Hui-Ling Zhen

LLM/Agent for Testing & Verification

Highlights

Built LLM-driven and agentic verification workflows for Verilog generation, bug finding, tool invocation, verifier feedback, and coverage-guided testbench generation.

Selected Projects

• Verilog-Evolve: Feedback-Driven and Skill-Evolving Verilog Generation, arXiv 2026 | paper | code
  A tool-grounded self-evolution framework for Verilog generation, using simulator, synthesis, timing, downstream feedback, and reusable RTL skills to improve generated RTL.
• Agentic ATPG | code
  The first agent-powered ATPG algorithm. Agent can first optimize and evaluate command-line configurations from outside the tool, then later make controlled white-box changes to the core algorithm once the evaluation loop is reliable.
• Agentic Verification | code
  LLM-powered Agentic Verification for hardware. The core bilevel architecture is constructed by CodeX and veriagent (our runtime). VeriAgent does not try to replace Codex as a general-purpose coding agent. Instead, it supervises Codex with explicit workflow stages, deterministic Checkers, verification-domain MCP tools, human checkpoints, and a measurable run_manifest.json.

Related Papers

1. BetterV: Controlled Verilog Generation with Discriminative Guidance, ICML 2024 | paper
   Developed a generator-discriminator training framework for 7B RTL code generation; even two years after release, it remains a same-size SOTA for correctness and outperforms several models larger than 14B.
2. DiLA: Enhancing LLM Tool Learning with Differential Logic Layer, KDD 2026 | paper
   Introduced a way to make the solver act as a differentiable layer that interacts with the language model, outperforming tool-call-only solver integration and improving success rates across multiple industrial scenarios.
3. SATformer: Transformer-Based UNSAT Core Learning, ICCAD 2023 | paper

AI for EDA (Formal Verification & Testing)

Highlights

Developed AI-driven SAT solving, preprocessing, conflict management, data augmentation, and learning-aided verification methods for industrial EDA workflows. Applications include ATPG, LEC, model checking, bug finding, and coverage improvement in real chip verification scenarios.

• First industrial-grade learning-driven solver enhancement: explored learning-driven ways to modify and enhance solver internals for industrial verification problems, improving practical solving performance beyond black-box solver selection.
• First industrial-grade GAN-based data synthesis for EDA: introduced generative data synthesis to improve the robustness of AI-driven verification methods under the data scarcity common in industrial EDA scenarios.
• First industrial-grade circuit-level SAT direction: moved beyond conventional CNF-centric modeling by constructing circuit-level SAT solving directly over circuits, enabling ATPG to serve as a solver for industrial LEC and contribute meaningfully to datapath proofs.

Related Papers

1. Accelerate SAT-Based ATPG via Preprocessing and New Conflict Management Heuristics, ASP-DAC 2021 | paper
2. Neural Fault Analysis for SAT-Based ATPG, ITC 2022 | paper
3. Conflict-Driven Structural Learning Towards Higher Coverage Rate in ATPG, ETS 2023 | paper
4. HardSATGEN: Understanding the Difficulty of Hard SAT Formula Generation and a Strong Structure-Hardness-Aware Baseline, KDD 2023 | paper
5. HardCore Generation: Generating Hard UNSAT Problems for Data Augmentation, NeurIPS 2024 | paper
6. DeepGate2: Functionality-Aware Circuit Representation Learning, ICCAD 2023 | paper
7. DiffSAT: Differential MaxSAT Layer for SAT Solving, ICCAD 2024 | paper
8. Logic Optimization Meets SAT: A Novel Framework for Circuit-SAT Solving, DAC 2025 | paper

LLM Reasoning & Agentic LLMs

Highlights

Worked on long-context training, post-training, agent self-evolution, reasoning data construction, verifier-based optimization, and training-inference co-design for agentic LLMs. Achieved up to 8x end-to-end speedup in deep research agents and up to 20% improvement on long-context and agent benchmarks.

Related Papers

1. LLMShare: Optimizing LLM Inference Serving with Hardware Architecture Exploration, DAC 2025 | paper
2. TrimR: Verifier-Based Training-Free Thinking Compression for Efficient Test-Time Scaling, ICLR 2025 | paper
3. Accelerating Large Language Model Reasoning via Speculative Search, ICML 2025 | paper
4. HyperTree Planning: Enhancing LLM Reasoning via Hierarchical Thinking, ICML 2025 | paper
5. DLLM Agent: See Farther, Run Faster, arXiv 2026 | paper
6. SCOPE: Prompt Evolution for Enhancing Agent Effectiveness, arXiv 2025 | paper
7. Towards Efficient Agents: A Co-Design of Inference Architecture and System, arXiv 2025 | paper
8. MOSS: Efficient and Accurate FP8 LLM Training with Microscaling and Automatic Scaling, arXiv 2025 | paper
9. CMoE: Fast Carving of Mixture-of-Experts for Efficient LLM Inference, ACL 2026 | paper

Learning to Optimize and Solver Intelligence

Highlights

Studied learning-to-optimize methods, solver intelligence, and machine learning techniques for large-scale industrial optimization and mixed-integer programming.

Related Papers

1. Pareto Multi-Task Learning, NeurIPS 2019 | paper
2. Learning to Select Cuts for Efficient Mixed-Integer Programming, Pattern Recognition 2022 | paper
3. A Survey for Solving Mixed Integer Programming via Machine Learning, Neurocomputing 2023 | paper
4. Bilevel Learning for Large-Scale Flexible Flow Shop Scheduling, Computers & Industrial Engineering 2022 | paper
5. Machine Learning Methods in Solving the Boolean Satisfiability Problem, Machine Intelligence Research 2023 | paper