Grounding the Score: Explicit Visual Premise Verification for Reliable Vision-Language Process Reward Models
EVPV gates PRM rewards by visual-premise reliability, reaching 67.46% Macro-F1 and improving Best-of-8 reranking.
Key Findings
Methodology
Explicit Visual Premise Verification (EVPV) separates perception from reasoning. A policy emits step-wise visual claims; a constraint extractor independently produces structured image facts C; type-aware matching yields claim supports p_j, aggregated into reliability r with a smoothed geometric mean. Visually dependent rewards are gated by α(r)=σ(β(r−τ)), attenuating unreliable evidence while preserving rewards when grounding is strong.
Key Results
- On VisualProcessBench, EVPV-PRM achieves 67.46% overall Macro-F1, exceeding VisualPRM at 62.00%, TIM-PRM at 61.70%, and QWEN-VL-PRM-7B at 58.60%.
- With InternVL2.5-8B, 26B, and 38B policies, EVPV-PRM obtains overall Best-of-8 accuracies of 41.67%, 46.75%, and 55.22%, corresponding to gains of 8.83, 9.52, and 9.78 points over Pass@1.
- Ablations show that removing structured facts reduces Macro-F1 to 63.38%; shuffling fact structure reduces it to 59.82%; deleting facts gives 62.11%; severe drop-facts corruption gives 35.77%, supporting a causal role for constraint fidelity.
Significance
The work addresses a central ambiguity in multimodal verification: a low step score may indicate faulty reasoning or faulty visual grounding. EVPV makes these sources separable, reducing both rewards for hallucinated premises and penalties for correct grounded claims. Because constraints are extracted once per image-question pair and reused across candidates, the framework is practical for Best-of-N reranking, search decoding, and error localization. Its plug-in form also improves black-box judges such as GPT and Gemini without retraining them.
Technical Contribution
The main contribution is judge-agnostic inference-time calibration rather than a replacement PRM architecture. The paper formalizes the visual-dependency indicator ν_t, constraint set C, support score p_j, geometric reliability r, and gated reward R_t. EVPV-PRM, based on Qwen2.5-VL-Instruct-7B, outputs probabilistic base rewards, while external binary judges can also be attached. The extractor uses a unified JSON schema for numeric readings, relations, and compositional structure, computed once and shared across traces.
Novelty
Unlike VisualPRM and ATHENA, which emphasize PRM training, and TIM-PRM, which performs per-step tool queries, EVPV explicitly exposes the visual premises on which steps depend and verifies them against reusable structured evidence. The fundamental novelty is therefore an interpretable reliability interface for reward calibration, not merely a stronger visual recognizer or a richer judge prompt.
Limitations
- The policy must report its visual dependencies. Omitted premises bypass gating, while over-reporting can suppress useful rewards; completeness is audited only on a human-verified subset.
- The constraint extractor is distilled from Qwen3-VL-235B-A22B-Instruct, so extraction errors or teacher bias directly distort reliability. Generalization to open-world images, fine-grained OCR, and non-geometric tasks remains uncertain.
- Matching tolerances and gating parameters β and τ require calibration; confidently incorrect constraints can preserve erroneous rewards.
Future Work
Promising directions include joint training of policy, extractor, and verifier; automatic discovery of visual dependencies; task-adaptive matching tolerances and gating; and probabilistic constraints with calibrated uncertainty. Broader evaluation should cover VLRMBench, multi-image and video reasoning, open-world perception, and comparisons with tool-augmented verifiers under standardized cost and latency measurements.
AI Executive Summary
Vision-language models can solve diagrams, tables, and mathematical figures, yet a single perceptual mistake may poison an otherwise fluent chain of reasoning. Conventional vision-language process reward models assign scores to intermediate steps but cannot tell whether a low score reflects faulty logic or faulty visual grounding. They may reward a confident claim about a nonexistent cylindrical hole, or reject a correct statement because the verifier misread the image. This ambiguity weakens Best-of-N selection and error diagnosis.
