An AI training method that treats an agent's choice between searching and answering as a question of confidence reported higher average accuracy than GiGPO across seven public question-answering benchmarks. MetaRAG reached 41.7% with Qwen2.5-3B-Instruct and 45.4% with Qwen2.5-7B-Instruct, 1.9 and 1.5 percentage points above GiGPO, respectively.
That approach is built around what the paper calls belief-action alignment: the quality of a Search-or-Answer decision is matched to the model's belief about whether it can answer. MetaRAG verifies proposed actions, probes that answerability belief and gates consistency credit by answer correctness. The belief probe is used during training only.
On the main tests, training used merged NQ and HotpotQA sets, while evaluation covered seven public QA benchmarks. The implementation used a 2018 Wikipedia dump and E5 retrieval of the top three passages for each query; training used GRPO with rollout groups of five, a maximum of four turns and eight A100 GPUs.
A retrieval trade-off, not a free gain
The accuracy comparison came with a retrieval trade-off. Relative to HiPRAG, MetaRAG was reported to have higher accuracy with fewer searches; relative to GiGPO, it had higher accuracy with more searches. In practical terms, the method occupies a different point on the trade-off between answer accuracy and retrieval use.
A post-hoc external validation examined under-search. An external LLM judge rated MetaRAG's under-search rate at 30.1%, compared with 35.6% for GiGPO, an absolute difference of 5.5 percentage points. The judge was not used for the reward, training or model-selection decisions.
The reported advantage also appeared in a separate zero-shot test on BrowseComp-Plus using BM25 retrieval. MetaRAG recorded 3.49% accuracy and 3.73% evidence recall, versus 2.65% and 3.10% for GiGPO. It was reported as the best among the compared 7B search agents on those measures.
What the component tests showed
In the component comparisons, the no-consistency-reward variant used 2.14 searches per question, compared with 1.60 for the default, while showing slightly higher accuracy. The no-verify-before-action variant had 40.4% accuracy. These were reported ablations of the method's components.
A separate analysis used two diagnostic scores, AUROC and PRR, to assess how well the model's internal signal marked the edge of its knowledge. MetaRAG's AUROC was 60.2 versus 59.5 for the base model, and its PRR was 21.8 versus 20.5. Both reported scores were higher for MetaRAG.
Additional settings produced similar reported comparisons. Under DAPO, MetaRAG scored 45.7% against GiGPO's 44.8% on Qwen2.5-3B-Instruct, while a PPO MetaRAG check scored 41.1%. The paper also reported higher performance than GiGPO on Llama3.2-3B and Qwen3-14B.
The approach was also combined with GiGPO's step-level credit assignment. Average accuracy was then 43.1% versus 39.8% for GiGPO alone on Qwen2.5-3B-Instruct, and 47.0% versus 43.9% on Qwen2.5-7B-Instruct, gains of 3.3 and 3.1 points.
The cost is still part of the story
MetaRAG's reported training time per step was 281 seconds, compared with 525 seconds for HiPRAG and 249 seconds for GiGPO. Its reported inference time was 0.1453 seconds per query. The belief probe is used during training, so these figures capture a method that adds training-time machinery while leaving a measurable query-time cost.
The evidence is a set of computational benchmark evaluations: seven public QA benchmarks plus the separate zero-shot BrowseComp-Plus test, run with the stated retrieval and training setup. The manuscript is an arXiv preprint, version 1, dated 25 Aug 2026. It reports support from National Natural Science Foundation of China grants 42394060 and 42394064 and the Ant Group Research Fund.
Paper data and sources
Original title: MetaRAG: Belief-Action Aligned Policy Optimization for Agentic RAG
Authors: Qiuyi Qi, Tian Liang, Jiamu Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text