Prakhar Gupta

Intern @ Databricks (Genie AI)
BSE/MSE CSE @ University of Michigan, Ann Arbor
AI Safety + Interpretability

I'm fortunate to have worked with and been mentored by Prof. Sarah-Jane Leslie (Princeton), Dr. Andrew Lee (Harvard), Dr. Henry Conklin (Princeton), Prof. Zhijing Jin (UToronto), Prof. Bernhard Schölkopf (MPI-IS), David Africa (UK AISI), Dr. Sohaib Imran (Lancaster), Prof. Yossi Gandelsman (OpenAI, TTIC), Jess Bergs (UK AISI), Prof. Amrita Roy Chowdhury (UMich), and Prof. Lin Ma (UMich).

I'm co-organizing the Trustworthy AI for Good workshop at ICML & NeurIPS 2026! (We're the largest workshop at both ICML & NeurIPS Paris!)

Publications

• Better World Models Can Lead to Better Post-Training Performance
  Prakhar Gupta, Henry Conklin, Sarah-Jane Leslie, and Andrew Lee
  EMNLP 2026 Findings, NeurIPS 2025 Workshop on Mechanistic Interpretability (Spotlight)
• How Does Alignment Tuning Shape Representations of Sycophancy and Related Cue-Induced Biases in LLMs?
  Prakhar Gupta, Terry Jingchen Zhang, Florent Draye, Bernhard Schölkopf, and Zhijing Jin
  EMNLP 2026 Findings, EMNLP 2026 Workshop on NLP for Positive Impact
• Many Biases, One Circuit: Post-Training Tunes a Pretrained Cue-Routing Mechanism
  Prakhar Gupta, Terry Jingchen Zhang, Florent Draye, Bernhard Schölkopf, and Zhijing Jin
  Preprint (under review), NeurIPS 2026 Workshop on Neural Network Artifacts as a New Data Modality (Spotlight)
• Consistency Training Suppresses Cue Following Without Teaching Cue Selectivity
  Prakhar Gupta, Terry Jingchen Zhang, and Zhijing Jin
  Preprint (under review), 2026
• Ask the Model: Counterfactual Cue Reliance Predicts Worst-Group Failure and Targets Repair
  Prakhar Gupta, Terry Jingchen Zhang, Florent Draye, Bernhard Schölkopf, and Zhijing Jin
  Preprint (under review), NeurIPS 2026 Workshop on Trustworthy AI Evaluation
• Context-Grounding Gains Are Mediated by Pre-existing Machinery: Auditing GRPO, SFT, and DPO
  Prakhar Gupta and Vaibhav Gupta
  Preprint (under review), NeurIPS 2026 Workshop on the Linguistic Principles for Foundation Models (Spotlight)
• Consistency Training Along the Transformer Stack
  Sukrati Gautam*, Neil Shah*, Arav Dhoot*, Bryan Maruyama*, Caroline Wei*, Rohan Kapoor, Robert Sidey, Prakhar Gupta, Zi Cheng Huang, and David Demitri Africa
  EMNLP 2026 Findings, ICML 2026 Workshop on Trustworthy AI for Good
• Consistency Training while Mitigating Obfuscation via Rate Matching
  Sohaib Imran, Prakhar Gupta, Jannes Elstner, and David Africa
  Preprint (under review), 2026
• Last Translation Benchmark
  Vilém Zouhar, Niyati Bafna, Mukund Choudhary, Maike Züfle, ..., Prakhar Gupta, et al.
  Preprint, 2026
• Self-Mined Hardness for Safety Fine-Tuning
  Prakhar Gupta, Garv Shah, and Donghua Zhang
  ICML 2026 Workshop on Failure Modes of Agentic AI
• Sparse Memory Finetuning as a Low-Forgetting Alternative to LoRA and Full Finetuning
  Prakhar Gupta*, Garv Shah*, Satyam Goyal*, and Anirudh Kanchi*
  ICML 2026 Workshop on Foundations of Deep Generative Models
• Bootstrapped Mixed Rewards for RL Post-Training: Injecting Canonical Action Order
  Prakhar Gupta and Vaibhav Gupta
  ICLR 2026 Workshop on Logical Reasoning of Large Language Models