Searcharxiv⌕ Search

arXiv subjects

Tong Zhu

Publications and source records attributed to Tong Zhu.

At least 37 records · Page 2Linked to original sources
arXiv2026

EvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments

Autonomous agents are increasingly expected to improve executable policies through feedback, yet existing evaluations often collapse this process into a final score or confound it with open-ended software-engineering progress. We introduce Autonomous Policy Evolution, a controlled evaluation setting in which a harness-model agent repeatedly edits an executable policy system under a fixed interaction budget. We instantiate this setting in EvoPolicyGym, a benchmark built from compact interactive RL environments that evaluates how agents iteratively improve explored policies. On the EvoPolicyGym suite, GPT-5.5 achieves the strongest aggregate rank score and top-two performance on all 16 environments. Beyond leaderboard results, EvoPolicyGym also provides trajectory-level diagnostics that distinguish how agents allocate budget, convert feedback into parametric tuning. These analyses show that strong autonomous policy evolution depends not only on isolated task wins, but on discovering task-appropriate mechanisms and refining policies under bounded feedback.

cs.AI↗
arXiv2026

CERES: A Cryogenic Experiment to Reconstruct Energy Systematics in TeO$_{2}$ bolometers

Cryogenic calorimetric detectors are a powerful tool in the search for rare events such as neutrinoless double beta decay ($0νββ$), due to their excellent energy resolution and low intrinsic background. The performance of these detectors depends critically on a precise understanding of their energy scale and energy resolution. Recent studies suggest that both energy scale and energy resolution may vary depending on the spatial location and topology of energy deposition within the detector, indicating the presence of previously uncharacterized systematic effects. The Cryogenic Experiment to Reconstruct Energy Systematics (CERES) is a dedicated experiment designed to directly measure the position dependence of calorimetric response in Tellurium Dioxide (TeO$_{2}$) crystals. This paper details the experimental design, current status, and future upgrade plans for CERES.

physics.ins-det↗
arXiv2026

$π$-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows

The rise of personal assistant agents, e.g., OpenClaw, highlights the growing potential of large language models to support users across everyday life and work. A core challenge in these settings is proactive assistance, since users often begin with underspecified requests and leave important needs, constraints, or preferences unstated. However, existing benchmarks rarely evaluate whether agents can identify and act on such hidden intents before they are explicitly stated, especially in sustained multi-turn interactions where user needs emerge gradually. To address this gap, we introduce $π$-Bench, a benchmark for proactive assistance comprising 100 multi-turn tasks across 5 domain-specific user personas. By incorporating hidden user intents, inter-task dependencies, and cross-session continuity, $π$-Bench evaluates agents' ability to anticipate and address user needs over extended interactions, jointly measuring proactivity and task completion in long-horizon trajectories that better reflect real-world use. Experiments show (1) proactive assistance remains challenging, (2) a clear distinction between task completion and proactivity, and (3) the value of prior interaction for proactive intent resolution in later tasks.

cs.AI↗
arXiv2026

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling

Recent progress in reasoning models has substantially advanced long-horizon mathematical and scientific problem solving, with several systems now reaching gold-medal-level performance on International Mathematical Olympiad (IMO) and International Physics Olympiad (IPhO) problems. In this paper, we introduce a simple and unified recipe for converting a post-trained reasoning backbone into a rigorous olympiad-level solver. The recipe first uses a reverse-perplexity curriculum for SFT to instill rigorous proof-search and self-checking behaviors, then scales these behaviors through a two-stage RL pipeline that progresses from RL with verifiable rewards to more delicate proof-level RL, and finally boosts solving performance with test-time scaling. Applying this recipe, we train a 30B-A3B backbone with SFT on around 340K sub-8K-token trajectories followed by 200 RL steps. The resulting model, SU-01, supports stable reasoning on difficult problems with trajectories exceeding 100K tokens, while achieving gold-medal-level performance on mathematical and physical olympiad competitions, including IMO 2025/USAMO 2026 and IPhO 2024/2025. It also demonstrates strong generalization of scientific reasoning to domains beyond mathematics and physics.

