Tim

Tim

Computational Linguistics PhD Student at Brandeis University

Posts by Collection

portfolio

projects

Mechanistic Interpretability of Deception in Language Models Trained on Social Deduction Games

Published:

Understanding how language models internally represent deception is critical for AI safety. This work investigates whether models fine-tuned on social deduction game transcripts develop interpretable deception representations amenable to linear probing and causal intervention. I fine-tune Llama 3.1 8B with QLoRA on Werewolf Among Us and SocialMaze data under a joint causal language modeling and binary deception classification objective. Layer-wise probing across five training checkpoints (200–800 steps) reveals that linear probes consistently achieve 94–97% deception detection accuracy against a 50–60% shuffled-label control baseline, confirming a genuine learned representation rather than a high-dimensional artifact. Analysis of the probe accuracy profile across layers shows that the deception feature is nonlinearly encoded at the embedding layer but becomes linearly separable by layer 5, thereafter persisting with near- constant accuracy through all subsequent layers. Contrastive activation steering provides causal evidence that this feature governs model behavior: perturbing activations along the identified direction monotonically shifts the deception probe prediction from 0% to 100%, while probe- direction steering confirms that the learned classifier aligns with a causally relevant subspace of the residual stream.

publications

GLAMR: Augmenting AMR with GL-VerbNet Event Structure

Published in LREC-COLING

This paper proposes GLAMR, an extension of AMR.

Jingxuan Tu, Timothy Obiso, Bingyang Ye, Kyeongmin Rim, Keer Xu, Liulu Yue, Susan Windisch Brown, Martha Palmer, and James Pustejovsky. 2024. GLAMR: Augmenting AMR with GL-VerbNet Event Structure. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 7746–7759, Torino, Italia. ELRA and ICCL.

HaRMoNEE at SemEval-2024 Task 6: Tuning-based Approaches to Hallucination Recognition

Published in Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024)

This paper presents the #1 (winning) and #10 submissions to the two SHROOM tasks at SemEval 2024.

Timothy Obiso, Jingxuan Tu, and James Pustejovsky. 2024. HaRMoNEE at SemEval-2024 Task 6: Tuning-based Approaches to Hallucination Recognition. In Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024), pages 1322–1331, Mexico City, Mexico. Association for Computational Linguistics.

Semantically Enriched Text Generation for QA through Dense Paraphrasing

Published in ICNLSP 2024

Timothy Obiso, Bingyang Ye, Kyeongmin Rim, and James Pustejovsky. 2024. Semantically Enriched Text Generation for QA through Dense Paraphrasing. In Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024), pages 279–286, Trento. Association for Computational Linguistics.

Dynamic Epistemic Friction in Dialogue

Published in CoNLL 2025 (ACL 2025)

Timothy Obiso, Kenneth Lai, Abhijnan Nath, Nikhil Krishnaswamy, and James Pustejovsky. 2025. Dynamic Epistemic Friction in Dialogue. In Proceedings of the 29th Conference on Computational Natural Language Learning (CoNLL 2025), Vienna. Association for Computational Linguistics.

talks

teaching