DeepDive vs OpenResearcher
THUDM recipe for deep-search agents: synthesize hard multi-hop QA from knowledge-graph random walks, then multi-turn GRPO RL — DeepDive-32B hits 14.8% BrowseComp; data feeds GLM-4.5/4.6. — versus — TIGER-AI-Lab's fully open deep-research recipe: 96K long-horizon trajectories (adopted by NVIDIA Nemotron), a 30B-A3B model hitting 54.8% BrowseComp-Plus, training code and eval harness.
Rival open recipes for training deep-research agents: DeepDive synthesizes hard QA from knowledge-graph walks and trains with multi-turn RL on slime (checkpoints still pending); OpenResearcher distills 96K long trajectories from GPT-OSS-120B and ships data, 30B model and eval framework complete.
| DeepDive | OpenResearcher | |
|---|---|---|
| Stars | 333 | 976 |
| Forks | 35 | 93 |
| Language | Python | Python |
| License | — | — |
| Last activity | 29 days ago | 1 months ago |
| Topics | training | training |
| Curated connections | 2 | 2 |
DeepDive — the curator's take
Read it for the data trick: random-walk paths through KILT/AMiner knowledge graphs, entity obfuscation into 'blurry entities', then difficulty-filtering that keeps only questions GPT-4o fails four times out of four. Strict binary rewards (format AND answer, else zero) resist reward hacking, and the test-time finding is counterintuitive gold — among 8 parallel trajectories, the answer reached with the FEWEST tool calls wins (24.8% vs 12.0% single-shot). Use it to train or study open-model search agents; NOT a runnable product — model checkpoints are still 'coming soon', and you bring Serper/Jina API keys plus slime training infra. The 4,108-entry dataset is open on HF and already went into GLM-4.5/4.6.
OpenResearcher — the curator's take
The open counterpoint to closed Deep Research products — and the data is the crown jewel: 100+-turn research trajectories distilled from GPT-OSS-120B over a self-built 11B-token retriever corpus (no search-API bills at generation scale), good enough that NVIDIA folded it into Nemotron 3 Ultra. The 30B-A3B model beats GPT-4.1, Claude-Opus-4 and Gemini-2.5-Pro on BrowseComp-Plus. Reproducing anything is a real commitment: the setup assumes 8×A100, training lives in a Megatron-LM fork, and the local retriever needs Java + tevatron. ⚠ No LICENSE file in the repo — clarify terms before commercial use. Pick DeepDive for the KG-synthesis + multi-turn-RL recipe; OpenResearcher for large-scale SFT distillation with everything — data, model, eval — actually released.