PaperlayerView exact revision on arXiv

Source-linked AI summary

MMVistaReason: Toward Open-Data and Post-Training Recipes for Multimodal Reasoning

Juekai Lin, Honglin Lin, Yuqian Yuan, Xiaolong Wu, Jie Cao, Liang Liang, Yunqi Cao, Yun Zhu, Wenqiao Zhang, Lijun Wu

arXiv:2610.01352v1cs.CV

TL;DR

Open multimodal reasoning post-training still faces uneven data quality, imbalanced capability coverage, and cross-domain interference. MMVistaReason constructs stage-specific supervision and trains complementary RL experts before integrating them with MOPD; its models achieve strong results across 15 benchmarks.

  • Problem

    Reliable multimodal post-training remains challenging because open data vary in quality and difficulty, real-world capabilities are underemphasized, and mixed-domain training can introduce negative transfer.

    Source: §1 · p. 2

  • Method

    MMVistaReason builds balanced Analytical and Real-World data, constructs capacity-aware SFT and scale-specific RL supervision, then integrates complementary RL experts through multi-teacher on-policy distillation.

    Source: §1 · pp. 2–3

  • Results

    72.8 average across 15 multimodal benchmarks for MVR-4B surpasses Qwen3.5-9B (Instruct) and MFR-8B, while 74.4 for MVR-9B surpasses Qwen3.5-35B-A3B (Instruct).

    Source: §1 · p. 2

  • Takeaways & Limitations

    Expert-based on-policy integration is the primary source of improvement, and the performance gap between KL directions remains small at both tested scales.

    Source: §4.4 · p. 16

    The KL-direction comparison is scoped to experts initialized from the same SFT generalist, whose output distributions remain relatively aligned.

    Source: §4.4 · p. 16

Abstract

from arXiv · show

Open multimodal reasoning models have benefited from large-scale reasoning supervision, yet reliable post-training remains challenging due to uneven data quality, inefficient supervision construction, imbalanced difficulty, and cross-domain interference. We introduce MMVistaReason (MVR), an open-data post-training recipe with three components: (1) broader capability coverage across complementary Analytical and Real-World reasoning groups, emphasizing structured reasoning versus visual perception and spatial grounding; (2) efficient SFT and RL data construction, standardizing heterogeneous open data through staged cleaning and annotation, combining difficulty-aware cascaded teacher distillation with answer-likelihood-based trajectory selection to construct MVR-SFT-528K, and applying scale-specific frontier filtering for MVR-RL-63K; and (3) specialize-then-integrate training, which trains complementary RL experts and consolidates their capabilities through multi-teacher on-policy distillation (MOPD). Our analyses reveal a capacity-dependent interaction between supervision difficulty, trajectory quality, and model capacity: smaller students benefit more from selected supervision, while larger students are robust to trajectory variation and mixed-domain interference. Mixed-domain RL introduces benchmark-level negative transfer, whereas MOPD provides consistent capability integration, with the preferred KL direction varying across model scales. Across 15 multimodal benchmarks, MVR-4B achieves an average score of 72.8, outperforming Qwen3.5-9B (Instruct) and MMFineReason-8B while using about 70% fewer samples than MMFineReason. Scaling to 9B improves the average to 74.4, surpassing Qwen3.5-35B-A3B (Instruct). Overall, MMVistaReason demonstrates that systematic open-data construction and capacity-aware post-training provide a practical and scalable path toward reliable multimodal reasoning.

1 Introduction

MMVistaReason addresses gaps in capability coverage, supervision construction, and cross-domain integration with a data and training recipe spanning Analytical and Real-World reasoning. Across 15 benchmarks, its 4B and 9B models outperform the cited larger baselines, while MOPD improves SFT generalists at both scales.

