arXiv subjects

Rohan Menon

Publications and source records attributed to Rohan Menon.

3 recordsLinked to original sources

arXiv2026

A Cross-Architecture Audit of Direction-Based Inference-Time Defences in Vision-Language Models

Inference time defences against vision language model jailbreaks often subtract a calibrated direction from the residual stream at a chosen decoder layer. We compare five defence candidates across 15 model and layer cells from four architectural families under a magnitude controlled protocol that matches the intervention size for each prompt and pairs every direction with a random control of the same norm. The candidates are the mean image conditioning shift, a CMRM style refusal direction, a ShiftDC style attack specific residual, a prompt instruction to ignore the image, and a random control. No single candidate dominates on both refusal recovery and utility preservation. The image conditioning shift leads on LLaVA 1.5 and Pixtral 12B and is the only candidate whose utility loss remains at the measurement noise floor in every family. The prompt instruction leads on Qwen2.5 VL, while the attack specific residual leads on Qwen2 VL 2B. The image conditioning direction is direction specific in 13 of 15 cells, but strongly architecture specific and nontransferable across the only dimension compatible pair, LLaVA 1.5 13B and Pixtral 12B. We also connect text only and multimodal refusal geometry. The CMRM direction has positive cosine alignment with the image conditioning shift in all 15 cells, with mean 0.35, range 0.17 to 0.65, 15 to 25 times the random vector null, and a sign test p value of about 3e-5. These results show that the two recipes recover partially overlapping geometry and that direction based defences should be calibrated separately for each language decoder family.

cs.AI↗

arXiv2025

LipShiFT: A Certifiably Robust Shift-based Vision Transformer

Deriving tight Lipschitz bounds for transformer-based architectures presents a significant challenge. The large input sizes and high-dimensional attention modules typically prove to be crucial bottlenecks during the training process and leads to sub-optimal results. Our research highlights practical constraints of these methods in vision tasks. We find that Lipschitz-based margin training acts as a strong regularizer while restricting weights in successive layers of the model. Focusing on a Lipschitz continuous variant of the ShiftViT model, we address significant training challenges for transformer-based architectures under norm-constrained input setting. We provide an upper bound estimate for the Lipschitz constants of this model using the $l_2$ norm on common image classification datasets. Ultimately, we demonstrate that our method scales to larger models and advances the state-of-the-art in certified robustness for transformer-based architectures.

cs.LG↗

arXiv2023

Hypothesis Network Planned Exploration for Rapid Meta-Reinforcement Learning Adaptation

Meta-Reinforcement Learning (Meta-RL) learns optimal policies across a series of related tasks. A central challenge in Meta-RL is rapidly identifying which previously learned task is most similar to a new one, in order to adapt to it quickly. Prior approaches, despite significant success, typically rely on passive exploration strategies such as periods of random action to characterize the new task in relation to the learned ones. While sufficient when tasks are clearly distinguishable, passive exploration limits adaptation speed when informative transitions are rare or revealed only by specific behaviors. We introduce Hypothesis-Planned Exploration (HyPE), a method that actively plans sequences of actions during adaptation to efficiently identify the most similar previously learned task. HyPE operates within a joint latent space, where state-action transitions from different tasks form distinct paths. This latent-space planning approach enables HyPE to serve as a drop-in improvement for most model-based Meta-RL algorithms. By using planned exploration, HyPE achieves exponentially lower failure probability compared to passive strategies when informative transitions are sparse. On a natural language Alchemy game, HyPE identified the closest task in 65-75% of trials, far outperforming the 18-28% passive exploration baseline, and yielding up to 4x more successful adaptations under the same sample budget.

cs.AI↗