Research map / Vision-language and multimodal / Multimodal LLMs

Bee: A High-Quality Corpus and Full-Stack Suite to Unlock Advanced Fully Open MLLMs

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Honey-Data-15M, a 15M-sample cleaned SFT corpus with dual-level (short/long) CoT enrichment built by the HoneyPipe/DataStudio pipeline, and trains Bee-8B, a fully open MLLM claimed to be SOTA among fully open models and competitive with semi-open InternVL3.5-8B.

Problem

Fully open MLLMs lag proprietary and semi-open models because open SFT datasets are noisy and lack complex long chain-of-thought reasoning data, and existing releases do not share adaptable curation pipelines.

Approach

A model-driven curation pipeline (rule and MLLM-based filtering, dual-level short/long CoT enrichment with fidelity verification by an LLM judge) that is released as a reusable framework rather than a static dataset, with the resulting corpus used to train an 8B fully open MLLM.

Key results

• Bee-8B-RL scores 67.0 on MathVerse (vision-only) vs 61.5 for InternVL3.5-8B, and 61.3 on LogicVista, 41.3 on DynaMath.
• Bee-8B-RL reports 84.8 on CharXiv DQ and 57.3 on CharXiv RQ, with an average of 70.2 over 29 benchmarks under the Qwen judge (69.9 under GLM judge).
• Ablation on 1M subsets: model trained on curated Honey-Data-1M outperforms Random-1M and surpasses Qwen2.5-VL-7B on nearly half of evaluated benchmarks.

Surprising findings

• Human evaluation shows 83.65% ties on accuracy, so enrichment mostly changes reasoning and style rather than factual content.
• Only 29 of 66,682 evaluation samples (0.043%) overlap with the training corpus at the tested thresholds.

Limitations

• Relies on a single Qwen2.5-VL-72B verifier and proprietary MLLMs for long CoT generation, with no quantified cost or proprietary model names.
• Bee-8B still trails InternVL3.5-8B on several benchmarks (e.g., MMMU 66.1 vs 73.4, MathVision 50.0 vs 56.8).

Experimental setup

Datasets: LLaVA-OneVision PixMo MAmmoth-VL VisualWebInstruct Vision-R1

Benchmarks: AI2D BLINK CountBench HallusionBench MMBench-CN MMBench-EN MMMU MMMU-Pro MMStar MMT-Bench MMVet MMVP

Metrics: accuracy average score across 29 benchmarks

Baselines: LLaVA-OneVision-7B Molmo-7B-D Qwen2.5-VL-7B Keye-VL-8B InternVL3.5-8B

Builds on: LLaVA-OneVision Qwen3 SigLIP2 GRPO VLMEvalKit

Closest papers by meaning

• HoneyBee: Data Recipes for Vision-Language Reasoners (CVPR 2026 · similarity 0.88)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.88)
• SEED-Bench: Benchmarking Multimodal Large Language Models (CVPR 2024 · similarity 0.88)
• MME-SCI: A Comprehensive and Challenging Science Benchmark for Multimodal Large Language Models (AAAI 2026 · similarity 0.87)
• metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.87)
• RedPajama: an Open Dataset for Training Large Language Models (NeurIPS 2024 · similarity 0.87)