cs.AI↗
arXiv2026

Automating Computational Chemistry Workflows via OpenClaw and Domain-Specific Skills

This work presents a decoupled framework for multi-step computational chemistry automation built on OpenClaw. OpenClaw serves as the general-purpose agent for task coordination and supervision. Planning skills externalize task descriptions into executable task specifications, domain skills provide computational chemistry procedures, and the DPDispatcher skill grounds computation in heterogeneous HPC environments. In a methane-oxidation reactive MD case study, the framework coordinated cross-tool execution, supported bounded recovery from runtime failures, and extracted reaction networks.

physics.chem-ph↗
arXiv2026

Persistent Visual Memory: Sustaining Perception for Deep Generation in LVLMs

While autoregressive Large Vision-Language Models (LVLMs) demonstrate remarkable proficiency in multimodal tasks, they face a "Visual Signal Dilution" phenomenon, where the accumulation of textual history expands the attention partition function, causing visual attention to decay inversely with generated sequence length. To counteract this, we propose Persistent Visual Memory (PVM), a lightweight learnable module designed to strengthen sustained, on-demand access to visual evidence. Integrated as a parallel branch alongside the Feed-Forward Network (FFN) in LVLMs, PVM establishes a distance-agnostic retrieval pathway that directly provides visual embeddings for enhanced visual perception, thereby structurally mitigating the signal suppression inherent to deep generation. Extensive experiments on Qwen3-VL models demonstrate that PVM brings notable improvements with negligible parameter overhead, delivering consistent average accuracy gains across both 4B and 8B scales, particularly in complex reasoning tasks that demand persistent visual perception. Furthermore, in-depth analysis reveals that PVM shows improved robustness in longer generations and accelerates internal prediction convergence.

cs.CV↗
arXiv2026

Common-agency Games for Multi-Objective Test-Time Alignment

Aligning large language models (LLMs) with human preferences is inherently multi-objective: different users and evaluation criteria impose heterogeneous and often conflicting requirements on model outputs. We propose CAGE (Common-Agency Games for Alignment), a training-free, game-theoretic framework for multi-objective test-time alignment. CAGE models alignment objectives as strategic principals that allocate token-level incentives to a shared LLM, inducing an equilibrium policy that captures the joint effect of competing objectives. We develop an efficient algorithm based on equilibrium problems with equilibrium constraints (EPEC) to compute this equilibrium, and establish theoretical guarantees including existence and uniqueness of the equilibrium policy, convergence and stability of the algorithm, and no-regret learning dynamics. Empirically, CAGE enables flexible and fine-grained trade-offs across objectives at inference time, consistently outperforming existing test-time alignment methods while requiring no retraining. It further supports weak-to-strong generalization, making multi-objective alignment practical in resource-constrained settings.

cs.GT↗
arXiv2026

Vib2Conf: AI-driven discrimination of molecular conformations from vibrational spectra

Retrieving or generating two-dimensional molecular structures on the basis of vibrational spectra has been well demonstrated via deep learning models. However, deciphering three-dimensional molecular conformations is still challenging, primarily due to spectral ambiguities caused by conformational heterogeneity, which are difficult to resolve. To address this limitation, we propose Vib2Conf, a deep learning model directly discriminating 3D molecular conformations from vibrational spectra. We implement an attentional resampler to distill conformation-sensitive features from sparse spectral signals, and integrate Mixture-of-Experts (MoE) to partition the conformational space for precise geometric mapping. These modules enable Vib2Conf to achieve state-of-the-art top-1 recall exceeding 95% on traditional spectrum-structure benchmarks, including QM9S, VB-Mols, and QMe14S. More importantly, Vib2Conf can discriminate near-isomeric conformers with a top-1 recall of 82.06% on VB-Confs test set, where conformational isomers differ by a root-mean-square deviation (RMSD) of only ~1 Å. In general, Vib2Conf is a promising method for fine-grained spectrum-to-conformation analysis.

