Obj-Disco decomposes LLM alignment reward signals into sparse weighted combinations of interpretable natural language objectives via iterative analysis of behavioral changes across checkpoints, capturing over 90% of observed reward behavior.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Discovering Implicit Large Language Model Alignment Objectives
Obj-Disco decomposes LLM alignment reward signals into sparse weighted combinations of interpretable natural language objectives via iterative analysis of behavioral changes across checkpoints, capturing over 90% of observed reward behavior.