Add files using upload-large-folder tool

README.md

@@ -0,0 +1,103 @@
+---
+library_name: transformers
+pipeline_tag: text-generation
+tags:
+- agent
+- deep-research
+- reasoning
+- tool-use
+- long-context
+- qwen3.5
+- mixture-of-experts
+language:
+- en
+---
+
+# AREX-Base
+
+<div align="center">
+  <img src="assets/arex-logo.svg" width="54%" alt="AREX" />
+</div>
+
+<div align="center">
+
+[Paper](#) | [GitHub / Project Page](https://545999961.github.io/AREX/web) | [AREX-Turbo](https://huggingface.co/BAAI/AREX-Turbo)
+
+</div>
+
+## Introduction
+
+AREX is a family of Recursively Self-Improving deep research agents developed by the AREX team at the Beijing Academy of Artificial Intelligence (BAAI). Deep research tasks are difficult because a correct answer often has to satisfy several coupled constraints at the same time: the agent must discover candidate answers, gather evidence from different sources, resolve conflicts, and verify that every required condition is supported.
+
+AREX is built around the observation that discovery and verification are asymmetric. Finding an answer that satisfies all constraints can require broad and expensive exploration, while checking a proposed answer can often be decomposed into more tractable constraint-wise verification steps. AREX uses this asymmetry as a control signal: it repeatedly turns partially verified answers into more targeted follow-up research problems.
+
+The AREX framework contains two interacting loops. The inner research loop searches, reads, integrates evidence, maintains candidates, and produces a provisional answer with supporting evidence and an answer-level confidence score. The outer self-improvement loop audits that provisional answer against the original constraints. High-confidence answers are accepted; low-confidence but recoverable trajectories are refined by preserving useful evidence and focusing on unresolved claims; uninformative trajectories can be restarted.
+
+To support long-horizon research, AREX learns to invoke an autonomous `update_context` tool. Instead of simply truncating or summarizing the dialogue, the model refreshes its research state around verified findings, source identifiers, current and rejected candidates, unresolved constraints, validity concerns, and the next research plan. This keeps the working context aligned with the agent's evolving beliefs and reduces repeated exploration of already rejected paths.
+
+This repository contains **AREX-Base**, the larger member of the AREX family. AREX-Base is a Mixture-of-Experts model with 122B total parameters and 10B activated parameters, built on the Qwen3.5-122B-A10B backbone and trained for long-horizon research, tool use, evidence-grounded answer construction, and recursive self-improvement.
+
+**Key Features**
+
+- Recursive self-improvement for multi-constraint deep research.
+- Inner research loop for evidence gathering, candidate tracking, and provisional answer construction.
+- Outer self-improvement loop for confidence-based acceptance, targeted refinement, or restart.
+- Autonomous context updating that preserves verified evidence, rejected candidates, unresolved constraints, and next-step plans.
+- 256K context configuration with long-horizon agent evaluation up to 300 inner research-loop turns and 5 outer self-improvement operations.
+
+## Model Family
+
+| Model | Scale | Backbone | Max Context | Positioning |
+| --- | ---: | --- | ---: | --- |
+| [AREX-Base](https://huggingface.co/BAAI/AREX-Base) | 122B total / 10B activated MoE | Qwen3.5-122B-A10B | 256K | Stronger deep research model with the best overall AREX performance. |
+| [AREX-Turbo](https://huggingface.co/BAAI/AREX-Turbo) | 4B dense | Qwen3.5-4B | 256K | Compact AREX model for lower-cost research and tool-use settings. |
+
+## Method Overview
+
+AREX is trained to perform research as an iterative process rather than a single search trajectory. Its training pipeline combines verified synthetic tasks, high-quality teacher trajectories, agentic mid-training, and long-horizon reinforcement learning.
+
+The training data emphasize tasks that require multi-hop evidence acquisition, reasoning over distributed information, source-grounded verification, and adaptive planning. Teacher trajectories are filtered for coherent research behavior, valid tool use, sufficient evidence support, and high-confidence final answers.
+
+During mid-training, AREX first learns browse-intensive multi-round tool use and evidence acquisition, then strengthens expert reasoning and hypothesis verification. A mixed-capability consolidation stage replays important long-horizon research steps and adds more challenging academic and knowledge-intensive tasks.
