Definition
Plain language
An extra training signal added alongside the main objective to teach a model a side skill.
As stated in the literature
A secondary training objective combined with the primary loss to inject additional supervision; in ECHO, next-token cross-entropy on environment-produced tokens layered onto GRPO with a small weight coefficient.
Also called: auxiliary objective
Why it matters: Auxiliary losses inject useful supervision into otherwise sparse training signals and often stabilize learning.
For example, alongside the main reward, a model might be trained with a small extra loss that just rewards predicting the next observation token correctly.
Heard on the show
“And there's a hyperparameter — the weight you put on the auxiliary loss, lambda — that has its own story.”Episode 084 — Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away