PaperlayerView exact revision on arXiv

Source-linked AI summary

HoneyBee: Data Recipes for Vision-Language Reasoners

Hritik Bansal, Devendra Singh Sachan, Kai-Wei Chang, Aditya Grover, Gargi Ghosh, Wen-tau Yih, Ramakanth Pasunuru

arXiv:2510.12225v2cs.CVcs.LG

TL;DR

The construction principles for effective VL reasoning datasets remain poorly understood, in part because many datasets and curation recipes are proprietary. The paper studies context sources, data interventions, and data scaling, then builds HoneyBee and reports improved benchmark performance, including a 28pp average gain for PLM-HoneyBee-1B over InternVL-3-1B-Instruct.

  • Problem

    VL reasoning CoT dataset design remains poorly understood, with proprietary datasets and recipes limiting study of context-source effects and other curation choices.

    Source: §1 · pp. 1–2

  • Method

    The paper compares context sources, tests targeted data interventions, studies scaling across data dimensions, and uses the findings to construct HoneyBee.

    Source: §1 · p. 2

  • Results

    HoneyBee-trained models outperform existing state-of-the-art instruction-tuned VLMs; PLM-HoneyBee-1B improves by 28 percentage points (pp) over InternVL-3-1B-Instruct, averaged across ten evaluation datasets.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    HoneyBee provides a large-scale CoT dataset and a foundation for further research into VL reasoning data curation.

    Source: §1 · p. 2, §7 · p. 16

    The study focuses on single-image data curation; extending the approach to reasoning over multiple images remains future work.

    Source: §7 · p. 16

Abstract

from arXiv · show

Recent advances in vision-language models (VLMs) have made them highly effective at reasoning tasks. However, the principles underlying the construction of performant VL reasoning training datasets remain poorly understood. In this work, we introduce several data curation approaches and study their impacts on VL reasoning capabilities by carefully controlling training and evaluation setups. We analyze the effects of context (image and question pair) sources, implement targeted data interventions, and explore scaling up images, questions, and chain-of-thought (CoT) solutions. Our findings reveal that (a) context source strategies significantly affect VLM performance, (b) interventions such as auxiliary signals from image captions and the inclusion of text-only reasoning yield substantial gains, and (c) scaling all data dimensions (e.g., unique questions per image and unique CoTs per image-question pair) consistently improves reasoning capability. Motivated by these insights, we introduce HoneyBee, a large-scale, high-quality CoT reasoning dataset with 2.5M examples consisting 350K image-question pairs. VLMs trained with HoneyBee outperform state-of-the-art models across model sizes. For instance, a HoneyBee-trained VLM with 3B parameters outperforms the SOTA model and the base model by 7.8% and 24.8%, respectively, on MathVerse. Furthermore, we propose a test-time scaling strategy that reduces decoding cost by 73% without sacrificing accuracy. Overall, this work presents improved strategies for VL reasoning dataset curation research. Data is available at https://huggingface.co/datasets/facebook/HoneyBee.

1 Introduction

Because VL reasoning CoT datasets and their curation recipes are often proprietary, the effects of context sources, data interventions, and scaling choices remain poorly understood. The paper studies these choices and introduces HoneyBee, reporting improved performance over existing instruction-tuned models and gains from scaling and selected interventions.

1 Introduction

  • The design space remains unclear because VL CoT datasets and their recipes are often proprietary, and prior work leaves context-source effects and data construction choices insufficiently examined.The paper calls for fair comparisons with fixed training and evaluation setups, plus robustness checks across model sizes and datasets.Source: §1 · pp. 1–2
  • HoneyBee contains 2.5M (image, question, CoT) tuples; VLMs sized 1B–8B trained on it show reasoning performance that strongly scales with training-data amount and outperform existing state-of-the-art instruction-tuned VLMs.Source: §1 · p. 2
  • The source of image-question contexts can shift average accuracy by up to 4% across evaluation datasets, motivating controlled comparisons of curation choices.The paper compares diverse context sources using 3B and 8B models across multiple evaluation datasets.Source: §1 · p. 2
  • Caption-and-solve and adding text-only reasoning yield major improvements across diverse benchmarks, while several other tested interventions do not outperform the baseline.The intervention set includes visual perturbation, text-rich images, perceptual redundancy, shallow perception, increased distractors, and length and difficulty filtering.Source: §1 · p. 2
  • 28 percentage points (pp) over InternVL-3-1B-Instruct, averaged across ten evaluation datasets, is PLM-HoneyBee-1B’s reported relative improvement; HoneyBee-trained models also outperform existing state-of-the-art instruction-tuned VLMs.PLM-HoneyBee-3B and PLM-HoneyBee-8B achieve relative gains of 8.4pp and 2.7pp over Qwen2.5-VL-3B-Instruct and Qwen2.5-VL-8B-Instruct, respectively.Source: §1 · pp. 2–3

2 Preliminaries

The paper represents synthetic VL reasoning data as image-question examples paired with generated chains of thought and optional reference answers. It uses this data to train VLMs with supervised fine-tuning, while positioning the work within teacher-student distillation and self-improvement approaches.

