LenGuard-GPC: Length Guarding with Guided-Prompt Consistency for Spatial Reasoning Reinforce Learning

2026-07-19

AI Summary

This work addresses the challenges in multi-view spatial reasoning, where chain-of-thought reasoning tends to be verbose and existing reinforcement learning approaches suffer from sparse rewards, difficulty in error localization, and poor control over generation length. To overcome these limitations, the authors propose LenGuard-GPC, a novel framework that introduces guided prompt consistency into spatial reasoning via reinforcement learning. Specifically, it constructs a dense reward signal based on the KL divergence between token-level prediction distributions under standard and guided prompts, complemented by a staged length-reward mechanism. This dual approach simultaneously enhances reasoning accuracy and effectively constrains response length. Extensive experiments across six benchmarks demonstrate that LenGuard-GPC significantly outperforms the original GRPO method, achieving higher accuracy while substantially reducing average response length.

Abstract

Multi-view spatial reasoning requires vision-language models to compare visual evidence across images, align object correspondences, and infer spatial relations over long visual contexts, a setting where chain-of-thought reasoning tends to grow verbose without becoming more accurate. Reinforcement learning with verifiable rewards is a natural fit for this task, but standard GRPO reward relies on sparse outcome-level feedback and gives no signal about where a reasoning trajectory goes wrong, nor any control over its length. We propose LenGuard-GPC, a dense reward framework that addresses both problems together. For each sampled trajectory, it compares the token-wise predictive distributions under a standard prompt and a guided prompt, and uses the resulting token-sum KL divergence as a dense reward signal. Since this KL penalty accumulates over tokens and would otherwise reward shorter responses regardless of their quality, we introduce a staged length bonus that keeps reasoning length within a controlled range without simply encouraging brevity. On six multi-view spatial reasoning benchmarks, LenGuard-GPC improves accuracy over vanilla GRPO while reducing average response length.

Problem

• spatial reasoning
• reinforcement learning
• reasoning length control
• dense reward
• multi-view vision-language models

Innovation

• dense reward
• guided-prompt consistency
• KL divergence
• length control
• spatial reasoning