step-deepresearch

name	step-deepresearch
title	Step-DeepResearch: Autonomous Research via Atomic Capabilities
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.20491
keywords	["research","agents","planning","information-seeking","reinforcement-learning"]
description	Enable autonomous AI research systems to move beyond web search toward true research through four composable atomic capabilities: planning/decomposition, deep search/synthesis, reflection/verification, and report generation. Progressive training across 32K→128K context with SFT and RL produces single ReAct agents matching proprietary systems with lower costs.

Overview

Step-DeepResearch addresses the fundamental gap between web search and true research. Search optimization merely retrieves documents, while research demands intent decomposition, multi-source verification, and coherent synthesis. This framework decompose research into four trainable atomic capabilities.

Core Technique

Rather than treating research as a monolithic task, decompose it into learnable, composable components.

Four Atomic Capabilities Framework: Each capability is independently trainable and composable into end-to-end workflows.

# Atomic capabilities for research automation
class ResearchCapabilities:
    def __init__(self):
        self.planner = PlanningAgent()           # Task decomposition
        self.searcher = DeepSearchAgent()        # Multi-hop synthesis
        self.verifier = ReflectionAgent()        # Error correction
        self.reporter = ReportGenerationAgent()  # Structured output

    def research_workflow(self, research_goal):
        """
        Orchestrate atomic capabilities into research process.
        """
        # Capability 1: Planning & Task Decomposition
        plan = self.planner.decompose(research_goal)
        # Returns: [subtask_1, subtask_2, ...]

        # Capability 2: Deep Search & Information Seeking
        evidence = self.searcher.multihop_search(plan)
        # Returns: {subtask: [sources, facts, citations]}

        # Capability 3: Reflection & Verification
        verified_evidence = self.verifier.cross_source_check(evidence)
        # Returns: {fact: confidence, citations: validated}

        # Capability 4: Report Generation
        report = .reporter.synthesize(verified_evidence)
        

         report