CompassOPD: Align relative teacher directions across model families

简体中文

L1 candidate-policy diagnostic, not full cross-tokenizer LLM training.

Paper information

Field	Details
Paper	arXiv 2609.10154
Institution	Institute of Information Engineering, Chinese Academy of Sciences (first author).
First public release	2026-09-09 (arXiv v1)
Author code	No public author repository found as of 2026-09-12.
Local adapter or method key	compass-opd
Local reproduction code	src/auto_research/post_training/latest_20260912.py

Paper method and evidence

Absolute probability scales differ across model families. Compare the post-trained teacher against its reference, then center the change within its family to remove a global offset while retaining the direction.

flowchart LR
  R[Teacher reference] --> D[Relative likelihood shift]
  T[Post-trained teacher] --> D
  D --> C[Within-family centering]
  C --> A[Student reference anchor]
  A --> O[OPD update]

Original paper figure

Figure 2 shows how relative teacher changes anchor the student update. From the original paper; copyright remains with the authors.

Core equation

Compute \(\Delta=z_{T_1}-z_{T_0}\) and center it as \(\Delta-\mathbb E[\Delta]\). The local diagnostic records the resulting correction as centered_cross_family_gap.

Local reproduction

Three-seed arithmetic diagnostic uses a shared finite action space. No cross-tokenizer mapping or full-parameter training.