I study machine cognition through adaptive inference. My work treats reasoning and generation as latent dynamical processes and develops inference-time mechanisms that evaluate the adequacy of ongoing computation, regulate it under uncertainty, and allocate additional deliberation when weighted. The goal is to extend existing foundation models with deployable control layers that improve stability and competence under distribution shift and long-horizon inference.

A recurring theme in my work is the role of invariants, understood as stable structural properties that preserve coherence as internal trajectories evolve. I study how such structure arises in learned systems, how it fails, and how lightweight signals derived from latent geometry can serve as internal diagnostics for monitoring and regulation.

My work draws on control theory, nonlinear dynamics, signal processing, structured multi-process coordination, and system neuroscience

Treat inference as a latent-space trajectory and design monitoring-and-feedback mechanisms that modulate inference dynamics online. This relaxes fixed-depth schedules with budgeted, conditional computation driven by intrinsic discrepancy signals and stability diagnostics.

Develop low-overhead inference-time signals based on Lyapunov-style proxy energies and trajectory geometry consisting of curvature proxies, angular velocity, jerk, divergence to detect structural stress before visible failure. These signals support bounded interventions such as damping, redirection, early stopping, abstention, and refinement.

Study how internal representations evolve under novelty, prediction error, and shift, and how slow-timescale adaptation can preserve manifold structure while maintaining task-relevant organization. The approach leans toward controlled adaptation over unconstrained parameter updates.

Model system cognition interacting inference processes coupled through shared geometric structure and coordination dynamics, inspired by recurrent cortical computation, multi-timescale plasticity, and modular control architectures.

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 impending structural collapse
• Uses bounded guidance damping, gain scheduling, and refractory control to prevent runaway instability
• Achieves large gains in stability under extreme guidance and adversarial prompts without retraining or added decoding cost
• Demonstrates that collapse is preceded by directional inconsistency, not merely large update magnitude

This work reframes diffusion inference as a regulated dynamical process, showing how internal proprioceptive signals can support robust, low-cost control.

→ Repository: error-360-

Modeled agents as dual-timescale dynamical systems coupling fast inference dynamics with slow structural adaptation that preserves latent invariants under distributional drift. Energy-based adaptation is regulated by internal control signals that bias updates toward invariant-supporting latent structure and gate plasticity based on internally detected novelty, formalized as Perceptual Gravity and Synthetic Dopamine.

→ Repository: lisa

Developing a deployable inference wrapper around standard backbones that separates fast proposal from slow verification. Uses stability checks and proxy energies as stopping criteria for refinement, enabling anytime behavior under explicit compute budgets and producing an Indulgence Score that measures deliberation required for convergence.

My long-term goal is a principled framework for system-level reasoning in which inference is treated as a regulated dynamical computation. The systems I am interested in allocate computation adaptively, maintain invariants under shift, and remain stable over long horizons through monitoring and feedback rather than retraining.

Concretely, I am interested in architectures that:

• Allocate compute based on intrinsic instability or insufficiency
• Enforce bounded behavior through damping, refinement, abstention, or early stopping
• Preserve latent structure under drift via slow-timescale adaptation
• Expose interpretable internal commitments that support governance and debugging

I draw inspiration from recurrent cortical circuits, neuromodulatory feedback, and multi-timescale plasticity. Using tools from nonlinear dynamics and control, I aim to design models that are self-monitoring, self-regulating, and robust by construction.

• Mathematics: nonlinear ODEs and discrete-time dynamical systems, stability of equilibria and invariant sets, Lyapunov-style energy methods, geometric analysis of latent-state trajectories
• Machine learning: adaptive and test-time inference, invariant representation learning, stability-aware reasoning diagnostics
• Systems & Tooling: PyTorch, NumPy/SciPy, Hugging Face Hub and Spaces, FastAPI, Jupyter, Docker, Linux/Bash
• Engineering Practice: closed-loop inference controllers, trajectory-level logging and analysis, diagnostic instrumentation, controlled experimental harnesses