Bankai applies XOR patches to 1-bit LLMs
Bankai is an open-source toolkit and paper about adapting true 1-bit LLMs by searching for sparse XOR masks over weight rows. The author reports that on Bonsai 8B, accepted bit flips can improve held-out behavior with extremely small patches, zero inference overhead, and instant apply/revert semantics. The project argues this is only practical on true binary models, where each weight is a single bit rather than a ternary encoding, and positions the approach as a lightweight alternative to adapter-based tuning for deployment on constrained devices.
Clever and unusually concrete, but the claim surface is wider than the evidence in the post. The interesting part is not just “bit flipping,” it is that the method turns post-training adaptation into a reversible, model-native edit primitive for true 1-bit weights.
- –The strongest idea is the XOR patch abstraction: if the model is truly binary, the patch is compact, exact, and reversible.
- –The reported results suggest high redundancy in the binary network, but the experiments sound narrow enough that broader robustness still needs validation.
- –The contrast with BitNet is important: ternary packed weights make naive XOR invalid, so this is genuinely architecture-specific rather than a generic compression trick.
- –The deployment angle is plausible: tiny patches and no per-token adapter overhead are attractive for edge and device-local serving.
- –The biggest open question is generality: whether sparse row flips scale to harder tasks, larger models, or stricter safety/behavior constraints without destructive side effects.
DISCOVERED
9d ago
2026-04-02
PUBLISHED
9d ago
2026-04-02
RELEVANCE
AUTHOR
Turbulent-Sky5396