1 Introduction

  • Existing data underemphasize real-world capabilities, while multimodal samples vary in difficulty and trajectory quality and SFT and RL need different supervision.Reliable SFT trajectories should match student capacity; RL prompts should remain learnable under the current policy.Source: §1 · p. 2
  • Mixed-domain objectives can introduce negative transfer, while integrating complementary specialist strengths into a unified generalist remains underexplored.On-policy distillation is presented as a promising direction for integration on student-generated trajectories.Source: §1 · p. 2
  • MVR balances Analytical reasoning in mathematics, science, visual logic, and charts with Real-World perception, counting, spatial understanding, and GUI grounding.The two groups emphasize structured, knowledge-intensive reasoning versus fine-grained visual perception and spatial grounding.Source: §1 · p. 2
  • For SFT, cascaded same-family teachers handle unresolved difficult samples and Answer-Likelihood Best-of-N selects verified trajectories; RL uses scale-specific frontier prompts and complementary experts.MOPD then routes expert token-level supervision to student-induced trajectories.Source: §1 · pp. 2–3
  • 72.8 average across 15 benchmarks for MVR-4B outperforms Qwen3.5-9B (Instruct) and MFR-8B; 74.4 for MVR-9B surpasses Qwen3.5-35B-A3B (Instruct).MVR-4B uses approximately 70% fewer SFT samples than MFR.Source: §1 · p. 2
  • Smaller students benefit more from broader selected supervision, whereas the 9B student matches full-data performance using a 124K moderately difficult subset.Answer-Likelihood Best-of-N selects more reliable trajectories than random or length-based strategies across model scales.Source: §1 · p. 2
  • MOPD improves SFT generalists from 70.8 to 72.8 at 4B and from 72.3 to 74.4 at 9B, with reduced cross-domain interference versus mixed-domain RL.The preferred KL direction varies with model scale.Source: §1 · p. 3

2 Related Work

Prior work moved from visual instruction tuning toward reasoning-enriched and deliberately curated multimodal data, while open recipes commonly trained a single generalist on mixed domains. Distillation research progressed from fixed teacher trajectories to student-conditioned, multi-teacher integration.

2.1 Multimodal Reasoning Recipes

  • Multimodal reasoning recipes progressed from broad visual instruction tuning to rationale-enriched supervision and data-centric curation of context, composition, quality, and difficulty.Existing open recipes typically optimize one generalist over mixed-domain data, despite evidence of both positive and negative cross-task transfer.Source: §2.1 · pp. 3–4

2.2 Knowledge Distillation

  • Distillation research spans sequence-level and generative objectives, while offline multimodal distillation uses fixed teacher traces that mismatch student-induced states.On-policy methods query teachers on student-generated prefixes; MOPD integrates multiple domain-specific RL teachers into a unified student.Source: §2.2 · p. 4

3 MMVistaReason Recipes

MMVistaReason provides a unified recipe from open-data cleaning through SFT and RL construction to specialist training and multi-teacher integration. It pairs broad, standardized supervision with scale-aware data selection and on-policy capability integration.

3.1 Data Cleaning Recipe

  • Staged cleaning combines rule-based and MLLM-assisted filtering, removes duplicate or visually invalid samples, and standardizes cleaned examples with domain and question-type annotations.The unified schema supports subsequent difficulty filtering, trace construction, balanced sampling, and SFT/RL training.Source: §3.1 · pp. 4–5

3.2 Data Construction Recipe

  • Cascaded SFT filters trivial samples, routes hard cases to stronger teachers, then selects verified valid trajectories by answer likelihood, forming MVR-SFT-528K at reduced inference cost.Answer-likelihood selection normalizes for answer length and favors trajectories that more confidently support the verified answer.Source: §3.2 · pp. 6–7
  • Scale-specific RL construction uses eight rollouts to retain frontier samples solved one to seven times, prioritizes those solved two to six times, and balances domains.Fully failed and fully solved samples are excluded because they are often too difficult or noisy, or offer limited additional learning signal.Source: §3.2 · p. 7
  • 355K contributes 64.1–73.2% across four reasoning-intensive domains, while 49K emphasizes GUI/Web (38.3%), 124K Spatial (27.0%), and 83.83% of RL data lies at mastery 0.375–0.625.The teacher-derived SFT subsets have complementary domain coverage; RL data concentrates on examples the corresponding SFT model solves inconsistently.Source: §3.2 · pp. 7–8

