The Work Behind the Models.
Peer-reviewed publications, preprints, and working papers from the TAL Corp research team — spanning alignment, interpretability, evaluation, reasoning, and agentic systems.
Every Claim Has a Citation.
Scalable Oversight via Recursive Constitutional Constraints in Large Language Models
J. Mercer · A. Krishnaswamy · D. Okafor · TAL Corp Research Team
Circuit-Level Analysis of Deceptive Alignment in Transformer-Based Agents
S. Park · L. Okonkwo · TAL Corp Research Team
SAFE-AGENT: A Comprehensive Benchmark for Evaluating Autonomous AI Safety
S. Park · R. Vasquez · TAL Corp Research Team
Capability-Invariant Corrigibility: Maintaining Human Oversight Across Capability Gains
J. Mercer · TAL Corp Research Team
Causal World Models with Structural Consistency Guarantees
A. Krishnaswamy · D. Okafor · TAL Corp Research Team
COORD-SAFE: A Safety Protocol for Multi-Agent Coordination Networks
D. Okafor · R. Vasquez · TAL Corp Research Team
Publish the Failures
We do not only publish results that confirm our hypotheses. When our methods fail, we document how and why. A research culture that hides failure is a research culture that cannot learn from it. Our preprints include null results, ablation failures, and benchmark regressions.
Theory Follows Evidence
We do not build elaborate theoretical frameworks and then look for data to support them. Every research agenda at TAL Corp begins with empirical observation — anomalies in model behavior, unexpected capability jumps, or alignment failures in deployment.
Interpretability is Non-Negotiable
We will not publish a result we cannot explain mechanistically. Statistical correlation between training conditions and model behavior is a starting point, not a finding. We require circuit-level or attention-level accounts of the phenomena we study.
Collaboration Over Competition
Frontier AI research is too important to be slowed by competitive secrecy. We share benchmarks, evaluation suites, and methodology with the broader research community — including the findings that make our own models look bad.
Long Timelines, Short Papers
We believe the 50-page paper is a relic of a slower era. Our publications are written to be read — dense with signal, free of padding. We would rather publish 28 pages that change how the field thinks than 60 pages that no one finishes.
Five Fronts. One Direction.
Constitutional constraints, corrigibility preservation, and value alignment that scales with capability.
Mechanistic analysis of transformer circuits, attention structures, and deceptive activation patterns.
Benchmark design for autonomous AI systems, adversarial evaluation suites, and OOD robustness testing.
Causal world modeling, structural invariance across distribution shift, and OOD generalization.
Multi-agent coordination protocols, goal drift prevention, and emergent misalignment in network contexts.
A sixth research area is currently in formation. Details forthcoming in Q2 2026.
Interested in Our Research?
We collaborate with academic institutions and independent researchers. We share benchmarks, evaluation suites, and findings openly. Get in touch to discuss joint work, data access, or visiting researcher positions.
Email us at: research@texasagilabs.com