Summary

DR Tulu is an open post-training recipe and framework for long-form deep research agents, with DR Tulu-8B trained via Reinforcement Learning with Evolving Rubrics (RLER) to directly optimize performance on open-ended, long-form research tasks such as multi-step literature synthesis and multi-document attributional answers.We introduce Deep Research Tulu (DR Tulu), an open post-training recipe and framework for long-form deep research agents.^[3]Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research.^[1]Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research.^[12]View PDF. Abstract:Deep research models perform multi-step research to produce long-form, well-attributed answers.^[1] By contrast, AI2 Asta is an agentic research platform by Allen Institute for AI designed to accelerate science with trustworthy agents, providing a production-grade set of tools for search, synthesis, and analysis, plus an evaluation mode to support rigorous, fair assessment of scientific workflows.Asta, an agentic tool for scientists · Find papers helps you discover relevant research using an LLM-powered search experience, like Google ...^[17]Asta Agents unite search, synthesis, and analysis to help scientists explore data, generate insights, and maintain scientific rigor.^[18]Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, ...^[20] In short, DR Tulu targets open-ended, long-form research as a primary objective, while Asta targets a broader ecosystem of scientific workflows and protocolized tasks with tool and evaluation support.

1) Scope and Target Tasks

DR Tulu is explicitly positioned for open-ended, long-form “deep research,” emphasizing multi-step research that produces long-form, well-attributed answers and synthesizes information across sources.We introduce Deep Research Tulu (DR Tulu), an open post-training recipe and framework for long-form deep research agents.^[3]View PDF. Abstract:Deep research models perform multi-step research to produce long-form, well-attributed answers.^[1] This framing centers on literature-grounded reasoning that accumulates citations and maintains attribution over extended reasoning traces.

AI2 Asta, by contrast, is an agentic platform aimed at accelerating scientific work, integrating functions for literature search, data exploration, synthesis, and analysis into a cohesive workflow oriented around protocols and research tasks.Asta, an agentic tool for scientists · Find papers helps you discover relevant research using an LLM-powered search experience, like Google ...^[17]Asta Agents unite search, synthesis, and analysis to help scientists explore data, generate insights, and maintain scientific rigor.^[18] Its design is broader—spanning discovery, data interrogation, and insight generation—rather than primarily long-form prose synthesis.

2) Training and Operational Paradigm

DR Tulu’s novelty lies in training with Reinforcement Learning with Evolving Rubrics (RLER), a curriculum and reward-learning approach intended to directly optimize long-form research behaviors, producing DR Tulu-8B as an open model trained for open-ended deep research.Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research.^[1]Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research.^[12] The project emphasizes an end-to-end, open recipe so others can reproduce and extend the training procedure.We introduce Deep Research Tulu (DR Tulu), an open post-training recipe and framework for long-form deep research agents.^[3]

AI2 Asta emphasizes building and evaluating agentic scientific workflows using an accessible set of tools and an evaluation mode to assess agent behavior and results, supporting search, synthesis, and analysis as first-class operations in a scientific context.Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, ...^[20]Asta Agents unite search, synthesis, and analysis to help scientists explore data, generate insights, and maintain scientific rigor.^[18] Unlike DR Tulu’s focus on training a model explicitly for open-ended deep research, Asta’s primary contribution is a framework, toolset, and evaluation infrastructure for scientific agents rather than a model trained directly on long-form research curricula.

3) Tools, Interaction Model, and Retrieval

DR Tulu positions itself as a research agent framework for long-form outputs, implying heavy use of planning, multi-step retrieval, and citation-grounded synthesis to produce well-attributed answers over long horizons.We introduce Deep Research Tulu (DR Tulu), an open post-training recipe and framework for long-form deep research agents.^[3]View PDF. Abstract:Deep research models perform multi-step research to produce long-form, well-attributed answers.^[1] The training method (RLER) is oriented to evolving behavioral norms for extended research traces, with rubrics shaping what constitutes a “good” multi-step solution.

Asta provides a broader toolbox for scientific workflows—unifying paper discovery, data analysis, and synthesis—along with an evaluation harness that supports fair and rigorous comparison of how agents perform across these functions, emphasizing protocolized tasks and maintaining rigor during automated exploration.Asta, an agentic tool for scientists · Find papers helps you discover relevant research using an LLM-powered search experience, like Google ...^[17]Asta Agents unite search, synthesis, and analysis to help scientists explore data, generate insights, and maintain scientific rigor.^[18]Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, ...^[20] This makes Asta more of an “agent platform with tools” than a model directly optimized for long-form prose.

