V-Zero: Zero-Annotation Multimodal Reasoning

• V-Zero is a naming motif for zero-label visual reasoning systems that eliminate external supervision through self-improving frameworks.
• It employs a co-evolving Questioner–Solver loop with majority voting and GRPO to enhance performance on vision-centric benchmarks.
• An alternative approach uses on-policy distillation with contrastive evidence gating to achieve fine-grained reasoning without answer labels.

Searching arXiv for papers relevant to “V-Zero” and closely related usages. V-Zero is a label used in recent arXiv literature for multiple, distinct multimodal-learning frameworks rather than a single canonical method. In one usage, it denotes a self-improving post-training framework for vision–LLMs that uses exclusively unlabeled images, zero human annotations, a co-evolving Questioner–Solver loop, and Group Relative Policy Optimization (GRPO) (Wang et al., 15 Jan 2026). In another, it denotes an answer-label-free framework for fine-grained visual reasoning that augments on-policy distillation with contrastive evidence gating derived from positive and negative visual views (Sun et al., 24 Jun 2026). Closely related naming conventions also appear in “Vision-Zero,” a gamified self-play framework for VLM self-improvement, while TEST-V explicitly notes that it does not introduce a method explicitly named “V-Zero” (Wang et al., 29 Sep 2025, Yan et al., 1 Feb 2025).

1. Nomenclature and scope

The current research usage of “V-Zero” is heterogeneous. The term appears in at least two 2026 multimodal-reasoning papers with different training regimes, supervision assumptions, and optimization targets, while adjacent papers use related names for distinct objectives.

Usage	Paper	Core setting
V-Zero	“V-Zero: Self-Improving Multimodal Reasoning with Zero Annotation” (Wang et al., 15 Jan 2026)	Zero-human-annotation post-training from unlabeled images
V-Zero	“V-Zero: Answer-Label-Free On-Policy Distillation with Contrastive Evidence Gating for Fine-Grained Visual Reasoning” (Sun et al., 24 Jun 2026)	Answer-label-free visual reasoning via gated on-policy distillation
Vision-Zero	“Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play” (Wang et al., 29 Sep 2025)	Strategic self-play from arbitrary image pairs

A separate paper titled “V-ZOR” concerns verifiable cross-blockchain communication via quantum-driven ZKP oracle relays and is unrelated to multimodal learning (Haider et al., 13 Sep 2025). TEST-V, although discussed in the broader context of zero-shot video classification and “V-zero shot,” explicitly states that it “does not mention a method explicitly named ‘V-Zero’” (Yan et al., 1 Feb 2025). This suggests that “V-Zero” currently functions more as a naming motif around zero-label or zero-annotation visual learning than as a stable technical designation.

2. Zero-annotation V-Zero: co-evolving Questioner and Solver

In “V-Zero: Self-Improving Multimodal Reasoning with Zero Annotation,” V-Zero is a self-improving post-training framework for vision–LLMs that needs zero human annotation and learns directly from raw, unlabeled images by instantiating two roles—Questioner and Solver—that co-evolve in a closed loop (Wang et al., 15 Jan 2026). The Questioner $Q_\theta$ perceives an input image $I$ and produces a visual description $d$, a multiple-choice question $q$ with four options, and an “intuitive” direct answer $a_{\text{fast}}$. The Solver $S_\theta$ receives $I$ and $q$, samples step-by-step chain-of-thought responses, and induces a pseudo-label by majority voting over its own sampled answers.

The framework alternates between Questioner training and Solver training. During Questioner training, a frozen Solver samples $m$ reasoning answers, majority voting yields a pseudo-label $\hat a$ and confidence $I$0, and the Questioner receives a dual-track reasoning reward that formalizes the contrast between intuition and reasoning. The reward is

$I$1

When $I$2, the reward peaks at $I$3, thereby targeting the Solver’s decision boundary. When $I$4, the reward favors cases in which intuition is wrong but careful reasoning confidently corrects it. Invalid multiple-choice formatting is penalized by setting $I$5.

During Solver training, the updated Questioner generates questions on new images, the Solver again samples $I$6 responses, majority vote yields $I$7 with vote proportion $I$8, and training keeps only the “mid-confidence” subset

$I$9

The Solver then uses binary accuracy rewards $d$0 and group-relative advantages via z-score normalization inside GRPO. The shared optimization template is a clipped GRPO objective with KL regularization,

$d$1

Implementation details are explicit. The backbone models are Qwen2.5-VL-3B-Instruct and Qwen2.5-VL-7B-Instruct; the Questioner and Solver are separate instances initialized from the same base weights; sampling uses temperature $d$2; majority vote uses $d$3; group sizes are $d$4 for the Questioner and $d$5 for the Solver; $d$6; the learning rate is $d$7; the global batch size is $d$8; vLLM is used as the rollout engine; FSDP with parameter offloading is used for memory; and each iteration takes $d$9 hours. Training uses 4,000 unlabeled images selected from OpenVLThinker GRPO-hard and GRPO-medium.

