Sameer Singh

Language Models Exhibit Systematic Certainty Inflation Bias, Particularly in Rewriting Tasks

This research quantifies "certainty distortion" in language models (LMs), where the expressed certainty of a statement changes despite preserved semantic content. Findings indicate a prevalent bias toward increasing certainty, affecting up to 75% of LM outputs and being systematically asymmetric in rewriting tasks. This distortion can compound with repeated paraphrasing, posing risks in high-stakes domains like scientific and medical communication due to a general inflation of expressed confidence.

Leveraging In-Context Learning for Language Model Agents

Characterizing Mamba's Selective Memory using Auto-Encoders

Parallel Token Prediction for Language Models

Entropy-Aligned Decoding of LMs for Better Writing and Reasoning

Lost in Simulation: LLM-Simulated Users are Unreliable Proxies for Human Users in Agentic Evaluations