Research map / Probabilistic and scientific modeling / Molecules and proteins
Rethinking Instruction Drift as a Sampling Error: SNR-Aware Power Distributions for Long-Horizon Robotic Planning
Kewei Chen, Yayu Long, Mingsheng Shang
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 power-distribution sharpening (alpha>1 instead of alpha=1) (ablation) | CAPS MCMC search · RoboTwin 1.0 | success rate (%): 38.5 -> 47.4 | Improved |
| use MCMC iterative refinement instead of independent Best-of-N (ablation) | CAPS · RoboTwin 1.0 | success rate (%): 40.2 -> 47.4 | Improved |
| add SNR gating instead of always-on search (ablation) | CAPS · RoboTwin 1.0 | latency (x): 8.50 -> 2.15 | Improved |
| add SNR gating instead of always-on search (ablation) | CAPS · RoboTwin 1.0 | success rate (%): 48.1 -> 47.4 | No clear effect |
| add CAPS inference-time search | pi0 · RoboTwin 1.0 | success rate (%): 32.2 -> 47.4 | Improved |
| use CAPS instead of TACO reranking | pi0 · RoboTwin 1.0 | success rate (%): 41.3 -> 47.4 | Improved |
| add CAPS inference-time search | pi0 · Simpler-WindowX | success rate (%): 48.0 -> 60.5 | Improved |
| use CAPS instead of TACO | pi0.5 · RoboTwin 2.0 | success rate (%): 64.0 -> 66.2 | Improved |
| add CAPS inference-time search | pi0.5 · RoboTwin 2.0 (Lift Pot) | success rate (%): 62.0 -> 76.0 | Improved |
| use CAPS instead of TACO | pi0.5 · Libero-long | success rate (%): 96.6 -> 97.6 | Improved |
| add CAPS inference-time search | pi0.5 · Libero-long (Moka Pots on Stove) | success rate (%): 68.0 -> 91.0 | Improved |
| use CAPS instead of TACO | pi0 · Real-world XLeRobot | success rate (%): 53.0% -> 71.0% | Improved |
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.89)
- PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking (NeurIPS 2026 · similarity 0.88)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.88)
- FASTER: Rethinking Real-Time Flow VLAs (NeurIPS 2026 · similarity 0.88)
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.87)
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.
