Steering Autoregressive Music Generation with Recursive Feature Machines

Controllable music generation remains a significant challenge, with existing methods often requiring model retraining or introducing audible artifacts. We introduce MusicRFM, a framework that adapts Recursive Feature Machines (RFMs) to enable fine-grained, interpretable control over frozen, pre-trained music models by directly steering their internal activations. RFMs analyze a model's internal gradients to produce interpretable "concept directions", or specific axes in the activation space that correspond to musical attributes like notes or chords. We first train lightweight RFM probes to discover these directions within MusicGen's hidden states; then, during inference, we inject them back into the model to guide the generation process in real-time without per-step optimization. We present advanced mechanisms for this control, including dynamic, time-varying schedules and methods for the simultaneous enforcement of multiple musical properties. Our method successfully navigates the trade-off between control and generation quality: we can increase the accuracy of generating a target musical note from 0.23 to 0.82, while text prompt adherence remains within approximately 0.02 of the unsteered baseline, demonstrating effective control with minimal impact on prompt fidelity. We release code to encourage further exploration on RFMs in the music domain.

Towards Interpretable and Inference-Optimal COT Reasoning with Sparse Autoencoder-Guided Generation

We propose a novel method that leverages sparse autoencoders (SAEs) and clustering techniques to analyze the internal token representations of large language models (LLMs) and guide generations in mathematical reasoning tasks. Our approach first trains an SAE to generate sparse vector representations for training tokens, then applies k-means clustering to construct a graph where vertices represent token clusters and weighted edges capture sequential token transitions. Using this graph, we define an edge-weight based reward function to quantify adherence to established reasoning traces, thereby identifying exploitative reasoning trajectories. Additionally, we measure generation diversity from clustering to assess the extent of exploration. Our findings indicate that balancing both exploitation and exploration is crucial for achieving high accuracy in mathematical reasoning tasks. During generation, the SAE can serve as a scalable reward model to guide generations, ensuring a balanced trade-off between exploitation and exploration. This prevents extreme behaviors in either direction, ultimately fostering a higher-quality reasoning process in LLMs.

Policy Gradients for Optimal Parallel Tempering MCMC

Parallel tempering is meta-algorithm for Markov Chain Monte Carlo that uses multiple chains to sample from tempered versions of the target distribution, enhancing mixing in multi-modal distributions that are challenging for traditional methods. The effectiveness of parallel tempering is heavily influenced by the selection of chain temperatures. Here, we present an adaptive temperature selection algorithm that dynamically adjusts temperatures during sampling using a policy gradient approach. Experiments demonstrate that our method can achieve lower integrated autocorrelation times compared to traditional geometrically spaced temperatures and uniform acceptance rate schemes on benchmark distributions.

Mapping "Brain Coral" Regions on Mars using Deep Learning

One of the main objectives of the Mars Exploration Program is to search for evidence of past or current life on the planet. To achieve this, Mars exploration has been focusing on regions that may have liquid or frozen water. A set of critical areas may have seen cycles of ice thawing in the relatively recent past in response to periodic changes in the obliquity of Mars. In this work, we use convolutional neural networks to detect surface regions containing "Brain Coral" terrain, a landform on Mars whose similarity in morphology and scale to sorted stone circles on Earth suggests that it may have formed as a consequence of freeze/thaw cycles. We use large images (~100-1000 megapixels) from the Mars Reconnaissance Orbiter to search for these landforms at resolutions close to a few tens of centimeters per pixel (~25--50 cm). Over 52,000 images (~28 TB) were searched (~5% of the Martian surface) where we found detections in over 200 images. To expedite the processing we leverage a classifier network (prior to segmentation) in the Fourier domain that can take advantage of JPEG compression by leveraging blocks of coefficients from a discrete cosine transform in lieu of decoding the entire image at the full spatial resolution. The hybrid pipeline approach maintains ~93% accuracy while cutting down on ~95% of the total processing time compared to running the segmentation network at the full resolution on every image. The timely processing of big data sets helps inform mission operations, geologic surveys to prioritize candidate landing sites, avoid hazardous areas, or map the spatial extent of certain terrain. The segmentation masks and source code are available on Github for the community to explore and build upon.