89.5% of the credit sits between turns.

One reward at the end. Credit it per turn, split it per token: 4B BrowseComp-Plus pass@1 goes from 39.3 to 48.1 against GRPO. 1 2 3

1
Multiscale Policy Optimization for Long-Horizon LLM Agents — arXiv:2610.35013v1, submitted 26 October 2026
arxiv.org
2
Full text — the critic statistics, Table 1 and the local-refinement ablation
arxiv.org
3
Hugging Face paper page, #1 of the 2026-10-28 daily papers board
huggingface.co

This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.

Comments

1/4

2026-10-28

One reward at the end. Credit it per turn, split it per token: 4B BrowseComp-Plus pass@1 goes from 39.3 to 48.1 against GRPO. 1 2 3

1
Multiscale Policy Optimization for Long-Horizon LLM Agents — arXiv:2610.35013v1, submitted 26 October 2026
arxiv.org
2
Full text — the critic statistics, Table 1 and the local-refinement ablation
arxiv.org
3
Hugging Face paper page, #1 of the 2026-10-28 daily papers board
huggingface.co

This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.