Ganesh Gopalakrishnan

Harvey Dam, Jonas Knochelmann, Vinu Joseph et al.

We introduce a method to reduce refusal rates of large language models (LLMs) on sensitive content without modifying model weights or prompts. Motivated by the observation that refusals in certain models were often preceded by the specific token sequence of a token marking the beginning of the chain-of-thought (CoT) block (<think>) followed by a double newline token (\n\n), we investigate the impact of two simple formatting adjustments during generation: suppressing \n\n after <think> and suppressing the end-of-sequence token after the end of the CoT block (</think>). Our method requires no datasets, parameter changes, or training, relying solely on modifying token probabilities during generation. In our experiments with official DeepSeek-R1 distillations, these interventions increased the proportion of substantive answers to sensitive prompts without affecting performance on standard benchmarks. Our findings suggest that refusal behaviors can be circumvented by blocking refusal subspaces at specific points in the generation process.

Harvey Dam, Tripti Agarwal, Ganesh Gopalakrishnan

Preserving topological features in learned latent spaces is a fundamental challenge in representation learning, particularly for topology-sensitive data. This paper introduces directional sign loss (DSL), an efficient, differentiable loss function that approximates the number of mismatches in the signs of finite differences between corresponding elements of two arrays. By penalizing discrepancies in critical points between input and reconstructed data, DSL encourages autoencoders and other learnable compressors to retain the topological features of the original data. We present the formulation and complexity analysis of DSL, comparing it to other non-differentiable topological measures. Experiments on multidimensional array data show that combining DSL with traditional loss functions preserves topological features more effectively than traditional losses alone. DSL serves as a differentiable, efficient proxy for common topology-based metrics, enabling topological feature preservation on previously impractical problem sizes and in a wider range of gradient-based optimization frameworks.

Tripti Agarwal, Harvey Dam, Dorra Ben Khalifa et al.

In response to the rapidly escalating costs of computing with large matrices and tensors caused by data movement, several lossy compression methods have been developed to significantly reduce data volumes. Unfortunately, all these methods require the data to be decompressed before further computations are done. In this work, we develop a lossy compressor that allows a dozen fairly fundamental operations directly on compressed data while offering good compression ratios and modest errors. We implement a new compressor PyBlaz based on the familiar GPU-powered PyTorch framework, and evaluate it on three non-trivial applications, choosing different number systems for internal representation. Our results demonstrate that the compressed-domain operations achieve good scalability with problem sizes while incurring errors well within acceptable limits. To our best knowledge, this is the first such lossy compressor that supports compressed-domain operations while achieving acceptable performance as well as error.

Mohammed S. Al-Mahfoudh, Ryan Stutsman, Ganesh Gopalakrishnan

Recent demand for distributed software had led to a surge in popularity in actor-based frameworks. However, even with the stylized message passing model of actors, writing correct distributed software is still difficult. We present our work on linearizability checking in DS2, an integrated framework for specifying, synthesizing, and testing distributed actor systems. The key insight of our approach is that often subcomponents of distributed actor systems represent common algorithms or data structures (e.g.\ a distributed hash table or tree) that can be validated against a simple sequential model of the system. This makes it easy for developers to validate their concurrent actor systems without complex specifications. DS2 automatically explores the concurrent schedules that system could arrive at, and it compares observed output of the system to ensure it is equivalent to what the sequential implementation could have produced. We describe DS2's linearizability checking and test it on several concurrent replication algorithms from the literature. We explore in detail how different algorithms for enumerating the model schedule space fare in finding bugs in actor systems, and we present our own refinements on algorithms for exploring actor system schedules that we show are effective in finding bugs.

Vinu Joseph, Shoaib Ahmed Siddiqui, Aditya Bhaskara et al.

With the rise in edge-computing devices, there has been an increasing demand to deploy energy and resource-efficient models. A large body of research has been devoted to developing methods that can reduce the size of the model considerably without affecting the standard metrics such as top-1 accuracy. However, these pruning approaches tend to result in a significant mismatch in other metrics such as fairness across classes and explainability. To combat such misalignment, we propose a novel multi-part loss function inspired by the knowledge-distillation literature. Through extensive experiments, we demonstrate the effectiveness of our approach across different compression algorithms, architectures, tasks as well as datasets. In particular, we obtain up to $4.1\times$ reduction in the number of prediction mismatches between the compressed and reference models, and up to $5.7\times$ in cases where the reference model makes the correct prediction; all while making no changes to the compression algorithm, and minor modifications to the loss function. Furthermore, we demonstrate how inducing simple alignment between the predictions of the models naturally improves the alignment on other metrics including fairness and attributions. Our framework can thus serve as a simple plug-and-play component for compression algorithms in the future.