Research
Some of my published and working papers.
Preprint / Published
- When Attention Repairs Itself: A Mechanistic Study of Backup Heads in GPT-2 This study introduces Frozen Attention Logit Intervention (FALI) to isolate whether backup attention heads in GPT-2 Small actively compensate after primary-head ablation, or whether observed Hydra Effect spikes are driven by LayerNorm scaling. On 200 Indirect Object Identification prompts, ablating L9H9 and L10H0 produces a robust, target-directed increase in L10H10 IO-directed attention that survives multiple ablation controls and causal mediation tests. Read the paper here !
Working papers
- The Sunglass Paradox: The Play of Perceptions in Consumer-Salesperson Dynamics This conceptual paper introduces the “Sunglass Paradox,” explaining how sunglasses can simultaneously enhance perceived confidence while weakening trust by obstructing eye contact in sales interactions.
- Just the Right Amount of Feelings: How Mixed Feelings Make Consumer Advice Most Helpful and When Credibility Amplifies the Effect This research shows that online reviews are most helpful when they contain a moderate mix of emotions, and that reviewer credibility strengthens consumers’ willingness to process emotionally nuanced advice.