Wang and colleagues introduce Explicit Visual Premise Verification, or EVPV. The policy first produces a visual checklist for each reasoning step. In parallel, a constraint extractor converts the image and question into structured JSON facts covering numbers, geometric relations, and compositional structure. EVPV matches checklist claims to these facts, computes a reliability score r, and applies α(r)=σ(β(r−τ)) to rewards for visually dependent steps. Constraints are extracted once and reused across candidates, avoiding expensive per-step tool calls. The authors instantiate the framework as EVPV-PRM, built on Qwen2.5-VL-Instruct-7B.
The system reaches 67.46% overall Macro-F1 on VisualProcessBench, versus 62.00% for VisualPRM. Across InternVL2.5-8B, 26B, and 38B policies, EVPV-PRM achieves 41.67%, 46.75%, and 55.22% Best-of-8 accuracy, improving over Pass@1 by 8.83, 9.52, and 9.78 points. Shuffling structured facts lowers Macro-F1 to 59.82%, while severe fact deletion reduces it to 35.77%, indicating that gains depend on evidence fidelity rather than prompt artifacts. EVPV also improves several external judges, suggesting a reusable inference-time calibration layer.
Deep Analysis
Background
Test-time scaling increasingly uses Best-of-N sampling, search, and process supervision. VisualPRM400K and VisualProcessBench established multimodal step verification, while ATHENA improved data efficiency and TIM-PRM used tool-based evidence queries. Nevertheless, perception errors in OCR, counting, diagrams, and geometry remain common. Existing PRMs collapse visual grounding and symbolic reasoning into one opaque score.
Core Problem
Given an image I, question q, and trace S, a verifier must judge each step and rank candidate solutions. Standard PRMs directly compute R_t, without knowing whether a step relies on an unreliable visual fact. This causes hallucinated premises to receive positive scores and correct grounded claims to be rejected. Per-step tool calls can help, but are expensive for long traces and large Best-of-N pools.
Innovation
EVPV contributes three linked ideas: explicit step-level visual declarations d_t; a one-shot extractor for numeric, relational, and compositional constraints; and a reliability gate that converts checklist-evidence consistency into reward calibration. It is judge-agnostic and can wrap a trained PRM or an external VLM/LLM. This differs from TIM-PRM’s repeated tool queries and from training-only improvements such as VisualPRM or ATHENA.
Methodology
- �� The policy generates steps s_t and declarations d_t; ν_t=I[d_t≠null], and non-null declarations form V.
- �� The extractor computes C=E_φ(I,q) under a unified JSON schema covering lengths, angles, table entries, parallelism, perpendicularity, equality, incidence, containment, and attachments.
- �� Type-aware matching gives p_j=m(v_j,C), followed by r=exp((1/M)Σlog(ε+p_j)).
- �� A judge supplies R_t^base. EVPV-PRM uses u_t=P_θ(y_t=1|I,q,s≤t) and R_t^base=2u_t−1.
- �� Visual steps receive R_t=α(r)R_t^base; non-visual steps remain unchanged. Candidates are ranked by the geometric mean of shifted positive rewards.
- �� E_φ is trained on 26,454 image-question pairs with SFT and 5,832 preference instances with DPO; V_θ uses 19,490 step-labeled trajectories.
Experiments
Step verification uses VisualProcessBench with Macro-F1 and accuracy. Downstream reranking uses LogicVista, MMMU, MathVerse-VO, MathVision, MathVista, and WeMath. InternVL2.5-8B, 26B, and 38B policies generate candidates; evaluation reports Pass@1, BoN@8, and gains. Baselines include VisualPRM, TIM-PRM, QWEN-VL-PRM-7B, and GPT, Gemini, and Qwen judges. Ablations remove structured facts, constraints, vision, JSON, or inject controlled corruption.
