| --- |
| title: MR-9B Scaffold — 0/7 Trajectories |
| emoji: 🔬 |
| colorFrom: red |
| colorTo: indigo |
| sdk: static |
| app_file: index.html |
| pinned: false |
| license: mit |
| short_description: How MR-9B explores SODA-2026 problems 0-step FA can't solve |
| --- |
| |
| # MR-9B RL scaffold on the FA-0-zero SODA-2026 problems |
|
|
| A static trace viewer for the **44 SODA-2026 `standalone_hard` (proof_writing)** |
| problems that the RL policy `meta-learning-llms/MR_9B_rl_215_step45` scored **0/7 on |
| all 8 samples** with a **0-step final answer** — after excluding 9 rows with broken |
| (`Total: 0`) rubrics that score 0 by construction. |
|
|
| Each trajectory is the **full 8-layer MR/E/FA scaffold** (MR/FA = the RL policy, |
| E = frozen `MR_midtrain_9B_v3`). It shows, in order: |
|
|
| 1. the **problem** + its **ground-truth rubric + reference solution**, |
| 2. the **final answer**, its **judge score** (gpt-oss-120b), and the **judge's |
| reasoning** for that score, |
| 3. then the **MR/E exploration** — per-layer directions, planner CoT, assembled |
| frontier, and each executed direction's finding + summary. |
|
|
| With one attempt the scaffold rescues **13/44** of these (vs 0/8 for FA-0). |
|
|
