GRPO with a strict reward for structured JSON output improves tool-call accuracy on small models, from 0.98%-6.1% to 22%-71% on the xLAM benchmark, though without baselines or variance.
Report on a General Problem-Solving Program,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
REJECT 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Advancing SLM Tool-Use Capability using Reinforcement Learning
GRPO with a strict reward for structured JSON output improves tool-call accuracy on small models, from 0.98%-6.1% to 22%-71% on the xLAM benchmark, though without baselines or variance.