AI

Semifactual Credit-Augmented Policy Optimization

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 03:18:25 UTC.

Briefings

Researchers developed SCAPO, a causally inspired variant of GRPO that uses semifactual stability to improve large language model reasoning. The study tested SCAPO on Qwen3-4B-Base and Qwen3-1.7B-Base, showing it increased AIME 2024-2026 accuracy by 5.63 percentage points on the 4B model and 4.17 percentage points on the 1.7B model. SCAPO reduces advantages for unstable tokens during early training without updating model weights, addressing limitations in standard RLVR methods.

Key points

• SCAPO improves AIME 2024-2026 accuracy by 5.63 percentage points on Qwen3-4B-Base compared to GRPO.
• SCAPO improves AIME 2024-2026 accuracy by 4.17 percentage points on Qwen3-1.7B-Base compared to GRPO.

Context

This research addresses limitations in reinforcement learning with verifiable rewards (RLVR) for large language models, specifically how high-drift token candidates affect reasoning accuracy. The study compares SCAPO against Group Relative Policy Optimization (GRPO) to evaluate the impact of semifactual stability on token-level credit assignment.

Limitations

Results are limited to two model scales (Qwen3-4B-Base and Qwen3-1.7B-Base) and specific benchmarks (AIME 2024-2026 and out-of-distribution benchmarks). The study does not provide causal claims beyond observed accuracy improvements, and the code is available on GitHub but lacks detailed experimental methodology in the provided text.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 16:38:22 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiesqq2ilwjioqyxgxu4yenwgabdt37kw6xwv46yuzpra53uzd6fia