---
title: 'Jr. AI Scientist: A Modular Research Platform'
url: https://www.emergentmind.com/topics/jr-ai-scientist-system
type: topic
---

# Jr. AI Scientist: A Modular Research Platform

A Jr. AI Scientist System is a modular, agentic software platform replicating the workflow of a human junior scientist, capable of proposing hypotheses, planning and running experiments, analyzing results, visualizing findings, and autonomously drafting scholarly manuscripts. Such systems implement a staged agent architecture, advanced tree-search methods for experiment scheduling, and automated review and refinement mechanisms—closely tracing the “AI Scientist-v2” system, which demonstrated the first AI-generated paper accepted in peer-review without template code or human orchestration [2504.08066].

## 1. Agentic System Architecture

The Jr. AI Scientist System comprises five dedicated agent modules, each fulfilling a core scientific role:

- **Hypothesis Generator (HG):** Synthesizes $k$ candidate research questions and experimental blueprints.
- **Experiment Manager (EM):** Executes and schedules code generation and experiment runs via progressive agentic tree search.
- **Data Analyzer (DA):** Applies statistical or ML methods to experiment outputs and creates summary tables/figures.
- **Manuscript Author (MA):** Crafts the scientific manuscript, integrating textual results, tables, and figures.
- **AI Reviewer (AR):** Evaluates manuscript drafts (text and figures) on clarity, novelty, rigor, reproducibility, and impact; invokes a Vision-Language Model (VLM) feedback loop for iterative figure refinement.

The agent interaction protocol follows a sequential cycle:
1. HG outputs $\{H_i\}$.
2. EM launches parallel tree searches for each $H_i$.
3. Each EM node generates code, runs experiments, passes results to DA.
4. DA analyzes outputs; figures are sent to VLM for critique.
5. EM marks nodes as buggy/non-buggy, prunes/expands per LLM evaluation.
6. End-of-budget, EM submits best checkpoints and DA summaries to MA.
7. MA drafts manuscript, embedding DA outputs.
8. AR reviews across five axes and VLM feedback.
9. MA revises to obtain final manuscript.

## 2. Progressive Agentic Tree-Search Methodology

The experiment manager implements a best-first parallel tree search, progressing through four explicit stages:

- **Preliminary:** Prototype experiment success.
- **Tuning:** Hyperparameter optimization.
- **Agenda:** Completion subject to compute constraints.
- **Ablation:** Systematic robustness analysis via component toggling.

Algorithmic flow:

```algorithm
\begin{algorithmic}[1]
\State \mathcal{T} \leftarrow \mathrm{Tree}(root: H)
\While{budget remains}
  \State \mathcal{C} \leftarrow \{\text{select nodes by scoring policy}\}
  \ForAll{n \in \mathcal{C} \textbf{ in parallel}}
    \If{n.status == buggy}
      \State n' \leftarrow \mathrm{DebugNode}(n)
    \Else
      \State n' \leftarrow \mathrm{RefineNode}(n)
    \EndIf
    \State \mathrm{ExecuteCode}(n')   \Comment{experiment run}
    \State DA.analyze(n')
    \State VLM.review_figures(n')
    \State n'.status \leftarrow buggy/non-buggy
    \State \mathrm{Insert\ child\ }n'\mathrm{\ under\ }n\mathrm{\ in\ }\mathcal{T}
  \EndFor
  \State budget \leftarrow budget - cost(\mathcal{C})
\EndWhile
\State \Return top-k non-buggy leaves
\end{algorithmic}
```

The state $s = (\text{code\_script}, \text{metrics}, \text{error\_log}, \text{visuals}, \text{VLM\_feedback})$ and available actions $\{ \mathtt{Debug}, \mathtt{Refine}, \mathtt{HyperparamSweep}, \mathtt{Ablation}, \mathtt{Replicate} \}$. Selection policy follows a UCT (upper confidence bound for trees):

\[
\pi(s) = \arg\max_a \left[ Q(s,a) + c \sqrt{\frac{\ln N(s)}{N(s,a)}} \right]
\]

with value backup and reward combination:

\[
Q(s,a) \leftarrow (1-\alpha) Q(s,a) + \alpha r(s,a)
\]
where $r(s,a)$ combines experiment metric (e.g., validation accuracy) and VLM figure score.

Hypothesis scoring and pruning employ:

\[
\text{score}(n) = \lambda_1 \text{perf}(n) + \lambda_2 \text{novelty}(H) - \lambda_3 \text{bugs}(n)
\]

Nodes with low score or repeated failures are pruned.

## 3. Experimental Design, Analysis, and Visualization

Experimental control encompasses dynamic node scheduling and rigorous Design-of-Experiments principles.

- Compute is dynamically allocated to nodes; buggy nodes are debugged prior to depth limit.
- Hyperparameter sweeps sample $\mathbf{x} \sim \mathcal{N}(\boldsymbol{\mu}, \mathbf{\Sigma})$ optimized via Bayesian strategies:

\[
\max_{\mathbf{x}} \text{perf}(\mathbf{x}) \quad \text{s.t.} \; \sum \text{runtime} \leq C_{max}
\]

Ablation studies toggle $c_i \in \{0,1\}$ for systematic robustness evaluation.

Data Analyzer processes outputs using regression and clustering:

- Regression: $\hat y = \beta_0 + \beta_1 x + \varepsilon$
- Clustering: $k$-means minimizes $\sum_{i=1}^{n} \min_j \|x_i - \mu_j\|^2$

Visual outputs are immediately checked by the VLM loop; image and text embeddings $f_{\mathrm{img}}, f_{\mathrm{txt}}$ are scored via

\[
s_{\mathrm{VLM}} = \sigma \left( f_{\mathrm{img}}^\top W f_{\mathrm{txt}} + b \right)
\]

Outputs below threshold $\tau$ are marked buggy and queued for revision.

## 4. Autonomous Manuscript Authoring and Review

Manuscript Author is tasked to:

1. Generate structured outline (sections, fig/table placeholders).
2. Fill text via LLM, conditioned on DA outputs.
3. Compile to PDF, check page constraints.
4. Self-review and style correction.

AI Reviewer evaluates with:

| Axis           | Metric Range      |
|----------------|------------------|
| Clarity        | 1–10             |
| Novelty        | 1–10             |
| Rigor          | 1–10             |
| Reproducibility| 1–10             |
| Impact         | 1–10             |

VLM iteratively refines figures for readability, color harmony, and aspect ratio, and can request code-level changes.

## 5. Empirical Results and Comparative Benchmarks

Evaluation involved three fully autonomous ICLR workshop submissions:

| Paper                          | Score   | Decision |
|--------------------------------|---------|----------|
| Compositional Regularization   | 6.33    | Accepted |
| Label Noise Calibration        | $\sim 3$| Rejected |
| Real-world Pest Detection      | $\sim 3$| Rejected |

- AI Scientist-v2: 1/3 accepted; surpassed average human acceptance threshold.
- Compared to v1 (linear pipeline, template code, 0/3 accepted), v2 achieved deeper domain-general code generation and peer-review success.
- Percentile achieved: $\approx45\%$ among workshop submissions.

## 6. Implementation, Reproducibility, and Limitations

Open-source code is structured with modular agent scripts and configuration files; rapid deployment and reproducibility are supported via:

```bash
git clone https://github.com/SakanaAI/AI-Scientist-v2.git
cd AI-Scientist-v2
pip install -r requirements.txt
python -m ai_scientist.main \
  --config configs/jr_scientist.json \
  --random-seed 42 \
  --max-time 6h
```

Core limitations:

- LLM hallucinations in citation and method sections.
- Shallow novelty vs. human experts.
- Compute- and resource-intensive tree search.
- Risks: Flawed code or misleading results may be generated; human audit advised prior to publication.

Future avenues include code snippet verification, leveraging domain-specific knowledge graphs, and a junior variant limiting tree-search branching and compute—tailored for educational labs and close oversight [2504.08066].

## 7. Jr. AI Scientist Variant Design Guidance

A scaled-down Jr. AI Scientist is recommended to:

- Reduce tree-search branching, restrict to two stages and ten total nodes.
- Maintain all agent modules, but employ stage caps for computational tractability.
- Encourage human auditing and close oversight in teaching and research environments.
- Retain modularity and VLM feedback for experimental robustness, with configurations specified in concise JSON for reproducibility.

This Jr. AI Scientist design inherits and translates all core architectural, algorithmic, and evaluative features of the AI Scientist-v2 platform to a domain-general, resource-conservative system suitable for broad adoption in educational, research, and computational science settings [2504.08066].

Source: https://www.emergentmind.com/topics/jr-ai-scientist-system