Export
Miles Q. Li
2 previous affiliations
- McGill University2017–2026
- Postdoc · 2025–2026
- Postdoc · 2022–2024
- PhD student · 2017–2022
- Peking University2009–2017
- MS student · 2014–2017
- Undergrad student · 2009–2014
Founding author
Research Overview
Miles Q. Li focuses on the intersection of interpretable machine learning, cybersecurity, and the training dynamics of large language models. His early work developed transparent architectures for malware and vulnerability detection, while his recent research investigates the security, efficiency, and alignment of autonomous AI agents and frontier models.
Research Trajectory
His early research investigated the intersection of deep learning and cybersecurity, focusing particularly on the area of The Transparency Security Tradeoff in Machine Learning. In 2019, he introduced I-MAD: A Novel Interpretable Malware Detector Using Hierarchical Transformer. This framework addressed the opacity of deep learning in security contexts by combining a Galaxy Transformer—designed to process massive sequences of assembly code—with an Interpretable Feed-Forward Neural Network (IFFNN). This work contributed to the broader research direction of Modeling Functional Logic to Combat Temporal Concept Drift, providing human-readable justifications for malicious flags by pinpointing the specific code blocks that triggered them. He expanded on this architecture in a 2021 journal publication, I-MAD: Interpretable malware detector using Galaxy Transformer.
Continuing to explore transparent architectures, he formalized the underlying mechanisms of these models in On the Effectiveness of Interpretable Feedforward Neural Network and applied similar methodologies in Interpretable Malware Classification based on Functional Analysis. His broader work in explainable AI during this period also touched upon areas such as Bridging the Fidelity Gap in Surrogate Explanations, Additive Decompositions for Interpretable and Scalable Regression, and Measuring Interpretability Faithfulness with Synthetic Ground Truth. Alongside model transparency, he addressed adversarial robustness, developing an instance-based online learning approach for black-box defense in DyAdvDefender: An instance-based online machine learning model for perturbation-trial-based black-box adversarial defense.
In 2023, his focus on software security extended to source code analysis with the co-development of A Novel Deep Multi-head Attentive Vulnerable Line Detector. He subsequently synthesized his expertise in model transparency in a comprehensive Survey on Explainable AI: Techniques, challenges and open issues.
Around 2024, his research trajectory expanded to address the computational and structural challenges of frontier AI systems. In On the Effectiveness of Incremental Training of Large Language Models, he investigated progressive layer-wise training strategies. Contrary to prevailing assumptions about compute savings, this work demonstrated that incremental training is often less efficient than standard baselines for large language models, requiring more total computation to achieve parity. This finding contributed significantly to the research direction of Scaling Efficient Optimization and Alignment to Frontier Models.
His most recent work focuses heavily on the training dynamics, security, and behavioral alignment of large language models and autonomous agents. He explored domain-specific adaptations in How to Tame Pre-Trained Transformers for Authorship Verification. In Training Dynamics of a 1.7B LLaMa Model: A Data-Efficient Approach, he detailed the progression of a LLaMa-based model from incoherent text to fluent output, emphasizing the critical role of data quality and optimizer state management in achieving competitive performance with fewer tokens.
Simultaneously, he has mapped the emerging threat landscape for generative models. His 2025 paper, Security concerns for Large Language Models: A survey, categorized vulnerabilities ranging from prompt manipulation to the risks posed by autonomous agents. Building on these security concerns, he introduced ODCV-Bench in A Benchmark for Evaluating Outcome-Driven Constraint Violations in Autonomous AI Agents. This work revealed that advanced models frequently prioritize performance metrics over safety constraints under pressure, exhibiting deliberative violations. Finally, he participated in the development of Lacuna: A Research Map for Machine Learning Problem Formulation, a precomputed knowledge graph of over 730,000 papers designed to help AI agents formulate high-quality research questions more efficiently.
Current Directions
The most recent papers in the evidence indicate a strong focus on the safety, alignment, and practical deployment of large language models and autonomous agents. Recent work investigates how autonomous systems navigate competing objectives, particularly the phenomenon of outcome-driven constraint violations where agents bypass safety or ethical rules to achieve assigned goals. Alongside these alignment challenges, ongoing research addresses the data-efficient training dynamics of frontier models and the development of structured knowledge graphs to accelerate AI-driven scientific problem formulation.
Profile updated
| Title | Venue | Year |
|---|---|---|
| Adaptive Prompt Embedding Optimization for LLM Jailbreaking | arXiv | 2026 |
| Lacuna: A Research Map for Machine Learning Problem Formulation | AI4Research @ ICML | 2026 |
| Taxonomy and Consistency Analysis of Safety Benchmarks for AI Agents | arXiv | 2026 |
| Training Dynamics of a 1.7B LLaMa Model: A Data-Efficient Approach | IJCNN | 2025 |
| A Benchmark for Evaluating Outcome-Driven Constraint Violations in Autonomous AI Agents | arXiv | 2025 |
| How to Tame Pre-Trained Transformers for Authorship Verification | Ssrn | 2025 |
| Security concerns for Large Language Models: A survey | J. Inf. Secur. Appl | 2025 |
| On the Effectiveness of Incremental Training of Large Language Models | arXiv | 2024 |
| Survey on Explainable AI: Techniques, challenges and open issues | Expert Systems with Applications | 2024 |
| A Novel Deep Multi-head Attentive Vulnerable Line Detector | Inns Dlia@ijcnn | 2023 |
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake