Research map / Vision-language and multimodal / Robot action and navigation
CofactVLA: Deconfounding Vision-Language-Action Models via Counterfactual Intervention
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.
Proposes CofactVLA, a causal deconfounding framework for VLA models that uses a language-masked counterfactual branch with action-level orthogonal projection guidance and feature-level counterfactual covariance reduction to counter visual shortcut overriding of language instructions.
Problem
Vision-language-action models suffer from "vision-override" causal confusion, where visual confounders dominate over language instructions, hurting generalization under distribution shift.
Approach
Formalizes VLA action generation as a Dual-path Deconfounding Graph; builds a language-masked counterfactual branch in one forward pass, projects the factual flow-matching velocity orthogonally away from the counterfactual velocity (OPG), and removes the top positive eigenspace of the factual-counterfactual covariance difference in attention KV features (CCR).
Key results
- CofactVLA achieves 98.5% average success on LIBERO vs. 96.9% for pi0.5 and 98.1% for X-VLA.
- On LIBERO-Plus zero-shot, CofactVLA reaches 69.1% total vs. 53.6% for pi0.
- Real-world OOD average success of 75.8% vs. 23.5% for pi0.5 (reported +52.3 points).
Surprising findings
- Plain additive and subtractive counterfactual interventions underperform the baseline ablations in some configurations (Add 94.0%, Sub 96.0% vs. baseline 97.0% average on LIBERO).
Limitations
- Bottlenecked by base VLM's zero-shot grounding; cannot synthesize unlearned concepts.
- Vulnerable to severe physical occlusions such as the arm blocking the camera.
- Failures under overlaid multiple perturbations in LIBERO-Plus.
Experimental setup
Datasets: LIBERO LIBERO-Plus custom real-world dataset (~400 trajectories)
Benchmarks: LIBERO LIBERO-Plus real-world robot tasks
Metrics: success rate
Baselines: OpenVLA-OFT pi0 pi0.5 DreamVLA X-VLA OpenVLA NORA WorldVLA UniVLA pi0-Fast CAG
Builds on: pi0.5 classifier-free guidance LIBERO LIBERO-Plus lerobot
Closest papers by meaning
- CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification (NeurIPS 2025 · similarity 0.90)
- DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge (NeurIPS 2025 · similarity 0.90)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · similarity 0.89)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 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.