective systems are strongly nonlinear, multiscale, and nonstationary in 3D. Radar is the backbone of nowcasting, yet existing methods struggle to predict extremes: physics-based extrapolation cannot capture growth and decay, deterministic learning tends to oversmooth and underestimate peaks, and purely generative models often lack physical consistency. Hybrid schemes help but are mostly limited to 2D composite reflectivity, collapsing the atmosphere into one layer and discarding vertical structure critical for height-dependent dynamics. We introduce Nowcast3D, a gray-box, fully 3D framework that works directly on volumetric radar reflectivity. The end-to-end model couples physically constrained neural operators (advection, local diffusion, and microphysics) with a conditional diffusion model to generate ensemble forecasts with quantified uncertainty. Trained on provincial-scale 3D volumes over a $10.24^\circ \times 10.24^\circ$ region and fine-tuned on a $2.56^\circ \times 2.56^\circ$ city region ($0.01^\circ \approx 1$ km), Nowcast3D provides near-real-time forecasts up to 3 h and outperforms competitive baselines in cross-region and temporal out-of-sample tests. It can also infer wind fields without labeled supervision, supporting physically plausible transport. In a nationwide blind evaluation by 160 meteorologists, Nowcast3D ranked first and was preferred in 57% of post-hoc assessments, surpassing the leading baseline (27%). These results highlight its reliability and operational value for extreme precipitation nowcasting.

cs.LG↗
arXiv2026

SSVP: Synergistic Semantic-Visual Prompting for Industrial Zero-Shot Anomaly Detection

Zero-Shot Anomaly Detection (ZSAD) leverages Vision-Language Models (VLMs) to enable supervision-free industrial inspection. However, existing ZSAD paradigms are constrained by single visual backbones, which struggle to balance global semantic generalization with fine-grained structural discriminability. To bridge this gap, we propose Synergistic Semantic-Visual Prompting (SSVP), that efficiently fuses diverse visual encodings to elevate model's fine-grained perception. Specifically, SSVP introduces the Hierarchical Semantic-Visual Synergy (HSVS) mechanism, which deeply integrates DINOv3's multi-scale structural priors into the CLIP semantic space. Subsequently, the Vision-Conditioned Prompt Generator (VCPG) employs cross-modal attention to guide dynamic prompt generation, enabling linguistic queries to precisely anchor to specific anomaly patterns. Furthermore, to address the discrepancy between global scoring and local evidence, the Visual-Text Anomaly Mapper (VTAM) establishes a dual-gated calibration paradigm. Extensive evaluations on seven industrial benchmarks validate the robustness of our method; SSVP achieves state-of-the-art performance with 93.0% Image-AUROC and 92.2% Pixel-AUROC on MVTec-AD, significantly outperforming existing zero-shot approaches.

cs.CV↗
arXiv2026

Unveiling and Bridging the Functional Perception Gap in MLLMs: Atomic Visual Alignment and Hierarchical Evaluation via PET-Bench

While Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in tasks such as abnormality detection and report generation for anatomical modalities, their capability in functional imaging remains largely unexplored. In this work, we identify and quantify a fundamental functional perception gap: the inability of current vision encoders to decode functional tracer biodistribution independent of morphological priors. Identifying Positron Emission Tomography (PET) as the quintessential modality to investigate this disconnect, we introduce PET-Bench, the first large-scale functional imaging benchmark comprising 52,308 hierarchical QA pairs from 9,732 multi-site, multi-tracer PET studies. Extensive evaluation of 19 state-of-the-art MLLMs reveals a critical safety hazard termed the Chain-of-Thought (CoT) hallucination trap. We observe that standard CoT prompting, widely considered to enhance reasoning, paradoxically decouples linguistic generation from visual evidence in PET, producing clinically fluent but factually ungrounded diagnoses. To resolve this, we propose Atomic Visual Alignment (AVA), a simple fine-tuning strategy that enforces the mastery of low-level functional perception prior to high-level diagnostic reasoning. Our results demonstrate that AVA effectively bridges the perception gap, transforming CoT from a source of hallucination into a robust inference tool and improving diagnostic accuracy by up to 14.83%. Code and data are available at https://github.com/yezanting/PET-Bench.

cs.CV↗
arXiv2026

Memory in the Age of AI Agents

Memory has emerged, and will continue to remain, a core capability of foundation model-based agents. As research on agent memory rapidly expands and attracts unprecedented attention, the field has also become increasingly fragmented. Existing works that fall under the umbrella of agent memory often differ substantially in their motivations, implementations, and evaluation protocols, while the proliferation of loosely defined memory terminologies has further obscured conceptual clarity. Traditional taxonomies such as long/short-term memory have proven insufficient to capture the diversity of contemporary agent memory systems. This work aims to provide an up-to-date landscape of current agent memory research. We begin by clearly delineating the scope of agent memory and distinguishing it from related concepts such as LLM memory, retrieval augmented generation (RAG), and context engineering. We then examine agent memory through the unified lenses of forms, functions, and dynamics. From the perspective of forms, we identify three dominant realizations of agent memory, namely token-level, parametric, and latent memory. From the perspective of functions, we propose a finer-grained taxonomy that distinguishes factual, experiential, and working memory. From the perspective of dynamics, we analyze how memory is formed, evolved, and retrieved over time. To support practical development, we compile a comprehensive summary of memory benchmarks and open-source frameworks. Beyond consolidation, we articulate a forward-looking perspective on emerging research frontiers, including memory automation, reinforcement learning integration, multimodal memory, multi-agent memory, and trustworthiness issues. We hope this survey serves not only as a reference for existing work, but also as a conceptual foundation for rethinking memory as a first-class primitive in the design of future agentic intelligence.

cs.CL↗
arXiv2026

Software-Hardware Co-optimization for Modular E2E AV Paradigm: A Unified Framework of Optimization Approaches, Simulation Environment and Evaluation Metrics

Modular end-to-end (ME2E) autonomous driving paradigms combine modular interpretability with global optimization capability and have demonstrated strong performance. However, existing studies mainly focus on accuracy improvement, while critical system-level factors such as inference latency and energy consumption are often overlooked, resulting in increasingly complex model designs that hinder practical deployment. Prior efforts on model compression and acceleration typically optimize either the software or hardware side in isolation. Software-only optimization cannot fundamentally remove intermediate tensor access and operator scheduling overheads, whereas hardware-only optimization is constrained by model structure and precision. As a result, the real-world benefits of such optimizations are often limited. To address these challenges, this paper proposes a reusable software and hardware co-optimization and closed-loop evaluation framework for ME2E autonomous driving inference. The framework jointly integrates software-level model optimization with hardware-level computation optimization under a unified system-level objective. In addition, a multidimensional evaluation metric is introduced to assess system performance by jointly considering safety, comfort, efficiency, latency, and energy, enabling quantitative comparison of different optimization strategies. Experiments across multiple ME2E autonomous driving stacks show that the proposed framework preserves baseline-level driving performance while significantly reducing inference latency and energy consumption, achieving substantial overall system-level improvements. These results demonstrate that the proposed framework provides practical and actionable guidance for efficient deployment of ME2E autonomous driving systems.

cs.AI↗