VBVR-Pro: Executable Visual Reasoning Verifiers¶
Core mechanism evidence. Paper and local evaluation are separate.
Paper information¶
| Field | Value |
|---|---|
| Paper | arXiv 2608.26105 |
| Affiliations | Nanyang Technological University / VBVR Community |
| First public release | 2026-08-26 (arXiv v1) |
| Authors' code | Author project and code. |
| Adapter / method | vbvr-pro |
| Local code | src/auto_research/reproductions/vbvr_pro/ |
Original paper¶
Background and contribution¶
A fluent response can mislead a general VLM judge. Per-task executable scorers check intermediate states, constraints, and final states to provide rewards for multimodal generators.
flowchart LR
I[Task and trajectory] --> C[Deterministic state and constraint checks] --> R[Verifiable reward]Key figure from the paper¶
Original Figure 2: VBVR-Pro. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\[
R(\tau)=\sum_j\lambda_j\mathbf 1[c_j(\tau)=\mathrm{true}],\qquad \max_\theta\mathbb E_{\tau\sim\pi_\theta}[R(\tau)].
\]
Paper results¶
Human vote agreement exceeds 0.60 versus GPT-5.5 at 0.54. Across nine trained models the paper reports overall mean gain 0.290. No industrial A/B.
Local reproduction¶
Three deterministic verifiers versus noisy scalar VLM-judge analogue on 300 procedural tasks: mean reward 0.5942→1.0000 (+68.30%). Metrics.
Reproduction limits¶
Verifies deterministic scoring and explainable failures, not seven external benchmarks or 14B/19B training.