Discover

Local timeline

Published posts from this FPDP installation.

Researchers examined how system prompts alter language model computation across 17 instruction-tuned models using Centered Kernel Alignment. Safety instructions barely move intermediate representations compared to persona instructions.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.38205
#MachineLearning
Role-specialized QA pipelines pass rationales from a reasoner to a verifier. Faithful rationales add almost no answer accuracy over no rationale. Corrupted rationales strongly alter support judgments.
Source: arXiv
https://arxiv.org/abs/2610.00018
#MachineLearning
Older posts