AI
SLCA-GRPO: Resolving Cross-Segment Credit Misattribution in Tool-Calling RL
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 23/09/2026, 20:00:00 UTC
Briefing created: 28/09/2026, 03:19:28 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/09/2026, 03:18:25 UTC.
Briefings
SLCA-GRPO addresses cross-segment credit misattribution in tool-calling RL by decoupling advantage estimation at the structural segment level. The framework constructs the Schema-Guided LLM Simulator (SGLS) as foundational training infrastructure and uses Hierarchical Rewards to route execution advantages to tool tokens while isolating preference advantages to summary tokens. On a 7B backbone, SLCA-GRPO achieved +2.53 points on in-domain evaluation, +1.36 points on the Berkeley Function-Calling Leaderboard, and +9.15 points on the 4-Bench under the same training budgets as standard GRPO.
Key points
- SLCA-GRPO decouples advantage estimation at the structural segment level to prevent gradient noise from summary generation from contaminating tool-decision tokens.
- The framework constructs the Schema-Guided LLM Simulator (SGLS) as foundational training infrastructure to enable scalable exploration without costly real APIs.
Context and limitations
Results are based on a 7B backbone model; evaluation metrics include in-domain performance, the Berkeley Function-Calling Leaderboard, and the 4-Bench under identical training budgets.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 28/09/2026, 03:19:28 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigfsvnkufifh3vzpa2yubtwrhqonwe2soiiww4syufapuqeooped4