Introduces G0.5, a pretrained autoregressive Vision-Language-Action model with a unified transformer decoder that emits reasoning and action tokens in a single stream. Features a Cross-Embodiment Action Codec, Native Chain-of-Thought, and Visual Memory. Surpasses state-of-the-art models across seven benchmarks, including real-world robot fine-tuning (76.7% vs. 53.3% for π0.5), BEHAVIOR-1K (31.4% task score), DROID (82.5%), LIBERO (98.9%), and RoboTwin 2.0 (93.3%).

My contribution: Contributed to G0.5 pretraining through training-data annotation, quality control, cleaning, and validation during my internship at Galaxea AI.

Tackles project-level code generation by introducing CodeProjectEval (a dataset of 18 real-world repositories averaging 12.7 files and ~2,389 lines per task) and ProjectGen, a multi-agent framework that decomposes generation into architecture design, skeleton generation, and code filling. Introduces Semantic Software Architecture Tree (SSAT) to bridge user requirements and code. Achieves 57% improvement on DevBench and ~10x improvement on CodeProjectEval over baselines.

Introduces CangjieBench, a benchmark of 248 manually translated tasks from HumanEval and ClassEval. Evaluates six LLMs using direct prompting, syntax-constrained prompting, retrieval over documentation, retrieval over code, and agent-based workflows. A concise syntax reference offers the best accuracy-cost trade-off among prompt-based methods; the best agent configuration reaches the highest Pass@1 at 10–120 times the token cost of a single prompt. Code-to-Code translation can reduce compile rates by introducing Python-specific patterns.

The arXiv preprint is titled CangjieBench: Benchmarking LLMs on a Low-Resource General-Purpose Programming Language.

Evaluates whether pre-trained code language models (CodeBERT, CodeT5, Codex, StarCoder, CodeLlama) can generalize to scientific computing programming languages (SCPLs). Finds that while SCPLs are more challenging than general-purpose languages, CLMs are nevertheless applicable and knowledge from general languages transfers effectively to SCPL analysis.

Proposes a low-cost pipeline for generating controllable synthetic data for cloth-changing person re-identification. Introduces the CCUP dataset with 6,000 IDs, ~1.18M images, 100 cameras, and 26.5 outfits per individual. A pretrain-finetune framework using CCUP significantly improves CC-ReID models, outperforming state-of-the-art methods on PRCC, VC-Clothes, and NKUP benchmarks.

Introduces AdaptiveLLM, a framework that dynamically selects the optimal cost-efficient LLM for code generation based on automatically assessed task difficulty using Chain-of-Thought length. Clusters tasks into three difficulty levels and uses XGBoost for model selection. Achieves 7.86% improvement in pass@1 while reducing resource consumption by 88.9% compared to ComplexityNet.