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

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.

💡 Research Summary

Paper Overview
The authors identify a critical weakness of compact large language models (LLMs): they often lack the ability to formulate a global plan before generating multi‑step reasoning, leading to error propagation in long‑horizon tasks. While teacher‑student approaches can provide external plans, they incur prohibitive latency and computational costs at inference time. Existing parameter‑efficient fine‑tuning (PEFT) methods such as LoRA or static steering vectors also fall short because they are instance‑agnostic and cannot adapt to the diverse logical demands of each query.

Key Idea – PILOT
PILOT (Planning via Internalized Latent Optimization Trajectories) introduces a non‑invasive framework that internalizes the strategic oversight of a larger “expert” model into a lightweight, query‑conditioned latent guidance vector. The method does not modify the backbone weights; instead, it injects a dynamically generated anchor into a chosen pivot layer of the frozen LLM.

Construction of Target Latent State (z*)

1. Construct‑and‑Verify Pipeline – For each training example, the base model’s failure is identified. An expert model (DeepSeek‑V3.1) generates a heuristic guidance sequence g_exp. A blind test discards any sample where g_exp alone solves the problem, ensuring the guidance encodes strategic intent rather than a shortcut.
2. Homogeneous Target Projection – The concatenated sequence `