Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CausalEvolve: Towards Open-Ended Discovery with Causal Scratchpad

Yongqiang Chen, Chenxi Liu, Zhenhao Chen, Tongliang Liu, Bo Han, Kun Zhang

Abstract

Evolve-based agent such as AlphaEvolve is one of the notable successes in using Large Language Models (LLMs) to build AI Scientists. These agents tackle open-ended scientific problems by iteratively improving and evolving programs, leveraging the prior knowledge and reasoning capabilities of LLMs. Despite the success, existing evolve-based agents lack targeted guidance for evolution and effective mechanisms for organizing and utilizing knowledge acquired from past evolutionary experience. Consequently, they suffer from decreasing evolution efficiency and exhibit oscillatory behavior when approaching known performance boundaries. To mitigate the gap, we develop CausalEvolve, equipped with a causal scratchpad that leverages LLMs to identify and reason about guiding factors for evolution. At the beginning, CausalEvolve first identifies outcome-level factors that offer complementary inspirations in improving the target objective. During the evolution, CausalEvolve also inspects surprise patterns during the evolution and abductive reasoning to hypothesize new factors, which in turn offer novel directions. Through comprehensive experiments, we show that CausalEvolve effectively improves the evolutionary efficiency and discovers better solutions in 4 challenging open-ended scientific tasks.

CausalEvolve: Towards Open-Ended Discovery with Causal Scratchpad

Abstract

Evolve-based agent such as AlphaEvolve is one of the notable successes in using Large Language Models (LLMs) to build AI Scientists. These agents tackle open-ended scientific problems by iteratively improving and evolving programs, leveraging the prior knowledge and reasoning capabilities of LLMs. Despite the success, existing evolve-based agents lack targeted guidance for evolution and effective mechanisms for organizing and utilizing knowledge acquired from past evolutionary experience. Consequently, they suffer from decreasing evolution efficiency and exhibit oscillatory behavior when approaching known performance boundaries. To mitigate the gap, we develop CausalEvolve, equipped with a causal scratchpad that leverages LLMs to identify and reason about guiding factors for evolution. At the beginning, CausalEvolve first identifies outcome-level factors that offer complementary inspirations in improving the target objective. During the evolution, CausalEvolve also inspects surprise patterns during the evolution and abductive reasoning to hypothesize new factors, which in turn offer novel directions. Through comprehensive experiments, we show that CausalEvolve effectively improves the evolutionary efficiency and discovers better solutions in 4 challenging open-ended scientific tasks.
Paper Structure (55 sections, 4 theorems, 44 equations, 2 figures, 3 tables)

This paper contains 55 sections, 4 theorems, 44 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. AI Scientist Agents.
  4. Causality for Scientific Discovery.
  5. Scientific Discovery via Objective Optimization
  6. Formulation of Scientific Discovery
  7. Scientific knowledge.
  8. POMDP process.
  9. Evaluation as intervention on SCM.
  10. Belief as a probability distribution over $\Theta$.
  11. Essentiality of Causal Knowledge for AI Scientists
  12. Evolutionary efficiency of Causal AI Scientist.
  13. Generalizability of Causal AI Scientist.
  14. Causal Scratchpad for Evolutionary Coding Agent
  15. Outcome-level Factor
  16. ...and 40 more sections

Key Result

Theorem 3.2

Under the given environment, there exists a policy $\pi_{\mathrm{causal}}$ such that with probability at least $1-\delta$, $F(\hat{p};\theta_\mathrm{sci})$ obtains less than $2\epsilon$ error than the optimal value, with $O(d\log(K))$ turns; In contrast, the black-box baseline needs $O(K)$.

Figures (2)

  • Figure 1: The iterative scientific discovery loop. Left: Conceptual flow of the agent. The agent maintains a scratchpad memory ($m$), proposes a program ($p$), and observes the outcome ($y$) which is constrained by the unknown world state ($\theta_{\mathrm{sci}}$). The outcome feeds back into the memory for the next step. Right: The diagram illustrates how the AI Scientist probes the unknown world state $\theta_{\mathrm{sci}}$. By proposing a candidate program $p_t$, the agent triggers an experiment yielding outcome $y_t$. This observation provides evidence about $\theta_{\mathrm{sci}}$, which is integrated into the agent's scratchpad memory $m_{t+1}$. Over time steps $t, t+1, \dots$, this recurrent process allows the agent to navigate the performance landscape and converge towards optimal programs despite the static but unknown nature of $\theta_{\mathrm{sci}}$.
  • Figure 2: The iterative scientific discovery loop. Left: Conceptual flow of the agent. The agent maintains a scratchpad memory ($m$), proposes a program ($p$), and observes the outcome ($y$) which is constrained by the unknown world state ($\theta_{\mathrm{sci}}$). The outcome feeds back into the memory for the next step. Right: The diagram illustrates how the AI Scientist probes the unknown world state $\theta_{\mathrm{sci}}$. By proposing a candidate program $p_t$, the agent triggers an experiment yielding outcome $y_t$. This observation provides evidence about $\theta_{\mathrm{sci}}$, which is integrated into the agent's scratchpad memory $m_{t+1}$. Over time steps $t, t+1, \dots$, this recurrent process allows the agent to navigate the performance landscape and converge towards optimal programs despite the static but unknown nature of $\theta_{\mathrm{sci}}$.

Theorems & Definitions (7)

  • Definition 3.1: Causal AI Scientist
  • Theorem 3.2: Informal
  • Theorem 3.3
  • Theorem B.1: Formal version of Theorem \ref{['thm:static-sample-efficiency-gap']}
  • proof
  • Theorem C.1: Non-identifiability barrier under shifts
  • proof