iCLP: Large Language Model Reasoning with Implicit Cognition Latent Planning

TL;DR

This work tackles the brittleness and inefficiency of explicit plan-based reasoning in large language models by introducing iCLP, which learns a latent plan space for guiding reasoning. Latent plans are obtained by distilling explicit plans from CoT traces, encoding them with a vector-quantized autoencoder using a codebook of size $K=2048$ and latent dimension $d_h=512$, and then fine-tuning LLMs to reason in language conditioned on discrete LP tokens. The approach yields notable improvements in accuracy and token efficiency on mathematical reasoning and code generation tasks, with strong cross-domain generalization and maintained interpretability of the reasoning process. By separating planning (latent space) from reasoning (textual CoT), iCLP enables generalizable, scalable guidance across diverse problems while preserving the transparency of chain-of-thought reasoning.

Abstract

Large language models (LLMs), when guided by explicit textual plans, can perform reliable step-by-step reasoning during problem-solving. However, generating accurate and effective textual plans remains challenging due to LLM hallucinations and the high diversity of task-specific questions. To address this, we draw inspiration from human Implicit Cognition (IC), the subconscious process by which decisions are guided by compact, generalized patterns learned from past experiences without requiring explicit verbalization. We propose iCLP, a novel framework that enables LLMs to adaptively generate latent plans (LPs), which are compact encodings of effective reasoning instructions. iCLP first distills explicit plans from existing step-by-step reasoning trajectories. It then learns discrete representations of these plans via a vector-quantized autoencoder coupled with a codebook. Finally, by fine-tuning LLMs on paired latent plans and corresponding reasoning steps, the models learn to perform implicit planning during reasoning. Experimental results on mathematical reasoning and code generation tasks demonstrate that, with iCLP, LLMs can plan in latent space while reasoning in language space. This approach yields significant improvements in both accuracy and efficiency and, crucially, demonstrates strong cross-domain generalization while preserving the interpretability of chain-of-thought reasoning.

Figures (6)

• Figure 1: Illustration of relations between explicit plans of questions from different categories. We extract $200$ samples from each category of the MATH dataset's 7 categories and prompt the LLM to decompose the answers into individual steps, followed by summarizing their explicit plans. (a)(b)(c) display the encoding distances between pairs of items: questions, explicit plans of Step 1 and Step 3, respectively. (d)(e) show the encoding clusters of the explicit plans of Step 1 and Step 3.
• Figure 2: Illustration of the overall pipeline of iCLP. The upper part shows the process of Plan Distillation using a sample from the Counting & Probability category of the MATH dataset. The right part depicts the encoder - quantizer - decoder structure used for Latent Plan Generation.
• Figure 3: Illustration of the number of distilled plans and the accumulation mode performance of LLMs with iCLP. (a) shows the number of explicit plans that can be distilled from each category of the MATH and GSM8K datasets. The x-axis labels represent the abbreviations of the seven category names (see appendix). (b)(c)(d) show the accuracy gain ('y-axis') over the base model after fine-tuning the LLM with plans accumulated from the MATH and GSM8K datasets. Accuracy is measured across four datasets, with abbreviated names provided in Table \ref{['tab:mainresults']}. (b)(c)(d) correspond to Qwen2.5 models with 0.5B, 3B, and 7B parameters, respectively.
• Figure 4: Illustration of the relations of the encoding distances between pairwise latent plans from different reasoning steps (1, 2, 3, and 4). We randomly sample 200 questions from the test set of the MATH dataset and extract the encodings of latent plans generated by Qwen2.5-7B with iCLP during problem solving. Subfigures (a), (b), (c), and (d) present the results for the 1st, 2nd, 3rd, and 4th latent plans, respectively.
• Figure 5: Illustration of the encodings of latent plans from different reasoning steps (1, 2, 3, and 4) in 2D space. We follow the same procedure as in Figure \ref{['fig:iplanrelations']} and visualize the latent plan encodings using t-SNE.