2 papers
cs.LG2026
Deterministic Decisions for High-Stakes AI. A Zero-Egress Pipeline with the Deployability of RAG and the Accuracy of Machine Learning
Craig Atkinson
We identify intervention bias as a previously unquantified failure mode of zero-shot large-language-model (LLM) educational advisory agents: without task-specific training, they re…
cs.LG2026
Tethered Reasoning: Decoupling Entropy from Hallucination in Quantized LLMs via Manifold Steering
Craig Atkinson
Quantized language models face a fundamental dilemma: low sampling temperatures yield repetitive, mode-collapsed outputs, while high temperatures (T > 2.0) cause trajectory diverge…