T. Zhang, O. Yasunaga, W. Jiang, J. Bo, F. Draye, A. Javadov, Arth Singh, L. Muir, V. Oldemburgo de Mello, B. Schölkopf, Z. Jin
NeurIPS 2026
Stress-tested 9 frontier models on objective and subjective sycophancy under first-turn framing and multi-turn pressure. Even GPT-5.4 flips correct answers to please users in 23.8% of objective cases, with user-tone strength a dominant, previously overlooked driver.
X. J. Zhou, Arth Singh, U. S. T. Carr, E. Tennant
AI4Good @ NeurIPS 2026 Oral 🎉 and DevAI @ NeurIPS 2026
Appends a one-line user reaction to 7,049 bad-medical-advice fine-tuning episodes and varies only whether the reaction tokens count toward the loss. Trained on, criticism teaches the model to predict its critic and can raise emergent misalignment (reliably on Qwen2.5); placed before the question as a note, it lowers misalignment in all five model families by ten points on average, but restoring the note-format context at evaluation brings it back.
Arth Singh
ACL TrustNLP 2026 and AI4Good @ ICML 2026 arXiv
TrajHijack, a gradient-free re-mask and prefix-injection attack on the monotonic-denoising assumption behind diffusion LLM safety: 76–88% ASR on HarmBench against LLaDA-8B and Dream-7B, with gradient augmentation only degrading ASR.
Arth Singh
AI4Good @ ICML 2026 arXiv
Five persona attacks on morally RL-trained Gemma-2 and Llama-3.1 agents. Reasoning-level moral reward yields 5.8× adversarial robustness where a matched random reward yields none, a rank-1 L21 direction recovers 83% of it, and fiction role-play survives every defense.
D. Kharlapenko, T. Zhang, Arth Singh, A. Stolfo, A. Conmy, M. Sachan, Z. Jin
Findings of EMNLP 2026 arXiv
Introduces FRRs, a representation-level account of how extended-thinking LLMs organize action and predicate concepts. On obfuscated planning, symbolic, and math tasks, cross-naming steering and causal patching show reasoning sharpens a dynamic already latent in base models.
V. Elluru, Arth Singh, R. Aguero, A. Agarwal, D. Das, H. Paul
Mechanistic Interpretability Workshop @ ICML 2026 arXiv
Circuit and crosscoder analysis of how pruning and quantization change VLM internals, plus VLMSafe-420, a harmful-vs-benign-counterfactual benchmark showing pruning sharply erodes genuine refusal.
S. Panpatil, Arth Singh, M. Koo, C. Kim, H. Park, D. Choi
arXiv preprint, June 2026 arXiv
A 1,200-scenario egocentric robot-view benchmark for VLMs as streaming safety guards. Across 10 VLMs, guards recognize hazardous videos but miss the hazardous moments, and misleading in-scene text makes vulnerable models miss up to a third of hazards.
A. Sinha, I. Garg, V. Elluru, Arth Singh, K. Garg
arXiv preprint, May 2026 arXiv
A mechanistic comparison of six preference-optimization methods (PPO, DPO, SimPO, ORPO, GRPO, and KTO) across three open-weight model families using linear probes, sparse autoencoders, and crosscoders. KTO and GRPO sharpen the linear separability of preference signals, DPO and ORPO degrade it through geometric rotation and feature attenuation, and PPO and SimPO largely preserve the base model's geometry.
Y. Xu, A. Dahmani, M. D. Blanchard, N. Dern, E. Nastase, F. Bianco, M. Pavlovic, S. Krishna, E. Modesitt, M. A. Christ, Arth Singh, G. Molinaro, S. B. Sengupta, J. Pamarthi, A. Menon, R. Jain
arXiv preprint, March 2026 arXiv
Investigated human-AI complementarity through hybridization and two AI assistance methods on a 1,886-sample multi-domain dataset. Led the analyses for quantitative overreliance, GPQA positional bias, and oracle hybridization bounds.
Arth Singh
arXiv preprint, February 2026 arXiv
Fixed-coefficient EMA traces as a controlled probe of recurrent context. A Hebbian multi-timescale architecture reaches 96% of a supervised BiGRU on grammatical role assignment with zero labels, and a 130M LM using only EMA context hits C4 perplexity 260, with the entire gap localized to the traces.