physics.chem-ph↗
arXiv2026

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.

cs.LG↗
arXiv2026

Constraints on light dark matter from primordial black hole evaporation at dark matter direct detection experiments

Primordial black holes (PBHs) are able to produce light dark matter (DM) particles via Hawking radiation, and yield a flux of boosted DM that can be probed at underground DM direct detection experiments. We analyze both galactic and extragalactic contributions to the differential flux of light DM from PBH evaporation, and then compute the expected event rate from PBH boosted DM scattering off electrons or nuclei after taking into account the attenuation effect. Using recent data from DM direct detection experiments XENONnT, PandaX-4T and LZ, we set constraints on both DM-electron and DM-nucleus scattering cross sections, as well as the fraction of DM composed of PBHs $f_{\rm PBH}$ for $9\times10^{14}-1\times10^{16}\,\mathrm{g}$ PBHs that are not fully evaporated today. We also investigate the spectral evolution induced by Hawking evaporation throughout the evaporation and post-evaporation regimes. The constraints on the PBH mass are then extended into the $1\times10^{13}-6\times10^{14}\,\mathrm{g}$ window for fully evaporated PBHs.

hep-ph↗
arXiv2026

GEMS: Agent-Native Multimodal Generation with Memory and Skills

Recent multimodal generation models have achieved remarkable progress on general-purpose generation tasks, yet continue to struggle with complex instructions and specialized downstream tasks. Inspired by the success of advanced agent frameworks such as Claude Code, we propose \textbf{GEMS} (Agent-Native Multimodal \textbf{GE}neration with \textbf{M}emory and \textbf{S}kills), a framework that pushes beyond the inherent limitations of foundational models on both general and downstream tasks. GEMS is built upon three core components. Agent Loop introduces a structured multi-agent framework that iteratively improves generation quality through closed-loop optimization. Agent Memory provides a persistent, trajectory-level memory that hierarchically stores both factual states and compressed experiential summaries, enabling a global view of the optimization process while reducing redundancy. Agent Skill offers an extensible collection of domain-specific expertise with on-demand loading, allowing the system to effectively handle diverse downstream applications. Across five mainstream tasks and four downstream tasks, evaluated on multiple generative backends, GEMS consistently achieves significant performance gains. Most notably, it enables the lightweight 6B model Z-Image-Turbo to surpass the state-of-the-art Nano Banana 2 on GenEval2, demonstrating the effectiveness of agent harness in extending model capabilities beyond their original limits.

cs.CV↗
arXiv2026

Suppressed excitonic effects enable high mobility, high-yield photoconductivity in a two-dimensional polymer crystal with axial pyridine coordination

Two-dimensional polymers (2DPs) and their layer-stacked covalent organic frameworks (2D COFs) offer modular, atomically precise platforms for organic optoelectronics, yet their photoconductive responses remain fundamentally constrained by strong excitonic effects and localized charge transport. Here, we demonstrate that a diyne-linked 2DP crystal with axial pyridine coordination overcomes this limitation, enabling simultaneous efficient free-carrier generation and band-like transport. Introducing pyridine ligands that axially coordinate to Cu-porphyrin nodes transforms weak van der Waals stacking into a pyridine-bridged architecture with pronounced interlayer band dispersion and substantially reduced carrier effective masses. The resulting strong interlayer electronic coupling suppresses the exciton binding energy to well below thermal energy, such that optical excitation directly populates delocalized electronic states. Time-resolved terahertz spectroscopy reveals Drude-type photoconductivity with room-temperature mobilities approaching 500 cm^2 V^-1 s^-1 and a photon-to-free-carrier conversion ratio of ~0.4, yielding a photoconductive response that exceeds that of state-of-the-art organic and many inorganic photoactive materials. These results establish interlayer coordination as a powerful strategy for mitigating excitonic effects and accessing inorganic-like charge transport in organic 2D crystals, opening a pathway toward highly efficient photo-to-electricity conversion in organic-based systems.

