On the Reasoning Gaps of Large Language Models

A Formal Characterization

NeurIPS 2026 Pre-Print

176,000 evaluations across 12 models and 9 benchmark tasks reveal systematic, predictable failure modes in LLM reasoning that correlate with computational complexity.

The Computational Complexity of Verifying LLM Outputs

ICLR 2027 In Progress

Formal complexity-theoretic framework for understanding when and why LLM outputs can be efficiently verified, with implications for scalable oversight.

A Taxonomy of Failure Modes in LLM-Based Autonomous Agents

ACL 2027 In Progress

Empirical taxonomy of how autonomous LLM agents fail in practice, drawn from 50+ real deployment incidents across research and production systems.

Impossibility Results for Unsupervised Self-Improvement in Language Models

ICLR 2027 Early Stage

Theoretical bounds on what language models can learn from their own outputs without external signal.