3.3 Model Training and Integration Recipe

  • Two domain-specific RL experts are trained from the SFT model, then routed to provide token-level supervision on student-generated prefixes through MOPD.Experts evaluate prefixes visited by the student rather than generating independent teacher trajectories.Source: §3.3 · pp. 8–9
  • GSPO uses sequence-level importance ratios for stable policy optimization, while MOPD tests forward- and reverse-KL matching on teacher-selected top-K token support.The study uses these KL directions to examine how distribution matching affects integration of complementary expert capabilities.Source: §3.3 · pp. 8–9

4 Experiments

MMVistaReason combines broader multimodal evaluation with capacity-aware data construction and specialist RL integration. Its final models outperform larger baselines overall, while analyses show that training-data difficulty, trajectory selection, and integration behavior vary with model scale.

4.1 Experimental Setup

  • Models receive SFT, complementary domain-specific RL expert training, and MOPD integration, then are evaluated with greedy decoding across 15 multimodal benchmarks.The benchmark suite is organized into Real-World and Analytical visual reasoning groups.Source: §4.1 · p. 9

4.2 Main Results

  • MVR-4B averages 72.8, beating MFR-8B by 1.9 points and Qwen3.5-9B (Instruct) by 0.7; MVR-9B averages 74.4, exceeding Qwen3.5-35B-A3B (Instruct) by 0.3.The comparison spans representative closed-source, open-weight, and open-source VLMs.Source: Table 2 · p. 10, §4.2 · p. 10
  • Real-World averages reach 82.0 at 4B and 83.8 at 9B, while Analytical averages reach 66.6 and 68.1, respectively.The reported gains span visual perception, spatial understanding, science, mathematics, logic, and chart reasoning.Source: §4.2 · p. 10

4.3 Performance across Training Stages

  • SFT raises overall averages from 66.6 to 70.8 at 4B and from 68.9 to 72.3 at 9B; MOPD then raises them to 72.8 and 74.4.The SFT models surpass their corresponding Qwen3.5-Instruct models; MOPD gains span both reasoning groups.Source: Table 3 · p. 11, §4.3 · p. 11

4.4 Further Analysis

  • The 2B and 4B students improve with the full 528K mixture, whereas the 9B student matches it with the 124K subset at 72.3.The 49K subset contains harder examples from a 122B teacher, while the 124K subset comes from a 27B teacher; results indicate that data volume alone does not determine effectiveness.Source: §4.4 · pp. 12–13
  • Answer-likelihood Best-of-N selection yields positive average gains over random selection at both scales, with gains declining from +0.29 at 4B to +0.24 at 9B.Longest-trajectory selection is less stable, particularly at 4B, though some tasks—especially at 9B—benefit from longer traces.Source: Figure 5 · p. 13, §4.4 · p. 13
  • The Real-World and Analytical RL experts specialize in their respective domains, and MOPD preserves most specialist gains while improving over SFT on all 15 benchmarks at both scales.MVR also surpasses both expert teachers on several benchmarks, including CountBenchQA, SFE, and MMMUval.Source: §4.4 · pp. 14–15
  • Mixed-RL regresses on 7 of 15 benchmarks at 4B and four at 9B, while the best MOPD variant improves or matches it on all 15 at each scale.Forward-KL performs better at 4B (72.8 vs. 72.3), whereas Reverse-KL is stronger at 9B (74.4 vs. 74.2); the experts’ shared SFT initialization keeps their distributions relatively aligned with the student.Source: Figure 6 · p. 15, §4.4 · pp. 15–16

5 Conclusion

MMVistaReason presents an open-data framework and adaptive training recipe for multimodal reasoning. Experiments at 4B and 9B show balanced benchmark performance, while analyses indicate that specialist RL followed by MOPD consolidates capabilities with reduced cross-domain interference.

5 Conclusion

  • MMVistaReason combines cleaned and difficulty-filtered open-data supervision, SFT generalists, complementary RL experts, and multi-teacher on-policy distillation.Experiments across 4B and 9B models report balanced performance across diverse multimodal benchmarks.Source: §5 · p. 16

A.1 SFT Data Statistics

The SFT corpus draws on varied open multimodal sources and includes subsets with differing difficulty and question formats. Its response-token statistics reveal substantial variation in reasoning length and computational intensity across domains.

