SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation

Problem
This paper addresses the interpretability gap in reasoning segmentation pipelines used in conjunction with large language models (LLMs). Existing methods, such as latent query alignment, operate as opaque “black boxes,” failing to provide transparent connections between reasoning processes and visual perception. Additionally, current textual localization methods are readable but lack true interpretability, often serving as unconstrained post-hoc analyses. The authors propose SegCompass, an end-to-end model that aims to create a clear and interpretable alignment pathway between reasoning and visual inputs. This work is presented as a preprint and has not yet undergone peer review.

Method
SegCompass employs a Sparse Autoencoder (SAE) architecture to facilitate an explicit and differentiable alignment between chain-of-thought (CoT) reasoning and visual tokens. The model processes an image-instruction pair by first generating a CoT trace. The SAE maps both the CoT and visual tokens into a shared high-dimensional sparse concept space. A query codebook is utilized to select salient concepts from this space, which are then spatially grounded by a slot mapper into a multi-slot heatmap. This heatmap guides the final mask decoder, enabling the model to produce segmentation masks. The training process integrates reinforcement learning for the reasoning path alongside standard segmentation supervision, allowing for a unified learning approach.

Results
SegCompass demonstrates competitive performance across five challenging benchmarks, matching or surpassing state-of-the-art methods. The authors report significant improvements in mask accuracy correlated with the quality of learned sparse concepts. While specific numerical results are not detailed in the abstract, the emphasis on strong correlations suggests that the model’s interpretability directly contributes to its performance. The paper includes extensive visual and quantitative analyses to support these claims, indicating that SegCompass provides a more coherent and traceable alignment than existing methods.

Limitations
The authors acknowledge that while SegCompass improves interpretability and performance, it may still be limited by the inherent challenges of aligning high-dimensional visual data with abstract reasoning. They do not discuss potential scalability issues or the computational overhead introduced by the SAE architecture. Additionally, the reliance on a joint training framework may complicate the model’s deployment in real-time applications, where interpretability and speed are critical.

Why it matters
The development of SegCompass has significant implications for the fields of computer vision and natural language processing, particularly in applications requiring interpretable AI systems. By providing a “white-box” approach to reasoning segmentation, this work paves the way for more transparent AI models that can be trusted in critical decision-making scenarios. The enhanced alignment between reasoning and visual perception could lead to advancements in areas such as autonomous systems, interactive AI, and human-AI collaboration, where understanding the rationale behind model decisions is essential.

Authors: Zhenyu Lu, Liupeng Li, Jinpeng Wang, Haoqian Kang, Yan Feng, Ke Chen, Yaowei Wang
Source: arXiv:2605.22658
URL: https://arxiv.org/abs/2605.22658v1