2 Preliminaries

  • The source dataset contains images, corresponding questions, and optional final answers; a generator produces a CoT for each image-question pair.The CoT combines reasoning steps with a predicted final answer, while the reference answer remains an optional source field.Source: §2 · p. 3
  • A VLM is trained by supervised fine-tuning to maximize the probability of the problem-solving CoT given its image and question.The paper frames this as training smaller VLMs on synthetic data from larger generator models for compute efficiency and comprehensive training runs.Source: §2 · pp. 3–4
  • The curation pipeline studies context sources, data interventions, and scaling along different data axes before constructing HoneyBee.The broader setup distinguishes a weaker student guided by a strong teacher (knowledge distillation) from using the generator itself as student (self-improvement).Source: Figure 2 · p. 4, §2 · p. 4

3 Data Curation Pipeline

The curation pipeline studies context sources, targeted interventions, and scaling across data dimensions to build stronger vision-language reasoning data. It combines findings from these stages in the HoneyBee dataset.

3.1 Context Curation

  • Context curation compares source datasets under controlled CoT generation and training, then tests equal-sized mixtures of the top two, top four, or all sources.Mixture quality is assessed against the best-performing individual source across downstream VL reasoning tasks.Source: §3.1 · pp. 4–5

3.2 Data Interventions

  • Interventions target perception and problem-solving, using replacement, augmentation, or filtering to alter the original VL reasoning data.The pipeline evaluates interventions against the original reasoning-data baseline under identical training and evaluation conditions across tasks and model scales.Source: §3.2 · pp. 5–6, §3.2.2 · p. 7

3.2.1 Perception Enhancement

  • Perception interventions include caption-augmented CoTs, image perturbations, text-rich image rendering, and filters for examples solvable without the image or from its caption alone.Perturbations include rotation, distractor concatenation, and dominance-preserving mixup; text-rich images embed the question over the original image.Source: §3.2.1 · p. 6

3.2.2 Problem-Solving Skill Enhancement

  • Problem-solving interventions add text-only reasoning, expand answer options, select longer CoTs, or balance examples across rubric-based difficulty levels.Text-only data is re-annotated with the same generator, while difficulty levels range from beginner through Olympiad-level problems.Source: §3.2.2 · p. 7

3.3 Scaling Diverse Data Axes

  • Scaling varies unique images, questions per image, and CoT traces per image-question pair; findings from these experiments are combined with curation and interventions for HoneyBee.The question and CoT experiments use subsets with 1, 2, 4, or 8 questions or CoTs, respectively.Source: §3.3 · p. 8

4 Experimental Setup

The experiments use a fixed data-generation and training setup to compare curation choices on five validation tasks, then evaluate HoneyBee on additional vision-language and text-only benchmarks.

Experimental data and training

  • Six source datasets were decontaminated against evaluation images, capped at 50K instances, and used for five-epoch training with the best checkpoint selected by five-task average performance.The sources were ViRL, Math-LLaVA, R1-OneVision, ThinkLite-VL-Hard, LLaVA-CoT, and MMK12.Source: §4 · p. 8

Generator model

  • CoTs were generated with fixed Llama4-Scout, a 109B-total-parameter model with 17B active parameters, avoiding time-varying API models.Source: §4 · p. 8

Models

  • PLM-3B and PLM-8B were used for curation experiments, with PLM-1B also trained after HoneyBee was created.PLMs serve as base models because they cannot generate vision-language CoTs without appropriate instruction data.Source: §4 · p. 8

Validation evaluation

  • Validation averaged accuracy across MathVerse, MathVista, MathVision, MMMU-Pro, and We-Math for PLM-3B and PLM-8B.Source: §4 · p. 9

Final evaluation

  • Final evaluation added DynaMath, LogicVista, HallusionBench, MATH500, and GPQA, using accuracy, boxed-answer prompts, and greedy decoding to 2048 tokens.Source: §4 · p. 9

5 Experiments

Across controlled experiments, context sources, data interventions, and scaling choices affect vision-language reasoning performance. HoneyBee-trained models lead average accuracy across model-size categories, while shared captions reduce test-time decoding costs and performance-based data curation choices correlate across model sizes.

