Most Current Model Organisms Leak: Perplexity Differencing Often Reveals Finetuning Objectives

·LessWrong··

Authors: Mohammad Abu Baker, Luca Baroni, Daniel WilhelmPaper: https://arxiv.org/abs/2605.00994Code: https://github.com/z3research/ppldiff-paperTwitter thread: https://x.com/m_shahoyi/status/2071892578476110136Top-ranked revealing completions can be inspected here: https://z3research.org/This post summarizes the paper and adds a few extra reflections in DiscussionTL;DRWe found that many current publicly available model organisms (MOs) "leak" instilled behaviorsWe present a simple contrastive met...

Read full article →

Related Articles

US sanctions force The Netherlands off Microsoft and toward alternative NixOS
mywacaday · Hacker News · 10h ago
500k facial scans at UK stations yield no arrests, 1 false positive
ilamont · Hacker News · 10h ago
How Delhi cut electricity loss from 50 to 5 percent
rbanffy · Hacker News · 9h ago
A Privacy Analysis of Web and Mobile Conversational AI Agents [pdf]
damaru2 · Hacker News · 12h ago
Does Reddit have an astroturfing problem? What the data suggests
p-s-v · Hacker News · 1d ago