As Large Language Models (LLMs) continue to advance, they exhibit certain cognitive patterns similar to those of humans that are not directly specified in training data. This study investigates this phenomenon by focusing on temporal cognition in LLMs. Leveraging the similarity judgment task, we find that larger models spontaneously establish a subjective temporal reference point and adhere to the Weber-Fechner law, whereby the perceived distance logarithmically compresses as years recede from this reference point. To uncover the mechanisms behind this behavior, we conducted multiple analyses across neuronal, representational, and informational levels. We first identify a set of temporal-preferential neurons and find that this group exhibits minimal activation at the subjective reference point and implements a logarithmic coding scheme convergently found in biological systems. Probing representations of years reveals a hierarchical construction process, where years evolve from basic numerical values in shallow layers to abstract temporal orientation in deep layers. Finally, using pre-trained embedding models, we found that the training corpus itself possesses an inherent, non-linear temporal structure, which provides the raw material for the model's internal construction. In discussion, we propose an experientialist perspective for understanding these findings, where the LLMs' cognition is viewed as a subjective construction of the external world by its internal representational system. This nuanced perspective implies the potential emergence of alien cognitive frameworks that humans cannot intuitively predict, pointing toward a direction for AI alignment that focuses on guiding internal constructions. Our code is available at https://TheOtherMind.github.io.

With large language models (LLMs) increasingly deployed as cognitive engines for AI agents, the reliability and effectiveness critically hinge on their intrinsic epistemic agency, which remains understudied. Epistemic agency, the ability to flexibly construct, adapt, and monitor beliefs about dynamic environments, represents a base-model-level capacity independent of specific tools, modules, or applications. We characterize the holistic process underlying epistemic agency, which unfolds in seven interrelated dimensions: prediction, decision-making, perception, memory, counterfactual thinking, belief updating, and meta-reflection. Correspondingly, we propose Reflection-Bench, a cognitive-psychology-inspired benchmark consisting of seven tasks with long-term relevance and minimization of data leakage. Through a comprehensive evaluation of 16 models using three prompting strategies, we identify a clear three-tier performance hierarchy and significant limitations of current LLMs, particularly in meta-reflection capabilities. While state-of-the-art LLMs demonstrate rudimentary signs of epistemic agency, our findings suggest several promising research directions, including enhancing core cognitive functions, improving cross-functional coordination, and developing adaptive processing mechanisms. Our code and data are available at https://github.com/AI45Lab/ReflectionBench.

We study the problem of Quality-of-Service (QoS)-Aware Personalized PageRank (PPR) computation. Existing studies mostly focus on improving the PPR query processing time. However, the query processing time alone may not reflect the service quality in real-world PPR-based systems. The query response time can be a more service-relevant measure in many applications such as the online game service of Tencent and the related-pin recommendation module of Pinterest. We make the first attempt at studying QoS-Aware PPR computation and present Quota, a system that adapts the state-of-the-art PPR algorithms to a given environment for minimizing query response time. Equipped with mathematical tools including queuing theory, algorithmic complexity analysis, and constrained optimization, Quota is designed to adapt itself to a wide spectrum of workloads. We conduct extensive experiments on real datasets and show that Quota can reduce the query response time compared with state-of-the-art PPR algorithms, often by a significant margin.

Resource optimization is commonly used in workload management, ensuring efficient and timely task completion utilising available resources. It serves to minimise costs, prompting the development of numerous algorithms tailored to this end. The majority of these techniques focus on scheduling and executing workloads effectively within the provided resource constraints. In this paper, we tackle this problem using another approach. We propose a novel framework D&A to determine the number of cores required in completing a workload under time constraint. We first preprocess a small portion of queries to derive the number of required slots, allowing for the allocation of the remaining workloads into each slot. We introduce a scaling factor in handling the time fluctuation issue caused by random functions. We further establish a lower bound of the number of cores required under this scenario, serving as a baseline for comparison purposes. We examine the framework by computing personalized PageRank values involving intensive computations. Our experimental results show that D&A surpasses the baseline, achieving reductions in the required number of cores ranging from$ 38.89\%$to$ 73.68\%$across benchmark datasets comprising millions of vertices and edges.