PaperBanana

Pipeline Mode

Select which agent pipeline to use

Pipeline Description

Retrieval Setting

How to retrieve reference diagrams

Number of Candidates

Aspect Ratio

Figure Size

Max Critic Rounds

1 5

Model Name

Model name to use for reasoning

Image Generation Model

Model for generating diagram images

Save Results

Load Example (Method)

Load Example (Caption)

Method Content

## Methodology: The PaperBanana Framework

In this section, we present the architecture of PaperBanana, a reference-driven agentic framework for automated academic illustration. As illustrated in Figure \ref{fig:methodology_diagram}, PaperBanana orchestrates a collaborative team of five specialized agents—Retriever, Planner, Stylist, Visualizer, and Critic—to transform raw scientific content into publication-quality diagrams and plots. (See Appendix \ref{app_sec:agent_prompts} for prompts)

### Retriever Agent

Given the source context $S$ and the communicative intent $C$, the Retriever Agent identifies $N$ most relevant examples $\mathcal{E} = \{E_n\}_{n=1}^{N} \subset \mathcal{R}$ from the fixed reference set $\mathcal{R}$ to guide the downstream agents. As defined in Section \ref{sec:task_formulation}, each example $E_i \in \mathcal{R}$ is a triplet $(S_i, C_i, I_i)$.
To leverage the reasoning capabilities of VLMs, we adopt a generative retrieval approach where the VLM performs selection over candidate metadata:
$$
\mathcal{E} = \text{VLM}_{\text{Ret}} \left( S, C, \{ (S_i, C_i) \}_{E_i \in \mathcal{R}} \right)
$$

### Planner Agent

The Planner Agent serves as the cognitive core of the system. It takes the source context $S$, communicative intent $C$, and retrieved examples $\mathcal{E}$ as inputs:
$$
P = \text{VLM}_{\text{plan}}(S, C, \{ (S_i, C_i, I_i) \}_{E_i \in \mathcal{E}})
$$

### Stylist Agent

The Stylist refines each initial description $P$ into a stylistically optimized version $P^*$:
$$
P^* = \text{VLM}_{\text{style}}(P, \mathcal{G})
$$

### Visualizer Agent

The Visualizer Agent leverages an image generation model:
$$
I_t = \text{Image-Gen}(P_t)
$$

### Critic Agent

The Critic provides targeted feedback and produces a refined description:
$$
P_{t+1} = \text{VLM}_{\text{critic}}(I_t, S, C, P_t)
$$
The Visualizer-Critic loop iterates for $T=3$ rounds.

Figure Caption

Status

Generated Candidates