简体中文 · English

Trajectory Learnability for Offline On-Policy Distillation

L1 core mechanism diagnostic. Per-token learnability aggregation and trajectory weighting run. The paper's GPU training budget was not reproduced.

Paper information

Field	Details
Paper	Trajectory Learnability for Offline On-Policy Distillation
Institution	National University of Singapore (first author affiliation)
First public release	2026-09-16 (arXiv v1)
Original source code	No public author repository found as of 2026-09-19.
Adapter/method	trajectory-learnability
Local reproduction code	src/auto_research/post_training/latest_20260919.py

Original paper summary

Background and main changes

Train a reference model on successful trajectories. Changes in per-token log likelihood between that reference and the current policy estimate learnability, which reweights offline OPD trajectories.

flowchart LR
  I[Public input and current state] --> M[Learnability and trajectory weighting]
  M --> A[Auditable intermediates]
  A --> O[Output or state update]

Key figure from the paper

Figure 2: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core mechanism

The local reference kernel retains defining gates, weights, state transitions or scheduling rules and writes intermediates to metrics. See implementation functions and test invariants for variables and formulas.

Offline and online results

The paper's deployment, benchmark, training-efficiency and inference-speed figures are author results. The local three-seed artifact checks mechanisms, shapes and state invariants, not paper-scale performance.

Scope

Per-token learnability aggregation and trajectory reweighting run locally. The paper's GPU budget was not reproduced.