Search papers, labs, and topics across Lattice.
Affiliation:
2
0
5
A checkpoint's static benchmark score cannot predict how it will respond to further training, but four tiny probe interventions can forecast downstream fine-tuning trajectories across entirely unseen model families with up to 78% lower error.
Arbitrarily splitting identical retrieved records into separate chunks swings an LLM's evidence weighting by up to 32 percentage points without altering a single word of context.
