Value Nets: Separating Judgment from Generation in LLM Agents

Current LLM agents collapse two different jobs into one model: generating candidate actions and judging whether those actions should be taken. Alignment training (RLHF, constitutional training) bakes values into the generator itself, where they cannot be inspected, debated, or changed per decision. We propose the opposite arrangement: keep the generator, and add a separate layer of value nets — independently trained scorers that rate every candidate thought or action against explicit value dimensions before anything is said or done. The winning candidate is selected by a stated rule over those scores, and every decision leaves an audit trail: which candidate won, what each value net scored it, and why the losers lost. Values become a replaceable, inspectable component of the agent rather than a property of its weights. A prototype, Bei, demonstrates the mechanism using LLM-as-judge scaffolding; we describe what it showed, what it got wrong, and what a production version requires.