Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better · HackerPapers discussionshttps://hackerpaper.kaipfister.de/p/2505.23705/Collected HN comments and submissions linking this paper.enMon, 02 Nov 2026 11:16:22 GMTKnowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better — mentioned by ACCount37https://news.ycombinator.com/item?id=49155448urn:hackerpapers:mention:2505.23705:49155448Mon, 03 Aug 2026 13:20:18 GMT<p>Read. The. Papers. https://arxiv.org/pdf/2505.23705 https://www.pi.website/download/pistar06.pdf https://www.pi.website/download/pi07.pdf The thing literally has a diffusion &quot;action expert&quot; sit in the same attention system as a pre-trained VLM. And the VLM itself is ALSO trained to generate raw actions as a part of the training recipe (the first paper) - it just doesn&#x27;t do it at inference time. What the &quot;action expert&quot; does is parallelize the action generation process - based on VLM&#x27;s internal states. It&#x27;s exactly the thing you claimed to be impossible. Described in detail in a paper from…</p><p><a href="https://hackerpaper.kaipfister.de/p/2505.23705/">Paper and discussion history</a>. A mention is not an endorsement.</p>cs.LG