File size: 4,938 Bytes
5275f2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
---
library_name: transformers
license: apache-2.0
base_model:
  - Qwen/Qwen3-VL-8B-Instruct
  - Accio-Lab/Metis-8B-ColdStart
tags:
  - multimodal
  - vision-language
  - reinforcement-learning
  - tool-use
  - agentic
  - qwen3_vl
  - HDPO
datasets:
  - Accio-Lab/Metis-RL
language:
  - en
pipeline_tag: image-text-to-text
---

# Metis-8B-RL

**Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models**

Metis-8B-RL is the final RL-trained checkpoint of the **Metis** framework, trained with **Hierarchical Decoupled Policy Optimization (HDPO)** on top of [Metis-8B-ColdStart](https://huggingface.co/Accio-Lab/Metis-8B-ColdStart). It is a strategic multimodal reasoning agent that selectively invokes code execution, text search, and image search tools during multi-turn reasoning.

[[Paper (arXiv)]](https://arxiv.org/abs/2604.08545) | [[GitHub]](https://github.com/Accio-Lab/Metis) | [[ColdStart Model]](https://huggingface.co/Accio-Lab/Metis-8B-ColdStart) | [[RL Data]](https://huggingface.co/datasets/Accio-Lab/Metis-RL) | [[ColdStart Data]](https://huggingface.co/datasets/Accio-Lab/Metis-ColdStart)

## Highlights

- **98% → 2% Tool Calls** — Reduces blind tool invocation by orders of magnitude.
- **SOTA Performance** — Best accuracy across 13 benchmarks among open-source 8B agentic models.
- **Meta-Cognitive Wisdom** — Learns *when* to use tools, not just *how*.

## Model Details

| Attribute | Value |
|---|---|
| Base model | [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) |
| SFT checkpoint | [Metis-8B-ColdStart](https://huggingface.co/Accio-Lab/Metis-8B-ColdStart) |
| RL algorithm | HDPO (Hierarchical Decoupled Policy Optimization) |
| Training data | [Metis-RL](https://huggingface.co/datasets/Accio-Lab/Metis-RL) (~5K prompts) |
| License | Apache-2.0 |

### HDPO Training Hyperparameters

| Hyperparameter | Value |
|---|---|
| Batch size | 128 |
| Rollouts per prompt (*G*) | 16 |
| Learning rate | 1e-6 |
| KL coefficient | 0 |
| Loss weights | w_acc = 1.0, w_tool = 0.15 |
| Max response length | 16,384 tokens |

## Method: Hierarchical Decoupled Policy Optimization (HDPO)

Current agentic multimodal models suffer from **blind tool invocation** — they reflexively call external tools even when queries are directly resolvable from the visual context. Existing RL methods attempt to fix this by coupling accuracy and tool-efficiency into a single scalar reward, but this creates an irreconcilable optimization dilemma.

HDPO resolves this through three key components:

1. **Dual Reward Design** — An accuracy reward (r_acc) and a tool-efficiency reward (r_tool) that is conditioned on correctness.
2. **Decoupled Advantage Estimation** — Accuracy advantages are computed over all rollouts; tool efficiency advantages are computed *exclusively* over correct rollouts (conditional GRPO).
3. **Hierarchical Policy Update** — Two independent clipped surrogate losses combined as `L_HDPO = w_acc · L_GRPO(A_acc) + w_tool · L_GRPO(A_tool)`.

This naturally induces an implicit curriculum: *first learn to be correct, then learn to be efficient*.

## Evaluation Results

### Perception and Document Understanding

| Model | V\*Bench | HR4K | HR8K | TreeBench | MME-RW | SEED2+ | CharXiv(DQ) | CharXiv(RQ) |
|---|---|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 86.4 | 78.9 | 74.6 | 40.7 | 61.9 | 71.0 | 83.0 | 46.3 |
| DeepEyesV2 | 81.8 | 77.9 | 73.8 | 42.5 | 64.9 | 70.5 | 78.6 | 48.9 |
| SenseNova-MARS-8B | **92.2** | 83.1 | 78.4 | - | 67.9 | - | - | - |
| Skywork-R1V4-30B-A3B | 88.0 | 82.8 | 79.8 | - | **71.4** | - | - | - |
| **Metis (Ours)** | 91.1 | **83.5** | **82.0** | **45.2** | 70.3 | **72.5** | **83.4** | **54.1** |

### Mathematical and Logical Reasoning

| Model | MathVista | MathVerse | WeMath | DynaMath | LogicVista | Avg. |
|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 76.3 | 61.3 | 38.8 | 65.5 | 54.9 | 59.4 |
| DeepEyesV2 | 71.9 | 52.7 | 38.1 | 57.2 | 48.7 | 53.7 |
| **Metis (Ours)** | **78.0** | **65.9** | **65.2** | **69.2** | **56.2** | **66.9** |

## Usage

Please refer to the [GitHub repository](https://github.com/Accio-Lab/Metis) for full installation and inference instructions.

### Installation

```bash
git clone https://github.com/Accio-Lab/Metis.git
cd Metis
pip install -e verl
pip install -e ".[vllm,search_tool,python_code_dep]"
```

## Citation

```bibtex
@article{yan2026metis,
  title={Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models},
  author={Yan, Shilin and Tong, Jintao and Xue, Hongwei and Tang, Xiaojun and Wang, Yangyang and Shi, Kunyu and Zhang, Guannan and Li, Ruixuan and Zou, Yixiong},
  journal={arXiv preprint arXiv:2604.08545},
  year={2026}
}
```

## Acknowledgments

Metis is built upon [verl](https://github.com/volcengine/verl), [verl-tool](https://github.com/TIGER-AI-Lab/verl-tool), and [Qwen3-VL](https://github.com/QwenLM/Qwen3-VL).