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>
Fix all mermaid diagrams in ML-Attention and ML-Normalization Quote every node and subgraph label in every mermaid block, and rewrite labels to avoid the characters that break GitHub's mermaid parser when used inside bare square brackets: - parentheses — interpreted as an alternate node shape - curly braces {N×d} — interpreted as rhombus nodes - colons — interpreted as classDef assignments (:::) - slashes / plus signs / ampersands / em-dashes — parser edge cases Affected diagrams: ML-Attention: §2 math baseline §3.1 MHA / MQA / GQA subgraphs §3.2 causal vs bidirectional vs block-causal §3.4 cross-attention §3.5 QK-Norm §5.2 Qwen3.5 hybrid block stack §5.5 GR00T AlternateVLDiT §5.6 π-series prefix-KV §5.7 DDVLA bidirectional action span ML-Normalization: §1 normalization taxonomy (also rename IN -> INORM, add shipped legend) §4 pre-norm vs post-norm §5 QK-Norm diagram §6 AdaLN diagram §8.5 GR00T N1 -> N1.7 vlln evolution Unicode math and special chars now live only in prose; diagrams use plain English equivalents (sqrt d_head, K_transposed, etc.).
Fix §1 attention taxonomy diagram in ML-Attention The original mermaid tree had two rendering issues: (1) unquoted parentheses in node labels (O(N²), Performer / Linformer) tripped GitHub's mermaid parser and (2) 9 flat children under SOFT made the TB layout cramped. Fix: group the variants into 5+4 categories with quoted labels, and highlight the shipped-in-2026 variants with a classDef fill. Same information, cleaner render.
Add ML Foundations section: Attention and Normalization reviews Two basic-ML building-block review pages with mermaid diagrams and concrete mapping to what Qwen3 / Qwen3.5 and the major VLAs actually use: - ML-Attention.md: full / causal / bidirectional / sliding-window / cross / GQA-MQA-MHA / linear / Gated DeltaNet / QK-Norm. Covers Qwen3 (GQA + QK-Norm), Qwen3.5 (3:1 Gated DeltaNet / Gated Attention hybrid, Feb 2026), Qwen3-VL (Interleaved-MRoPE, DeepStack), plus pi-series prefix-KV, GR00T AlternateVLDiT, DDVLA bidirectional action span, Fast-in-Slow embedded blocks. - ML-Normalization.md: BN / LN / RMSNorm / GN / IN / AdaLN(-Zero) / adaptive RMSNorm / QK-Norm / FiLM, pre vs post-norm. Covers Qwen3 per-head QK-RMSNorm (V not normalized, eps=1e-6), GR00T N1 -> N1.7 vlln + vl_self_attention progression, pi0.7 adaptive-RMSNorm timestep injection, RDT-1B's rejection of AdaLN. - ML.md: hub page linking the two reviews. - Home.md and _Sidebar.md: new ML foundations section wired in.