Source-verification audit: fix fabrications, rename PI slugs, link hygiene, add foundational pages - Re-verified all 195 pages vs original sources; removed 47 confirmed fabricated tables/numbers, restored 19 false-positive deletions (full-PDF re-check) - Renamed PI tech-report slugs ICLR-2026-pi07/pi06/RECAP -> PI-pi07/PI-pi06/PI-RECAP (these are PI technical reports, not ICLR 2026 papers); updated 171 wikilinks - Fixed 60 broken wikilinks -> 0 dangling across the wiki - Added foundational pages: OpenVLA, ReKep, AgiBot World Colosseo, RoboBrain 2.0; linked from Home + sidebar + venue indexes Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add in-depth review for Fast-in-Slow (NeurIPS 2025) Review-Fast-in-Slow.md is a long-form companion to the existing NeurIPS 2025 Fast-in-Slow summary page. Sources are arXiv 2506.01953 and the project site. Covers: - Dual-system VLA context table comparing GR00T N1 (cascaded), Fast-in-Slow (embedded), ChatVLA-2 (MoE-routed), ThinkAct (RL-visual-plan) — four distinct architectural patterns at one conference - Figure 2 (FiS-VLA framework) embedded from the paper with attribution + a mermaid reconstruction showing blue = shared- parameter blocks - Full method: "last 2 transformer blocks of the LLM repurposed as System 1" via partial parameter sharing; 1:4 S2:S1 frequency ratio; heterogeneous inputs (S2 sees only text+2D; S1 also sees robot state + 3D point clouds + noised actions); dual-aware co-training L_fast (diffusion MSE) + L_slow (autoregressive CE) - Results: RLBench 69% (vs CogACT 61%, π0 55%); AgileX 68% (vs π0 59%); AlphaBot 74% (vs π0 61%); largest single-task gain fold-towel 40 -> 60%; headline +8% sim / +11% real - Control frequency: 117.7 Hz on NVIDIA 4090 @ chunk=8, ~10x faster than cascaded dual-system baselines - Ablations: 2 shared blocks optimal (saturates); 1:4 frequency optimal; removing L_slow drops RLBench 69 -> 62%; each modality (robot state, 3D point clouds) contributes substantially - Limitations (authors' + reviewer): statically-configured sharing and frequency ratio; 19-29% OOD performance drop; no head-to-head vs pi0.6; no LIBERO numbers; fixed sharing depth raises transferability questions Assets: - assets/fis_fig2_framework.png (extracted from arXiv 2506.01953 page 4, cropped to the framework figure) Navigation: - NeurIPS-2025-Fast-in-Slow.md adds "In-depth review" link - _Sidebar.md per-paper reviews section adds Fast-in-Slow - Home.md per-paper reviews table + "What's new" surface the addition Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>