BAAI reports benchmark and ablation results for agentic academic research tasks

BAAI benchmarks more than 13 frontier models and end-to-end research agents on academic research tasks, using a ReAct framework with the DeepXiv full-text search environment.

Claude-Opus-4.6 leads Deep Research accuracy at 9.39%, followed by Gemini-3.1-Pro (7.93%), GPT-5.4 (7.44%), and open-source leader Qwen3.5-397B (6.97%). On Wide Research IoU, Gemini-3.1-Pro-Preview peaks at 9.31%, with most baselines below 5%, according to BAAI.

In separate findings, BAAI says explicit reasoning mode gives no consistent gains on these tasks and often adds redundant searches and latency on Wide tasks. Replacing DeepXiv full-text retrieval with open-web search lowers Deep accuracy from 5.42% to 3.97%. BAAI also reports that scaling inference compute substantially improves Deep Research but gives only marginal gains on Wide tasks, which it attributes to recall bottlenecks.