AutoDiscovery: Open-ended Scientific Discovery via Bayesian Surprise

1. Problem Statement

The paper addresses the challenge of Autonomous Scientific Discovery (ASD), specifically moving beyond "goal-driven" settings (where a human provides a specific research question) to open-ended discovery (where the AI system must independently generate, test, and refine hypotheses without explicit guidance).

Key Challenges Identified:

• Search Space Complexity: The space of potential scientific hypotheses is vast. Simple diversity heuristics often fail to navigate this space meaningfully.
• Subjectivity of "Interestingness": Existing methods often rely on human proxies (e.g., "interestingness" or "novelty") which are subjective, inconsistent among experts, and difficult to automate reliably.
• Lack of Direction: Without a clear goal, AI agents struggle to balance exploration (trying new, unknown hypothesis spaces) and exploitation (refining promising leads) effectively.

2. Methodology: AUTODISCOVERY

The authors propose AUTODISCOVERY, a framework that drives open-ended scientific exploration using Bayesian Surprise as the primary reward signal. The system operates within a Data-Driven Discovery (DDD) setting, where the agent has access to datasets and a Python execution environment.

A. Core Concept: Bayesian Surprise

Instead of asking "Is this hypothesis interesting?", the system asks "How much did this new data change the model's belief?"

• Belief Elicitation: The system treats the Large Language Model (LLM) as a Bayesian observer. It estimates the LLM's prior belief ($\theta_H$) about a hypothesis $H$ before seeing data, and its posterior belief after seeing experimental results ($V_D$).
• Surprisal Calculation: Surprise is quantified as the Kullback-Leibler (KL) divergence between the prior and posterior distributions:
  $$BS(H, V_D) = D_{KL}(P(\theta_H | V_D) \parallel P(\theta_H))$$
• Belief Shift: A "surprise" is formally defined as a significant directional shift in the expected belief (e.g., moving from "likely true" to "likely false" or vice versa) across a decision threshold (typically 0.5).

B. Search Algorithm: MCTS with Progressive Widening

To navigate the combinatorial space of hypotheses, AUTODISCOVERY employs Monte Carlo Tree Search (MCTS) guided by Bayesian surprise.

• Tree Structure: Nodes represent hypotheses; edges represent the transition to new experiments/hypotheses.
• Progressive Widening: Since the branching factor (number of possible next hypotheses) is infinite, the algorithm dynamically increases the number of children a node can have based on its visitation count. This prevents the search from getting stuck in local optima while ensuring diverse exploration.
• Selection Policy (UCT): The algorithm uses the Upper Confidence Bound (UCT) formula to balance exploration and exploitation:
  $$UCT(H) = \underbrace{\frac{\sum S(h)}{N(H)}}_{\text{Exploitation}} + C \cdot \underbrace{\sqrt{\frac{2 \log N(H_{parent})}{N(H)}}}_{\text{Exploration}}$$
  Where $S(h)$ is the surprisal reward.
• Deduplication: To handle semantic redundancy, the system uses LLM-based Hierarchical Agglomerative Clustering (HAC). It groups semantically similar hypotheses and merges them if an LLM confirms they are equivalent, ensuring the search focuses on unique insights.

C. Multi-Agent Architecture

The discovery process is managed by a collaborative multi-agent system:

1. Hypothesis Generator: Proposes new assertions.
2. Experiment Programmer: Writes Python code to test the hypothesis.
3. Code Executor: Runs the code in a sandboxed environment.
4. Analyst/Reviewer: Analyzes results, critiques code, and updates the belief state.
5. Belief Elicitor: Samples the LLM multiple times to estimate the prior and posterior probability distributions required for the surprise calculation.

3. Key Contributions

1. Formal Definition of Surprise: The paper provides the first formal definition of "surprise" in the context of ASD, grounding it in Bayesian belief shifts rather than subjective human metrics.
2. Novel Search Framework (AutoDS): It introduces a method combining Bayesian surprise with MCTS and progressive widening to perform open-ended hypothesis search without a predefined goal.
3. Empirical Validation: The authors demonstrate that optimizing for Bayesian surprise is a superior proxy for human "interestingness" compared to other automatic metrics.

4. Experimental Results

The system was evaluated across 21 real-world datasets spanning biology, economics, finance, and behavioral science (e.g., hurricane damage, BMI studies, Alzheimer's data).

• Performance vs. Baselines: AUTODISCOVERY (MCTS) significantly outperformed strong baselines, including:
  • Repeated Sampling: Independent, context-free generation.
  • Linear Search: Sequential, single-path generation.
  • Greedy Search: Always picking the highest-value node (no exploration).
  • Beam Search: Breadth-first with pruning.
  • Result: AUTODISCOVERY produced 5–29% more surprising discoveries under a fixed budget of 500 hypotheses.
• Human Evaluation: In a study with domain experts (STEM PhDs):
  • 67% of the discoveries deemed "surprising" by the LLM were also rated as surprising by human experts.
  • Bayesian surprise showed a much higher correlation with human surprise (0.67) compared to other automatic rewards like "LLM interestingness" (0.11–0.22).
• Search Efficiency: The MCTS approach maintained high search efficiency over time, whereas greedy and beam search strategies degraded as the search space was exhausted.

5. Significance and Implications

• Paradigm Shift: The paper moves ASD from "answering specific questions" to "discovering what questions to ask." This mimics the workflow of human scientists who explore data to find unexpected patterns.
• Objective Metric for Novelty: By using Bayesian surprise, the system bypasses the subjectivity of "interestingness," providing a mathematically grounded metric for novelty that correlates well with human intuition.
• Scalability: The combination of MCTS and progressive widening offers a scalable solution for navigating massive hypothesis spaces, a critical bottleneck in automated science.
• Future Directions: The authors note that while the system is promising, it requires guardrails (e.g., human oversight) to prevent the generation of scientifically spurious findings. Future work includes integrating literature grounding to refine the "prior" beliefs of the LLM.

In summary, AUTODISCOVERY represents a significant step toward fully autonomous scientific agents that can not only execute experiments but also autonomously identify the most scientifically valuable and surprising avenues of inquiry.