Xiaoshu Chen

Xiaoshu Chen, Sihang Zhou, Ke Liang et al.

Compressing long chains of thought (CoT) into compact latent tokens is crucial for efficient reasoning with large language models (LLMs). Recent studies employ autoencoders to achieve this by reconstructing textual CoT from latent tokens, thus encoding CoT semantics. However, treating textual CoT as the reconstruction target forces latent tokens to preserve surface-level linguistic features (e.g., word choice and syntax), introducing a strong linguistic inductive bias that prioritizes linguistic form over reasoning structure and limits logical abstraction. Thus, we propose ImgCoT that replaces the reconstruction target from textual CoT to the visual CoT obtained by rendering CoT into images. This substitutes linguistic bias with spatial inductive bias, i.e., a tendency to model spatial layouts of the reasoning steps in visual CoT, enabling latent tokens to better capture global reasoning structure. Moreover, although visual latent tokens encode abstract reasoning structure, they may blur reasoning details. We thus propose a loose ImgCoT, a hybrid reasoning that augments visual latent tokens with a few key textual reasoning steps, selected based on low token log-likelihood. This design allows LLMs to retain both global reasoning structure and fine-grained reasoning details with fewer tokens than the complete CoT. Extensive experiments across multiple datasets and LLMs demonstrate the effectiveness of the two versions of ImgCoT.

Xiaoshu Chen, Sihang Zhou, Ke Liang et al.

Chain of thought finetuning (cot-finetuning) aims to endow small language models (SLM) with reasoning ability to improve their performance towards specific tasks by allowing them to imitate the reasoning procedure of large language models (LLM) beyond simply predicting the answers. Most existing cot-finetuning methods adopt a pre-thinking mechanism, allowing the SLM to generate a rationale before providing an answer. This mechanism enables SLM to analyze and think about complex questions, but it also makes answer correctness highly sensitive to minor errors in rationale. Therefore, we propose a robust post-thinking mechanism to generate answers before rationale. Thanks to this answer-first setting, 1) the answer can escape from the adverse effects caused by minor errors in the rationale; 2) the rationale serves as an error amplifier to the answer, which makes the SLM focus on learning hard samples; 3) the inferring efficiency can also benefit from the setting since users can stop the generation right after answers are outputted when inference is conducted. However, although the post-thinking mechanism brings many advantages and improves the overall performance of SLM on specific tasks, it may lose the ability to think about the questions and decompose complex questions into simple sub-questions compared to pre-thinking mechanism. Therefore, a plug-and-play adaptive-thinking mechanism is proposed with the aid of the soft prompt tuning to integrate the merits of the pre-thinking mechanism and post-thinking mechanism, in which a perception module is introduced to adaptively prompt SLM answer or think first based on perceiving the complexity of the questions. Extensive experiments are conducted across 12 reasoning tasks and 2 representative language models to demonstrate the effectiveness of the proposed mechanism.

Yanwen Chong, Congchong Nie, Yulong Tao et al.

Global context information is vital in visual understanding problems, especially in pixel-level semantic segmentation. The mainstream methods adopt the self-attention mechanism to model global context information. However, pixels belonging to different classes usually have weak feature correlation. Modeling the global pixel-level correlation matrix indiscriminately is extremely redundant in the self-attention mechanism. In order to solve the above problem, we propose a hierarchical context network to differentially model homogeneous pixels with strong correlations and heterogeneous pixels with weak correlations. Specifically, we first propose a multi-scale guided pre-segmentation module to divide the entire feature map into different classed-based homogeneous regions. Within each homogeneous region, we design the pixel context module to capture pixel-level correlations. Subsequently, different from the self-attention mechanism that still models weak heterogeneous correlations in a dense pixel-level manner, the region context module is proposed to model sparse region-level dependencies using a unified representation of each region. Through aggregating fine-grained pixel context features and coarse-grained region context features, our proposed network can not only hierarchically model global context information but also harvest multi-granularity representations to more robustly identify multi-scale objects. We evaluate our approach on Cityscapes and the ISPRS Vaihingen dataset. Without Bells or Whistles, our approach realizes a mean IoU of 82.8% and overall accuracy of 91.4% on Cityscapes and ISPRS Vaihingen test set, achieving state-of-the-art results.