PILOT: Planning via Internalized Latent Optimization Trajectories for Large Language Models

TL;DR

PILOT is a framework that internalizes strategic planning into large language models using latent guidance, improving multi-step reasoning without modifying core weights.

Contribution

It introduces a lightweight Hyper-Network to synthesize internal guidance, enabling large models to perform better in reasoning tasks without external dependencies.

Findings

PILOT outperforms baselines by +8.9% on MATH500.

It stabilizes reasoning trajectories in mathematical and coding benchmarks.

PILOT achieves these improvements with negligible inference latency.

Abstract

Strategic planning is critical for multi-step reasoning, yet compact Large Language Models (LLMs) often lack the capacity to formulate global strategies, leading to error propagation in long-horizon tasks. Our analysis reveals that LLMs possess latent reasoning capabilities that can be unlocked when conditioned on explicit plans from a teacher model; however, runtime reliance on external guidance is often impractical due to latency and availability constraints. To bridge this gap, we propose PILOT (Planning via Internalized Latent Optimization Trajectories), a non-invasive framework designed to internalize the strategic oversight of large models into intrinsic Latent Guidance. Instead of altering backbone weights, PILOT employs a lightweight Hyper-Network to synthesize a query-conditioned Latent Guidance vector. This vector acts as an internal steering mechanism, guiding the model's representations toward optimal reasoning paths. Extensive experiments on mathematical and coding benchmarks demonstrate that PILOT effectively stabilizes reasoning trajectories, consistently outperforming strong baselines (e.g., +8.9% on MATH500) with negligible inference latency.