What latent features are encoded in language model (LM) representations? Recent work on training sparse autoencoders (SAEs) to disentangle interpretable features in LM representations has shown significant promise. However, evaluating the quality of these SAEs is difficult because we lack a ground-truth collection of interpretable features which we expect good SAEs to identify. We thus propose to measure progress in interpretable dictionary learning by working in the setting of LMs trained on Chess and Othello transcripts. These settings carry natural collections of interpretable features—for example, “there is a knight on F3”—which we leverage into metrics for SAE quality. To guide progress in interpretable dictionary learning, we introduce a new SAE training technique, $p$-annealing, which demonstrates improved performance on our metric.

Metaverse can obscure the boundary between the physical and virtual worlds. Specifically, for the human-centric Metaverse in vehicular networks, i.e., the vehicular Metaverse, vehicles are no longer isolated physical spaces but interfaces that extend the virtual worlds to the physical world. Accessing the human-centric Metaverse via autonomous vehicles (AVs), drivers and passengers can immerse in and interact with 3D virtual objects overlaying views of streets on head-up displays (HUD) via augmented reality (AR). The seamless, immersive, and interactive experience rather relies on real-time multi-dimensional data synchronization between physical entities, i.e., AVs, and virtual entities, i.e., Metaverse billboard providers (MBPs). However, mechanisms to allocate and match synchronizing AV and MBP pairs to roadside units (RSUs) in a synchronization service market, which consists of the physical and virtual submarkets, are vulnerable to adverse selection. In this paper, we propose an enhanced second-score auction-based mechanism, named EPViSA, to allocate physical and virtual entities in the synchronization service market of the vehicular Metaverse. The EPViSA mechanism can determine synchronizing AV and MBP pairs simultaneously while protecting participants from adverse selection and thus achieving high total social welfare. We propose a synchronization scoring rule to eliminate the external effects from the virtual submarkets. Then, a price scaling factor is introduced to enhance the allocation of synchronizing virtual entities in the virtual submarkets. Finally, rigorous analysis and extensive experiments demonstrate EPViSA can achieve at least 96% of the social welfare compared to the omniscient benchmark while ensuring strategy-proof and adverse selection free through a simulation testbed.

In this paper we draw on the analysis of sound in film theory in order to explore the potential that sound offers cybercartography. We first argue that the theoretical body developed in film studies is highly relevant to the study of sound/image relationships in mapmaking. We then build on this argument to develop experimental animated and interactive sound maps for the Cybercartographic Atlas of Antarctica that further explore the potential of sound for integrating emotional, cultural and political dimensions in cartography. These maps have been designed to recreate cinematic soundscapes, to provide contrapuntal perspectives on the cartographic image and to generate an aural identity of the atlas. As part of this experimental mapping, an innovative sound infrastructure is being developed to allow complex sound designs to be transmitted over the Internet as part of atlas content. Through this infrastructure the user can select as well as contribute his own sounds. The overall cartographic message is becoming less predictable, thus opening new perspectives on the way we design, interact with, and modify sounded maps over the Internet.