5.1 Impact of Context Curation

  • Across source datasets, ViRL has the highest average performance at 40.1%, compared with MMK12, the lowest at 36.0%; mixing top-2, top-4, or all sources does not improve over ViRL.Source: §5.1 · p. 10

5.2 Impact of Data Interventions

  • Caption-and-solve, which adds image descriptions to reasoning CoTs, improves accuracy averaged across downstream tasks and training runs by 3.3pp.The reported intervention results use each strategy’s best-performing configuration.Source: §5.2 · p. 10
  • On MATH500, mixing VL and text-only reasoning data raises accuracy from 39.2% to 59.7%, while VL reasoning alone reaches 39.2%.The base PLM-3B and PLM-8B score 32.2% and 34.0%, respectively, on this text-only task.Source: Table 5 · p. 10, §5.2 · p. 11

5.3 Scaling to Million Samples

  • Reasoning performance improves as images, questions per image, and CoTs per image-question pair scale for both PLM-3B and PLM-8B.HoneyBee’s VL subset is built by generating CoTs and new questions from ViRL contexts, then using majority voting to filter CoTs for new questions.Source: §5.3 · pp. 11–12, Figure 4 · p. 12
  • HoneyBee contains 2.5M instances, 28K unique images, and 350K unique questions; its average caption-plus-solution CoT is approximately 600 words (780 tokens).The dataset combines 1.5M VL reasoning instances with 1M text-only reasoning instances.Source: §5.3 · p. 12, Figure 5 · p. 12

5.4 Training VL Reasoners with HoneyBee

  • HoneyBee-trained models achieve the highest average accuracy in each size category; 1B gains 28pp over InternVL-3-1B-Instruct, while 3B and 8B gain 8.4pp and 2.7pp over the corresponding Qwen2.5-VL models.On RL training, HoneyBee-SFT plus GRPO gains 9.2pp in average accuracy over OpenVLThinker-v1.2-3B; against the SFT model, gains are 27.5pp on MathVision and 10.0pp on We-Math.Source: §5.4 · pp. 13–14, Table 6 · p. 14

5.5 Efficient test-time scaling with Shared captions

  • With 64 MathVista solution attempts, shared captions reduce generated tokens and FLOPs by 73% while achieving similar performance.The naive approach generates 42.6K tokens; shared captions generate 24.5K by reusing the first attempt’s image caption for subsequent solutions.Source: §5.5 · p. 14, Figure 8 · p. 14

5.6 Correlation Analysis

  • PLM-3B and PLM-8B show strong correlation in average accuracy across data curation experiments, suggesting similar curation impacts across model scales.The authors suggest small-VLM data selection may be likely to remain optimal for larger VLMs.Source: §5.6 · p. 15, Figure 9 · p. 15

6 Related Work

Prior work established synthetic chain-of-thought reasoning for language models, while visual-language reasoning requires integrating image and text. HoneyBee studies data sources, interventions, and scaling for this multimodal setting.

Chain-of-Thought Reasoning Datasets

  • HoneyBee addresses visual-context reasoning, where models must integrate images and text, unlike text-centered CoT datasets such as OpenThoughts [19].It examines data sources, filtering strategies, and scaling at multiple model sizes.Source: §6 · p. 15

VL Reasoning Datasets

  • Existing VL datasets either lack CoTs, focus on narrow visual domains, or leave unclear whether gains come from data sources, teachers, or filtering.HoneyBee re-annotates CoTs with a given teacher model and tests context mixing, interventions, and scaling across images, questions, and CoTs.Source: §6 · p. 16

VL Data Curation

  • Prior VL curation interventions were tested under differing training paradigms, model families, sizes, and evaluations, complicating comparisons of their reliability.The paper compares datasets with fixed training and evaluation setups.Source: §6 · p. 16

7 Conclusion

The paper presents HoneyBee, a large-scale VL reasoning dataset with CoTs, as a foundation for further data-curation research. Its scope is single-image reasoning, with broader VL skills and multiple-image reasoning left for future study.

7 Conclusion

  • HoneyBee is a high-quality, large-scale dataset of visual-language reasoning examples with chain-of-thoughts.The authors describe it as a foundation for VL reasoning data research.Source: §7 · p. 16
  • Future work can assess whether these curation insights apply to general-purpose VL training and skills beyond reasoning, such as VQA.Source: §7 · p. 16
  • The study covers single-image data curation; extending the approach to reasoning over multiple images remains future work.Source: §7 · p. 16

A Additional Context Dataset Details and Results

The appendix documents context-dataset sampling and filtering, then reports how answer-correctness filtering affected trained reasoners. Effects varied by source dataset.

