Add files using upload-large-folder tool

README.md

@@ -14,7 +14,14 @@ tags:
   - safetensors
 ---
 
-# AREX-2
+<div align="center">
+  <strong>AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks</strong><br>
+  <a href="https://arxiv.org/abs/2607.21461"><img src="https://img.shields.io/badge/-Paper-B31B1B?style=for-the-badge&amp;logo=arxiv&amp;logoColor=white" alt="Paper"></a>
+  <a href="https://github.com/VectorSpaceLab/AREX-2"><img src="https://img.shields.io/badge/-Homepage-24292F?style=for-the-badge&amp;logo=github&amp;logoColor=white" alt="Homepage"></a>
+  <a href="https://huggingface.co/collections/BAAI/arex-2"><img src="https://img.shields.io/badge/-Collection-FFD21E?style=for-the-badge&amp;logo=huggingface&amp;logoColor=black" alt="Model collection"></a>
+</div>
+
+## Introduction
 
 AREX-2 is a 27B-parameter long-horizon agent model from the Beijing Academy of
 Artificial Intelligence (BAAI). It learns to improve a solution over multiple
@@ -25,53 +32,95 @@ verifiable feedback, together with the existing AREX deep-research data. The
 learned self-improvement behavior transfers to deep research without adding new
 search trajectories.
 
-[Project](https://github.com/VectorSpaceLab/AREX-2) ·
-[Paper](https://arxiv.org/abs/2607.21461) ·
-[Model collection](https://huggingface.co/collections/BAAI/arex-2)
-
-## Model summary
-
-- **Backbone:** Qwen3.8-27B, dense multimodal model
-- **Context:** 262,144 tokens
-- **Weights:** BF16 Safetensors, 11 shards
-- **Inputs:** text, images, and videos
+- **Architecture:** Dense Qwen3.8-compatible multimodal model
+- **Parameters:** 27B
+- **Context length:** 262,144 tokens
+- **Tensor type:** BF16
+- **Inputs:** Text, images, and videos
 - **Tool calls:** XML (`<tool_call>...</tool_call>`)
 
-## Method
-
-AREX-2 treats self-improvement as test-time scaling: more rounds should produce
-more useful progress. Its training pipeline is:
-
-1. **Build environments.** Teacher models turn machine-learning repositories
-   and online-judge problems into sandboxes with tasks and executable scorers.
-   Environments must have a valid reference solution and room to improve over a
-   simple baseline.
-2. **Generate trajectories.** A teacher agent works for many rounds, reading,
-   searching, editing, running experiments, submitting solutions, and using
-   feedback to choose the next step. Failures and recoveries remain in context.
-3. **Select and supervise.** Keep trajectories with a strong final score and a
-   valid process. Apply training loss to decisions that diagnose failures,
-   change strategy, or improve the solution; ignore polling, duplicates, and
-   environment messages.
-4. **Fine-tune.** Mix the selected trajectories with the unchanged AREX
-   deep-research data and fine-tune Qwen3.8-27B.
+## Key features
 
-## Evaluation
-
-| Benchmark | Metric | AREX-2 |
-| --- | --- | ---: |
-| Frontier-CS (188-task Agent Track) | Accuracy | **70.7** |
-| MLE-bench Lite | Any Medal (3-seed average) | **81.8** |
-| BrowseComp | Accuracy | **84.0** |
-| HLE (text-only subset) | Accuracy | **52.6** |
-| GAIA | Accuracy | **92.2** |
-| DeepSearchQA | F1 | **93.8** |
+- **Long-horizon self-improvement:** turns additional test-time rounds into
+  continued solution refinement.
+- **Feedback-driven reflection:** uses scores, logs, errors, and timings to
+  decide what to try next.
+- **Cross-domain transfer:** learns from verifiable coding and ML tasks, then
+  applies the behavior to deep research.
+- **Long-horizon tool use:** supports multi-round reasoning, tool calls, and
+  multimodal inputs. External tools are provided by the host application.
 
-The coding results use the official agent settings: five hours per Frontier-CS
-task and 12 hours per MLE-Lite task. Non-coding evaluations allow up to 300
-inner turns and 1,500 total turns.
+## Evaluation
 
