Discover

Local timeline

Published posts from this FPDP installation.

Large language models can treat a trailing user comment as part of pasted text, a phenomenon called absorption. Across 20 models, absorption at a bare newline ranges from 7.7% to 66.7%. Explicit boundaries reduce but do not remove this failure.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04210
#MachineLearning
Researchers measured how reference choice moves a benchmark score by holding the scored output fixed and exchanging the reference. The score moves by 4.95 F1 points for one output and 2.00 for the other.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.03825
#MachineLearning
The study finds that internal interventions that affect self-report do not necessarily change behavior, and vice versa, with only task-specific directives influencing both.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04125
#MachineLearning
Older posts