Open-Data Source Composition

  • MVR-SFT-528K draws on diverse open multimodal datasets covering science, mathematics, spatial reasoning, charts, visual logic, and general visual understanding.Table 5 reports sample and response-token counts for each source subset.Source: §A.1 · p. 22, Table 5 · p. 22

Difficulty and Question-Type Statistics

  • Subsets distilled from stronger teachers have lower average rollout pass rates, indicating harder supervision signals.The corpus also spans numeric, short-answer, multiple-choice, and entity-based questions, with complementary distributions across teacher-derived subsets.Source: §A.1 · p. 22, Table 6 · p. 22, Table 7 · p. 23

Domain-wise Token Statistics

  • Mathematics supplies 49.70% of response tokens from 24.76% of samples, with a mean response length of 3,970.5 tokens and P95 of 12,188.Spatial responses have a long-tailed distribution, while General and GUI/Web responses are dominated by short-form outputs.Source: §A.1 · pp. 22–23, Table 8 · p. 23

A.2 RL Data Statistics

The scale-specific RL expert subsets have stable difficulty profiles, with Analytical RL more challenging than Real-World RL. Their task compositions differ by domain and contain answer types suited to automatic verification.

Difficulty Distribution of RL Experts

  • Analytical RL mean pass rates are 0.4502 at 4B and 0.4503 at 9B, below Real-World RL’s 0.4981 at both scales; both groups have median pass rates of 0.5.The overall difficulty profiles remain highly consistent across model scales.Source: §A.2 · p. 23, Table 9 · p. 24

Question Type Distribution of RL Experts

  • Analytical RL is dominated by multiple-choice and numeric tasks, while Real-World RL contains more counting and position-related questions.These task-type patterns remain consistent across scales and generally provide well-defined answer spaces for automatic verification.Source: §A.2 · p. 24, Table 10 · p. 24

B Experimental Details

The cascaded distillation setup routes difficult samples to progressively larger teachers to reduce generation cost. The section specifies the teacher deployment and trajectory-generation configuration.

B.1 Distillation Details

  • Qwen3.5-9B, 27B, and 122B-A10B serve as cascaded teachers, with difficult samples routed to larger teachers to reduce distillation cost.They use 1, 2, and 4 NVIDIA A100 GPUs, respectively.Source: §B.1 · p. 24
  • Each input receives four candidate trajectories, with generation capped at 16,384 tokens and teachers run in instruct mode with thinking disabled.The stated throughput is approximately 200, 120, and 80 samples per hour for the 9B, 27B, and 122B-A10B teachers, respectively.Source: §B.1 · p. 24
  • The corresponding generation rates are roughly 800, 480, and 320 trajectories per hour for the 9B, 27B, and 122B-A10B teachers.The teachers are deployed on NVIDIA A100 GPUs with 80 GB memory.Source: §B.1 · p. 24

B.2 Training Details

The training details specify the SFT configuration, the RL expert setup, and the MOPD student–teacher arrangement. Additional hyperparameters are referenced in Table 11.

B.2 Training Details

  • SFT uses LLaMA-Factory on eight NVIDIA A100 GPUs, with full-parameter fine-tuning, packed sequences, FlashAttention-2, a 16,384-token cutoff, and global batch size 32.Other SFT hyperparameters are summarized in Table 11.Source: §B.2 · pp. 24–25, Table 11 · p. 25
  • RL experts are trained with VeRL on eight NVIDIA H200 GPUs using GSPO, eight rollouts per prompt, and an answer-only judge reward.Maximum prompt and response lengths are 4,096 and 16,384 tokens, respectively; other RL hyperparameters appear in Table 11.Source: §B.2 · p. 25, Table 11 · p. 25
  • For MOPD, the student generates on-policy responses and receives top-K token-level supervision with K = 32 from each teacher.The student uses four GPUs and each teacher uses two; other settings are reported in Table 11.Source: §B.2 · p. 25, Table 11 · p. 25

B.3 Evaluation Details

This section details a standardized protocol for preparing and evaluating multimodal reasoning data, covering data cleaning, validation, refinement and annotation, domain labeling, SFT trajectory generation and verification, and grading that prioritizes final-answer correctness.

