Idea Ripper // ArchiveCards: 1555 [Live] | Updated: 2026-10-14
Card #c26IR // #c26Synthesis

Capturing all of these dimensions in a single scalar reward is challenging; in practice, institutions often train separate reward models for different preference dimensions and combine their signals.

— In our words · Source: Large Language Models in Finance (Miquel Noguer i Alonso)

[Use when]

When reviewing an RLHF setup for a compliance-sensitive task, ask how many reward dimensions are trained separately — one scalar for everything is a red flag.

[Why it matters]

A single scalar forces rigor, calibration, tone, compliance, and readability to trade off against each other inside one number — so the model learns whichever dimension is easiest to game. Separate reward models per dimension keep the trade-offs visible and auditable.

Related principles

#c18The central purpose of governance is not to prevent the use of LLMs. It is to make valuable uses repeatable while keeping failure within…Use when a governance or model-risk proposal is being dismissed as bureaucracy that slows the business down.

More from Large Language Models in Finance

#c13Prediction and agency should not be conflated. A credit model can estimate the probability that a borrower breaches a covenant. An agentic…Use when a team proposes a stronger model as the fix for an agent that lacks memory, permissioning, or an… #c14The technically and institutionally mature objective is bounded autonomy : the agent can plan, retrieve, calculate, and propose actions…Use when deciding what an AI agent is permitted to do in production: which actions stay read-only, which need… #c15This principle prevents a common category error. A model that says it will respect a mandate has not yet respected the mandate. A model…Use when a vendor demo, benchmark score, or red-team chat log is being offered as proof that an agent is safe…