Minghang Deng

Combined column retrieval with Arctic embeddings and model-based schema linking by extracting SQL-generated fields to compress Snowflake semantic model schemas. Achieved 90% perfect and 96% average column recall with 90% context reduction, and 99% perfect and 99% average table recall with 50% context reduction, given a 32k-token budget on Snowflake internal datasets. Maintained efficiency with per-example latency under 10 seconds. Contributed to the Snowflake Cortex Analyst product by significantly reducing semantic model sizes. The generation schema linking technique could also help find low quality golden query & SQL pairs.

Text-to-SQL

Proposed ReFoRCE, a Text-to-SQL framework that combines (a) pattern-based table compression and LLM-guided schema linking to mitigate long-context challenges (b) self-refinement for correcting syntax and semantic errors across SQL dialects (c) majority-vote consensus to identify high-confidence predictions while deferring ambiguous cases (d) execution-guided iterative column exploration to resolve them. Achieved state-of-the-art performance on the Spider 2.0-Snow and Lite leaderboards.

Trained Text-to-SQL Reasoning Models with Execution Feedback: Fine-tuned Qwen-Coder and DeepSeek-R1-Distill models via the VeRL framework. Parsed and executed SQL outputs with the vLLM inference engine, appended execution results to model inputs, and optimized performance using the GRPO algorithm.

Collaborated with Snowflake AI Research to propose a scalable RL-based model family combined with carefully curated data, strong supervised initialization, and effective training practices with Group Relative Policy Optimization (GRPO) algorithm. Achieved state-of-the-art accuracy on 6 Test-to-SQL benchmarks, ranking first on the BIRD leaderboard.

Contributed to the CLLM repository by fixing bugs in trajectory augmentation and repetitive pattern detection. Proposed Gist CLLM by introducing gist tokens and modifying attention masks to inject context into special tokens. Extended sequence length and increased batch size to speed up speculative decoding via Jacobi iterations.

QEMU Simulation: Demonstrated proficiency in virtualization technology through an in-depth study of QEMU source code and virtual machines. Successfully simulated various virtual machines, including the LC-3 and NES6502, by leveraging the capabilities of the QEMU emulator.

Developed a pioneering framework to generate two equal-size models with the same MD5 checksum.