Research
Research areas.
Our work examines how systems learn from observations, use evidence, adapt to changing conditions, and execute models within practical constraints.
The research spans statistical learning, large language models (LLMs), adaptive systems, and inference engineering. Financial markets provide the primary application setting.
Questions we study
-
Statistical Learning
What can be learned from observations, and how well does it generalize?
We study methods for estimation, prediction, and uncertainty analysis. This includes machine learning, time-series modeling, and sequential inference, where estimates are updated as observations accumulate. Research questions concern generalization, changing data distributions, and the calibration of predictions. Predictive relationships are distinguished from causal explanations.
- Statistical inference
- Machine learning
- Sequential methods
- Time-series modeling
- Uncertainty and calibration
-
LLMs and Agents
How can models turn information into useful, verifiable work?
We investigate how LLMs interpret documents, use supporting evidence, and produce structured outputs. Agent systems extend these capabilities through tool use and sequences of observations and actions. Research questions include checking extracted claims against sources, coordinating tasks, and identifying failures that may be obscured by plausible language.
- LLMs
- Evidence grounding
- Structured outputs
- Tool use
- Agent coordination
-
Adaptive Systems
How can systems acquire new capabilities while retaining useful ones?
We study adaptation across changing data and tasks. Continual learning concerns the retention of useful knowledge as new information is learned. Related research includes meta-learning and neuro-inspired approaches to memory and adaptation. Governed recursive self-improvement is a research direction within this area. We investigate systems that propose changes to their methods or components, with a separate evaluation process assessing those changes before adoption. Consequential modifications require human authorization.
- Continual learning
- Knowledge retention
- Meta-learning
- Neuro-inspired learning
- Governed recursive self-improvement
-
Inference Engineering
How can trained models operate effectively within computational and operational constraints?
Inference engineering concerns the execution of trained models. We examine trade-offs between task quality, latency, throughput, memory use, and computational cost. Research questions include selecting appropriate execution configurations, measuring performance under representative workloads, and detecting failures in model-serving systems. Execution efficiency and the correctness of outputs are evaluated separately.
- Model execution and serving
- Latency and throughput
- Memory and computational cost
- Workload measurement
- Operational reliability
How results are judged
- Separate exploration from evaluation
- Use exploratory analysis to develop questions, then assess proposed methods on held-out observations under a defined procedure.
- Compare with baselines
- Determine whether a method improves on simpler alternatives and whether the improvement justifies its complexity.
- Measure across conditions
- Assess performance, uncertainty, efficiency, and sensitivity to changes in inputs or workloads.
- Evaluate changes independently
- Separate the process proposing a change from the process judging whether it should be adopted.