I study intelligence as a dynamical, interactive process rather than a static feed-forward computation.
My research focuses on adaptive inference, stability-aware reasoning, and latent geometric structure in learned systems.

I investigate how models can monitor their own inference trajectories, detect early signs of instability, and regulate computation in closed loop, drawing on ideas from control theory, nonlinear dynamics, signal processing, structured multi-agent coordination, and systems neuroscience.

Viewing inference as a latent space trajectory that can be monitored, regulated, and adjusted online using intrinsic discrepancy signals and stability diagnostics, rather than fixed schedules or static heuristics.

Developing lightweight inference-time signals (Lyapunov-style energies, curvature and jerk indicators, off-manifold detection) that anticipate failure before visible collapse, enabling preventative control rather than post-hoc correction.

Studying how internal representations and geometry evolve under novelty, distribution shift, and prediction error, and how slow-timescale adaptation supports robustness in changing environments.

Modeling cognition as interacting reasoning processes coupled through shared geometric fields and interaction dynamics, inspired by recurrent cortical circuits, multi-timescale plasticity, and coordinated control systems.

A lightweight, inference-time control mechanism that treats diffusion sampling as a dynamical system and stabilizes it using intrinsic trajectory signals.

• Introduces a latent jerk signal (angular acceleration of update directions) as an early-warning indicator of structural collapse
• Uses bounded guidance damping with gain scheduling and refractory control to prevent runaway instability
• Achieves large gains in survival under extreme guidance and adversarial prompts without retraining or added decoding cost
• Demonstrates that instability is preceded by directional inconsistency, not just large update magnitude

This work positions diffusion inference as a regulated dynamical process, rather than a fixed numerical procedure, and shows how internal “proprioceptive” signals can enable robust, low-cost control.

→ Repository: error-360-

My long-term goal is to help develop intelligent systems that go beyond static feed-forward pipelines and instead support dynamically regulated, structurally adaptive reasoning.

I am particularly interested in systems that:

• Allocate more computation when instability or uncertainty is detected
• Know when to stop, damp, or redirect inference
• Adapt internal structure under distribution shift
• Maintain coherent organization over long horizons

I take inspiration from recurrent cortical circuits, neuromodulatory feedback, and multi-timescale plasticity, using tools from dynamical systems and control theory to design models that are self-monitoring, self-regulating, and robust by construction.

• Mathematics: nonlinear dynamics, Lyapunov stability, time-series analysis, geometric signals
• Machine learning: adaptive inference, representation learning, stability diagnostics, test-time control
• Systems: PyTorch, NumPy/SciPy, Hugging Face stack, FastAPI, Jupyter, Docker, Linux/Bash
• Engineering: inference-time controllers, trajectory logging, diagnostic instrumentation, experimental harnesses