Selected work

Representative Papers

SmileyLlama: Modifying Large Language Models for Directed Chemical Space Exploration

Sep 03, 2024arXiv.org

This work addresses the challenge of transforming general-purpose large language models (LLMs) into attribute-controllable molecular generators. We propose a lightweight adaptation paradigm that converts open-source Llama models into chemical language models (CLMs) via supervised fine-tuning (SFT) and direct preference optimization (DPO), enabling direct SMILES string generation conditioned on multidimensional physicochemical properties (e.g., logP, aqueous solubility). To our knowledge, this is the first empirical demonstration that an adapted general LLM achieves performance on multi-objective molecular generation tasks comparable to or exceeding that of domain-specific chemically pretrained models. The approach enables a paradigm shift from “chemical knowledge question-answering” to “property-directed molecular design,” significantly enhancing controllability, interpretability, and interactive exploration of chemical space.

7 citationsRead paper

A practical guide to machine learning interatomic potentials – Status and future

Mar 01, 2025Current opinion in solid state & materials science

To address the practical challenge that non-expert researchers face in applying machine learning interatomic potentials (MLIPs), this work introduces the first industrial-grade, end-to-end MLIP practice framework. Methodologically, it systematically integrates state-of-the-art models—including GAP, M3GNet, NequIP, and Allegro—and unifies active learning, uncertainty quantification, and physics-informed constraint embedding, while establishing standardized protocols for data generation, model selection, interpretability validation, and cross-platform deployment. Its key contributions are: (i) the first formal definition of a practical MLIP construction paradigm and evaluation benchmark; (ii) the open release of a fully reproducible toolchain and implementation guidelines; and (iii) substantial improvements in generalizability and computational efficiency of MLIPs within molecular dynamics simulations—demonstrated successfully in alloy design and catalytic modeling. This framework effectively bridges the gap between ML research and applied computational materials science.

2 citationsRead paper

Black-box unadjusted Hamiltonian Monte Carlo

Dec 12, 2024

Unadjusted Hamiltonian Monte Carlo (HMC) and underdamped Langevin algorithms suffer from asymptotic bias in high-dimensional sampling due to numerical integration errors and lack automatic step-size adaptation. Method: This paper establishes, for the first time, a quantitative relationship between Hamiltonian energy error and asymptotic bias, and proposes the first black-box, provably bounded step-size adaptation scheme that eliminates the need for Metropolis–Hastings (MH) correction. Contribution/Results: The method rigorously controls asymptotic bias within any user-specified tolerance. Theoretical analysis—validated on Gaussian and canonical Bayesian models—confirms strong bias controllability. Empirical evaluation demonstrates several-fold speedup over MH-adjusted samplers in high dimensions, alongside markedly improved stability and overall performance. This breakthrough overcomes a key practical barrier to deploying unadjusted samplers in real-world applications.

2 citationsRead paper

PRISM: Distribution-free Adaptive Computation of Matrix Functions for Accelerating Neural Network Training

Jan 29, 2026

This work addresses the computational inefficiency of matrix functions—such as square roots, inverse roots, and orthogonalization—in neural network training, which stems from traditional iterative methods’ reliance on prior spectral information and their inability to adapt to dynamically changing matrix spectra. The authors propose PRISM, a novel framework that enables adaptive computation of matrix functions without requiring any prior knowledge of the spectrum. PRISM constructs, at each iteration, a polynomial surrogate of the current spectrum using random sketching and relies predominantly on GPU-friendly matrix multiplications. This approach automatically adapts to spectral shifts during training, substantially reducing computational overhead. When integrated into Shampoo and Muon optimizers, PRISM maintains optimization accuracy while significantly decreasing both iteration counts and wall-clock runtime.

1 citationsRead paper

AlphaSyndrome: Tackling the Syndrome Measurement Circuit Scheduling Problem for QEC Codes

Jan 18, 2026

This work addresses the lack of a general automated method for scheduling stabilizer measurements in non-surface-code quantum error correction, which often leads to significant fluctuations in logical error rates. We propose the first optimization framework tailored to generic swap-based stabilizer codes, formulating measurement scheduling as an optimization problem that controls error propagation pathways. By integrating Monte Carlo Tree Search (MCTS) with feedback from heuristic decoders, our approach automatically discovers optimal measurement orderings and parallelization strategies that steer error propagation away from logical operators while keeping it within the correctable range of the decoder. Evaluated across diverse code families, system sizes, and decoders, the method reduces logical error rates by 80.6% on average—up to 96.2% in the best case—matching the performance of Google’s hand-optimized surface code schedules and surpassing IBM’s existing strategy for Bivariate Bicycle codes.

