Research map / Vision-language and multimodal / Robot action and navigation
BeTTER: Diagnose the Illusion of Embodied Reasoning in Vision-Language-Action Models
Haiweng Xu, Sipeng Zheng, Hao Luo, Wanpeng Zhang, Zongqing Lu
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 |
|---|---|---|---|
| scale InternVL-3.5 backbone up from 2B to 8B (ablation) | InternVL-3.5 · RefSp. | score: 6.00 -> 27.50 | Improved |
| scale InternVL-3.5 backbone up from 2B to 8B (ablation) | InternVL-3.5 · EgoPlan | score: 33.38 -> 40.80 | Improved |
| add VLM-VLA co-training data to 2B model (ablation) | InternVL-3.5 2B · RefSp. | score: 6.00 -> 29.50 | Improved |
| add VLM-VLA co-training data to 2B model (ablation) | InternVL-3.5 2B · EgoPlan | score: 33.38 -> 31.95 | No clear effect |
| use multi-patch high-res input instead of single 224px image (ablation) | InternVL-3.5 2B + VLM + VLA data · RefSp. | score: 7.50 -> 29.50 | Improved |
| add VLM pre-training to VLA-only pre-training (ablation) | VLA policy · CALVIN ABC->D | avg length: 3.86 -> 4.09 | Improved |
| add VLM pre-training to VLA-only pre-training (ablation) | VLA policy · LIBERO | success rate: 96.3 -> 97.2 | No clear effect |
| add VLA (OXE) pre-training vs no pre-training (ablation) | VLA policy · LIBERO | success rate: 97.4 -> 96.3 | No clear effect |
| add VLA (OXE) pre-training vs no pre-training (ablation) | VLA policy · CALVIN ABC->D | avg length: 3.67 -> 3.86 | Improved |
| test on unseen composition B->C | GR00T-N1.6 · Packing Fast Food Order | success (%): 75.0/40.0 -> 15.0 | Hurt |
| recombine spatial and semantic instructions under counterfactual layouts | GR00T-N1.6 · coffee task | target success (%): spatial 100.0%; semantic 5.0% | Mixed |
Closest papers by meaning
- Robobench: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models as Embodied Brain (ECCV 2026 · similarity 0.92)
- Thinking from the Robot’s View: The CoT-HRC Benchmark for Human Intent Reasoning in Embodied Collaboration (ECCV 2026 · similarity 0.91)
- Guide, Think, Act: Interactive Embodied Reasoning for Vision-Language-Action Model (ECCV 2026 · similarity 0.91)
- EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents (ICML 2025 · similarity 0.90)
- Seeing Across Views: Benchmarking Spatial Reasoning of Vision-Language Models in Robotic Scenes (ICLR 2026 · similarity 0.90)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
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.
