Supervisors: Prof. Daniel Fried, Prof. Graham Neubig — CMU, LTI
- Improving web agents during inference with auto-eval feedback to deduce granular sub-trajectories and induce memory.
- Studying failure modes across open-source and closed-source model families to tune for specific inabilities.
Supervisor: Prof. Jun-Yan Zhu — CMU, LTI · Report
- Improving binding and context in diffusion models via attention-weighted encoding to improve downstream generations for ambiguous and long-range prompts during inference.
Supervisor: Prof. Graham Neubig — CMU, LTI · Report
- Analyzed the robustness of diverse language models in using relevant information in long context regardless of position, examining the effects of positional encodings, attention, and architecture.
- Empirically demonstrated that most long-context adaptation methods (e.g. RoPE, attention-based retrieval) remain susceptible to the lost-in-the-middle problem.
- Extended this analysis to RepoEval (repository-level code generation), showing a significant drop in function-calling performance as context length increases.
Supervisor: Prof. Maarten Sap — CMU, LTI · Report
- Improving factual correctness in summarization models to mitigate intrinsic hallucination with automatic feedback.
- Developing fine-grained reward models to capture factual contradictions and unsupported spans in generated summaries.
CLSNet Lab, BITS Goa · Supervisor: Prof. Swati Agarwal
- Built classification models to detect hate speech, trolling, and offensive language in tweets.
- Built a federated learning setup to jointly learn from diverse code-switched, code-mixed Hindi-English tweet datasets.
- Built knowledge graphs linking connected concepts in English and Hindi offensive content, enabling better representations via graph embeddings.