Results
EVPV-PRM obtains Macro-F1 scores of 69.57, 68.86, 67.09, 65.27, and 69.11 on DynaMath, MMMU, MathVerse, MathVision, and WeMath, averaging 67.46%. On InternVL2.5-38B, it reaches 55.22% BoN@8 versus 50.70% for VisualPRM. Adding EVPV to Qwen3-VL-235B raises overall judge Macro-F1 from 58.51% to 65.45%. Removing vision and JSON reduces overall Macro-F1 to 48.23%, confirming the importance of explicit visual evidence.
Applications
The framework suits diagram mathematics, table question answering, chart analysis, and educational answer selection. A single extracted evidence set can be reused across many candidates, making it attractive for batch reranking. Organizations can also attach EVPV to black-box GPT, Gemini, or internal VLM judges to reduce visual hallucination effects without retraining the judge.
Limitations & Outlook
Performance depends on complete policy declarations and accurate extraction. The training distribution emphasizes geometry and tables, while open-world, multi-image, video, and difficult OCR settings remain under-tested. Constraint extraction still incurs inference cost, and matching tolerances plus β and τ require tuning. Future work should develop joint calibration, automatic dependency discovery, probabilistic constraints, and broader benchmarks with standardized latency and cost analysis.
Plain Language Accessible to non-experts
Imagine a teacher grading a mathematics worksheet. A student writes a polished solution, but secretly relies on claims such as “the picture shows three apples” or “this line is perpendicular to that one.” A conventional teacher mainly checks whether the later reasoning sounds coherent. If the student confidently invents an apple, the answer may receive too much credit.
EVPV asks the student to list every important fact taken from the picture. A second teacher independently inspects the original image and writes down confirmed facts: numbers, positions, connections, and shapes. The system compares the two lists. If a crucial claim is unsupported, the whole solution becomes less trustworthy; if the claims agree, the normal reasoning score is preserved. One serious visual mistake matters because it can contaminate every later step.
EVPV does not automatically declare the entire answer wrong. It mainly reduces the influence of steps that depend on questionable picture facts, while leaving pure arithmetic or language steps alone. Since the image is checked once, the same evidence can be reused for many candidate answers. In the paper, this approach reaches 67.46% Macro-F1 on VisualProcessBench and improves Best-of-8 selection across three InternVL2.5 sizes.
The method still depends on the second teacher seeing correctly and on the student reporting relevant facts. If the evidence list is wrong, the reliability judgment can also be wrong. The corruption experiments are important because performance falls when facts are shuffled or removed, showing that accurate evidence—not merely a new instruction—is responsible for the gains.
ELI14 Explained like you're 14
Imagine playing a puzzle game with a screenshot full of buttons, arrows, and numbers. You write a walkthrough: “There are three buttons on the right, so press the middle one.” But what if you actually saw two buttons? Every later move might sound clever while starting from the wrong map. A normal scorekeeper may be fooled because the walkthrough is written confidently.
EVPV is like bringing a careful teammate into the game. First, you mark which steps depend on the screenshot: the number of buttons, their positions, or which lines touch. Then your teammate independently checks the screenshot and makes a fact list. If your claim matches the list, the step keeps its influence. If it does not, the system turns down that step’s score instead of blindly trusting it.
Why is that useful? Suppose the model creates eight possible walkthroughs. EVPV helps prevent a confidently invented map from winning just because the writing sounds smooth. In the experiments, EVPV-PRM scored 67.46% Macro-F1 on VisualProcessBench. With InternVL2.5 models, Best-of-8 performance improved over a single attempt by 8.83 to 9.78 percentage points.
It is not magic, though! If the teammate misunderstands the screenshot, or if you forget to mention an important visual fact, the check can fail. The researchers deliberately shuffled and deleted facts, and performance dropped sharply. That is good scientific evidence: the system works because accurate visual facts guide the score, not simply because someone added a longer prompt.