cond-mat.mtrl-sci↗
arXiv2026

ExpertWeaver: Unlocking the Inherent MoE in Dense LLMs with GLU Activation Patterns

Mixture-of-Experts (MoE) effectively scales model capacity while preserving computational efficiency through sparse expert activation. However, training high-quality MoEs from scratch is prohibitively expensive. A promising alternative is to convert pretrained dense models into sparse MoEs. Existing dense-to-MoE methods fall into two categories: \textbf{dynamic structural pruning} that converts dense models into MoE architectures with moderate sparsity to balance performance and inference efficiency, and \textbf{downcycling} approaches that use pretrained dense models to initialize highly sparse MoE architectures. However, existing methods break the intrinsic activation patterns within dense models, leading to suboptimal expert construction. In this work, we argue that the Gated Linear Unit (GLU) mechanism provides a natural blueprint for dense-to-MoE conversion. We show that the fine-grained neural-wise activation patterns of GLU reveal a coarse-grained structure, uncovering an inherent MoE architecture composed of consistently activated universal neurons and dynamically activated specialized neurons. Leveraging this discovery, we introduce ExpertWeaver, a training-free framework that partitions neurons according to their activation patterns and constructs shared experts and specialized routed experts with layer-adaptive configurations. Our experiments demonstrate that ExpertWeaver significantly outperforms existing methods, both as a training-free dynamic structural pruning technique and as a downcycling strategy for superior MoE initialization.

cs.CL↗
arXiv2026

Safe-SDL:Establishing Safety Boundaries and Control Mechanisms for AI-Driven Self-Driving Laboratories

The emergence of Self-Driving Laboratories (SDLs) transforms scientific discovery methodology by integrating AI with robotic automation to create closed-loop experimental systems capable of autonomous hypothesis generation, experimentation, and analysis. While promising to compress research timelines from years to weeks, their deployment introduces unprecedented safety challenges differing from traditional laboratories or purely digital AI. This paper presents Safe-SDL, a comprehensive framework for establishing robust safety boundaries and control mechanisms in AI-driven autonomous laboratories. We identify and analyze the critical ``Syntax-to-Safety Gap'' -- the disconnect between AI-generated syntactically correct commands and their physical safety implications -- as the central challenge in SDL deployment. Our framework addresses this gap through three synergistic components: (1) formally defined Operational Design Domains (ODDs) that constrain system behavior within mathematically verified boundaries, (2) Control Barrier Functions (CBFs) that provide real-time safety guarantees through continuous state-space monitoring, and (3) a novel Transactional Safety Protocol (CRUTD) that ensures atomic consistency between digital planning and physical execution. We ground our theoretical contributions through analysis of existing implementations including UniLabOS and the Osprey architecture, demonstrating how these systems instantiate key safety principles. Evaluation against the LabSafety Bench reveals that current foundation models exhibit significant safety failures, demonstrating that architectural safety mechanisms are essential rather than optional. Our framework provides both theoretical foundations and practical implementation guidance for safe deployment of autonomous scientific systems, establishing the groundwork for responsible acceleration of AI-driven discovery.

cs.RO↗
arXiv2026

ALIGN: Aligned Delegation with Performance Guarantees for Multi-Agent LLM Reasoning

LLMs often underperform on complex reasoning tasks when relying on a single generation-and-selection pipeline. Inference-time ensemble methods can improve performance by sampling diverse reasoning paths or aggregating multiple candidate answers, but they typically treat candidates independently and provide no formal guarantees that ensembling improves reasoning quality. We propose a novel method, Aligned Delegation for Multi-Agent LLM Reasoning (ALIGN), which formulates LLM reasoning as an aligned delegation game. In ALIGN, a principal delegates a task to multiple agents that generate candidate solutions under designed incentives, and then selects among their outputs to produce a final answer. This formulation induces structured interaction among agents while preserving alignment between agent and principal objectives. We establish theoretical guarantees showing that, under a fair comparison with equal access to candidate solutions, ALIGN provably improves expected performance over single-agent generation. Our analysis accommodates correlated candidate answers and relaxes independence assumptions that are commonly used in prior work. Empirical results across a broad range of LLM reasoning benchmarks consistently demonstrate that ALIGN outperforms strong single-agent and ensemble baselines.

