COLD: Causal reasOning in cLosed Daily activities

TL;DR

COLD introduces a closed, world-grounded framework for evaluating causal reasoning in LLMs using script-based daily activities. By deriving observational graphs $\,\mathcal{G}_o$ and causal graphs $\mathcal{G}_c$ from the DeScript corpus, it generates ~ $9.52$ million causal query triplets and analyzes LM performance through MCQA prompts and backdoor-adjusted estimands, revealing that causal reasoning remains challenging for open-weight models. The approach combines causal inference concepts ($do$-operator, $d$-separation, backdoor criterion) with large-scale data generation, achieving a near mini-Turing test scale for causal queries and enabling rigorous LM validation. The findings underscore the gap in current LLM causal understanding and establish COLD as a versatile platform for future causal-reasoning research and potential training objectives.

Abstract

Large Language Models (LLMs) have shown state-of-the-art performance in a variety of tasks, including arithmetic and reasoning; however, to gauge the intellectual capabilities of LLMs, causal reasoning has become a reliable proxy for validating a general understanding of the mechanics and intricacies of the world similar to humans. Previous works in natural language processing (NLP) have either focused on open-ended causal reasoning via causal commonsense reasoning (CCR) or framed a symbolic representation-based question answering for theoretically backed-up analysis via a causal inference engine. The former adds an advantage of real-world grounding but lacks theoretically backed-up analysis/validation, whereas the latter is far from real-world grounding. In this work, we bridge this gap by proposing the COLD (Causal reasOning in cLosed Daily activities) framework, which is built upon human understanding of daily real-world activities to reason about the causal nature of events. We show that the proposed framework facilitates the creation of enormous causal queries (~ 9 million) and comes close to the mini-turing test, simulating causal reasoning to evaluate the understanding of a daily real-world task. We evaluate multiple LLMs on the created causal queries and find that causal reasoning is challenging even for activities trivial to humans. We further explore (the causal reasoning abilities of LLMs) using the backdoor criterion to determine the causal strength between events.

Figures (13)

• Figure 1: U denotes the unobserved variables, confounding all events present in a real-world activity. In an activity, some events cause other events to happen. For example, in "traveling by an airplane", the event of "check-in luggage" causes events like "taking back luggage."
• Figure 2: Left: the figure represents the closed nature of daily real-world activities (capturing commonsense, commonly understood by humans), start and end given the context of the task, i.e., the pre-activity world and post-activity world activities marginalize out the dependence of event occurring during the activity with the rest of the world. Right: Causal Graph for "going grocery shopping." Notice the collider (red nodes) makes the independent set of nodes (highlighted in different colors) unconditionally independent in the causal graph. In contrast, when given a condition on a collider ("put bags in cart", the two clusters (yellow and blue) become dependent (if collider is observed, both yellow and blue clusters may have been observed as well).
• Figure 3: The proposed COLD framework for evaluating LLMs for causal reasoning. The human-written Event Sequence Descriptions (ESDs) are obtained from crowdsource workers and include a telegrammic-style sequence of events when performing an activity. The Observational Graph and the Causal Graph for an activity are used to create causal query triplets (details in Algorithm \ref{['alg:dataset_creation']}), shown towards the right. Using counterfactual reasoning, “going to the kitchen” is possible without going to the market (if the ingredients are already available), making “come home with the ingredients.” a more plausible effect among the given choices. Similarly, in the second example, the event “going to market” has no direct relation with the event “heating the oven”.
• Figure 4: Causal Graphical Model of Events. $E_1$ temporally precedes $E_2$, and $z$ is trajectory variable, which assumes a values $t$ where $t \in$ All trajectories from start to $E_1$
• Figure 5: Input prompt formats for the MCQA-based evaluation of autoregressive open-weight models (e.g., llama(-2), GPT-J, etc.). The black text is the templated input. The orange text is the input from the created causal query triplets, where the activity name denotes the description of the activity like baking a cake. The next-token prediction probabilities of the option IDs at the red text is used as the observed prediction distribution.