Samuel Denton leads platform research at Applied Compute. His work examines how models can learn from enterprise experience while limiting unrelated changes to their behavior.[2]
Start with Bringing Continual Learning into Enterprises →
Education
Links
Current professional profile Conference biography Platform Research Lead · Applied Compute
Career
Denton’s public professional account identifies earlier work at Scale AI and Amazon. His conference biography records study at Columbia and a move to Applied Compute in 2026. His research connects model adaptation with the operational question of learning from an organization’s work before it has perfect answer keys or a complete replay environment.[1][2]
Contributions
Teach the behavior that needs to change
His continual-learning talk organizes systems by whether traces are historical or generated online, and whether corrective hints are fixed or built from an agent’s rollout. A teacher can receive extra information about a failed decision and demonstrate a better response. Distillation then transfers that behavior to the student. Selecting relevant steps or tokens narrows the update, reducing the tendency to copy unrelated teacher preferences. The examples include meeting a submission deadline and fixing hyperlink format.[3]
Test whether post-training learns the wrong shortcut
His coauthored paper compares supervised fine-tuning with DPO and KTO, two preference-based training methods, on synthetic tasks containing misleading correlations. It varies the prevalence and kind of shortcut in reasoning, instruction-following and document-grounded questions. Preference training was relatively robust on some mathematical tasks, while supervised fine-tuning performed better on complex context-heavy tasks. The result makes robustness depend on the task and data artifact rather than declaring one method universally best.[4]
Tools, writing & talks
Writing, research & talks
Start here
Bringing Continual Learning into Enterprisesai.engineerA taxonomy and demonstrations of learning from production traces, corrective hints and targeted distillation.
Start with the trace-and-hint taxonomy, then follow one failed-decision example through the teacher and student. Familiarity with model training helps; identify what extra information the teacher receives before comparing the update methods.
Presenter[3]Applied Compute pushed a Qwen model's SWE-bench submissions from 80 turns toward 40, lifting the submit-tool call rate from 22% to 60% with pass rate flat, and the teacher never touched the tool-call tokens, moving the reasoning path instead.
Catalog workOriginal sourceA study of supervised and preference training under controlled shortcut patterns.
Coauthor[4]