Research map / Vision-language and multimodal / Multimodal LLMs
ProactiveBench: Benchmarking Proactiveness in Multimodal Large Language Models
Thomas De Min, Subhankar Roy, Stéphane Lathuilière, Elisa Ricci, Massimiliano Mancini
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 |
|---|---|---|---|
| scale model size from InternVL3-1B to 8B (ablation) | InternVL3 · ProactiveBench MCQA | accuracy: 27.1% vs 12.7% | Hurt |
| use newer larger LLaVA-OV-72B instead of LLaVA-1.5-7B (ablation) | LLaVA · ProactiveBench MCQA | accuracy: 24.8% vs 13.0% | Hurt |
| use Vicuna instead of Mistral LLM (ablation) | LLaVA-NeXT-7B · ProactiveBench MCQA | accuracy: 4.5% -> 19.3% | Improved |
| add proactive hints to the prompt (ablation) | MLLMs · ProactiveBench MCQA | proactive suggestion rate: +1.9 | Improved |
| add proactive hints to the prompt (ablation) | MLLMs · ProactiveBench MCQA | accuracy: 25.8% (+8.3%) | Mixed |
| condition on conversation history (ablation) | MLLMs · ProactiveBench MCQA | accuracy: -7% accuracy; psr 0.5 -> 1.8 | Hurt |
| add 1 or 3 few-shot in-context examples (ablation) | MLLMs · ProactiveBench ROD | accuracy: 6.7% with one sample, 12.0% with three | Hurt |
| replace valid proactive options with random invalid ones (ablation) | LLaVA-NeXT-Vicuna-7B · ProactiveBench MCQA | proactive action probability: 37% -> 49% | No clear effect |
| GRPO post-training with proactive reward r_p=0.5 | LLaVA-NeXT-Mistral-7B · ProactiveBench MCQA | accuracy: 4.5 -> 40.4 | Improved |
| GRPO post-training with proactive reward r_p=0.75 | Qwen2.5-VL-3B · ProactiveBench MCQA | accuracy: 11.0 -> 38.6 | Improved |
| raise proactive reward to r_p=1.0 (ablation) | Qwen2.5-VL-3B · ProactiveBench MCQA | accuracy: 38.6 -> 17.2 | Hurt |
| raise proactive reward to r_p=1.0 equal to correct reward (ablation) | Qwen2.5-VL-3B · ProactiveBench MCQA | proactive suggestion rate: 0.4 to 15.1 | Hurt |
Closest papers by meaning
- ProAct: A Benchmark and Multimodal Framework for Structure-Aware Proactive Response (ICML 2026 · similarity 0.88)
- EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents (ICML 2025 · similarity 0.87)
- MMBench-Live: A Continuously Evolving Benchmark for Multimodal Models (ICML 2026 · similarity 0.87)
- Grounding Multimodal Large Language Models in Actions (NeurIPS 2024 · similarity 0.87)
- PhysVLM-AVR: Active Visual Reasoning for Multimodal Large Language Models in Physical Environments (NeurIPS 2025 · similarity 0.86)
- TPRU: Advancing Temporal and Procedural Understanding in Large Multimodal Models (ICLR 2026 · similarity 0.86)
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.