cs.LG↗
arXiv2026

A Graph Neural Network for the Era of Large Atomistic Models

Foundation models, or large atomistic models (LAMs), aim to universally represent the ground-state potential energy surface (PES) of atomistic systems as defined by density functional theory (DFT). The scaling law is pivotal in the development of large models, suggesting that their generalizability in downstream tasks consistently improves with increased model size, expanded training datasets, and larger computational budgets. In this study, we present DPA3, a multi-layer graph neural network founded on line graph series (LiGS), designed explicitly for the era of LAMs. We demonstrate that the generalization error of the DPA3 model adheres to the scaling law. The scalability in the number of model parameters is attained by stacking additional layers within DPA3. Additionally, the model employs a dataset encoding mechanism that decouples the scaling of training data size from the model size within its multi-task training framework. When trained as problem-oriented potential energy models, the DPA3 model exhibits superior accuracy in the majority of benchmark cases, encompassing systems with diverse features, including molecules, bulk materials, surface and cluster catalysts, two-dimensional materials, and battery materials. When trained as a LAM on the OpenLAM-v1 dataset, the DPA-3.1-3M model exhibits lowest overall zero-shot generalization error across 12 downstream tasks spanning a diverse array of research domains. This performance suggests superior accuracy as an out-of-the-box potential model, requiring minimal fine-tuning data for downstream scientific applications.

physics.comp-ph↗
arXiv2026

Vib2Mol: from vibrational spectra to molecular structures-a unified deep learning framework

There will be a paradigm shift in chemical and biological research, to be enabled by autonomous, closed-loop, real-time self-directed decision-making experimentation. Spectrum-to-structure correlation, which is to elucidate molecular structures with spectral information, is the core step in understanding the experimental results and to close the loop. However, current approaches usually divide the task into either database-dependent retrieval and database-independent generation and neglect the inherent complementarity between them. In this study, we proposed Vib2Mol, a unified deep learning framework designed to flexibly handle diverse spectrum-to-structure tasks according to the available prior knowledge by bridging the retrieval and generation. Empowered by our coarse-to-fine retrieval and generate-then-rerank strategies, Vib2Mol not only achieves state-of-the-art performance in analyzing theoretical Infrared and Raman spectra, but also outperform previous models on experimental data. Moreover, our model demonstrates promising capabilities in predicting reaction products and sequencing peptides, enabling vibrational spectroscopy a potential guide for autonomous scientific discovery workflows.

physics.chem-ph↗
arXiv2025

DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models

While recent Multimodal Large Language Models (MLLMs) have attained significant strides in multimodal reasoning, their reasoning processes remain predominantly text-centric, leading to suboptimal performance in complex long-horizon, vision-centric tasks. In this paper, we establish a novel Generative Multimodal Reasoning paradigm and introduce DiffThinker, a diffusion-based reasoning framework. Conceptually, DiffThinker reformulates multimodal reasoning as a native generative image-to-image task, achieving superior logical consistency and spatial precision in vision-centric tasks. We perform a systematic comparison between DiffThinker and MLLMs, providing the first in-depth investigation into the intrinsic characteristics of this paradigm, revealing four core properties: efficiency, controllability, native parallelism, and collaboration. Extensive experiments across four domains (sequential planning, combinatorial optimization, constraint satisfaction, and spatial configuration) demonstrate that DiffThinker significantly outperforms leading closed source models including GPT-5 (+314.2\%) and Gemini-3-Flash (+111.6\%), as well as the fine-tuned Qwen3-VL-32B baseline (+39.0\%), highlighting generative multimodal reasoning as a promising approach for vision-centric reasoning.

cs.CV↗