framework
active
framework:sfr-deepresearch

SFR-DeepResearch

The paper's core contribution: an RL-based framework for training autonomous single-agent LLMs to perform deep research with web search, browsing, and code execution.

Neighborhood — ranked by edge-count

Thinkers (7)

thinker

Methods (2)

method
  • Agentic Inference Scaffolding
    implements
    The paper's inference framework that reformulates multi-turn tool-calling as single-turn contextual QA for Qwen models and implements context memory management.
  • Training Data Synthesis Pipeline
    implements
    Iterative approach to construct challenging synthetic multi-hop QA pairs, long-form report writing tasks, and math/code reasoning tasks that exceed difficulty of existing datasets.

Concepts (2)

concept
  • Deep Research
    about
    The task of autonomously reasoning through complex questions requiring extensive search and reasoning over multiple sources, combining web browsing and code execution.
  • Single-Agent vs Multi-Agent Systems Design
    about
    Core architectural choice in the paper: single agents determine actions dynamically based on context, while multi-agents follow pre-defined workflows with specialized roles.

Datasets (6)

dataset
  • GAIA
    mentions
    General assistant benchmark focusing on browsing tasks; SFR-DR-20B achieves 66.0% on text-only evaluation set.
  • Qwen3-8B
    mentions
    Reasoning-optimized base model used for training SFR-DR-8B variant.
  • QwQ-32B
    mentions
    Reasoning-optimized base model used for training SFR-DR-32B variant.
  • FRAMES
    mentions
    Browsing-focused benchmark for multi-hop reasoning QA; SFR-DR-20B achieves 82.8%.
  • gpt-oss-20b
    mentions
    Reasoning-optimized base model used for training SFR-DR-20B, the paper's best-performing variant.
  • Humanity's Last Exam
    mentions
    Reasoning-focused benchmark covering math and science domains; SFR-DR-20B achieves 28.7% on the full text-only subset.

Frameworks (1)

framework
  • REINFORCE
    implements
    Classical RL algorithm adapted by the paper with modifications including clipped-surrogate losses and length-normalized advantages for agentic training.

Findings (1)

finding

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

  • Explicit ESRconcept0.711
    The form of ESR focused on in this paper, measured by verbal self-interruption phrases as segment boundaries
  • Air Force Office Of Scientific Research Afosrinstitute0.700
  • Deep Reinforcement Learningmethod0.699
    AI training method inspired by behaviorism, used for autonomous cars and drones; cited as bioinspired success
  • Deep Structureconcept0.683
  • Implicit ESRconcept0.682
    Form of ESR occurring without explicit verbal self-interruption markers, not captured by current metrics
  • Deep Likingconcept0.682
    Synonym for real liking or liking from the heart; the authentic, convergent preference that tracks living structure.
  • Deep adaptationconcept0.681
    The ultimate goal of participation: an environment so deeply fitted to its users that genuine satisfaction and life emerge
  • SimCLRmethod0.679
    Self-supervised contrastive learning method cited as instance of NCE-type objectives that converge to PMI kernel