AREX

Efficient Long-Horizon Research Agents

Overview

AREX is a family of foundation models for research agents, designed for long-horizon deep research. We believe a real research agent should be not only capable, but also efficient: it should gather the most useful information within limited interaction steps, inference budgets, and deployment costs.

AREX focuses on agentic research workflows that require deep search, multi-step reasoning, context management, and evidence synthesis. It can browse and retrieve information across sources, keep track of intermediate findings, decide when to compress or preserve context, and produce grounded answers for open-ended research problems.

Table of Contents

• News
• Introduction
• Quick Start
• Model Variants
• Key Capabilities
• Use Cases
• Benchmark Evaluation
• Resources
• Citation
• Star History

News

• [2026-06-30] The AREX project page is online.

Introduction

Many long-context and agent systems rely on hard truncation, fixed summarization, or simply larger inference budgets. AREX instead treats research as an active decision process. At each step, the model learns to choose whether it should search, read, verify, write notes, compress old reasoning, retrieve key evidence, or stop and synthesize.

This efficiency-first design helps reduce unnecessary tool calls and inference latency while preserving the reasoning and evidence-tracking behavior required by deep research tasks.

Quick Start

The quick start runs a concise AREX research workflow. AREX calls search, google_scholar, and visit in XML tool-call format; the runner executes those tools and feeds <tool_response> observations back to the model until AREX produces a final plain-text answer.

The examples do not ship with default endpoint or key values. Set the model endpoint plus the URL and key for each tool before running them.

pip install -e .

export AREX_BASE_URL="https://your-arex-model-endpoint.example/v1"
export AREX_API_KEY="your-model-api-key"
export AREX_MODEL="your-served-model-name"

export AREX_SEARCH_URL="https://your-search-service.example/serp_search_v1"
export AREX_SEARCH_API_KEY="your-search-tool-api-key"
export AREX_SCHOLAR_URL="https://your-scholar-service.example/serp_search_v1"
export AREX_SCHOLAR_API_KEY="your-scholar-tool-api-key"
export AREX_VISIT_URL="https://your-visit-service.example/visit_pages_v1"
export AREX_VISIT_API_KEY="your-visit-tool-api-key"

export AREX_PROMPT="Find recent evidence on how retrieval-augmented research agents trade off latency and answer quality."

Optional controls:

export AREX_SUMMARY_BASE_URL="$AREX_BASE_URL"
export AREX_SUMMARY_API_KEY="$AREX_API_KEY"
export AREX_SUMMARY_MODEL="$AREX_MODEL"
export AREX_MAX_ROUNDS=600
export AREX_RESPONSE_MAX_TOKENS=4096
export AREX_VISIT_PAGE_MAX_CHARS=180000

Command Line

bash examples/quickstart.sh

You can also pass values as flags. After pip install -e ., arex-quickstart is equivalent to python src/arex_react.py.

python src/arex_react.py \
  --base-url "$AREX_BASE_URL" \
  --api-key "$AREX_API_KEY" \
  --model "$AREX_MODEL" \
  --search-url "$AREX_SEARCH_URL" \
  --search-api-key "$AREX_SEARCH_API_KEY" \
  --scholar-url "$AREX_SCHOLAR_URL" \
  --scholar-api-key "$AREX_SCHOLAR_API_KEY" \
  --visit-url "$AREX_VISIT_URL" \
  --visit-api-key "$AREX_VISIT_API_KEY" \
  --prompt "$AREX_PROMPT"

Python SDK

from arex_client import AREXReActClient

client = AREXReActClient(
    base_url="https://your-arex-model-endpoint.example/v1",
    api_key="your-model-api-key",
    model="your-served-model-name",
    search_url="https://your-search-service.example/serp_search_v1",
    search_api_key="your-search-tool-api-key",
    scholar_url="https://your-scholar-service.example/serp_search_v1",
    scholar_api_key="your-scholar-tool-api-key",
    visit_url="https://your-visit-service.example/visit_pages_v1",
    visit_api_key="your-visit-tool-api-key",
    summary_base_url="https://your-summary-endpoint.example/v1",
    summary_model="your-summary-model-name",
)

result = client.run("Find recent evidence on how retrieval-augmented research agents trade off latency and answer quality.")
print(result.answer)

Or run the environment-variable based Python example:

python examples/quickstart.py

Model Variants

Model	Deployment Profile	Intended Use
AREX-Turbo	Dense, low-latency option	Resource-constrained serving, rapid research loops, and cost-sensitive agent deployments
AREX-Base	Sparse-expert option	Stronger long-horizon reasoning and tool-use performance at controlled active compute cost

Key Capabilities

• Efficiency-First Search: Optimizes each interaction for useful information gain, reducing unnecessary browsing, inference cost, and response latency.
• Agentic Context Management: Treats context as an active workspace. AREX can decide when to preserve notes, compress old reasoning, and retrieve important evidence.
• Robust Tool Use: Supports complex browsing, retrieval, data analysis, and expert-style verification workflows across multiple rounds.
• Long-Horizon Reasoning: Decomposes complex questions, follows intermediate findings, reconciles evidence, and synthesizes final answers.

Use Cases

• Automated scientific literature review and evidence synthesis.
• Multi-round commercial, market, and competitor research across websites.
• Question answering systems that require complex reasoning chains and tool calls.
• Autonomous analysis of long documents spanning dozens or hundreds of pages.

Benchmark Evaluation

AREX is evaluated on public benchmarks covering browsing, retrieval, long-horizon reasoning, tool use, and open-domain deep search.

Covered benchmarks include:

• BrowseComp and WideSearch for browsing and retrieval.
• HLE and GAIA for long-horizon reasoning and tool interaction.
• DeepSearchQA and xbench-2510 for open-domain deep search.

Resources

Resource	Link
Project website	vectorspacelab.github.io/arex-model
Paper	Coming soon
AREX-Turbo	Coming soon
AREX-Base	Coming soon

Citation

Citation information will be added when the paper is available.

Star History