+
+For long-horizon reinforcement learning, AREX uses outcome rewards together with step-level shaping for decision-critical research moments, such as discovering decisive evidence, rejecting an incorrect path, redirecting the search, or updating context after meaningful progress. This helps the model learn the intermediate actions that make recursive research effective.
+
+## Evaluation
+
+AREX is evaluated with a unified long-horizon search-agent interface. The agent can use search, visit, `update_context`, and `finish` tools; HLE with tools also includes a python tool. The reported setup allows up to 300 inner research-loop turns and up to 5 outer self-improvement operations.
+
+<div align="center">
+  <img src="assets/arex-benchmark-overview.svg" width="100%" alt="AREX benchmark results" />
+</div>
+
+The figure above highlights a curated subset of the comparison. The table below reproduces the main experiment table from the AREX paper.
+
+| Category | Model | BrowseComp | GAIA | xbench-2510 | DeepSearchQA | WideSearch-en | HLE (tool) |
+| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Frontier models | GPT-5.4 | 82.7 | - | - | 88.5 | 77.5 | 52.1* |
+| Frontier models | Opus-4.6 | 83.7 | - | - | 91.3 | 77.5 | 53.0* |
+| Frontier models | Gemini-3.1-Pro | 85.9 | 80.6 | 53.0 | 93.3 | 66.4 | 51.4* |
+| Open-source models | GLM-5 | 75.9 | 70.0 | - | - | 69.8 | 50.4 |
+| Open-source models | Kimi-K2.6 | 83.2 | 80.6 | 90.0 | 92.5 | 80.8 | 54.0* |
+| Open-source models | DeepSeek-V4-Flash | 73.2 | - | 69.0 | 90.6 | 76.4 | 45.1 |
+| Open-source models | DeepSeek-V4-Pro | 83.4 | - | 80.0 | 88.7 | 78.0 | 48.2 |
+| Open-source models | Tongyi-DeepResearch-30B | 43.4 | 70.9 | 55.0 | - | - | 32.9 |
+| Open-source models | Qwen3.5-35B | 61.0 | 80.0 | 50.3 | 68.5 | 57.1 | 47.4 |
+| Open-source models | Qwen3.5-122B | 63.8 | 81.6 | - | - | 60.5 | 47.5 |
+| Open-source models | Qwen3.5-397B | 78.6 | 83.5 | 61.0 | 82.1 | 74.0 | 48.3 |
+| Open-source models | MiroThinker-1.7-mini | 67.9 | 80.3 | 57.2 | 67.9 | - | 36.4 |
+| Open-source models | MiroThinker-1.7 | 74.0 | 82.7 | 62.0 | 72.1 | - | 42.9 |
+| Open-source models | MiroThinker-H1 | 88.2 | 88.5 | 72.0 | 80.6 | - | 47.7 |
+| Open-source models | Quest-35B | 64.6 | 80.8 | - | - | 60.6 | 37.2 |
+| Ours | AREX-Turbo | 70.7 | 81.6 | 57.0 | 78.5 | 68.5 | 40.6 |
+| Ours | AREX-Base | 82.5 | 85.4 | 71.0 | 89.9 | 82.0 | 52.4 |
+
+DeepSearchQA is reported with Item-F1; the other benchmarks are reported with accuracy. `*` denotes results reported on the full HLE; unmarked HLE results are reported on the text-only subset.
+
+AREX-Base provides the strongest AREX performance across the evaluated deep search, wide search, agentic reasoning, and tool-use benchmarks. Its results show that recursive preservation of verified progress, targeted refinement of unresolved constraints, and autonomous context updating can improve deep research performance without relying only on larger active parameter counts.
+
+## Intended Use
+
+AREX-Base is intended for research-agent scenarios that require long-horizon information seeking, evidence aggregation, multi-constraint verification, and tool-augmented reasoning. It is especially suited for experiments on deep research workflows where the agent must maintain and improve an evolving research state across many steps.
+
+The model card currently focuses on the model introduction, method, family relationship, and evaluation summary. Detailed usage examples will be added in a later revision.