Somshubhra Roy

Chiung-Yi Tseng, Somshubhra Roy, Maisha Thasin et al.

There is a substantial body of literature examining the mathematical reasoning capabilities of large language models (LLMs), particularly their performance on precise arithmetic operations in autoregressive architectures. However, their ability to perform approximate reasoning in informal, fast-paced mathematical operations has received far less attention, especially among non-autoregressive decoder models. Our work addresses this gap by introducing StreetMath, a benchmark designed to evaluate models' approximation abilities under real-world approximation scenarios. We conduct extensive evaluations across different LLM architectures: Qwen3-4B-Instruct-2507, Qwen3-4B-Thinking-2507, Dream-v0-Instruct-7B, Falcon-Mamba-7B-Instruct, and Mamba-GPT-3B. Furthermore, we apply mechanistic interpretability techniques to probe their internal computational states. Our analysis reveals that LLMs generally attempt to compute exact values or invoke external tools even in tasks that call for approximation. Moreover, while models sometimes reach the correct answer in early layers or steps, they still consume more tokens when solving approximation tasks. Additional experiments indicate that exact and approximate arithmetic operations rely on largely separate neural components. Drawing upon research on cognitive psychology, we argue that LLMs do not exhibit cognitive miserliness in the same way humans do in street math settings. We open source our work https://github.com/ctseng777/StreetMath

Effiong Blessing, Chiung-Yi Tseng, Somshubhra Roy et al.

Memory-augmented spiking neural networks (SNNs) promise energy-efficient neuromorphic computing, yet their generalization across sensory modalities remains unexplored. We present the first comprehensive cross-modal ablation study of memory mechanisms in SNNs, evaluating Hopfield networks, Hierarchical Gated Recurrent Networks (HGRNs), and supervised contrastive learning (SCL) across visual (N-MNIST) and auditory (SHD) neuromorphic datasets. Our systematic evaluation of five architectures reveals striking modality-dependent performance patterns: Hopfield networks achieve 97.68% accuracy on visual tasks but only 76.15% on auditory tasks (21.53 point gap), revealing severe modality-specific specialization, while SCL demonstrates more balanced cross-modal performance (96.72% visual, 82.16% audio, 14.56 point gap). These findings establish that memory mechanisms exhibit task-specific benefits rather than universal applicability. Joint multi-modal training with HGRN achieves 94.41% visual and 79.37% audio accuracy (88.78% average), matching parallel HGRN performance through unified deployment. Quantitative engram analysis confirms weak cross-modal alignment (0.038 similarity), validating our parallel architecture design. Our work provides the first empirical evidence for modality-specific memory optimization in neuromorphic systems, achieving 603x energy efficiency over traditional neural networks.

Elias Hossain, Swayamjit Saha, Somshubhra Roy et al.

Even when prompts and parameters are secured, transformer language models remain vulnerable because their key-value (KV) cache during inference constitutes an overlooked attack surface. This paper introduces Malicious Token Injection (MTI), a modular framework that systematically perturbs cached key vectors at selected layers and timesteps through controlled magnitude and frequency, using additive Gaussian noise, zeroing, and orthogonal rotations. A theoretical analysis quantifies how these perturbations propagate through attention, linking logit deviations to the Frobenius norm of corruption and softmax Lipschitz dynamics. Empirical results show that MTI significantly alters next-token distributions and downstream task performance across GPT-2 and LLaMA-2/7B, as well as destabilizes retrieval-augmented and agentic reasoning pipelines. These findings identify cache integrity as a critical yet underexplored vulnerability in current LLM deployments, positioning cache corruption as a reproducible and theoretically grounded threat model for future robustness and security research.