Papers by Susmit Jha
Task-Agnostic Detector for Insertion-Based Backdoor Attacks (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods for textual backdoor detection are task-specific and less effective beyond sentence classification. |
| Approach: | They propose a task-agnostic method for backdoor detection that leverages final layer logits and an efficient pooling technique. |
| Outcome: | TABDet can jointly learn from diverse task-specific models, demonstrating superior detection efficacy over traditional methods. |
NSF-CoT: Neuro-Symbolic Formal Verification of Chain-of-Thought Faithfulness in Contextual Question Answering (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing faithfulness checks only test whether changing the written chain changes the answer, without verifying whether the steps are truly supported by the given evidence. |
| Approach: | They propose a method that checks CoT faithfulness step by step for contextual question answering. |
| Outcome: | The proposed method outperforms causal mediation, perturbation probes, and behavioral monitoring in OpenBookQA, QASC, and HotpotQA. |
Polysemantic Dropout: Conformal OOD Detection for Specialized LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | specialized large language models (LLMs) are capable of generating coherent, contextually relevant and often human-level language, but they remain vulnerable to errors when confronted with data that falls outside the scope of their domain. |
| Approach: | They propose an inference-time out-of-domain (OOD) detection algorithm that leverages the Inductive Conformal Anomaly Detection framework and a new non-conformity measure based on the model’s dropout tolerance. |
| Outcome: | The proposed method improves detection while maintaining theoretical false alarm bounds from ICAD while maintaining false alarms bounds. |