OmniVerifier-C1: When Verifiers Stop Saying Right/Wrong and Start Scoring
Imagine teaching a child chess. If you only say "that move was right" or "that move was wrong," they learn little beyond rigid rules. But if you say "good move, but your knight blocked your rook," they learn far more. That's the core problem OmniVerifier-C1 addresses: in multimodal AI, verifiers act as judges of whether generated images, videos, or scenes meet requirements — and traditionally they only flash red or green cards: True or False.
From Binary Judge to Graded Scorer
OmniVerifier-C1's key innovation is upgrading the verifier from binary judgment to graded scoring, via two mechanisms:
1. Automatic decomposition. Given a complex multimodal prompt — e.g., "generate a living room with a red sofa, blue curtains, and light coming from the left" — OmniVerifier doesn't output a single True/False. It splits the request into sub-conditions (red sofa? blue curtains? left-side lighting?) and scores each one separately.
2. Decoupled training. The team introduces a "decoupled reinforcement learning" strategy optimizing two objectives simultaneously: sub-question-level binary judgment and global alignment scoring. Local feedback grounds fine-grained verification; the global score provides a continuous rating on top.
It's like an essay grader: first checking each paragraph for errors (local binary judgment), then assigning an overall score (global continuous rating). The two feedback levels are complementary, not substitutive.
Why Binary Verification Isn't Enough
The paper analyzes this theoretically: with reward signals of only 0 and 1, a model can't distinguish "almost right" from "completely off track" — like telling a student only "fail" without knowing if they scored 59 or 9. Graded rewards provide dense gradient signals. Experiments confirm this: in traditional text-to-image tasks, graded rewards significantly outperform binary rewards and other continuous reward models.
Rule-Governed World Models
Text-to-image is only the warm-up. OmniVerifier's real ambition is "rule-governed generative world modeling" — AI that not only generates content but understands how the world works (gravity, collision, support relations). The verifier must check not just "does it look good" but "are the physics and causal rules correct." Two verification modes are introduced:
- Pointwise Verification: validating each rule individually
- Pairwise Verification: comparing which of two generations better satisfies the rules
- Text-to-image: graded rewards > binary rewards > other continuous rewards
- World modeling: pairwise > pointwise verification
- Out-of-domain generalization: structured outputs significantly boost generalization
- SFT provides domain-specific execution priors, helping verifiers adapt quickly to new scenarios
The paper shows pairwise verification outperforms pointwise verification in this setting — mirroring human cognition: we struggle to rate a dish 8.7/10 but can easily say which of two restaurants is better.
Cross-Domain Generalization
The most striking finding: structured, interpretable verification outputs improve out-of-domain generalization. Once a verifier learns to decompose complex requirements and score sub-conditions, it gives reasonable judgments even on novel task types — suggesting a generalizable verification capability, like a student who tackles any new question type by "decompose first, then judge."
Technical Details
Training uses both text-to-image and rule-governed scenario data, built on standard RL (GRPO) with the decoupled strategy — local binary judgments and global scores optimized separately, then jointly. Experimental findings:
Commentary
The core insight is simple: finer feedback enables more effective learning. The contribution is turning that into an operational framework — decompose, score, decouple training.
The "binary vs. graded" framing applies beyond multimodal verification: AI safety alignment is often judged with binary safe/unsafe labels, but decomposing safety requirements into scored sub-conditions could surface subtler risks.
More deeply, OmniVerifier hints at a philosophy: judging "how good" is closer to real intelligence than judging "right or wrong." The world isn't black-and-white. A truly rule-understanding AI should say "this scene satisfies 90% of the rules but violates one physical detail" — not simply "wrong."
From binary judge to graded coach: the necessary evolution of verifiers.
---
Paper: https://arxiv.org/abs/2610.22355 Code: https://github.com/Cominclip/OmniVerifier