Evidence · DiggingBeagle record
LLMON: instruction and data channel separation by post-training
IBM Granite model post-training experiments distinguishing instruction-bearing and lower-trust data tokens.
- Published
- Oct 30, 2026
- Source role
- primary disclosure
Evidence record
IBM Granite model post-training experiments distinguishing instruction-bearing and lower-trust data tokens.
Claim-level citations (1)
- supportsTyped instruction/data model training reduced tested injection but left other failure classes: The LLMON authors report substantial indirect-injection reductions after typed instruction/data post-training, while acknowledging weaker results on other instruction conditions.
LLMON paper, structured-token training method and benchmark comparison tables