[KongchangAI]
Conceptmonotone mapping

Monotone Mapping

An order-preserving functional transformation that remaps inputs to outputs according to a monotone relation. In safety alignment it is used to calibrate systematic mean bias in scorers, requiring only order preservation so as to remain invariant to arbitrary order-preserving distortions.

Timeline (last 90 days)

Nov 13

In the dual-channel constrained GRPO, the base reward channel retains the original scores for in-group comparison, the penalty term and constraint estimate are passed through a monotonic mapping fitted on judge-labeled training responses, and the multiplier is retained in the advantage function, enabling the safety constraint to affect gradient updates.

Unverified50%
Nov 13

A monotonic mapping is an order-preserving functional transformation that only requires monotonicity, and thus is invariant to any order-preserving distortion of the scorer; it is fitted on judge-labeled training responses to correct the scorer's systematic mean bias.

Unverified50%
Nov 13

After applying order-preserving distortion to the scorer, the predicted controller state was reproduced under every random seed; the results also held under a mapping fitted to another judge's labels, an image-reading judge, and manual auditing.

Unverified50%

All Facts (3)

Source Articles