HoneyBee: Data Recipes for Vision-Language Reasoners

AI summary

Overview

Research area: Vision-language model (VLM) post-training, specifically the data-centric study of how synthetic chain-of-thought (CoT) datasets for multimodal reasoning should be constructed.

Technical level: Advanced. The paper is an empirical, multi-stage data curation study requiring familiarity with supervised finetuning, chain-of-thought distillation, and multimodal benchmark evaluation.

Scope: The paper systematically isolates the effects of context source, dataset mixing, targeted data interventions, and scaling along image, question, and CoT axes, then uses the findings to build HoneyBee, a 2.5M-example VL reasoning CoT dataset covering 350K image-question pairs.

What This Paper Is About

Vision-language models have become strong reasoners, but the recipes used to build their multimodal CoT training data are largely proprietary, so it is unclear which data design choices actually cause the gains. This paper separates those choices and tests them one at a time — where the images and questions come from, which data interventions help, and which data dimensions are worth scaling — while holding the training algorithm and CoT generator fixed. The output is HoneyBee, a large curated CoT dataset, plus an efficient test-time decoding strategy for models trained on it.

Key Contributions

1. A controlled context curation study. Six VL reasoning source datasets (ViRL, Math-LLaVA, R1-OneVision, ThinkLite-VL-Hard, LLaVA-CoT, MMK12) are decontaminated against evaluation images with perceptual hashing, capped at 50K instances each, annotated with CoTs from a single fixed generator, and ranked by training PLM-3B and PLM-8B and evaluating on five downstream VL reasoning tasks. The paper also tests whether mixing the top-2, top-4, or all sources helps.

2. A systematic set of data interventions with replacement, augmentation, and filtering variants. Nine strategies targeting perception (visual perturbation, text-rich images, perceptual redundancy, shallow perception, caption-and-solve) and problem-solving (text-only reasoning, increased distractors, length filtering, difficulty filtering) are compared against the same baseline under identical training and evaluation.

3. A scaling study across three data axes. The paper constructs matched subsets that scale the number of images (N/8, N/4, N/2, N), the number of questions per image (n_q = 1, 2, 4, 8), and the number of CoTs per image-question pair (n_c = 1, 2, 4, 8).

4. HoneyBee, a 2.5M-example dataset with 350K image-question pairs, and a test-time scaling strategy. VLMs of 1B to 8B parameters trained on HoneyBee are compared against PLM, InternVL-2.5, InternVL-3-Instruct, and Qwen2.5-VL baselines on ten evaluation datasets, and a decoding strategy is proposed that cuts decoding cost by 73% without sacrificing accuracy.

Main Findings

• Context source matters a lot. Training on CoTs generated over different (image, question) sources produced a spread from MMK12 at 36.0% average accuracy to ViRL at 40.1%, a gap the paper reports as 11.4 percentage points between lowest and highest performers (the introduction describes the source-level differences as up to a 4% difference in average accuracy). Full ranking: ViRL 40.1, Math-LLaVA 37.7, R1-OneVision 37.3, ThinkLite-VL-Hard 37.1, LLaVA-CoT 36.3, MMK12 36.0. ViRL and ThinkLite-VL-Hard benefited from filtering CoTs whose final answer was incorrect, while the others performed better unfiltered.

• Mixing sources did not beat the best single source. Top-2, Top-4, and All mixtures averaged 38.6, 38.2, and 38.5 respectively, versus 40.1 for the best individual source (Top-1). The paper concludes that mixing can degrade VL reasoning performance relative to the strongest single dataset.

• Most interventions did not beat the baseline. Against the Original Data baseline of 40.1 average, the following all underperformed: Visual Perturb (38.5), Text-Rich Images (38.8), Perceptual Redundancy (36.5), Shallow Perception (35.6), Increased Distractors (34.6), Length filtering (36.4), and Uniform Difficulty filtering (34.6).

• Caption-and-solve gave a solid gain. Augmenting the CoT with an image caption from the generator raised the average to 41.4, a reported +3.3 percentage points, with MathVerse improving from 33.7 to 37.0 and We-Math from 53.9 to 56.0.

• Text-only reasoning data produced the largest single gain. Augmenting the VL data with a high-performing text-only reasoning dataset raised the average to 43.1, a reported +7.5 percentage points, with MathVerse at 39.7, MathVista at 61.8, and We-Math at 58.3.

• Scaling all three dimensions helped. The paper reports that performance improved with scaling images, unique questions per image, and unique CoTs per image-question pair, and that reasoning performance scales strongly with the amount of HoneyBee training data for model sizes from 1B to 8B.

• HoneyBee-trained models led across model sizes on the ten-dataset average. PLM-HoneyBee-1B reached 36.2 versus 28.3 for InternVL-3-1B-Instruct, 27.6 for InternVL-2.5-1B, and 25.9 for PLM-1B; PLM-HoneyBee-3B reached 46.2 versus 42.6 for Qwen2.5-VL-3B-Instruct, 41.5 for InternVL-2.5-4B, and 33.8 for PLM-3B; PLM-HoneyBee-8B reached 49.8 versus 48.5 for Qwen2.5-VL-7B-Instruct, 45.1 for InternVL-3-8B-Instruct, 41.4 for InternVL-2.5-8B, and 34.6 for PLM-8B.

• Reported gains relative to strong baselines. The paper states PLM-HoneyBee-1B improves by 28 percentage points over InternVL-3-1B-Instruct averaged across ten datasets, and PLM-HoneyBee-3B and PLM-HoneyBee-8B improve by 8.4 and 2.7 percentage points over Qwen2.5-VL-3B-Instruct and Qwen2.5-VL-8B-Instruct respectively.