-## Inference with Transformers
+AREX-2 is evaluated on algorithmic programming, machine-learning engineering,
+deep research, and general agentic reasoning. Results follow the protocols
+reported in the AREX-2 paper.
+
+### Coding and machine-learning engineering
+
+| Model | Params | Frontier-CS | MLE-Lite |
+| --- | ---: | ---: | ---: |
+| GPT-5.6 Sol | - | 76.4 | 72.7 |
+| Claude Opus 4.8 | - | 74.5 | 63.6 |
+| GPT-5.5 | - | 72.1 | 68.2 |
+| Gemini-3.1-Pro | - | 68.9 | - |
+| Qwen3.7-Max | - | 61.9 | - |
+| Kimi-K3 | 2.8T | - | 72.7 |
+| Naive-N0.5-Flash | 309B | - | 73.7 |
+| DeepSeek-V4-Pro | 1.6T | 44.7* | 54.5 |
+| DeepSeek-V4-Flash | 284B | 39.1* | 51.5 |
+| Kimi-K2.7-Code | 1T | 54.7* | - |
+| GLM-5.3-Flash | 320B | 50.4* | - |
+| Kimi-K2.6 | 1T | 46.9 | 66.7 |
+| Frontis-MA1-35B | 35B | - | 71.2 |
+| BigBang-V1 | 35B | - | 59.1 |
+| Qwen3.6-35B-A3B | 35B | 23.4* | 39.4 |
+| **AREX-2** | **27B** | **70.7** | **81.8** |
+
+Frontier-CS is the 188-task Agent Track. `*` marks results reproduced by the
+AREX team. MLE-Lite reports Any Medal averaged over three seeds.
+
+### General agentic reasoning and deep research
+
+| Model | Params | BrowseComp | HLE | GAIA | DeepSearchQA |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| GPT-5.6 Sol | - | 90.4 | 58.0* | - | - |
+| GPT-5.6 Terra | - | 87.5 | - | - | - |
+| GPT-5.6 Luna | - | 83.3 | - | - | - |
+| Kimi-K3 | 2.8T | 91.2 | 56.0* | - | 95.0 |
+| Claude Fable 5 | - | 88.0 | 64.5* | - | 94.2 |
+| Claude Opus 4.8 | - | 84.3 | 57.9* | - | 93.1 |
+| GPT-5.5 | - | 84.4 | 52.2* | 87.4 | - |
+| Gemini-3.1-Pro | - | 85.9 | 51.4* | 80.6 | 93.3 |
+| GLM-5 | 744B | 75.9 | 50.4 | 70.0 | - |
+| Kimi-K2.6 | 1T | 83.2 | 54.0* | 80.6 | 92.5 |
+| GLM-5.3-Flash | 320B | - | 55.3* | 78.8 | - |
+| DeepSeek-V4-Flash | 284B | 73.2 | 45.1 | 57.5 | 90.6 |
+| DeepSeek-V4-Pro | 1.6T | 83.4 | 48.2 | 71.1 | 88.7 |
+| MiroThinker-1.7 | 235B | 74.0 | 42.9 | 82.7 | 72.1 |
+| XYZ-Aquila-pro | 397B | 84.8 | 53.3 | - | 92.5 |
+| Iris-pro | 397B | 88.6 | 56.4 | - | 92.9 |
+| **AREX-Base** | **122B** | **82.5** | **52.4** | **85.4** | **89.9** |
+| Tongyi-DeepResearch-30B | 30B | 43.4 | 32.9 | 70.9 | - |
+| Qwen3.5-35B | 35B | 61.0 | 47.4 | 80.0 | 68.5 |
+| XYZ-Aquila-mini | 35B | 78.8 | 51.1 | 97.1 | 89.5 |
+| BigBang-V1 | 35B | 76.5 | 50.3 | - | - |
+| Quest-35B | 35B | 64.6 | 37.2 | 80.8 | - |
+| Apodex-1.0-mini | 35B | 71.5 | 46.8 | - | 82.2 |
+| Agents-A1 | 35B | 75.5 | 47.6 | 96.0 | - |
+| MiroThinker-1.7-mini | 30B | 67.9 | 36.4 | 80.3 | 67.9 |
+| Iris-mini | 35B | 82.2 | 52.3 | - | 86.9 |
+| **AREX-Turbo** | **4B** | **70.7** | **40.6** | **81.6** | **78.5** |
+| **AREX-2** | **27B** | **84.0** | **52.6** | **92.2** | **93.8** |
+
+Metrics are accuracy except DeepSearchQA, which uses F1. HLE values marked `*`
+are from the full HLE set; unmarked values use the text-only subset.
+
+For non-coding benchmarks, evaluation allows up to 300 inner turns and 1,500
+total turns. Frontier-CS uses a five-hour budget per task, and MLE-Lite uses the
+OpenMLE protocol with a 12-hour budget per task.
+
+## Inference
 
 Use a recent Transformers release with Qwen3.8 support.
 
@@ -94,11 +143,8 @@ messages = [{
     "content": "Propose a solution and explain how you would improve it over several rounds.",
 }]
 inputs = processor.apply_chat_template(
-    messages,
-    add_generation_prompt=True,
-    tokenize=True,
-    return_dict=True,
-    return_tensors="pt",
+    messages, add_generation_prompt=True, tokenize=True,
+    return_dict=True, return_tensors="pt",
 ).to(model.device)
 
 with torch.inference_mode():
@@ -109,11 +155,11 @@ print(processor.decode(
 ```
 
 For images or videos, use the standard Qwen multimodal content format. For
-function calling, pass tool schemas to `apply_chat_template`; the host must
-execute the generated XML call and return the result as a `tool` message.
-AREX-2 does not include external search, browsing, or code-execution tools.
+tool use, pass tool schemas to `apply_chat_template`; the host must execute the
+XML call and return the result as a `tool` message. AREX-2 does not include
+external search, browsing, or code-execution tools.
 
-## Intended use and limitations
+## Intended use
 
 AREX-2 is intended for research on long-horizon agents, iterative problem
 solving, machine-learning engineering, algorithmic coding, and tool-augmented