Long Grounded Thoughts: Synthesizing Visual Problems and Reasoning Chains at Scale

1. Introduction

• The framework first generates diverse questions from images, then composes them into harder problems and synthesizes reasoning traces for both stages.Its stated motivation is that open vision-centric reasoning datasets with complex problems and traces are often limited in scale or focused on visual math.Source: §1 · p. 1, Figure 1 · p. 2
• Grounded object metadata lets the framework scale problem synthesis from LPT’s 30K to over 1M+ problems without performance saturation.The framework targets scale, complexity, and richer reasoning traces.Source: §1 · pp. 1–2
• Composition hardening reduces trivially solvable questions from 36.7% in baselines to 3.3%, approximately a 10× reduction.The framework combines simpler questions into more difficult visual reasoning problems.Source: §1 · p. 2, Figure 1 · p. 2
• The synthesized traces increase the frequency of complex cognitive behaviors by 206%, yielding traces 3× richer than the baseline.The paper identifies verification, backtracking, and subgoal setting as examples of complex reasoning structures.Source: §1 · pp. 1–2
• Finetuning a 7B VLM on the data outperforms open-data baselines on several vision-centric benchmarks and matches or outperforms MiMO-VL-7B-RL on 3 of 5 benchmarks.The data also transfers positively to text-only and audio reasoning, and improves open-ended embodied QA despite containing no embodied-QA data.Source: §1 · p. 2
• The 1M+ study of VLM post-training finds that cognitive-behavior SFT supports online RL, offline SFT→DPO can be competitive with online RL, and base-model online RL plateaus quickly.High-quality SFT alone also improves out-of-domain, cross-modality transfer.Source: §1 · p. 2

2.1. Scale and Diversity

• Grounded metadata and dense captions support diverse MCQ generation; object grounding yields 3.2× wider semantic spread and 0.61 versus 0.82 average pairwise cosine similarity.Grounded object metadata also sustains positive downstream performance slope beyond 1M visual problems.Source: §2.1 · p. 3

2.2. Compositionality for Complexity

• Composition hardening merges same-image MCQs into multi-step problems, which are harder and elicit significantly more subgoal setting than Stage 1 questions.The initial grounded synthesis is harder than caption-only methods, but many questions remain solvable by base VLMs.Source: §2.2 · pp. 3–4

2.3. Synthesizing Reasoning Traces

• Long reasoning traces pair an initial VLM rationale with a reasoning model's continuation; only expanded Long CoT traces yield consistent gains over Short CoT or no CoT.Regex-guided decoding addresses explicit references to image descriptions during expansion; captions are not shown to the training model or at inference.Source: §2.3 · p. 4

2.4. Offline and Online Synthesis for RL

• For offline RL, preference pairs compare correctness and compactness; online GRPO retains generated visual problems and verifiable answers.Trace correctness is inferred from the associated answer rather than verified directly.Source: §2.4 · pp. 4–5

3.1. Scale, Complexity and Richness of CoT

• Beyond 1M+ visual problems, the framework maintains a positive SFT-scaling slope under identical training conditions.The comparison is against LPT.Source: §3.1 · p. 5
• Composition hardening lowers the perfect solve rate from 36.7% to 3.3% and the average pass rate from 66.1% to 38.4%.Lower pass/solve rates indicate greater problem complexity.Source: §3.1 · p. 5, Table 3 · p. 6
• Stage 2 traces average 1.99 cognitive behaviors versus LPT's 0.65, a +206% increase and roughly 3× richer structure.Richness is measured by the average count of distinct cognitive behaviors per trace.Source: §3.1 · p. 5, Table 3 · p. 6

3.2. Setup and Main Results

• Experiments fine-tune Qwen2.5-VL-7B-Instruct on benchmarks covering visual search, 2D/3D spatial reasoning, fine-grained attribution, and scene understanding.Evaluation includes V*Bench, CV-Bench, MMVP, RealWorldQA, and vision splits of general-knowledge datasets.Source: §3.2 · p. 5
• Our models consistently improve on the base and achieve the best results, outperforming on 3/5 vision-centric benchmarks.Table 2 compares against open- and closed-source VLMs across five benchmarks.Source: Table 2 · p. 6, §3.2 · p. 7