• On MathVerse specifically, the 3B model gains were large. PLM-HoneyBee-3B scored 42.8, which the abstract describes as outperforming the SOTA model and base model by 7.8% and 24.8% respectively (matching 35.0 for Qwen2.5-VL-3B-Instruct and 18.0 for PLM-3B in the results table).

• Test-time scaling cut cost without losing accuracy. A proposed decoding strategy generates multiple solutions from HoneyBee-trained reasoners using 73% fewer inference tokens with no performance loss. The provided content does not give the detailed accuracy breakdown of this strategy beyond that figure.

Methodology in Plain English

The researchers treated dataset construction as a controlled experiment. They started from six published VL reasoning datasets containing (image, question, final answer) tuples, removed any images that matched evaluation images via perceptual hashing, and capped each source at 50K instances so that data quantity would not confound data quality. A single strong open-weights model, Llama4-Scout (109B total parameters, 17B active), generated the chain-of-thought for every image-question pair, and the training algorithm was held fixed as supervised finetuning. They trained two model scales, PLM-3B and PLM-8B, for 5 epochs each and picked the best of 5 checkpoints based on average accuracy across five benchmarks: MathVerse (testmini, vision-only, 788 examples), MathVista (testmini, 1000), MathVision (testmini, 304), MMMU-Pro (vision, 1730), and We-Math (testmini, 1740). That average became the yardstick for every subsequent comparison.

With that yardstick in place, they tested three families of changes. First, context curation: score the individual sources and then mix the top-2, top-4, and all of them. Second, data interventions, grouped by whether they target perception or problem-solving and by whether they replace data, augment it, or filter it — for example, perturbing images with rotations, distractor concatenation, or dominance-preserving mixup; re-rendering questions as text inside the image; filtering out examples the model could answer without the image or with only an image caption; prepending an image caption to the CoT; mixing in text-only reasoning data; rewriting questions to have ten options; keeping only above-median-length CoTs; and balancing difficulty levels. Third, scaling: holding images fixed at N/8 and varying questions per image (1, 2, 4, 8) or CoTs per image-question pair (1, 2, 4, 8), and separately varying images (N/8, N/4, N/2, N).

The best configurations were combined into HoneyBee, and models from 1B to 8B were trained on it. These were then evaluated on the five hill-climbing benchmarks plus five held-out ones — DynaMath (1000 examples), LogicVista (448), HallusionBench, the text-only MATH500 (500), and GPQA (198) — using accuracy, an identical prompt requiring boxed answers, and greedy decoding with a maximum generation length of 2048.

Why This Matters

Impact on research. The paper reframes multimodal reasoning data design as an empirically testable question rather than a matter of intuition. By fixing the training algorithm, the generator model, and the evaluation protocol, it isolates the contribution of data choices — and shows that many intuitively appealing interventions (visual perturbation, distractor scaling, difficulty balancing) actually hurt relative to a well-chosen baseline. It also shows that the largest single win came from an unimodal source: text-only reasoning data. Because the authors release code at github.com/facebookresearch/HoneyBee_VLM, others can reproduce and extend the comparisons.

Real-world applications:

• Visual data analysis, such as interpreting charts, tables, plots, and documents, which the paper cites as a motivation and which appears directly in the evaluation mix.
• Education, where step-by-step solutions to math and geometry problems shown alongside figures are useful for tutoring.
• Scientific discovery, where figures and quantitative reasoning must be combined.
• General-purpose assistants that need both image-grounded and text-only reasoning, since the text-only augmentation makes the model more capable outside vision tasks as well.

Industry relevance. State-of-the-art VL reasoners are trained on proprietary CoT recipes, so organizations that cannot reproduce those recipes are at a disadvantage. This paper supplies an open, reproducible alternative and shows that smaller models (1B to 8B) trained on curated data can beat larger or comparable instruction-tuned baselines on the reported evaluation set. The 73% reduction in decoding tokens for equivalent accuracy is directly relevant to inference cost. The pipeline also relies on an open-weights generator that can be hosted on a single A100 node with vLLM, making the recipe accessible without API dependence.

Future Directions

• The paper reports that mixing source datasets degraded performance relative to the best single source, which contradicts intuitions from text-only LLM data work; understanding when and why mixing helps remains unresolved.
• Most interventions failed to beat the baseline, suggesting the intervention list is not exhaustive. What other kinds of transformations could reliably improve data quality is left open.
• Scaling images, questions, and CoTs all helped in the reported experiments, but the optimal trade-off among the three axes under a fixed training budget is not resolved in the content provided.
• The test-time scaling result is reported only as a 73% reduction in inference tokens at no accuracy loss; the mechanism and its behavior across model scales are not detailed in the provided content.
• The HoneyBee-trained models are evaluated on ten datasets, several of which were unseen during curation, but how the recipe generalizes to domains beyond math-centric VL reasoning is not established.

Target Audience

Researchers and engineers working on multimodal model post-training, synthetic data generation, and dataset curation will get the most from this paper, particularly those deciding what to include in a vision-language reasoning SFT mixture. It is also relevant to practitioners who train small-to-mid-size VLMs (1B to 8B) and need practical guidance on which data investments pay off, and to evaluation-oriented researchers interested in controlled comparisons of data curation choices. Readers should be comfortable with benchmark tables, chain-of-thought supervision, and terminology such as replacement, augmentation, and filtering.

Authors’ abstract

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.