3. Empirical behavior of the zero-annotation framework

The zero-annotation V-Zero is evaluated on general vision-centric benchmarks—MMMU and MMStar—and visual mathematical reasoning benchmarks—MathVision (Test-mini), MathVerse (Vision-Only), MathVista (Test-mini), and LogicVista (Wang et al., 15 Jan 2026). On Qwen2.5-VL-7B-Instruct, the reported average scores are Base $q$0, Supervised GRPO $q$1, V-Zero Iter 1 $q$2, and V-Zero Iter 2 $q$3. The abstract summarizes the headline gains as $q$4 on visual mathematical reasoning and $q$5 on general vision-centric evaluation. Per-dataset numbers at Iter 2 are MMMU $q$6 $q$7 vs Base$q$8, MMStar $q$9 $a_{\text{fast}}$0, MathVision $a_{\text{fast}}$1 $a_{\text{fast}}$2, MathVerse $a_{\text{fast}}$3 $a_{\text{fast}}$4, MathVista $a_{\text{fast}}$5 $a_{\text{fast}}$6, and LogicVista $a_{\text{fast}}$7 $a_{\text{fast}}$8.

The 3B setting exhibits a different scaling profile. Qwen2.5-VL-3B-Instruct moves from Base Avg. $a_{\text{fast}}$9 to V-Zero Iter 1 $S_\theta$0 and Iter 2 $S_\theta$1, with notable early gains on MMMU $S_\theta$2 and MathVision $S_\theta$3, followed by faster saturation. The paper attributes this to smaller capacity and higher sensitivity to exploration noise.

Ablations identify three structural dependencies. Freezing the Questioner causes severe degradation, with Math Avg. $S_\theta$4 and General Avg. $S_\theta$5, indicating that dynamic co-evolution is essential. Removing the dual-track reward and using only uncertainty reduces performance by $S_\theta$6 on general and $S_\theta$7 on math average, indicating that explicit intuition–reasoning divergence is key to eliciting hard questions. Removing difficulty-guided filtering causes $S_\theta$8 on general and $S_\theta$9 on math average, indicating that the $I$0 filter stabilizes RL and focuses learning on “mid-confidence” samples.

The framework also tracks question quality evolution. The valid format rate rises from Base $I$1 to Iter 1 $I$2 to Iter 2 $I$3, while a difficulty score judged by Qwen3-VL-32B rises from $I$4 to $I$5 to $I$6. The authors interpret this as a self-generated curriculum that becomes harder across iterations. Limitations remain explicit: smaller models saturate faster; noisy pseudo-labels are not eliminated by filtering; very complex scenes or under-specified questions can lead to low confidence; and broader domain generalization may require more diverse unlabeled images.

4. Answer-label-free V-Zero: on-policy distillation with contrastive evidence gating

In “V-Zero: Answer-Label-Free On-Policy Distillation with Contrastive Evidence Gating for Fine-Grained Visual Reasoning,” the central problem is fine-grained visual reasoning, where an MLLM must identify task-relevant visual evidence and ground reasoning in local image regions rather than global gist (Sun et al., 24 Jun 2026). The paper revisits On-Policy Distillation (OPD) as “negative-free stop-gradient alignment.” In that formulation, a student policy $I$7 aligns to a frozen teacher distribution on the student’s own prefixes, which yields dense token-level correction but no trajectory-level discrimination.

The proposed V-Zero augments OPD with trajectory-level discrimination derived from visual evidence while remaining answer-label-free. For each student-sampled trajectory, the teacher is replayed under a positive view $I$8 and a negative view $I$9. The positive view is a question-relevant regional crop provided by the training set curated by Zooming without Zooming (ZwZ). The negative view is constructed by downsampling the full image by $q$0 and randomly sampling an equal-size crop from a region that does not overlap the positive region. The teacher-side token scores are

$q$1

with token-level evidence gap

$q$2

and trajectory-level evidence score

$q$3

V-Zero samples $q$4 sibling rollouts for the same prompt, normalizes the evidence scores within the sibling group, and defines the gate

$q$5

with $q$6 and $q$7 in experiments. Larger gates amplify OPD updates for trajectories more strongly supported by the positive evidence relative to their siblings; smaller gates suppress weakly grounded rollouts.

The resulting objective is a trajectory-weighted positive-view reverse-KL:

$q$8

The sampled-token surrogate likewise multiplies per-token OPD terms by the stop-gradient gate. No explicit contrastive loss is used; the negative view only enters through $q$9 and the resulting $m$0.

The main implementation uses a Qwen3.5-4B student and a frozen Qwen3.5-27B teacher, with Qwen3.5-9B used in ablations. Training is built on VeRL (HybridFlow) with its OPD reverse-KL estimator and sampled-token surrogate. No additional localization module is introduced, and the crops are used only for teacher-side replay during training, not at inference.

