LLM Reddit 3h ago 2 min read
Bankai, highlighted in LocalLLaMA, proposes post-training adaptation for true 1-bit LLMs by applying sparse XOR patches directly to binary weights. According to the GitHub repo and paper, patches around 1 KB changed Bonsai 8B behavior with zero inference overhead, fixed 4 of 17 held-out failures without breaking 13 already-correct cases, and could be applied or reverted with the same XOR operation in microseconds.