SPaRK, an offline RL method that rewards diverse tool use, reports 40.8% MMLU-Pro accuracy versus 33.0% without diversity, but missing controls and an underspecified reward weaken the claim.
https://blog.google/technology/google-deepmind/ gemini-model-thinking-updates-march-2025/
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Step-wise Policy for Rare-tool Knowledge (SPaRK): Offline RL that Drives Diverse Tool Use in LLMs
SPaRK, an offline RL method that rewards diverse tool use, reports 40.8% MMLU-Pro accuracy versus 33.0% without diversity, but missing controls and an underspecified reward weaken the claim.