Skip to content

History / NeurIPS 2025

Revisions

  • Fix broken tables: escape unescaped pipes inside wikilinks in table rows ICLR.md (and 12 other pages) had [[label|Page]] wikilinks with raw pipes inside GFM table cells, which the GitHub-wiki renderer reads as column separators — mangling the table. Escaped the wikilink-internal pipes to \| (matching the convention already used by CVPR/NeurIPS/CoRL pages); table column separators left intact. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

    @Heungwoo Heungwoo committed Jun 11, 2026
  • Add NeurIPS 2025 survey + 15 standout VLA paper pages NeurIPS 2025 (San Diego + Mexico City, Dec 2-7 2025) accepted 5,290 papers (24.52% of 21,575 submissions). This adds a retrospective survey mirroring the ICLR 2026 and CoRL 2025 pattern: New venue / year pages: - NeurIPS.md, NeurIPS-2025.md (venue + year index) - NeurIPS-2025-VLA-Manipulation-Survey.md (main survey w/ CoRL 2025 -> NeurIPS 2025 -> ICLR 2026 lineage map) 15 standout per-paper pages: - Spotlights: Knowledge Insulation (PI, the published pi0.5 -> pi0.6 / pi0.7 bridge), SafeVLA (CMDP safety alignment), Latent Policy Barrier (OOD recovery via expert manifold), Robo2VLM-1 (684,710 VQA from 176k robot trajectories) - Architecture / Dual-system: ChatVLA-2 (dynamic MoE), DreamVLA (multi-modal world-knowledge forecasting), VLA-OS (controlled planning-representation ablation), Fast-in-Slow (embedded S1-in-S2), ThinkAct (NVIDIA RL visual plans), Chain-of-Action (ByteDance backward trajectory AR) - Training / inference: Real-Time Chunking (PI+Berkeley async chunk inpainting) - RL for VLA: What-Can-RL-Bring (PPO > DPO/GRPO empirical), ReinFlow (learnable noise enables online RL on flow matching) - World models: VideoVLA (joint action + video DiT as VAM) Key framings captured: - Knowledge Insulation is THE published bridge CoRL 2025 -> ICLR 2026. pi0.5 had it informally; KI formalizes it; pi0.6 and pi0.7 adopt it directly. - Six trends: dual-system default, VLM/expert reconciliation, routine RL post-training, world models as conditioning signal, safety/failure becomes first-class, humanoid+dex data via LLM sim synthesis. - Correction: HPT (2409.20537) is NeurIPS 2024, not 2025. Navigation: - Home.md: adds NeurIPS row to the venues table and NeurIPS 2025 survey to "Latest surveys". "What's new" surfaces the additions. - _Sidebar.md: adds NeurIPS section + NeurIPS 2025 subsection with all 15 paper pages. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

    @Heungwoo Heungwoo committed Apr 17, 2026