Research map / Vision-language and multimodal / Robot action and navigation
Counterfactual VLA: Self-Reflective Vision-Language-Action Model with Adaptive Reasoning
Zhenghao Peng, Wenhao Ding, Yurong You, Yuxiao Chen, Wenjie Luo, Thomas Tian, Yulong Cao, Apoorva Sharma, Danfei Xu, Boris Ivanovic, Boyi Li, Yan Wang and 1 more
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 |
|---|---|---|---|
| add meta-actions as intermediate representation (ablation) | traj-only VLA · Internal driving val (no route) | minADE: 0.9283 -> 0.8411 | Improved |
| add language reasoning supervision to meta-act (ablation) | meta-act · Internal driving val (no route) | minADE: 0.8411 -> 0.8021 | Improved |
| add counterfactual reasoning (round1) on top of meta-act | meta-act · Internal driving val (no route) | minADE: 0.8411 -> 0.7650 | Improved |
| add counterfactual reasoning (round1) on top of meta-act | meta-act · Internal driving val (with route) | minADE: 0.7263 -> 0.6712 | Improved |
| add counterfactual reasoning (round1) to meta-act | meta-act · Internal driving val (no route) | collision rate: 0.0224 -> 0.0191 | Improved |
| add second round of counterfactual training (ablation) | CF-VLA (with route) · Internal driving val (with route) | avgADE: 1.4574 -> 1.3898 | Mixed |
| force reasoning on every sample (ablation) | CF-VLA (adaptive) · Internal driving val (no route) | minADE: 0.7650 -> 0.9319 | Hurt |
| let model decide when to think instead of never thinking (ablation) | CF-VLA · Internal driving val (no route) | minADE: 0.7897 -> 0.7650 (minADE); avgADE 1.4890 -> 1.5606 | Mixed |
| filter CF data by trajectory disagreement instead of using whole dataset (ablation) | CF-VLA (with route) · Internal driving val (with route) | minADE: 0.6811 -> 0.6712 (minADE); avgADE 1.4185 -> 1.4574 | Mixed |
| use CF editing of meta-actions instead of reusing filtered data as fixed meta-act labels (ablation) | meta-act · Internal driving val (no route) | minADE: 0.7906 -> 0.7650 | Improved |
| pre-fill ground-truth meta-actions (oracle) (ablation) | meta-act · Internal driving val (no route) | minADE: 0.8411 -> 0.4831 | Improved |
Closest papers by meaning
- AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.92)
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.91)
- OmniDrive: A Holistic Vision-Language Dataset for Autonomous Driving with Counterfactual Reasoning (CVPR 2025 · similarity 0.90)
- OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation (NeurIPS 2026 · similarity 0.90)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.89)
- DriveAgent-R1: Advancing VLM-based Autonomous Driving with Active Perception and Hybrid Thinking (ICLR 2026 · 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.