3.3. Analysis of VLM Post-training Stages

• Starting from base instruct, GRPO peaks at 0.695 (10K), below our SFT's 0.716; structured complex reasoning traces were not elicited without reasoning-SFT initialization.The figure's example shows the baseline missing a partially obscured dog while the SFT-initialized model self-corrects.Source: §3.3 · p. 7, Figure 5 · p. 8
• Staged offline SFT→DPO reaches 0.740, within 1.7 points of online SFT→GRPO at 0.757, while decoupling compute demands.The best offline configuration uses SFT 750K + DPO 129K; the best online configuration uses SFT 750K + GRPO 70K.Source: §3.3 · p. 7
• At comparable scale, grounded-MCQ SFT scores 0.716 versus 0.682 for LPT SFT (+3.4 points on average); atop SFT, GRPO reaches 0.757 at 70K, then falls to 0.746 at 129K.Source: §3.3 · p. 7

• Open-source reasoning datasets have advanced language-based reasoning, while open-source multimodal reasoning efforts appear to lag behind.The related-work discussion notes that innovations in data curation and distillation have advanced open-source language models.Source: §4 · p. 8
• Multimodal datasets have largely remained linear or reused existing questions rather than synthesizing new problems across complexity levels for the VLM post-training pipeline.The cited examples of work on complex multimodal reasoning include Virgo and VLAA-Thinking.Source: §4 · p. 8
• Building on LPT, this work synthesizes diverse, verifiable visual problems and complex reasoning traces across scale, complexity, and reasoning depth.It differs from fixed-stage or linear CoTs and from approaches that reuse existing visual questions.Source: §4 · p. 8

• The framework combines grounded metadata and composition hardening to generate diverse, verifiable visual problems and complex reasoning structures in over 1M examples.The framework targets scale, complexity, and cognitive richness.Source: §5 · p. 8
• Experiments show improved downstream performance on vision-centric benchmarks and positive transfer across modalities and out-of-distribution tasks.Source: §5 · p. 8
• The conclusion characterizes the resulting data as high-quality and the generated visual problems and reasoning structures as surpassing prior efforts.Source: §5 · p. 8

• The data is synthesized from frontier models, creating a potential risk of propagating biases present in those teachers.Source: Impact Statements · p. 9
• The authors do not foresee immediate negative societal consequences beyond general dual-use concerns applicable to visual understanding and reasoning systems.Source: Impact Statements · p. 9
• The work presents open datasets and reproducible training recipes as contributing to democratizing visual reasoning capabilities.Source: Impact Statements · p. 9

A.1. Stage 1: Object-centric Problem Generation

• Stage 1 combines dense image captions with Grounded SAM-2 object labels and bounding boxes to prompt an LLM to generate object-centric multiple-choice questions.The pipeline uses a 0.9 detection confidence cutoff and limits same-category instances to nine per image.Source: §A.1 · p. 13
• Stage 1 filtering uses an LLM verifier for factual and logical correctness, then semantic deduplication based on question, answer, and category similarity, with a deduplication threshold of 0.82.The composite similarity weights question-stem and answer cosine similarity together with category-tag Jaccard similarity.Source: §A.1 · p. 13

A.3. Vision-Centric Prompt Templates

• The vision-centric prompt grounds questions in bounded regions and object context; adding a “computer vision expert” role raised Stage 1 instruction-following F1 from 96.3 to 99.2.Its four guiding principles emphasize object focus, spatial precision, contextual relationships, and multi-level analysis, motivated by spatial and contextual failure patterns.Source: §A.3 · p. 15

A.4. Stage 2: Compositional Problem Generation

• Stage 2 hardens simpler questions by composing up to five image-grounded question-answer pairs and the global caption into a multi-hop multiple-choice problem, retaining only answers with consistency τ ≥0.8.This stage addresses the fact that many Stage 1 questions are simple enough for the base VLM to solve directly.Source: §A.4 · p. 17
• Reasoning traces are distilled in two steps: a VLM first creates an in-distribution Simple CoT, which primes a stronger reasoning LLM to expand it with more complex problem-solving strategies.The method scales model choice with task complexity, using efficient models for Stage 1 and frontier models for Stage 2.Source: §A.4 · p. 18

A.5. Local Qwen CoT Verification Prompt

• A local Qwen-32B-A3B-2507-Instruct verifier checks whether an expanded reasoning trace logically and factually supports the known correct answer, filtering inconsistencies at scale.The verifier derives an answer from the reflection without relying on the supplied answer, then checks for consistency.Source: §A.5 · p. 18

