Large Language Models Can Take False First Steps at Inference-time Planning

TL;DR

This paper explains why large language models often seem short-sighted during inference despite their training on sequence planning, attributing it to the influence of self-generated context on their planning behavior.

Contribution

It introduces a Bayesian model that accounts for the gap between training and inference planning behaviors in LLMs, supported by controlled experiments.

Findings

Self-generated context influences planning behavior during inference.

Planning strength increases as self-generated context accumulates.

Reduced initial bias when conditioning on self-generated sequences.

Abstract

Large language models (LLMs) have been shown to acquire sequence-level planning abilities during training, yet their planning behavior exhibited at inference time often appears short-sighted and inconsistent with these capabilities. We propose a Bayesian account for this gap by grounding planning behavior in the evolving generative context: given the subtle differences between natural language and the language internalized by LLMs, accumulated self-generated context drives a planning-shift during inference and thereby creates the appearance of compromised planning behavior. We further validate the proposed model through two controlled experiments: a random-generation task demonstrating constrained planning under human prompts and increasing planning strength as self-generated context accumulates, and a Gaussian-sampling task showing reduced initial bias when conditioning on self-generated sequences. These findings provide a theoretical explanation along with empirical evidence for characterizing how LLMs plan ahead during inference.