Language Agents Mirror Human Causal Reasoning Biases. How Can We Help Them Think Like Scientists?

TL;DR

This work probes whether language-model agents can discover and reason about causal structure, using a text-based Blicket Test to elicit active exploration and inference. It reveals a robust disjunctive bias across model families, with performance dropping as task complexity increases and conjunctive rules become more challenging. The authors quantify exploration via information gain and demonstrate that LM exploration is inefficient compared to an information-maximizing Oracle, and that LM inferences are biased toward disjunctive explanations even with uncertain data. To address this, they introduce inference-time hypothesis sampling to flatten priors and encourage elimination of competing hypotheses, which substantially reduces bias and improves scientific-like causal reasoning. The findings highlight both the limitations of current LM agents in causal exploration and a practical path toward more principled, scientifically grounded reasoning in autonomous systems.

Abstract

Language model (LM) agents are increasingly used as autonomous decision-makers which need to actively gather information to guide their decisions. A crucial cognitive skill for such agents is the efficient exploration and understanding of the causal structure of the world -- key to robust, scientifically grounded reasoning. Yet, it remains unclear whether LMs possess this capability or exhibit systematic biases leading to erroneous conclusions. In this work, we examine LMs' ability to explore and infer causal relationships, using the well-established Blicket Test paradigm from developmental psychology. We find that LMs reliably infer the common, intuitive disjunctive causal relationships but systematically struggle with the unusual, yet equally (or sometimes even more) evidenced conjunctive ones. This "disjunctive bias" persists across model families, sizes, and prompting strategies, and performance further declines as task complexity increases. Interestingly, an analogous bias appears in human adults, suggesting that LMs may have inherited deep-seated reasoning heuristics from their training data. To this end, we quantify similarities between LMs and humans, finding that LMs exhibit adult-like inference profiles (but not child-like). Finally, we propose a test-time sampling method which explicitly samples and eliminates hypotheses about causal relationships from the LM. This scalable approach significantly reduces the disjunctive bias and moves LMs closer to the goal of scientific, causally rigorous reasoning.

Figures (16)

• Figure 1: The Blicket Test.
• Figure 2: Quiz accuracy of various models during the question-and-answering phase of the Blicket Test. The accuracy measures the proportion of trials where the model correctly identifies all Blickets.
• Figure 3: Correlation analysis of factors contributing to model performance. Each point is a unique model + prompt + environment rule combination. The Spearman's rank correlation along with the p-value is reported.
• Figure 4: Hypothesis elimination efficiency.
• Figure 5: Evaluating LM's ability to infer causal relationship when the same exploration data is given as context in the 8 objects setting. Error bar denote standard error of mean.

Theorems & Definitions (2)

• Remark 1
• proof