Latent Logic Tree Extraction for Event Sequence Explanation from LLMs

TL;DR

The paper addresses explainability for high-volume event sequences by learning latent logic trees that explain observed histories. It proposes LaTee, an amortized EM framework that uses an LLM-derived prior over logic trees and a temporal point process likelihood, with a GFlowNet-based E-step to sample diverse trees and an M-step to update likelihood and prior parameters. Key contributions include (i) amortized EM for latent logic-tree inference, (ii) GFlowNet-guided diverse structure generation, and (iii) strong performance on real datasets with semantic event information, approaching or surpassing state-of-the-art TPP baselines while enabling online adaptation. The approach enhances interpretability and robustness in settings like healthcare and robotics by combining symbolic explanations with probabilistic modeling, leveraging local LLMs for privacy-preserving reasoning and knowledge-driven event prediction.

Abstract

Modern high-stakes systems, such as healthcare or robotics, often generate vast streaming event sequences. Our goal is to design an efficient, plug-and-play tool to elicit logic tree-based explanations from Large Language Models (LLMs) to provide customized insights into each observed event sequence. Built on the temporal point process model for events, our method employs the likelihood function as a score to evaluate generated logic trees. We propose an amortized Expectation-Maximization (EM) learning framework and treat the logic tree as latent variables. In the E-step, we evaluate the posterior distribution over the latent logic trees using an LLM prior and the likelihood of the observed event sequences. LLM provides a high-quality prior for the latent logic trees, however, since the posterior is built over a discrete combinatorial space, we cannot get the closed-form solution. We propose to generate logic tree samples from the posterior using a learnable GFlowNet, which is a diversity-seeking generator for structured discrete variables. The M-step employs the generated logic rules to approximate marginalization over the posterior, facilitating the learning of model parameters and refining the tunable LLM prior parameters. In the online setting, our locally built, lightweight model will iteratively extract the most relevant rules from LLMs for each sequence using only a few iterations. Empirical demonstrations showcase the promising performance and adaptability of our framework.

Figures (9)

• Figure 1: GPT can help last event type prediction. We found replacing the semantic meaningful event names by numerical event ids in event history degrades the performance of event prediction.
• Figure 2: The Architecture of Proposed Framework. The presented event history represents a typical human trajectory: beginning with relocating to a new place, opening a box, picking up an object, and culminating in placing the retrieved item in a specific location. In the training phase, we first convert this explicit event history into a textual format. Subsequently, we employ a LLM to execute conditional sampling and perform backward reasoning, starting from the predetermined goal (E-Step). The resultant reasoning pathway is then transformed into a symbolic logic tree, which aids in updating the event probabilities (M-Step). In this context, the campfire icon signifies that the model is being updated, while the snowflake icon indicates that the model is in a 'frozen' state. We use the thickness of a path in the symbolic logic tree to represent its posterior probability.
• Figure 3: Empirical rule distributions sampled from the Language Model fine-tuned by three different approaches. We use 10,000 samples to depict the frequency distribution of a complete logic rules search space with support $|\mathcal{R}|=2500$. The x-axis represents these logic rules in a nominal 1-D format, where each point corresponds to a specific rule. The ordering of these points is not indicative of any inherent sequence.
• Figure 4: (a) Illustration of Scalability on the number of event types on four synthetic datasets. (b) Illustration of the performance of using semantic and not using semantic information on two real-world datasets.
• Figure 5: Illustration of learned symbolic logic tree structures from event histories on two real-world datasets containing semantic information. Fig. (a)-(d) are learned from EPIC-100 and Fig. (e)-(f) are learned from MIMIC-3. We use the thickness of the edges to represent posterior probability and the color to represent the weights corresponding to each rule (Black color stands for activation and Red color stands for inhibition, i.e., low blood pressure will facilitate low urine while normal blood pressure suppresses low urine).