Work & Research

First author of a regularization-based study of Vision-Language-Action (VLA) models, now available as a preprint on arXiv. The project introduces a regularization term into VLA training to improve the behavior and robustness of learned policies, with both a principled motivation and empirical evaluation.

My contribution covers the method design, the theoretical motivation, and the experimental evaluation.

Keywords / skills involved: vision-language-action models, robot learning, regularization, imitation learning, policy training.

I work on the theoretical analysis of randomised search heuristics and genetic programming, under the supervision of Prof. Pietro Simone Oliveto and in collaboration with Mingxuan Yin. The current direction focuses on runtime analysis and generalisation bounds, with a full paper targeting AAAI 2027.

My contribution includes deriving runtime bounds, formalising proof arguments, and exploring how algorithmic parameters affect generalisation in evolutionary search.

Keywords / skills involved: probabilistic methods, drift theorems, runtime analysis, evolutionary algorithms, mathematical proofs.

Second author of a small-target recognition project, now available as a preprint on arXiv. The work focuses on improving detection and representation learning for low-resolution or weakly salient objects under challenging visual conditions.

My contribution covers the model design choices, training-pipeline engineering, and extensive benchmark experiments validating the proposed approach.

Keywords / skills involved: object detection, low-resolution vision, attention mechanisms, representation learning, PyTorch.

First author of a double-blind submission targeting ICML 2027, on expected policy gradients through hard events (guards, impacts, contact-mode switches) in differentiable simulation and model-based RL. The work combines certified handling of moving, policy-dependent event boundaries with structure-exploiting probabilistic inference, so that automatic differentiation is applied only to checked, scope-aware programs.

My contribution covers the method design, the formal analysis, and the experimental evaluation. Specifics are kept anonymous while the paper is under review.

Keywords / skills involved: differentiable simulation, hard events, measure transport, variable elimination, automatic differentiation, policy gradients.

Co-first author of a work on context engineering, now available as a preprint on arXiv. The study explores how structured context design affects reasoning and decision-making in large language models, with an emphasis on theoretical understanding and reproducible empirical protocols.

My contribution focuses on the initial problem formulation, validation of the core idea, and the large-scale empirical design. I also propose a theoretical framing of how context structure influences model behavior, to be developed in future work.

Keywords / skills involved: empirical protocol design, LLM evaluation, reproducible experiments.

Co-first author of a safe reinforcement learning project, now available as a preprint on arXiv. The work approaches safety through a risk-sensitive framework inspired by quantitative finance, using tail-risk estimation (e.g., CVaR-style measures) to bound rare but catastrophic failures without relying on dense reward engineering.

My contribution centers on the algorithmic design, convergence analysis, and theoretical justification of safety guarantees under tail-risk constraints.

Keywords / skills involved: reinforcement learning theory, safety constraints, tail risk estimation, CVaR, risk-sensitive RL.

Co-first author of a context-engineering work currently under review at ICLR 2027. The study introduce a framework for evaluating RAG semantic cascades, together with a series of pratical coarse-grained reranker that improves reranking cascades.

My contribution focused on the initial problem formulation, core idea validation, and large-scale empirical design. We perform comprehensive experiments on our argument.

Keywords / skills involved: empirical evaluation protocol design, LLM evaluation, Long-context QA.

Co-first author of a post-training work currently under review at ICML 2027. The study explores efficient self-play RLVR for LLMs.

My contribution focused on the initial problem formulation and proposed some novel ideas for the solution of this problem.

Keywords / skills involved: post-training, RLVR, gradient computation.

Co-first author of a context-engineering work currently under preparation for ICML 2027. The study reports the training phenomena of a special network structure, which may lead to better network design.

My contribution focused on the initial problem formulation, core idea validation, and large-scale empirical design. We design a thoughtful 4-stage experiment to fulfil a theoretical closure.

Keywords / skills involved: Isometry in Neural Network, Explainable AI.