# Beacon: Knowing When and How to Perform Agentic Visual Reasoning

> Beacon uses reinforcement learning to teach vision-language models when to invoke tools and when text-alone suffices, achieving the largest net benefit from tool use.

- **Source:** [arXiv](https://arxiv.org/abs/2607.28595)
- **Published:** 2026-08-01
- **Permalink:** https://picx.dev/p/fwZ0m9
- **Whiteboard:** https://picx.dev/p/fwZ0m9/image

## Summary

## Summary (Overview)

- **Novel Framework:** Proposes Beacon, an agentic visual reasoning model that learns *when* to use tools (Mode Adaptiveness) and *how* to use them effectively (Tool Effect) via reinforcement learning.
- **Key Empirical Finding:** Existing agentic models (Thyme, CodeV, DeepEyesV2, Metis) show limited adaptiveness; their tool-induced gains on hard problems are largely offset by errors introduced on easy problems.
- **Core Technical Innovation:** Introduces a Necessity-Aware Adaptive Reward (NAAR) that encourages tool use only when text-only reasoning fails, and a Hint-Guided Capability Expansion (HCE) mechanism that uses expert-generated hints to recover useful learning signals from otherwise unsolvable examples.
- **State-of-the-Art Performance:** Beacon-8B achieves the best average performance across 13 benchmarks, ranking first on 11 of them, and demonstrates the largest positive gap between tool gain and tool harm (+3.14%).
- **Robust Adaptiveness:** Beacon exhibits strong adaptive tool-invocation behavior, invoking tools more frequently on text-hard samples and avoiding them on text-easy samples, unlike prior models that favor a single reasoning mode.

## Introduction and Theoretical Foundation

The paper argues that the fundamental goal of agentic visual reasoning is to *improve the success rate* of multimodal large language models (MLLMs) on complex tasks, not merely to equip them with a sophisticated but inefficient reasoning paradigm. The authors identify two critical but often overlooked dimensions of tool use:

1.  **Mode Adaptiveness (MA):** Whether a model can recognize when tools are truly necessary and invoke them accordingly, avoiding unnecessary computational overhead while improving performance on challenging problems.
2.  **Tool Effect (TE):** Whether tool use genuinely extends the model’s capabilities on problems unsolvable through text-only reasoning (Tool Gain), while avoiding additional errors on problems the model can already solve without tools (Tool Harm).

The paper systematically analyzes existing agentic visual reasoning models (Thyme, CodeV, DeepEyesV2, Metis) and reveals that they exhibit low Mode Adaptiveness (often favoring a single reasoning mode) and that their Tool Gain is largely offset by Tool Harm, resulting in negligible net improvement over text-only reasoning.

## Methodology

### Proposed Metrics
The paper defines precise metrics to quantify the two dimensions:

- **Mode Adaptiveness (MA):**
  - $MA_{\text{text}} = \frac{1}{5|E|} \sum_{x \in E} \sum_{i=1}^{5} (1 - u_{x,i})$ (avoids tools on text-easy samples)
  - $MA_{\text{tool}} = \frac{1}{5|H|} \sum_{x \in H} \sum_{i=1}^{5} u_{x,i}$ (invokes tools on text-hard samples)
  - Where $E$ and $H$ are sets of text-easy and text-hard samples (determined by 5 text-only rollouts), and $u_{x,i}$ indicates tool use.

- **Tool Effects (TE):**
  - $Tool\text{-}Gain = \frac{1}{5N} \sum_{x \in H} \sum_{i=1}^{5} u_{x,i} c_{x,i}$ (correct tool-assisted answers on text-hard samples)
  - $Tool\text{-}Harm = \frac{1}{5N} \sum_{x \in E} \sum_{i=1}^{5} u_{x,i} (1 - c_{x,i})$ (incorrect tool-assisted answers on text-easy samples)

### Beacon Model Architecture

Beacon adopts an SFT-then-RL training paradigm on top of Qwen3-VL-8B-Instruct.

**1. Training Data Construction:**
- **Source Data:** 16 open-source benchmarks covering STEM, chart understanding, perception, agentic tasks, and general visual reasoning.
- **SFT Data:** Challenging examples (answered correctly ≤2/5 times by the base model) are used to generate code-assisted trajectories via Gemini 3.1 Pro, followed by a refinement step to ensure quality.
- **RL Data:** Examples that remain challenging for the SFT model (answered correctly ≤3/5 times) are retained.

**2. Cold-Start SFT:** Standard cross-entropy loss on synthesized trajectories, with code output masked. Pure-text trajectories from the base model are injected to mitigate forgetting.

**3. Reinforcement Learning (RL):** Uses GRPO as the base algorithm with two key innovations:

- **Necessity-Aware Adaptive Reward (NAAR):**
  $$R_{\text{adaptive}}(y_i; G) = \begin{cases} 1, & I_{\text{text}}(G) = 1, y_i \text{ is text-only and correct} \\ 0.25, & I_{\text{text}}(G) = 1, y_i \text{ uses code and is correct} \\ 1, & I_{\text{text}}(G) = 0, y_i \text{ uses code and is correct} \\ 0, & \text{otherwise} \end{cases}$$
  - Where $I_{\text{text}}(G)$ indicates if the group contains at least one correct text-only response.
  - This encourages text-only reasoning when sufficient, but still assigns partial credit to correct code-based solutions.

- **Hint-Guided Capability Expansion (HCE):**
  - For groups where all rollouts are incorrect, a strong expert model (Gemini 3.1 Pro) generates an answer-free hint containing intermediate reasoning steps and expected subgoals.
  - A hinted rollout is performed with the hint appended to the prompt.
  - During policy optimization, the hint is removed from the model input while retaining the hint-generated trajectories, transferring the capability back to the hint-free policy.

