Aniket Sharma

Caleb DeLeeuw, Gaurav Chawla, Aniket Sharma et al.

We investigate strategic deception in large language models using two complementary testbeds: Secret Agenda (across 38 models) and Insider Trading compliance (via SAE architectures). Secret Agenda reliably induced lying when deception advantaged goal achievement across all model families. Analysis revealed that autolabeled SAE features for "deception" rarely activated during strategic dishonesty, and feature steering experiments across 100+ deception-related features failed to prevent lying. Conversely, insider trading analysis using unlabeled SAE activations separated deceptive versus compliant responses through discriminative patterns in heatmaps and t-SNE visualizations. These findings suggest autolabel-driven interpretability approaches fail to detect or control behavioral deception, while aggregate unlabeled activations provide population-level structure for risk assessment. Results span Llama 8B/70B SAE implementations and GemmaScope under resource constraints, representing preliminary findings that motivate larger studies on feature discovery, labeling methodology, and causal interventions in realistic deception contexts.

Subham Banga, Ujjwal Upadhyay, Piyush Agarwal et al.

Speech emotion analysis is an important task which further enables several application use cases. The non-verbal sounds within speech utterances also play a pivotal role in emotion analysis in speech. Due to the widespread use of smartphones, it becomes viable to analyze speech commands captured using microphones for emotion understanding by utilizing on-device machine learning models. The non-verbal information includes the environment background sounds describing the type of surroundings, current situation and activities being performed. In this work, we consider both verbal (speech commands) and non-verbal sounds (background noises) within an utterance for emotion analysis in real-life scenarios. We create an indigenous dataset for this task namely "Indian EmoSpeech Command Dataset". It contains keywords with diverse emotions and background sounds, presented to explore new challenges in audio analysis. We exhaustively compare with various baseline models for emotion analysis on speech commands on several performance metrics. We demonstrate that we achieve a significant average gain of 3.3% in top-one score over a subset of speech command dataset for keyword spotting.