← News

Agents & MCPSources: arXiv · 18 Nov 2026

RICE-PO turns retrieval interactions into training signals for reasoning agents

The arXiv paper proposes a critic-free policy optimization framework that converts agent-retriever interactions into localized learning signals, reporting gains on BRIGHT and BEIR benchmarks.

In plain words

AI agents that search documents often reason in hidden steps before acting, and it is hard to know which reasoning actually helped. RICE-PO is a training method that uses the agent's own search interactions to figure out which reasoning steps deserve credit. The authors report it beats comparable training approaches on two retrieval benchmarks.

What was announced

Training language agents that search iteratively poses a credit-assignment problem, the paper's abstract explains: executable actions like queries or summaries can be evaluated directly by the retriever, but latent reasoning steps are not observable and only influence future actions. The authors argue this asymmetry makes outcome-level rewards unreliable, since the same final reward may credit reasoning that never shaped retrieval success.

RICE-PO addresses this by picking high-uncertainty executable actions as anchors, testing local counterfactual branches with retrieval metrics, and propagating credit to reasoning steps only when reasoning-to-action influence is strong and future residual effects are stable. On BRIGHT and BEIR, the framework reportedly outperforms prompt-based agents and group-based RL baselines under the same retriever setting, suggesting agent-environment interaction itself can supply useful supervision.

What it does not show

  • The evidence is the paper's abstract only; benchmark claims are author-reported and not independently confirmed, and the excerpt does not cover methodology details, baselines' identities, or reproducibility.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.