DR-Venus: Towards Frontier Edge-Scale Deep Research Agents with Only 10K Open Data

2026-03-01 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, long

Summary

DR-Venus is a 4B parameter deep research agent designed for edge-scale deployment, trained exclusively on approximately 10K open data points. It significantly outperforms existing agentic models under 9B parameters on deep research benchmarks like BrowseComp, BrowseComp-ZH, and xBench-DS-2510, while also reducing the performance gap to much larger 30B-class systems. The training methodology involves a two-stage process: initial agentic supervised fine-tuning (SFT) with strict data cleaning and long-horizon trajectory resampling, followed by agentic reinforcement learning (RL) using an Information-Gain Policy Optimization (IGPO) algorithm. This RL stage incorporates turn-level rewards based on information gain and format-aware regularization to enhance supervision density and credit assignment, crucial for small models tackling long-horizon tasks. The project releases its models, code, and key recipes to foster reproducible research.

Key takeaway

For NLP engineers developing edge-scale deep research agents, DR-Venus demonstrates that strong performance is achievable with 4B models and limited open data. You should prioritize rigorous data cleaning and long-horizon trajectory resampling during SFT, and implement turn-level reinforcement learning with information gain and format penalties to stabilize tool use and execution. Consider exploring test-time scaling techniques to further enhance your agent's capabilities.

Key insights

Small models can achieve strong deep research capabilities through high-quality data and effective utilization.

Principles

• Data quality and utilization are critical for small agent training.
• Turn-level RL improves long-horizon task reliability.
• Test-time scaling can unlock small model potential.

Method

DR-Venus uses a two-stage training: SFT with data cleaning and trajectory resampling, then RL with IGPO, turn-level information gain, and format-aware rewards for dense supervision.

In practice

• Oversample long-horizon trajectories for SFT.
• Implement turn-level format penalties in RL.
• Use LLM-judges for information gain rewards.

Topics

• DR-Venus
• Edge-Scale Deep Research Agents
• Small Language Models
• Agentic Supervised Fine-Tuning
• Information-Gain Policy Optimization

Code references

• inclusionAI/DR-Venus

Best for: NLP Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Mind DeepResearch Technical Report — MindDR achieves state-of-the-art deep research with smaller models via specialized multi-agent architecture and staged training.
• TLMs: Tiny LLMs and Agents on Edge Devices with LiteRT-LM — Cormac Brick, Google — Edge AI enables low-latency, private, and cost-effective on-device model deployment, leveraging tiny LLMs and agent skills.
• SuperIntelligence: Why the Future of AI is a File System (CORAL) — Coral enhances AI system intelligence by optimizing external file systems and multi-agent coordination, not by training LLMs.
• Google’s new Deep Research and Deep Research Max agents can search the web and your private data — Google's new Deep Research agents fuse web and private data, generating visual reports for enterprise workflows.
• Multi-Turn Evaluation of Deep Research Agents Under Process-Level Feedback — Deep Research Agents show initial gains from process-level feedback but struggle with sustained multi-turn improvement due to regression.
• The Sequence Opinion #782: The New Gradient: Research Directions That Will Ship in 2026 — AI development is shifting from generative scale to verifiable correctness, aiming for "System 2" reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.