4) Benchmarks and Evaluation Emphasis

DR Tulu’s evaluation centers on long-form, multi-source research benchmarks where models must generate multi-step plans, retrieve and synthesize evidence, and provide attribution—i.e., deep research tasks rather than narrowly scoped lab protocols.We introduce Deep Research Tulu (DR Tulu), an open post-training recipe and framework for long-form deep research agents.^[3]View PDF. Abstract:Deep research models perform multi-step research to produce long-form, well-attributed answers.^[1] Public discussions of the work report that DR Tulu-8B beats prior open deep research models across long-form benchmarks by double-digit points on average, indicating a focus on long-form synthesis quality as a primary metric.Across four long-form benchmarks, DR Tulu-8B beats all existing open deep research models by 13.7–53.4 points on average, including larger 30B ...^[6]

Asta’s evaluation emphasizes its toolset and the ability to execute scientific workflows with an embedded evaluation mode for fair comparison, spanning discovery (literature search), synthesis, and analysis over structured tasks and protocols rather than purely prose benchmarks.Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, ...^[20]Asta Agents unite search, synthesis, and analysis to help scientists explore data, generate insights, and maintain scientific rigor.^[18] The platform is oriented to measurable progress on science-specific tasks, including data-driven discovery, rather than to maximizing unstructured long-form text quality.

5) Openness, Maintenance, and Ecosystem

DR Tulu is presented as an open, end-to-end training recipe and framework, including the 8B model, with linked repositories, model cards, and documentation aimed at reproducibility and extension by the community.We introduce Deep Research Tulu (DR Tulu), an open post-training recipe and framework for long-form deep research agents.^[3]Official repository for DR Tulu: Reinforcement Learning with Evolving Rubrics for Deep Research Paper • Data & Models • Blogpost • Video• Static Demo (Our ...^[4]DR Tulu: Reinforcement Learning with Evolving Rubrics for Deep Research Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research.^[2] This promotes a cycle where evolving rubrics and training data can be iteratively improved for long-form research.

Asta is part of AllenAI’s AI for Science ecosystem and emphasizes a production-grade resource set and evaluation support; it also aims for trustworthiness and transparency across scientific workflows, providing tools and an evaluation mode to facilitate rigorous, fair agent development and auditing.Asta, an agentic tool for scientists · Find papers helps you discover relevant research using an LLM-powered search experience, like Google ...^[17]Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, ...^[20] Its openness focuses on a modular, extensible agent framework and toolset rather than on a single model trained for long-form synthesis.

When to Use Which (and Where They Overlap)

• Choose DR Tulu when your primary need is long-form, multi-source research synthesis with attribution and multi-step planning, and you want a model and training recipe explicitly optimized for open-ended research tasks.Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research.^[1]View PDF. Abstract:Deep research models perform multi-step research to produce long-form, well-attributed answers.^[1]
• Choose AI2 Asta when you need a rigorous, tool-rich agent platform to automate scientific workflows—covering literature discovery, data analysis, and synthesis—with built-in evaluation for protocolized tasks and experiments.Asta Agents unite search, synthesis, and analysis to help scientists explore data, generate insights, and maintain scientific rigor.^[18]Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, ...^[20]

The overlap is greatest in agentic research: both target scientific literature understanding and synthesis, but they differ in emphasis—DR Tulu prioritizes long-form, attribution-rich outputs trained with evolving rubrics; Asta prioritizes a production-grade, evaluated suite of tools and workflows for scientific tasks, with an emphasis on rigor and fair evaluation across research agents.We introduce Deep Research Tulu (DR Tulu), an open post-training recipe and framework for long-form deep research agents.^[3]Asta Agents unite search, synthesis, and analysis to help scientists explore data, generate insights, and maintain scientific rigor.^[18]Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, ...^[20]

Notable Reported Performance Evidence

Early reports on DR Tulu-8B indicate substantial gains over prior open deep research models on long-form benchmarks, highlighting the impact of training directly for open-ended research with evolving rubrics rather than general-purpose instruction following.Across four long-form benchmarks, DR Tulu-8B beats all existing open deep research models by 13.7–53.4 points on average, including larger 30B ...^[6] This supports the thesis that DR Tulu’s training paradigm is tuned for multi-step, long-form synthesis quality.

Bottom Line

• DR Tulu: open, model- and recipe-focused system trained with RLER for open-ended, long-form deep research, producing attribution-rich, multi-step answers; emphasizes long-form benchmarks and rubric-driven training evolution.Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research.^[1]View PDF. Abstract:Deep research models perform multi-step research to produce long-form, well-attributed answers.^[1]
• AI2 Asta: agentic research platform with a production-grade toolset spanning search, synthesis, and analysis, plus a dedicated evaluation mode to fairly assess scientific workflows; oriented to protocolized scientific tasks and trustworthy automation rather than maximizing long-form prose quality.Asta Agents unite search, synthesis, and analysis to help scientists explore data, generate insights, and maintain scientific rigor.^[18]Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, ...^[20]

In practice, DR Tulu is to long-form research synthesis as Asta is to protocolized scientific workflow automation: the former optimizes for high-quality, multi-source long-form outputs and rubric-driven training; the latter optimizes for reliable, evaluable execution across a broad toolkit for scientific discovery and analysis.