1 citationsRead paper

Recent publications

Latest Papers

Equivariant learning of a transferable three-dimensional classical density functional

Aug 13, 2026

This work addresses the longstanding challenge of efficiently constructing liquid density functionals applicable across diverse thermodynamic conditions, interfacial settings, and confinement environments. The authors propose a three-dimensional, transferable density functional learning framework that requires no explicit labels for free energy or chemical potential. By integrating equivariant neural networks with variational self-consistency constraints, the method directly learns the functional from equilibrium density field data alone, rigorously preserving spatial symmetries and thermodynamic consistency. For the first time, this approach achieves generalization across temperature, system size, and statistical ensembles, accurately predicting unseen macroscopic phenomena—including structure factors, equations of state, vapor–liquid coexistence, interfacial broadening, non-monotonic solvation forces in complex geometries, and adsorption behavior in bicontinuous helical pore channels—demonstrating exceptional accuracy and transferability.

0 citationsRead paper

Arithmetic Variable LogLog: Advancing the Memory-Variance Frontier

Aug 12, 2026

This work proposes a high-density register structure for data stream cardinality estimation that leverages arithmetic coding, relative NLZ (number of leading zeros) storage, and a shared offset mechanism to fully utilize 64-bit words and eliminate low-frequency states. Integrated with a non-iterative, four-component hybrid estimator (HLDLC) and an early-exit strategy, the approach achieves significantly improved accuracy and speed while maintaining O(B + log log C) memory scalability. Experimental results demonstrate that, under a 1 KB memory budget, the method attains a weighted mean absolute error of 1.63%, outperforming ExaLogLog by 4.7%; it also reduces the memory-variance product to 3.4 (surpassing ExaLogLog’s 3.78) and accelerates processing by 2.7–4.5×, marking the first technique to comprehensively dominate ExaLogLog across all tested memory configurations.

0 citationsRead paper

Autonomous discovery of accelerator commissioning algorithms

Aug 07, 2026

Traditional accelerator tuning relies on manually crafted procedures, which are difficult to reuse efficiently after lattice modifications, thereby hindering rapid early-stage design iteration. This work proposes the first end-to-end autonomous algorithm discovery framework driven by a language model agent, integrating large language models, a particle accelerator simulation platform, and an automated feedback loop to iteratively generate and refine tuning algorithms starting from minimal initial code. For the first time, an intelligent agent directly participates in exploring accelerator tuning strategies, successfully producing 16 non-dominated algorithms in the ALS-U accumulator ring model. These algorithms exhibit diverse physical trade-offs between beam capture speed and error correction performance and significantly outperform expert-designed solutions.

0 citationsRead paper

Breaking Memory Bottlenecks in Quantum Control Systems for More Precise Experiments and Higher Throughput Computing

Aug 06, 2026

Quantum control systems face memory bottlenecks due to limited on-chip BRAM capacity and unpredictable DRAM access latency, hindering high-precision experiments and high-throughput quantum circuit execution. To address this, this work proposes Ant-Q, a novel memory hierarchy that uniquely integrates heterogeneous BRAM/DRAM storage, pipelined scheduling, and deterministic timing control to simultaneously achieve high throughput and strict timing guarantees. Evaluation across 26 real-world quantum circuits demonstrates that Ant-Q reduces the overhead of circuit loading and readout—from previously accounting for 22.90% to 1417.05% of total execution time—to nearly zero, substantially improving system efficiency. The design has been successfully integrated into the QubiC 3.0 quantum control system.

0 citationsRead paper

Contrast-invariant deep ptychography neural networks

Aug 03, 2026

Deep ptychographic imaging neural networks often suffer from illumination-induced scale inconsistencies during out-of-distribution generalization, undermining their practical reliability. To address this issue, this work proposes a scale-decoupled factorization strategy that reformulates object representation from the conventional amplitude–phase domain to a real–imaginary formulation. Additionally, a synthetic object sampling scheme is introduced to reduce the phase distribution discrepancy between synthetic and experimental data. This approach effectively disentangles texture from measurement scale, yielding substantial improvements in generalization across five cross-illumination experimental datasets. Compared to the PtychoPINN-torch baseline, the proposed method reduces Fourier error by up to fivefold.

0 citationsRead paper