Additional Context Dataset Details

  • Context experiments randomly sampled at most 50K examples per dataset after removing images matching evaluation images by pHash.Filtered versions retained instances based on final-answer correctness; Table 7 summarizes dataset sizes, contamination, and original versus filtered counts.Source: A Additional Context Dataset Details… · p. 22, Table 7 · p. 22

Additional Context Dataset Results

  • ViRL and ThinkLite-VL-Hard benefit from final-answer-correctness filtering, while the original unfiltered source datasets produce better reasoners overall.The appendix reports average accuracy across five evaluation datasets for PLM-3B and PLM-8B; Table 8 presents the comparison.Source: A Additional Context Dataset Details… · p. 22, Table 8 · p. 22

B Additional Data Intervention Details and Results

The appendix details intervention implementations and outcomes, including visual perturbation, text-rich images, caption-and-solve, text-only reasoning, and data-selection strategies.

Visual Perturbation

  • Augmenting original data with visually perturbed versions worsens reasoning performance, possibly because of redundancy and perturbation noise.Perturbations include rotation, distractor concatenation, and dominance-preserving mixup; augmentation concatenates perturbed and original data.Source: B Additional Data Intervention Details… · p. 23, Table 9 · p. 23

Text-Rich Images

Perceptual Redundancy and Shallow Perception

  • Perceptual-redundancy filtering removes instances solvable without the image, while shallow-perception filtering also removes those solvable using image captions.These strategies remove 42% and 55% of the data, respectively.Source: B Additional Data Intervention Details… · p. 23

Caption and Solve

  • Generating captions and solutions independently produces the strongest caption-and-solve reasoners and allows the two outputs to be generated in parallel.Across variants, students receive the image and question as input and jointly generate the caption and solution; Table 10 describes the variants.Source: B Additional Data Intervention Details… · pp. 23–24, Table 10 · p. 24

Text-Only Reasoning

Increased Distractors

Length

Difficulty

C Self-Improvement

Finetuning Llama-4-Scout on HoneyBee yields a 3.7pp relative gain over its base model, with improvements across most evaluation datasets.

C Self-Improvement

D Additional Training and Evaluation Setup

The appendix specifies PLM and lightweight GRPO training and evaluation setups, alongside prompts and code used to construct and augment training examples.

D Additional Training and Evaluation Setup

  • PLMs use a Llama-3.2 backbone with full-parameter finetuning for five epochs, with learning rates and batch sizes varying by model scale.Experiments use 8 or 16 A100 80GB nodes.Source: D Additional Training and Evaluation… · p. 25
  • A one-epoch GRPO run with four generations per prompt demonstrates lightweight RL training as a cold start for SFT models.The run uses a verifiable multimodal dataset.Source: D Additional Training and Evaluation… · p. 25
  • Evaluation uses boxed final answers and a rule-based verifier, with an LLM extracting answer values for MathVerse questions containing units.The extraction step addresses verification difficulty for free-form answers with numerics and units.Source: D Additional Training and Evaluation… · p. 25

E Prompts

  • The CoT prompt requests detailed image-grounded reasoning with planning, evaluation, reflection, and exploration before a boxed final answer.For multiple-choice questions, it asks for the option letter.Source: E Prompts · p. 26
  • A question-generation prompt asks for a new solvable problem based directly on the image, while a separate prompt requests a detailed image caption.The generated question should not answer the original problem.Source: E Prompts · p. 26
  • Distractor rewriting expands questions to ten options and rewrites the provided answer while retaining its step-by-step structure.The prompt instructs the model to shuffle existing options when present and not solve the question.Source: E Prompts · pp. 26–27
  • The text-rich image script wraps a question on a blank canvas, overlays a resized image, and crops and saves the result.It selects background colors, text colors, and fonts from provided lists.Source: E Prompts · pp. 27–28

F Qualitative Examples

The qualitative examples illustrate diverse, reasonable synthetic questions and show that HoneyBee CoTs concatenate captions with problem solutions in raw text.

F Qualitative Examples

  • Figures 11 and 12 show diverse, reasonable synthetic questions generated for a given image.The section also describes the examples as newly generated questions.Source: F Qualitative Examples · p. 28, Figure 11 · p. 29, Figure 12 · p. 29
  • Figures 13 and 14 provide HoneyBee examples whose CoTs concatenate image captions and problem solutions in raw text.The passages identify the examples but do not specify further figure contents.Source: F Qualitative Examples · p. 28, Figure 13 · p. 31, Figure 14 · p. 32
  • Figure 10 plots scaling trends across individual evaluation datasets, with the first plot showing average performance across ten datasets.The caption says the models range from 1B to 8B and are trained on diverse HoneyBee subsets.Source: Figure 10 · p. 30

Related papers

Loading 2510.12225v2…