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.

Some researchers speculate that intelligent reinforcement learning (RL) agents would be incentivized to seek resources and power in pursuit of the objectives we specify for them. Other researchers point out that RL agents need not have human-like power-seeking instincts. To clarify this discussion, we develop the first formal theory of the statistical tendencies of optimal policies. In the context of Markov decision processes, we prove that certain environmental symmetries are sufficient for optimal policies to tend to seek power over the environment. These symmetries exist in many environments in which the agent can be shut down or destroyed. We prove that in these environments, most reward functions make it optimal to seek power by keeping a range of options available and, when maximizing average reward, by navigating towards larger sets of potential terminal states.

This paper presents a radar sensor package specifically developed for wide-coverage sounding and imaging of polar ice sheets from a variety of aircraft. Our instruments address the need for a reliable remote sensing solution well-suited for extensive surveys at low and high altitudes and capable of making measurements with fine spatial and temporal resolution. The sensor package that we are presenting consists of four primary instruments and ancillary systems with all the associated antennas integrated into the aircraft to maintain aerodynamic performance. The instruments operate simultaneously over different frequency bands within the 160 MHz-18 GHz range. The sensor package has allowed us to sound the most challenging areas of the polar ice sheets, ice sheet margins, and outlet glaciers; to map near-surface internal layers with fine resolution; and to detect the snow-air and snow-ice interfaces of snow cover over sea ice to generate estimates of snow thickness. In this paper, we provide a succinct description of each radar and associated antenna structures and present sample results to document their performance. We also give a brief overview of our field measurement programs and demonstrate the unique capability of the sensor package to perform multifrequency coincidental measurements from a single airborne platform. Finally, we illustrate the relevance of using multispectral radar data as a tool to characterize the entire ice column and to reveal important subglacial features.

Information on the bedrock topography below the Greenland and Antarctic ice sheets is vital to developing models of future sea-level rise. To measure the topography, advanced data acquisition and processing techniques, including Synthetic Aperture Radar (SAR), are required. This work investigates the optimal beamwidth that would enable SAR processing to maximize the signal to noise ratio of the target. Platform height above the ice surface and bedrock roughness determine the optimal beamwidth. We found that for data collected at a “typical” altitude of 867 m, the optimal beamwidth is about 8°. In the high-altitude case, we found that beamwidth did not have a significant effect on the signal-to-noise ratio. This is probably related to scattering from the ice surface.

Sea-level rise will affect populations worldwide with considerable and lasting consequences in the not-too-distant future. Accurate measurement of fast flowing outlet glaciers in Greenland and Antarctica are of vital importance to ice sheet models that predict the course of sea-level rise. The Center for the Remote Sensing of Ice Sheets (CReSIS) has developed a suite of tools designed for use with data collected by CReSIS depth sounding radar platforms. This suite includes algorithms for removing clutter and noise from coherent radar data, and the results show successful sounding of some of these fast-flowing glaciers for the first time.