Research map / Vision-language and multimodal / Multimodal LLMs
Eagle 2.5: Boosting Long-Context Post-Training for Frontier Vision-Language Models
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.
Eagle 2.5 is a tiling-based VLM (Qwen2.5-based, 8B) trained with information-first sampling (Image Area Preservation, Automatic Degradation Sampling), progressive long-context post-training (32K→64K→128K), and a new Eagle-Video-110K long-video dataset.
Problem
Long-context vision-language modeling lags: existing long-context VLMs underperform proprietary models, fail to improve consistently with more visual input, and their training strategies and data recipes are under-explored, especially for long video and high-resolution inputs.
Approach
Combines an information-first sampling strategy (area-preserving tiling and a budget-based degradation that keeps all text and maximizes visual tiles/frames) with progressive increase of max context length during post-training, plus a dual-level (story-level and clip-level) annotated long-video dataset.
Key results
- Eagle2.5-8B reaches 72.4 on Video-MME (w/o subtitles) and 66.4 on LongVideoBench, surpassing InternVL2.5-78B on LongVideoBench (63.6) and being close to the 72B model on Video-MME.
- Eagle2.5-8B scores 75.6 average across 13 image benchmarks, with 94.1 DocVQA and 869 OCRBench, on par with Qwen2.5-VL-8B (75.6 average).
- Progressive 32K→64K training with Eagle-Video-110K raises Video-MME (w/o subtitle) from 67.9-68.1 to 68.8 and MLVU from 74.0-74.5 to 75.1 in ablations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Progressively train from 32K to 64K instead of direct 64K mixed training (ablation) | Eagle2.5-8B · MVBench | score: 71.3 -> 73.0 | Helped |
| Progressively train from 32K to 64K instead of direct 64K mixed training (ablation) | Eagle2.5-8B · MLVU Val | score: 74.0 -> 74.5 | Helped |
| Progressively train from 32K to 64K instead of direct 64K mixed training (ablation) | Eagle2.5-8B · Video-MME w/o subtitle | score: 67.9 -> 68.1 | Helped |
| Add Eagle-Video-110K to progressive 32K→64K Open-Data training (ablation) | Eagle2.5-8B · MVBench | score: 73.0 -> 73.9 | Helped |
| Add Eagle-Video-110K to progressive 32K→64K Open-Data training (ablation) | Eagle2.5-8B · MLVU Val | score: 74.5 -> 75.1 | Helped |
| Add Eagle-Video-110K to progressive 32K→64K Open-Data training (ablation) | Eagle2.5-8B · Video-MME w/o subtitle | score: 68.1 -> 68.8 | Helped |
| Add long-context training with Lmax=32K (ablation) | Eagle2.5-S2 · Image benchmarks | Avg Score: 74.8 -> 75.3 | Helped |
| Increase long-context training from Lmax=32K to Lmax=64K (ablation) | Eagle2.5-S2 · Image benchmarks | Avg Score: 75.3 -> 75.6 | Helped |
| Increase long-context training from Lmax=64K to Lmax=128K (ablation) | Eagle2.5-S2 · Image benchmarks | Avg Score: 75.6 -> 75.7 | Helped |
| Remove Image Area Preservation (ablation) | Eagle2.5-8B · InfoVQA and Perception-test | score | Hurt |
| Remove Automatic Degradation Sampling (ablation) | Eagle2.5-8B · image and video tasks | performance | Hurt |
| Add extensive image pre-training (ablation) | Eagle2.5-8B · MVBench and MLVU | score | Helped |
Surprising findings
- Adding long-context video data under the recipe does not harm and slightly benefits short-context image benchmarks.
- Extensive image pre-training notably boosts short video benchmarks like MVBench and MLVU but gives smaller gains on Video-MME.
Limitations
- Video-MME gains from progressive training and data are described as less pronounced in some comparisons.
- Some Eagle2.5 results and baselines rely on evaluation settings (frame limits, tiling choices) that differ across benchmarks.
- Eagle-Video-110K relies on GPT-4o/GPT-4 annotations, so label quality is tied to proprietary models.
Experimental setup
Datasets: ALLaVA Open-Data (Kinetics710, Something-Something-v2, ActivityNet, COIN, LLaVA-Video, SlideVQA, etc.) Eagle-Video-110K Eagle2.5-Image-SFT DenseFusion-1M Object365
Benchmarks: MVBench Perception_test EgoSchema MMB-Video MLVU LVBench Video-MME CG-Bench HourVideo Charade-STA DocVQA ChartQA
Metrics: accuracy mIoU average score
Baselines: GPT-4o Gemini-1.5-Pro Claude-3.5-Sonnet Qwen2.5-VL-8B InternVL2.5-8B LongVILA-8B VideoChat-Flash-8B LLaVA-Video-8B MiniCPM-V2.6-8B
Builds on: LLaVA Eagle-2 InternVL Qwen2.5 SigLIP
Closest papers by meaning
- CoVFT: Context-aware Visual Fine-tuning for Multimodal Large Language Models (CVPR 2026 · similarity 0.90)
- LongVILA: Scaling Long-Context Visual Language Models for Long Videos (ICLR 2025 · similarity 0.90)
- MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly (NeurIPS 2025 · similarity 0.89)
- Unleashing Hour-Scale Video Training for Long Video-Language Understanding (NeurIPS 2025 · similarity 0.89)
- VideoICL: Confidence-based Iterative In-context Learning for Out-of-Distribution Video Understanding (CVPR 2025 · similarity 0.89)
- LION: Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge (CVPR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.