ROSS: Relearning from Self-Generated Rollouts through Selective Supervision¶
L1 core-mechanism diagnostic. Checks the full-context/selective-loss boundary. A fixture mask is not the paper's selector.
Paper information¶
| Field | Details |
|---|---|
| Paper | ROSS: Relearning from Self-Generated Rollouts through Selective Supervision |
| Organization | AllSpark Team; the paper lists the team byline without a separate institutional affiliation. |
| First public release | 2026-09-28 (arXiv v1) |
| Authors' code | No public author repository was found as of 2026-09-30. |
| Adapter / method | ross |
| Local reproduction code | src/auto_research/post_training/latest_20260930_closure.py |
Original paper¶
Motivation and method¶
Keep the full context of historical self-rollouts but compute loss only on selected model continuations, avoiding imitation of erroneous or redundant steps.
flowchart LR
I[Public input and current state] --> M[ross mechanism]
M --> A[Auditable intermediates]
A --> O[Output or state update]Key figure from the paper¶
Figure 4: Overview of the paper's method and its key components. From the original paper; copyright remains with the authors. Click the image for the source.
Core computation¶
The reference kernel retains the paper's decisive gates, masks, recursion, credit weights, or promotion conditions. Tests also check gradient isolation, type boundaries, and determinism.
Paper results¶
Benchmarks, speed, and training curves belong to the paper. The local three-seed mini-suite checks mechanisms and invariants, not paper-scale results or online gains.
Local reproduction¶
See the three-seed diagnostic, marked diagnostic_only=true and ineligible for formal capability ranking.
Reproduction boundaries¶
The local run covers the full-context/selective-loss boundary only. Its selection mask is a fixture, not a reproduction of the paper's selector.