Publications

Dhananjay Ashok Ashutosh Chaubey, Hirona J. Arai, Jonathan May, Jesse Thomason

TLDR: While VLMs can recall factual associations when provided a textual reference to an entity, their ability to do so is halved when the reference is visual instead.

Dhananjay Ashok, Jonathan May

TLDR: Performance declines associated with replacing human generated data with synthetic data is most chronic only after crossing 90% replacement.

Dhananjay Ashok, Jonathan May

TLDR: Conformal probes trained on the hidden states of LMs can preemptively detect a variety of downstream behaviors, including susceptibility to jailbreaking, low confidence responses and alignment failures.

Dhananjay Ashok, Barnabas Poczos

TLDR: We show that prior methods for controlling text generation of base Language Models perform worse than Instruction-Tuning. We also release ConGenBench, a testbed of more difficult controllable text generation problems.

Dhananjay Ashok, Atharva Kulkarni, Hai Pham, Barnabas Poczos

TLDR: We combine synthetic data generation and score guided decoding to outperform GPT3 on Scientific Factual Error Correction.

Matthew Barker, Emma Kallina, Dhananjay Ashok, Katherine Collins, Ashley Casovan, Adrian Weller, Ameet Talwalkar, Valerie Chen, Umang Bhatt

TLDR: Introduces FeedbackLogs, an addenda to existing documentation of ML pipelines that tracks the input of multiple stakeholders.

Dhananjay Ashok, Zachary Chase Lipton

TLDR: We set the state-of-the-art in several FewShot and CrossDomain NER benchmarks with a Prompting approach.

Dhananjay Ashok, Vineel Nagisetty, Christopher Srinivasa, Vijay Ganesh

TLDR: Using Mixed Integer Linear Programming Solvers to fine-tune Neural Networks and escape local minima during optimization.

Dhananjay Ashok, Joseph Scott, Sebastian J Wetzel, Maysum Panju, Vijay Ganesh

TLDR: A data augmentation approach that uses prior knowledge to accelerate equation discovery.