arXiv subjects

Prithviraj Singh Shahani

Publications and source records attributed to Prithviraj Singh Shahani.

3 recordsLinked to original sources

arXiv2025

Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models

Capability evaluations play a crucial role in assessing and regulating frontier AI systems. The effectiveness of these evaluations faces a significant challenge: strategic underperformance, or ``sandbagging'', where models deliberately underperform during evaluation. Sandbagging can manifest either through explicit developer intervention or through unintended model behavior, presenting a fundamental obstacle to accurate capability assessment. We introduce a novel sandbagging detection method based on injecting noise of varying magnitudes into model weights. While non-sandbagging models show predictable performance degradation with increasing noise, we demonstrate that sandbagging models exhibit anomalous performance improvements, likely due to disruption of underperformance mechanisms while core capabilities remain partially intact. Through experiments across various model architectures, sizes, and sandbagging techniques, we establish this distinctive response pattern as a reliable, model-agnostic signal for detecting sandbagging behavior. Importantly, we find noise-injection is capable of eliciting the full performance of Mistral Large 120B in a setting where the model underperforms without being instructed to do so. Our findings provide a practical tool for AI evaluation and oversight, addressing a challenge in ensuring accurate capability assessment of frontier AI systems.

cs.AI↗

arXiv2025

Noise Injection Systemically Degrades Large Language Model Safety Guardrails

Safety guardrails in large language models (LLMs) are a critical component in preventing harmful outputs. Yet, their resilience under perturbation remains poorly understood. In this paper, we investigate the robustness of safety fine-tuning in LLMs by systematically injecting Gaussian noise into model activations. We show across multiple open-weight models that (1) Gaussian noise raises harmful-output rates (p < 0.001) by up to 27%, (2) that deeper safety fine-tuning affords no extra protection, and (3) that chain-of-thought reasoning remains largely intact. The findings reveal critical vulnerabilities in current safety alignment techniques and highlight the potential of reasoning-based and reinforcement learning approaches as promising direction for developing more robust AI safety systems. These results have important implications for real-world deployment of LLMs in safety-critical applications as these results imply that widely-deployed safety tuning methods can fail even without adversarial prompts.

cs.CL↗

arXiv2025

Probing a Vision-Language-Action Model for Symbolic States and Integration into a Cognitive Architecture

Vision-language-action (VLA) models hold promise as generalist robotics solutions by translating visual and linguistic inputs into robot actions, yet they lack reliability due to their black-box nature and sensitivity to environmental changes. In contrast, cognitive architectures (CA) excel in symbolic reasoning and state monitoring but are constrained by rigid predefined execution. This work bridges these approaches by probing OpenVLA's hidden layers to uncover symbolic representations of object properties, relations, and action states, enabling integration with a CA for enhanced interpretability and robustness. Through experiments on LIBERO-spatial pick-and-place tasks, we analyze the encoding of symbolic states across different layers of OpenVLA's Llama backbone. Our probing results show consistently high accuracies (> 0.90) for both object and action states across most layers, though contrary to our hypotheses, we did not observe the expected pattern of object states being encoded earlier than action states. We demonstrate an integrated DIARC-OpenVLA system that leverages these symbolic representations for real-time state monitoring, laying the foundation for more interpretable and reliable robotic manipulation.

cs.RO↗