publications
- Tail-Influence Sampling for CVaR Policy Evaluation2026Tail-Influence Sampling (TIS) allocates a fixed evaluation budget across queryable components of a stochastic workflow toward those that matter most for lower-tail CVaR, attaining oracle asymptotic variance and cutting MSE by 41% versus learned occupancy and 76% versus complete rollouts on CliffWalking.
- VoiceAgentGuard: Contract-Gated Tool Release for Production Voice AgentsIn EMNLP, 2026VoiceAgentGuard prevents unsafe private-read and state-changing tool calls in voice agents by forcing every proposed action through an auditable contract gate before execution.
- LeanPolish: Verified Supervision for Lean Proof Compression2026A neurosymbolic Lean 4 proof-compression method that records every verified edit, exposes the selection effects of search-generated supervision, and shows when learning from it helps beyond symbolic search.
- Coverage Cliffs in Learning from Logged World FeedbackIn ICML Workshop on RL from World Feedback, 2026We study when logged world feedback can certify tail safety, giving a PAC-Bayes off-policy VaR certificate and identifying a quantile-specific coverage cliff caused by insufficient weighted CDF mass.
- CovCal: Risk-Controlled Lean-as-Judge for Natural-Language Mathematical ReasoningIn ICML AI for Math Workshop, 2026COVCAL shows when Lean proofs can be safely trusted as partial evidence for math-answer selection by calibrating coverage diagnostics and abstaining when formal evidence is insufficient.
- Information-Geometric Neural Granger CausalityIn ICML Workshop on High-dimensional Learning Dynamics, 2025We introduce Information-Geometric Neural Granger Causality (IG-NGC), a framework that provides a partial theoretical explanation for neural causality methods. Our key insight is that when assuming Gaussian output distributions and diagonal Fisher Information approximation, causality can be interpreted as directional information flow measured by a simplified Fisher metric on statistical manifolds.
- FrEVL: Leveraging Frozen Pretrained Embeddings for Efficient Vision-Language UnderstandingIn ICCV Workshop on Safe and Trustworthy Multimodal AI Systems, 2025Oral PresentationFrEVL is a vision-language understanding by freezing pretrained CLIP embeddings and training only a lightweight fusion network. This approach delivers: 3× faster inference than ALBEF/BLIP, 70% lower deployment costs, 68.4M trainable parameters (vs 200M+ in SOTA models), 850 images/sec throughput on single V100, production-ready with <25ms p99 latency.
- Kernel-Based Anomaly Detection Using Generalized Hyperbolic ProcessesIn International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025We integrate Generalized Hyperbolic processes into kernel-based anomaly detection through a provably valid GH kernel used within KDE and one-class SVMs, better capturing heavy-tailed and skewed data and improving detection on imbalanced, non-Gaussian distributions.
- Kimina-Prover Preview: Towards Large Formal Reasoning Models with Reinforcement LearningarXiv preprint arXiv:2504.11354, 2025Kimina-Prover is a large formal reasoning model trained with reinforcement learning that uses a structured reasoning pattern to iteratively generate and refine Lean 4 proof steps, reaching 80.7% on the miniF2F benchmark (pass@8192) with strong sample efficiency and clear performance scaling with model size.
- Multi-Modal Information Bottleneck Attribution with Cross-Attention GuidanceIn British Machine Vision Conference (BMVC), 2024A multi-modal attribution method that leverages the information bottleneck principle and cross-attention guidance to identify the most relevant features across different modalities.
- Quaternion Recurrent Neural Network with Real-Time Learning and Maximum CorrentropyIn International Joint Conference on Neural Networks (IJCNN), 2024Oral PresentationWe develop a robust quaternion recurrent neural network (QRNN) for real-time processing of 3D and 4D data with outliers. This is achieved by combining the real-time recurrent learning (RTRL) algorithm and the maximum correntropy criterion (MCC) as a loss function. While both the mean square error and maximum correntropy criterion are viable cost functions, it is shown that the non-quadratic maximum correntropy loss function is less sensitive to outliers, making it suitable for applications with multidimensional noisy or uncertain data. Both algorithms are derived based on the novel generalised HR (GHR) calculus, which allows for the differentiation of real functions of quaternion variables and offers the product and chain rules, thus enabling elegant and compact derivations.
- Convex Quaternion Optimization for Signal Processing: Theory and ApplicationsIEEE Transactions on Signal Processing, 2023Convex optimization methods have been extensively used in communications and signal processing. However, the theory of quaternion optimization is currently not as fully developed and systematic as that of complex and real optimization. To this end, we establish the convex optimization theory in quaternion variables based on the generalized Hamilton-real (GHR) calculus. This is achieved in a way that conforms with traditional complex and real optimization theory. We present several discriminant theorems for convex quaternion functions analogous to their complex counterparts. We also provide several discriminant criteria for strongly convex functions by the theorems of convex quaternion functions. Furthermore, we prove that the quaternion Newton method can converge in one step for positive definite quadratic quaternion functions and provide two applications in quaternion signal processing. These results provide a solid theoretical foundation for convex quaternion optimization and open avenues for further developments in quaternion signal processing applications.
- The HR-Calculus: Enabling Information Processing with Quaternion AlgebraarXiv preprint arXiv:2311.16771, 2023In this work, the foundations of the HR-calculus are revised and the required tools for deriving adaptive learning techniques suitable for dealing with quaternion-valued signals, such as the gradient operator, chain and product derivative rules, and Taylor series expansion are presented. This serves to establish important applications of adaptive information processing in the quaternion domain for both single-node and multi-node formulations.
- Timing of Hypoxia PET/CT Imaging after 18F-Fluoromisonidazole Injection in NSCLC PatientsNature Scientific Reports, 2022
-
Dataset