Bringing Continual Learning into Enterprises — Samuel Denton, Applied Compute

A reported training experiment increased how often a coding agent submitted its work while leaving its test-pass rate unchanged. The conference description reports a rise in submit-tool calls from 22% to 60%, in work aimed at getting the agent to finish earlier.

The training used a teacher model given corrective guidance. Because the historical trace contained no submission call, the teacher could not directly teach that action at its original position. Instead, it changed the reasoning leading toward submission. The result illustrates a targeted behavioral improvement: changing how an agent completes its workflow without reporting an increase in coding accuracy.

“Applied Compute wanted it wrapping up by turn 40 and got the submit tool call rate from 22% to 60% with test pass rate flat.”

AI Engineer

“because the rollout was conditioned on an old production trace that never called the tool, the teacher never touched the tool call tokens at all. It moved the reasoning path toward the call instead, and the call followed.”

AI Engineer

“rewarding the format directly and finetuning on correct examples both degraded coding ability, while a hint written against each rollout took correct formatting from 15% to 80%.”

AI Engineer

“Let a judge pick where in the rollout the hint goes and distill only the next few steps, since the learning signal decays with distance from the hint.”

Sam Denton

“mask which tokens you learn from, because the teacher has strong opinions about connector words that have nothing to do with the lesson.”

Sam Denton