A.6.1. MULTIMODAL AUDIO UNDERSTANDING AND OMNI MODEL

• For out-of-domain MMAU audio reasoning, the experiment fine-tunes only Qwen2.5-Omni-7B-Instruct’s thinker, freezes its audio encoder, and merges the tuned thinker back with the original talker.This setup aims to improve reasoning without directly altering the talker’s audio-specific knowledge.Source: §A.6.1 · p. 19
• The 71.00 MMAU baseline surpasses Gemini-2-Flash at 67.03, while Virgo fine-tuning degrades performance to 62.60, showing reasoning-data transfer can be negative.LGT data, in contrast, shows positive scaling behavior.Source: §A.6.1 · p. 19
• At 1M LGTs, the model achieves a 72.32 MMAU average, versus 71.80 with 500k LGTs, a new state-of-the-art among tested configurations.The result is reported for the omni-model audio-understanding evaluation.Source: §A.6.1 · p. 19, Table 6 · p. 20
• MMAU-Sound rises from 76.77 to 78.30 and MMAU-Music from 67.33 to 70.20 with 1M LGTs, driving gains in non-speech acoustic reasoning.The passage attributes these gains to improvements on sound and music tasks.Source: §A.6.1 · p. 19
• Vision-centric LGT reasoning traces also improve OOD audio and dual-modal vision-audio reasoning, supporting the finding that high-quality reasoning text can enhance audio reasoning (Rouditchenko et al., 2025).The appendix gives a qualitative example of reasoning about complex, compositional sound and speech events.Source: §A.6.1 · p. 20

A.6.2. TEXT-ONLY MMLU-PRO

• MMLU-Pro covers 14 domains and over 12,000 reasoning-intensive questions; the appendix points to an extended results table but supplies no scores in these passages.MMLU-Pro extends the original MMLU with more reasoning-intensive questions and a larger answer-choice set.Source: §A.6.2 · p. 20, Table 7 · p. 20

A.2. Additional Analysis of Data Diversity

• The authors hypothesize that dense-caption conditioning repeatedly targets salient image features, while bounding-box conditioning explores less prominent visual concepts.They link avoiding saturation in supervision to continuous performance scaling, unlike caption-based methods that reach diminishing returns.Source: §A.2 · p. 14
• The comparison samples 1,000 questions per method from three shared seed images and estimates semantic-space density with Gaussian KDE.Questions are encoded using SentenceBERT all-MiniLM-L6-v2.Source: §A.2 · p. 14
• LGT yields a 3.2× wider distribution than the baseline, measured as mean Euclidean distance to the centroid.The KDE comparison shows the baseline concentrated in a narrow semantic region, while LGT covers a wider semantic space.Source: §A.2 · p. 14, Figure 7 · p. 14
• Average pairwise cosine similarity is 0.61 for LGT versus 0.82 for LPT, indicating lower question redundancy in LGT.The authors conclude object-level metadata prevents synthesis collapse and maintains diversity at scale.Source: §A.2 · p. 14

A.6.3. EMBODIED OPEN-ENDED QA BENCHMARK

• The modified NiEH evaluation focuses on the single-evidence setting, where one frame in a time-ordered trajectory suffices to answer a question.The test set contains 829 image–question pairs; inputs use a 2,048-token window centered on the answer-bearing frame while preserving temporal order.Source: §A.6.3 · p. 20
• The evaluation prompt asks models to answer from images in time order with a single word or phrase and no explanation.The baseline uses the original Qwen-7B-VL prompt; the authors’ models use a fixed system prompt and task instruction.Source: §A.6.3 · p. 20
• Table 8 summarizes modified NiEH single-evidence results using exact match, with higher scores indicating better performance.The supplied passages do not provide the table’s score values.Source: §A.6.3 · p. 20, Table 8 · p. 21

A.7. Experiments on MathVision and VisuLogic Reasoning

• Despite containing no math questions, the fine-tuned model shows meaningful improvements on both MathVision and VisuLogic.The authors interpret this as evidence that strengthening core reasoning abilities transfers to out-of-domain tasks; Table 9 presents the benchmark results.Source: §A.7 · p. 21, Table 9 · p. 21
• Figure 8 provides an additional qualitative comparison of a reasoning trace from the post-trained model and the base model.The caption does not state the outcome of that comparison.Source: Figure 8 · p. 23