Mechanics of Learned Reasoning 1: TempoBench, A Benchmark for Interpretable Deconstruction
of Reasoning System Performance on Temporal Reasoning.

Under submission at OOPSLA 2026. We created a benchmark for interpretable deconstruction of reasoning system performance on temporal reasoning. We show how this can be used to understand the behavior of large language models and other learned reasoning systems.

OOPSLA 2026 Submission

AI Generated Podcast For Paper

AI-generated podcast discussion of the TempoBench paper

SNNs Are Not Transformers (Yet)

Spiking Neural Networks are a biologically plausible architectures with much greater energy efficiency then current deep learning architectures. Despite this, these models have faced slow adoption. We derive a sample complexity for these models and show how this scales with respect to longer input sequences. This demonstrates some of the weakness in these models in being adopted for language modeling. This work was presented at the SHDA Workshop in conjunction with the International Conference on Supercomputing and is currently under review for the Neuromorphic Computing and Engineering Journal.

Neuromorphic Computing and Engineering Journal Submission

Multi-Agent Path Planning for Cops And Robbers Via Reactive Synthesis

We are using Temporal Stream Logic (TSL) to create correct by construction coordinated systems for generating solutions to the Graph Theory Game Cops and Robbers

Cops And Robbers Reactive Synthesis PrePrint

Conference Talks

Amherst College AI in Liberal Arts A Better Path Forward: Reconciling AI's Drive for Scale With the Reality of Its Energy Costs View Slides
FMCAD Automata Learning Meets State Space Machines View Slides
SHDA Workshop ICS SNNs Are Not Transformers (Yet) View Slides