Samuel Denton leads platform research at Applied Compute, developing AI systems that learn from enterprise work without requiring perfect answer keys or sacrificing existing capabilities. His work spans enterprise continual learning, organizational memory, model post-training, and the infrastructure needed to make specialized models practical.
Denton studied at Columbia University and previously worked at Amazon and Scale AI before joining Applied Compute in 2026. At Scale, he coauthored research on spurious correlations in language-model post-training, examining how misleading training patterns affect supervised fine-tuning and preference-based methods. Different approaches proved more robust under different tasks and biases, sharpening a question central to his subsequent work: how to improve specific model behaviors without inadvertently changing others.
At Applied Compute, Denton coauthored research on the Context Engine and Contextbases, which capture institutional knowledge, refine it into procedural memory, and retrieve relevant context for agents. This supports reasoning-effort amortization: converting expensive reasoning into reusable knowledge instead of repeatedly paying for the same computation. In company experiments, access to that memory helped lower-reasoning configurations match stronger settings operating without it.
He also coauthored research introducing ACL-Wiki and ACLBench, systems for extracting reusable knowledge from coding-agent interactions and evaluating it against internal engineering tasks. Their workflow collects traces from Cursor, Claude Code, and Codex; removes stale or redundant information; and delivers relevant memories to coding agents through an MCP server.
- Learning without definitive answer keys. Denton organizes enterprise continual learning around two variables: whether production traces are historical or generated online, and whether behavioral hints are fixed or constructed from an agent’s actual rollout. Historical traces can improve existing agents immediately; online traces and tailored hints enable continuing adaptation.
- Behavioral correction without capability regression. Experiments with Qwen3.5-Thinking on SWE-bench distinguished task completion from underlying test performance, measuring whether a coding agent could finish sooner without becoming worse at coding. Separate experiments used rollout-specific hints to teach unusual customer-required hyperlink formatting while avoiding regressions associated with reward shaping and supervised fine-tuning.
- Relevance-masked self-distillation. Denton and collaborators use model-based judges to identify where corrective hints belong and which teacher-generated tokens contain useful instruction. Targeting the relevant step reduces training on incidental stylistic preferences and limits unwanted changes elsewhere in the model.
Denton has also identified automated research for specialized models as an active priority: lowering the barriers to building custom enterprise systems that improve through accumulated organizational experience.
