V-Rubrics: Local credit from multidimensional visual rubrics¶
Deterministic mechanism mini-suite; not paper benchmark reproduction.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2608.25580 |
| Institution | S-Lab, Nanyang Technological University (first-listed first-author affiliation). |
| First public release | 2026-08-26 (arXiv v1) |
| Author code | Author project. |
| Local adapter or method key | v-rubrics |
| Local reproduction code | src/auto_research/post_training/latest_20260827.py |
Paper method and evidence¶
Decompose a reference answer into atomic rubric criteria for visual fidelity, reasoning consistency and instruction following. Attribute evidence to the relevant prefix so a terminal scalar does not hide local hallucinations.
flowchart LR
A[Reference answer] --> R[Atomic visual, reasoning and instruction criteria]
R --> L[Localize supporting evidence]
L --> C[Prefix-level credit]Original paper figure¶
The original figure shows localized rubric credit. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
The authors report +1.79 average points over answer-only GRPO and +4.00 visual reasoning points over SFT. No industrial A/B.
Local reproduction¶
auto-research post-train --algorithm v-rubrics --dataset arithmetic-smoke --steps 100 --seed 42 --offline
Arithmetic-smoke, 100 steps: accuracy 0.1250→0.2188 (+75.00%). Only multidimensional prefix-localized candidate credit is tested. Metrics, batch index. No original checkpoint, private data, external service or full benchmark.