5. Empirical profile of contrastive evidence gating

The answer-label-free V-Zero is evaluated on VStar, HR-Bench (HR-4K and HR-8K), ZoomBench, MME-RealWorld, and MMStar (Sun et al., 24 Jun 2026). Under the authors’ re-evaluation, the Qwen3.5-4B backbone records VStar $m$1, HR-4K $m$2, HR-8K $m$3, ZoomBench $m$4, MME-RW $m$5, MMStar $m$6, and Avg. $m$7. V-Zero-4B improves these to VStar $m$8, HR-4K $m$9, HR-8K $\hat a$0, ZoomBench $\hat a$1, MME-RW $\hat a$2, MMStar $\hat a$3, and Avg. $\hat a$4, corresponding to gains of $\hat a$5, $\hat a$6, $\hat a$7, $\hat a$8, $\hat a$9, $I$00, and $I$01 on average.

Training data consists of 23K high-quality examples from ZwZ, each providing a full image, a question, and a question-relevant regional crop $I$02. The method adds a single negative crop $I$03 per example using the $I$04 downsample plus equal-size random crop outside $I$05. No annotated answers or reasoning traces are used.

The efficiency profile is emphasized quantitatively. V-Zero trains on $I$06 RTX PRO 6000 96G in $I$07 hours. The paper reports ZwZ on $I$08 H100 for approximately one day and DeepEyes on $I$09 H100 for approximately two days, yielding the claim that V-Zero is more than $I$10 faster than previous supervised fine-tuning methods and more than $I$11 faster than reinforcement learning baselines. A central reason is that there is no reward design, no external answer labels, and no separate textual supervision; evidence gating is computed from teacher log-probabilities under two fixed crops and acts only as a scalar per trajectory.

Ablations show that the gate matters. On Perception Avg. over VStar, HR-4K, HR-8K, and ZoomBench, “None (no gate)” gives $I$12, “Random crops for both pos/neg” gives $I$13, and V-Zero with true $I$14 and random $I$15 gives $I$16. Rollout group size also matters: $I$17 gives $I$18, while $I$19 gives $I$20. Training-step sensitivity is non-monotone: step $I$21 gives $I$22, step $I$23 gives $I$24, step $I$25 gives $I$26, step $I$27 gives $I$28, step $I$29 gives $I$30, and step $I$31 gives $I$32.

The limitations are specific. The method assumes training-time access to question-relevant crops $I$33; if those crops are noisy or misaligned, gating weakens. The gate inherits teacher preferences, so teacher miscalibration can up- or down-weight trajectories suboptimally. A single random negative crop may be too easy or too hard for some cases. HR-8K shows smaller gains, which the paper associates with diminishing returns when the full image already carries rich detail.

6. Related methods, misconceptions, and broader significance

A recurrent misconception is to treat “V-Zero” as a single technical framework. The evidence does not support that reading. The zero-annotation V-Zero of (Wang et al., 15 Jan 2026) is a Questioner–Solver co-evolution system trained by GRPO on unlabeled images; the answer-label-free V-Zero of (Sun et al., 24 Jun 2026) is a gated OPD method for fine-grained visual reasoning. Their commonality lies in eliminating external supervision, but their optimization geometry, data assumptions, and training signals are different.

The naming field also includes adjacent but distinct methods. “Vision-Zero” is a domain-agnostic framework enabling VLM self-improvement through competitive visual games generated from arbitrary image pairs, with a “Who Is the Spy”-style game, verifiable rewards, and Iterative Self-Play Policy Optimization alternating between Self-Play and reinforcement learning with verifiable rewards (Wang et al., 29 Sep 2025). On Qwen2.5-VL-7B, it reports reasoning/math average improvements from Base $I$34 to $I$35 on CLEVR and Real-World training and to $I$36 on Chart training, together with chart/OCR and vision-centric gains. TEST-V, by contrast, is a training-free test-time support-set tuning method for zero-shot video classification; it explicitly states that it does not introduce a method explicitly named “V-Zero,” even though it situates itself within the broader context of “V-zero shot” or “V-ZS” (Yan et al., 1 Feb 2025).

A plausible implication is that the current “V-Zero” lineage marks a broader post-training trend in multimodal research: replacing externally labeled answers, reasoning traces, or human-authored curricula with internally verifiable signals, self-generated curricula, or contrastive evidence. Within that trend, the 2026 V-Zero papers occupy complementary positions. One uses a co-evolutionary curriculum built from MCQ generation, majority voting, and GRPO; the other inserts trajectory-level evidence discrimination into OPD without answer labels. The term is therefore best understood not as a single method family with a unified formalism, but as a convergent naming convention for zero-label visual reasoning systems.