publications

2026

1. PFW at SemEval-2026 Task 6: Multi-Seed DeBERTa Ensembles for Political Response Clarity and Evasion Classification

   Taleef Tamsal and Jonathan Rusert

   In Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026), 2026

   Abs PDF Code

   A multi-seed DeBERTa ensemble for political response clarity and evasion classification. The system ranked 18/41 on the clarity subtask and 12/33 on the evasion subtask, while analysis showed that seed diversity transferred more reliably than post-hoc calibration under limited evaluation data.

2. PFW Task 8 at SemEval-2026 Task 8: Lightweight Tri-Fusion Retrieval with Prompt-Engineered Faithful Generation for Multi-Turn RAG

   Taleef Tamsal and Jonathan Rusert

   In Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026), 2026

   Abs PDF Code

   A lightweight multi-turn RAG system combining tri-fusion retrieval with prompt-engineered faithful generation. The system ranked 20/38 on Task A, 6/26 on Task B, and 14/29 on Task C.

3. Do Efficient Adaptations Reduce Safety? Jailbreak Robustness of PEFT vs. Full Fine-Tuning on Consumer-Accessible LLMs

   Taleef Tamsal

   Purdue University Fort Wayne, 2026

   Abs PDF Code

   A systematic study of 25 model configurations evaluated against four jailbreak attacks and two safety benchmarks, producing nearly 600 attack-success-rate measurements calibrated against a 750-sample human annotation study. The thesis examines how full fine-tuning, LoRA, QLoRA, and quantization change black-box safety and how evaluator choice affects conclusions.

4. Your Judge Is a Confound: Evaluator and Attack Choice Distort Jailbreak-Safety Measurement for Fine-Tuned LLMs

   Taleef Tamsal and Jonathan Rusert

   In AdvML-Frontiers × CoTMA Workshop at COLM 2026, 2026

   Accepted

   Abs

   An empirical study showing how evaluator and attack choice can distort jailbreak-safety measurement for fine-tuned language models.