Glossary
Vision-Language Process Reward Model
A model that evaluates intermediate steps in multimodal reasoning rather than only the final answer. Technically, it produces step-level correctness rewards for process supervision.
EVPV calibrates VL-PRM scores when steps depend on uncertain visual premises.
Explicit Visual Premise Verification
A framework that exposes visual facts used by each step and checks them against independently extracted image evidence. It produces a scalar grounding-reliability signal.
EVPV is the paper’s central inference-time framework.
Structured Constraints
Machine-readable visual facts represented in a unified JSON schema, including values, relations, and composition. They are easier to match and reuse than free-form descriptions.
The extractor generates C once per image-question instance.
Reliability Gating
A mechanism that scales rewards according to evidence reliability. Low reliability moves visually dependent rewards toward neutrality, while high reliability preserves the base reward.
The paper uses α(r)=σ(β(r−τ)).
Best-of-N
A test-time strategy that samples N candidate solutions and selects one using a verifier. It separates candidate generation quality from selection quality.
The main downstream setting is Best-of-8 reranking.
Constraint Fidelity
The degree to which extracted constraints match the actual visual facts in an image. High fidelity is necessary for reliable checklist matching.
Corruption experiments test whether EVPV gains depend on fidelity.
Open Questions Unanswered questions from this research
- 1 Reliable discovery of every visual dependency remains unsolved. Omission bypasses gating, while over-reporting can suppress useful rewards; stronger supervision or joint learning is needed.
- 2 Constraint fidelity in open-world images, video, multi-image reasoning, and difficult OCR is unknown. Probabilistic evidence and broader calibration benchmarks are needed.
- 3 Fixed β, τ, and matching tolerances may not transfer across tasks. Automatic task-adaptive gating remains an open direction.
Applications
Immediate Applications
Multimodal mathematics and chart reranking
Education systems can generate multiple diagram or table solutions, extract shared visual facts once, and gate steps that depend on them. This supports inexpensive batch selection and highlights likely perception errors.
Calibration for black-box visual judges
Companies can attach EVPV to GPT, Gemini, or internal VLM judges using structured evidence. The judge need not be retrained; unsupported visual premises simply contribute less to reranking and diagnosis.
Long-term Vision
Evidence-grounded multimodal agents
Future agents could maintain reusable visual evidence across planning, robotics, and video reasoning. The key obstacles are cross-frame consistency, uncertainty representation, and trustworthy provenance for extracted constraints.
Abstract
Vision-language process reward models (VL-PRMs) are increasingly used to score intermediate reasoning steps and rerank candidates under test-time scaling. However, they often function as black-box judges: a low step score may reflect a genuine reasoning mistake or simply the verifier's misperception of the image. This entanglement between perception and reasoning leads to systematic false positives (rewarding hallucinated visual premises) and false negatives (penalizing correct grounded statements), undermining both reranking and error localization. We introduce Explicit Visual Premise Verification (EVPV), a lightweight verification interface that conditions step scoring on the reliability of the visual premises a step depends on. The policy is prompted to produce a step-wise visual checklist that makes required visual facts explicit, while a constraint extractor independently derives structured visual constraints from the input image. EVPV matches checklist claims against these constraints to compute a scalar visual reliability signal, and calibrates PRM step rewards via reliability gating: rewards for visually dependent steps are attenuated when reliability is low and preserved when reliability is high. This decouples perceptual uncertainty from logical evaluation without per-step tool calls. Experiments on VisualProcessBench and six multimodal reasoning benchmarks show that EVPV improves step-level verification and consistently boosts Best-of-N reranking accuracy over strong baselines. Furthermore, injecting controlled corruption into the extracted constraints produces monotonic performance degradation, providing causal evidence that the gains arise from constraint fidelity and explicit premise verification rather than incidental prompt effects. Code is available at: https://github.com/Qwen-Applications/EVPV-PRM