Research map / Vision-language and multimodal / Visual question answering
Large Language Models are Visual Reasoning Coordinators
Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, Ziwei Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| coordinate two VLMs with instruction-tuned LLM instead of averaging ensemble | BLIP+OFA · A-OKVQA val (MC) | accuracy: 56.6 -> 77.7 | Improved |
| instruction-tune the coordinator LLM (Cola-FT) vs 2-shot in-context (Cola-Zero) | Cola · A-OKVQA val (MC) | accuracy: 70.4 -> 77.7 | Improved |
| instruction-tune coordinator vs ensemble on e-SNLI-VE | BLIP+OFA · e-SNLI-VE val | accuracy: 48.8 -> 81.6 | Improved |
| use Cola-Zero in-context coordination vs ensemble on OK-VQA | BLIP+OFA · OK-VQA val | accuracy: 39.2 -> 39.4 | No clear effect |
| add a second VLM (use both BLIP and OFA instead of one) (ablation) | Cola-FT · A-OKVQA val (MC) | accuracy | Improved |
| include VLM plausible answers in addition to captions (ablation) | Cola-FT · A-OKVQA val (MC) | accuracy | Improved |
| keep correct VLM answer labels at tuning and evaluation vs perturbed labels (ablation) | Cola-FT · A-OKVQA val (MC) | accuracy | Improved |
| train LLM to output rationale before answer (ablation) | Cola-FT · A-OKVQA val (MC) | accuracy: 77.7 -> 74.3 | Hurt |
| use three different OFA sizes instead of identical models, Cola-FT vs ensemble gap | Cola-FT · A-OKVQA val (MC) | accuracy: 46.62 -> 54.26 | Improved |
| scale number of OFA-base VLMs beyond 5 (Cola-FT) or 3 (Cola-Zero) (ablation) | Cola · A-OKVQA val (MC) | accuracy: 49.77% at 5 VLMs | No clear effect |
| use a FLAN-T5 80M coordinator (small) vs baselines (ablation) | Cola-FT · A-OKVQA val (MC) | accuracy: 55% -> 65% | Improved |
Closest papers by meaning
- Prism: A Framework for Decoupling and Assessing the Capabilities of VLMs (NeurIPS 2024 · similarity 0.92)
- Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.92)
- Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.91)
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.91)
- Latent Visual Reasoning (ICLR 2026 · similarity 0.91)
- VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs (CVPR 2026 · similarity 0.91)
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.