B.3 Evaluation Details

  • Evaluations use CompassVerifier, deterministic decoding, and manual checks for benchmarks with unusually large performance deviations.The setup uses VLMEvalKit on NVIDIA H200 GPUs, with a 16,384-token response limit; responses without a valid final answer within the limit are incorrect.Source: §B.3 · p. 25

C.1 MLLM-Assisted Cleaning and Annotation

  • The prompts standardize questions in English, remove irrelevant content, filter unanswerable samples, and refine low-quality instructions without changing task meaning or solving problems.Answerability checks reject questions that require unavailable context, external actions, or unverifiable responses.Source: §C.1 · pp. 26–28
  • Further prompts extract and canonicalize existing answers, then classify each sample by its primary reasoning domain.Canonicalization preserves answer meaning and important details, while domain labels include science, math, chart/table/doc, logic/game/puzzle, and general visual question answering.Source: §C.1 · pp. 28–30
  • Spatial labeling covers image-grounded relations such as positions, layouts, viewpoints, and navigation; GUI/web labeling covers interface screenshots and their controls and layouts.Source: §C.1 · p. 30
  • Domain labels follow the main skill needed to answer, directing chart/document, science or math, spatial, and GUI/web tasks to their specialized categories rather than general.General is reserved for tasks without specialized scientific, mathematical, logical, spatial, chart/document, or GUI/web reasoning.Source: §C.1 · pp. 30–31

C.2 SFT Trajectory Construction

  • Trajectory prompts request concise, image-grounded reasoning that uses relevant visual evidence before a format-compliant final answer enclosed in <answer> tags.The final-answer content should contain only the requested answer, in the specified format when one is given.Source: §C.2 · pp. 31–32
  • Verification evaluates semantic correctness of the final answer first, requiring all parts to be correct and rejecting incorrect answers regardless of preceding reasoning.Incomplete, severely truncated, repetitive, and refusal responses are also rejected.Source: §C.2 · pp. 32–33
  • When a final answer is correct, verification checks reasoning for consistency with the question, visual evidence, and answer, rejecting major errors that undermine the solution.Minor omissions or imprecision may be tolerated, but a substantive reasoning error can invalidate an otherwise matching final answer.Source: §C.2 · p. 33

D Data Examples

The representative examples illustrate MMVistaReason’s range of visual reasoning tasks, from geometry and logic puzzles to chart reading, perception, spatial ordering, and interface questions. They show varied question formats and answer types.

D Data Examples

  • A chemistry-image question asks for the hydrogenation product of furfuralacetone, and the response identifies it as a saturated alcohol.The example describes reduction of the side-chain double bond and carbonyl while preserving the furan ring.Source: D Data Examples · p. 34
  • The geometry example asks for a highlighted area and gives 50.3 square inches as the answer, while a maze example asks which moves are available and answers “G.”The geometry passage explains the result as 16π, approximately 50.3; the maze response identifies down, left, and right as available.Source: D Data Examples · pp. 35–38
  • A chart example asks for the x-coordinate of the tallest bar in subplot (b), and the response gives x = 3.The response identifies the tallest bar as near the center-right of the chart.Source: D Data Examples · pp. 38–39
  • A bird-identification example names the bird as a Red-legged Honeycreeper (Cyanerpes cyaneus).The response cites its coloration and red legs as identifying features.Source: D Data Examples · p. 39
  • A spatial example asks for the depth order of three marked regions and gives 2, 1, 3, from closest to farthest.The response places the foreground lawn before the lamppost and the building.Source: D Data Examples · pp. 40–41
  • A GUI/web example asks how many customized sections follow the generic template on the M&S study site, and the response gives five.The response lists Your modules, Your subjects, Free learning, News, and Upcoming events.Source: D Data Examples · pp. 41–42

Related papers

Loading 2610.01352v1…