Shota Takashiro

Kohsei Matsutani, Shota Takashiro, Gouki Minegishi et al.

Large language models (LLMs) are typically trained by reinforcement learning (RL) with verifiable rewards (RLVR) and supervised fine-tuning (SFT) on reasoning traces to improve their reasoning abilities. However, how these methods shape reasoning capabilities remains largely elusive. Going beyond an accuracy-based investigation of how these two components sculpt the reasoning process, this paper introduces a novel analysis framework that quantifies reasoning paths and captures their qualitative changes under each training process (with models of 1.5B, 7B, and 14B parameters on mathematical domains). Specifically, we investigate the reasoning process at two levels of granularity: the trajectory-level, which examines complete reasoning outputs, and the step-level, which analyzes reasoning graphs whose nodes correspond to individual reasoning steps. Notably, clustering of unique reasoning trajectories shows complementary effects: RL compresses incorrect trajectories, whereas SFT expands correct ones. Step-level analysis reveals that RL steepens (about 2.5 times), while SFT flattens (reduced to about one-third), the decay rates of node visitation frequency, degree, and betweenness centrality distributions in the reasoning graph. This indicates that RL concentrates reasoning functionality into a small subset of steps, while SFT homogenizes it across many steps. Furthermore, by evaluating the reasoning graph topologies from multiple perspectives, we delineate the shared and distinct characteristics of RL and SFT. Our work presents a novel reasoning path perspective that explains why the current best practice of two-stage training, with SFT followed by RL, is successful, and offers practical implications for data construction and more efficient learning approaches.

Fumiya Uchiyama, Rintaro Yanagi, Shohei Taniguchi et al.

Contrastive learning has the capacity to model multimodal probability distributions by embedding and aligning visual representations with semantics from captions. This approach enables the estimation of relational semantic similarity; however, it remains unclear whether it can also represent absolute semantic informativeness. In this work, we introduce a semantic informativeness metric for an image calculated from text samples via a contrastive learning model; similarly, the informativeness of a text is calculated from image samples. We propose a redefinition of the concept of Information Gain, a concept previously explored in natural language processing, extending its application to the domains of vision and language. Our metric quantifies how conditioning on an image distorts the distribution of associated texts, and vice versa for text conditioning on image distributions. In OpenCLIP's empirical results, we observe that images with the lowest Information Gain scores often correspond to placeholder icons such as "image not found." Furthermore, we propose to measure a norm-based metric of the embedding to estimate the Information Gain, following the theoretical results for Skip-Gram with Negative Sampling (SGNS) word embedding. Information Gain can be measured using either CLIP or SigLIP, and the results demonstrate a strong correlation with a coefficient of determination ranging from 0.98 to 1.00. After obtaining the mean and the covariance of the sample embedding, the computational cost of this method is independent of the sample size, and it is compatible with publicly available, open-weight models.