Deep Research

Deep Research is the episode’s frame for agentic search, planning, evidence gathering, synthesis, and tool use over longer research tasks. In E242｜最快半年AI跑通自进化？与陈天桥首席科学家聊聊硅谷模型必争之地, Du Shaolei says Apodex started with Deep Research because search is a basic capability for post-training and self-improvement: a model that wants to improve itself must find relevant code, papers, examples, failure cases, and candidate answers.

The source treats Deep Research as part of a larger loop rather than only a user-facing report product. Apodex’s post-trained Qwen-based model is described as improving planning, search, and agent-team work, and benchmark results are discussed cautiously because search benchmarks can be contaminated by public answers.

Episode 17: 向量模型工程师：AI 的隐藏瓶颈与新时代的信息迷宫 adds the retrieval-engineering version. N 同学 / N Student argues that open-ended research is hard to evaluate because the quality of a broad report depends on retrieved evidence, multi-hop search paths, expert rubrics, and synthesis standards. This connects Deep Research to Retrieval-Augmented Generation, Reranking Models, and AI Search Evaluation rather than treating it as only a model-planning problem.

Key Claims

• Deep Research combines search, planning, long-horizon reasoning, and synthesis.
• Search is not a side feature; it supplies material for post-training task creation and verification.
• Agent-team design can strengthen Deep Research by letting different agents search, solve, cross-check, and preserve intermediate memory.
• Benchmark success is useful but fragile when answers can be found directly online or leak into training/evaluation.
• Deep Research becomes more valuable when connected to AI Verification and Recursive Self-Improvement, not just report generation.
• Deep Research quality depends on retrieval quality and relevance standards, not only on how well an agent writes the final report.

Connections

• Apodex, Du Shaolei, Li Beibin, and Qwen — company, speakers, and model context in the source.
• Recursive Self-Improvement, Discovery Model, and AI Verification — broader loop Deep Research supports.
• Multi-Agent Collaboration, Agent Harness, and Persistent Agent Memory — agent-system mechanisms needed for longer research tasks.
• Deerflow — existing open-source project in the wiki that began as a Deep Research-like capability.
• Retrieval-Augmented Generation, Reranking Models, and AI Search Evaluation — retrieval and evaluation layer added by the Fuyou Tiandi vector-model episode.