**4. Training Objective:**
$$L_{\text{GRPO}}(\theta) = -\frac{1}{Z} \sum_{b \in \{\text{n},\text{h}\}} \sum_{y_i \in \mathcal{B}^b} \sum_{t \in \mathcal{T}_i} \min\left(\rho^b_{i,t}(\theta) A_i, \text{clip}\left(\rho^b_{i,t}(\theta), 1-\epsilon, 1+\epsilon\right) A_i\right)$$

## Empirical Validation / Results

### Main Results (Tables 1 & 2)

| Model | HR-Visual Search Avg | Spatial & Perceptual Avg | Quantitative & Diagrammatic Avg | Compositional & Agentic Avg | **Overall Avg** |
|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 60.57 | 60.57 | 43.97 | 43.97 | 52.27 |
| Thyme-7B | 59.91 | 59.91 | 33.76 | 33.76 | 46.84 |
| DeepEyesV2-7B | 55.00 | 55.00 | 37.40 | 37.40 | 46.20 |
| CodeV-7B | 58.44 | 58.44 | 37.07 | 37.07 | 47.76 |
| Metis-8B | 64.57 | 64.57 | 47.46 | 47.46 | 56.02 |
| **Beacon-8B (Ours)** | **66.05** | **66.05** | **50.73** | **50.73** | **58.98** |

Beacon achieves the highest average performance among all open-source models, outperforming its base model by +6.07 points on average.

### Mode Adaptiveness and Tool Effect Analysis (Table 3)

| Model | $MA_{\text{tool}} \uparrow$ | $MA_{\text{text}} \uparrow$ | $MA_{\text{mean}} \uparrow$ | Tool-Gain $\uparrow$ | Tool-Harm $\downarrow$ | $\Delta_{TE} \uparrow$ |
|---|---|---|---|---|---|---|
| Thyme | 4.28 | 92.95 | 48.61 | 0.19 | 0.13 | +0.06 |
| DeepEyesV2 | 99.71 | 0.63 | 50.17 | 7.85 | 6.11 | +1.74 |
| CodeV | 14.84 | 77.54 | 46.19 | 1.25 | 1.19 | +0.07 |
| Metis | 32.61 | 72.71 | 52.66 | 1.14 | 1.10 | +0.04 |
| **Beacon (Ours)** | **94.75** | 22.91 | **58.83** | **9.29** | 6.15 | **+3.14** |

Key observations:
- Beacon has the highest $MA_{\text{mean}}$ (58.83%), indicating the best adaptive mode selection.
- Beacon shows the largest $\Delta_{TE}$ (+3.14%), meaning tool use provides genuine net benefits.
- Prior models (e.g., Metis) have $\Delta_{TE}$ near zero, indicating tool gains are offset by harms.

### Ablation Study (Table 4)

| Configuration | Overall Acc. $\uparrow$ | $MA_{\text{mean}} \uparrow$ | $\Delta_{TE} \uparrow$ |
|---|---|---|---|
| SFT only | 56.91 | – | – |
| SFT + GRPO | 57.10 | 56.30 | +1.40 |
| SFT + GRPO + NAAR | 57.75 | **59.68** | +2.54 |
| SFT + GRPO + HCE | 57.62 | 58.36 | +2.96 |
| **SFT + GRPO + NAAR + HCE** | **58.98** | 58.83 | **+3.14** |

## Theoretical and Practical Implications

- **Rethinking Agentic Evaluation:** The paper demonstrates that aggregate task accuracy is insufficient for evaluating agentic visual reasoning. It proposes a more nuanced evaluation framework that separately considers adaptiveness (MA) and genuine capability extension (TE).
- **Addressing a Fundamental RLVR Limitation:** The HCE mechanism directly addresses a key limitation of reinforcement learning with verifiable rewards (RLVR): hard examples often provide no useful learning signal because all rollouts fail. By injecting expert hints, the model can discover successful trajectories on these examples.
- **Balancing Efficiency and Capability:** The NAAR provides a principled way to balance computational efficiency (avoiding unnecessary tool calls) with task success (using tools when needed), avoiding the extremes of always or never using tools.
- **Practical Deployment:** Beacon’s adaptive behavior means it can be deployed in resource-constrained environments where unnecessary tool calls are costly, while still maintaining high performance on complex tasks.

## Conclusion

This work revisits agentic visual reasoning from the perspective of whether tools are used both *appropriately* and *effectively*. The key contributions are:

1.  **Systematic Analysis:** A comprehensive evaluation framework (MA and TE metrics) reveals that existing models lack adaptive tool-use behavior, and tool-induced gains are largely offset by harms.
2.  **Beacon Model:** A novel agentic visual reasoning model trained with:
    - A Necessity-Aware Adaptive Reward that encourages adaptive tool invocation.
    - A Hint-Guided Capability Expansion mechanism that recovers useful learning signals from hard examples.
3.  **State-of-the-Art Results:** Beacon achieves the best average performance across 13 diverse benchmarks, exhibits the strongest reasoning-mode adaptiveness, and demonstrates the largest positive gap between tool-induced gain and harm.

The paper concludes that progress in agentic visual reasoning should be measured not merely by how frequently models use tools, but by whether they know *when* tools are necessary and *how* to use them to produce genuine capability gains.

---

_Markdown view of https://picx.dev/p/fwZ0m9, served by PicX — AI-generated visual whiteboard summaries of research papers._