# RESEARCHER_PROFILE

name:        "Maheep Chaudhary"
role:        "Machine Learning Engineer"
affiliation: "Algoverse AI Research" [link](https://algoverseairesearch.org)
email:       maheepchaudhary.research@gmail.com
links:       [GitHub](https://github.com/MaheepChaudhary) [Twitter](https://twitter.com/MaheepChaudhary) [Scholar](https://scholar.google.com/citations?user=YOUR_ID)

## EDUCATION

- Aug 2023 — May 2024
  Nanyang Technological University, Singapore
  M.Sc. Artificial Intelligence, Computer Science

- Aug 2018 — May 2022
  Bundelkhand Institute of Engineering and Technology, India
  B.E. Electronics and Communication Engineering

## EXPERIENCE

- Jun 2025 — Present  Algoverse
  → Machine Learning Engineer

- Jan 2025 — May 2025  AI Safety Camp
  → Machine Learning Researcher

- Nov 2024 — May 2025  University of Oxford
  → Research Intern (Supervisor: Fazl Barez)

- Jul 2024 — Sep 2024  WhiteBox Research
  → Research Mentor

- Oct 2023 — Sep 2024  Pr(Ai)²R Group, Stanford
  → Alignment Research Engineer (Supervisor: Atticus Geiger)

- Sep 2021 — Jan 2024  UIUC
  → Research Collaboration (Supervisor: Haohan Wang)

- Oct 2019 — Mar 2021  IIT Indore
  → Research Collaboration (Supervisor: Chandresh Maurya)

## RESEARCH_TAXONOMY

primary_focus:
  [AI Safety] [Mechanistic Interpretability]

secondary:
  [Chain-of-Thought] [Causal Abstraction]
  [Deception Detection] [Attention Mechanisms]

methods:
  [White-box Analysis] [Activation Steering]
  [Probing] [Anomaly Detection] [Causal Intervention]

models:
  [Llama] [Gemma] [Qwen] [Mistral] [GPT-2]

## METRICS

journal_conference:  5
workshop_papers:     7
preprints:           6
as_project_lead:     8

## PUBLICATIONS

### Journal & Conference

1. "Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability"
   A. Geiger, D. Ibeling, A. Zur, Maheep Chaudhary, et al.
   [JMLR 2024] [paper](http://jmlr.org/papers/v26/23-0058.html)

2. "MemeCLIP: Leveraging CLIP Representations for Multimodal Meme Classification"
   S. B. Shah, S. Shiwakoti, Maheep Chaudhary, H. Wang
   [EMNLP 2024] [paper](https://aclanthology.org/2024.emnlp-main.959/)

3. "An Intelligent Recommendation cum Reminder System"
   R. Saxena, Maheep Chaudhary, C.K. Maurya, S. Prasad
   [ACM CODS-COMAD 2022]

4. "CQFaRAD: Collaborative Query-Answering Framework"
   M. Singh, S. Pandey, R. Saxena, Maheep Chaudhary, N. Lal
   [Springer IJIT 2021]

### Workshop Papers (NeurIPS 2025)

1. "Evaluation Awareness Scales Predictably in Open-Weights LLMs"
   Maheep Chaudhary†, I. Su, N. Hooda, et al.
   [Responsible FM] LEAD

2. "FRIT: Using Causal Importance to Improve CoT Faithfulness"
   A. Swaroop, et al., Maheep Chaudhary†
   [FoRLM] LEAD

3. "Optimizing CoT Confidence via Topological and Dirichlet Risk Analysis"
   A. More, et al., Maheep Chaudhary†
   [Responsible FM] LEAD

4. "Alignment-Constrained Dynamic Pruning for LLMs"
   D. Patel, et al., Maheep Chaudhary†
   [Responsible FM] LEAD

5. "SALT: Steering Activations towards Leakage-free Thinking"
   S. Batra, et al., Maheep Chaudhary†
   [Responsible FM] LEAD

6. "Amortized Latent Steering: Low-Cost Test-Time Optimization"
   N. Egbuna, et al., Maheep Chaudhary†
   [Efficient Reasoning] LEAD

7. "Hydra: A Modular Architecture for Efficient Long-Context Reasoning"
   S. Chaudhary, D. Patel, Maheep Chaudhary, B. Browning
   [Efficient Reasoning]

### Preprints

1. "SafetyNet: Detecting Harmful Outputs in LLMs"
   Maheep Chaudhary, F. Barez
   LEAD [arXiv](https://arxiv.org/abs/2505.14300)

2. "Evaluating Sparse Autoencoders on Disentangling Factual Knowledge"
   Maheep Chaudhary, A. Geiger
   LEAD [arXiv](https://arxiv.org/abs/2409.04478)

3. "PALADIN: Self-Correcting LM Agents to Cure Tool-Failure Cases"
   S. V. Vuddanti, et al., Maheep Chaudhary†
   LEAD

4. "Modular Training of Neural Networks aids Interpretability"
   S. Golechha, Maheep Chaudhary, et al.
   [arXiv](https://arxiv.org/abs/2502.02470)

5. "Punctuation and Predicates in Language Models"
   S. Chauhan, Maheep Chaudhary, et al.
   [arXiv](https://arxiv.org/abs/2504.05110)

6. "Towards Trustworthy ML: A Data-centric Survey with Causality Perspectives"
   Maheep Chaudhary*, H. Liu*, H. Wang
   Co-first [arXiv](https://arxiv.org/abs/2307.16851)

## KEY_FINDINGS

[1] Evaluation awareness scales with model size
    → standard eval may fail for AGI-level systems

[2] Harmful content produces distinct attention signatures
    → ~95% detection accuracy

[3] Information flows left→right to last token
    → punctuation/predicates serve as storage anchors

[4] Obfuscation shifts info from non-linear to linear spaces
    → white-box monitoring can detect this

## SERVICE

reviewer:
  - ICML 2025 (Actionable Interpretability)
  - NeurIPS 2024 (MATH+AI, CALM)

mentoring:
  - UNESCO-India-Africa Program (20+ countries)
  - Research Mentor at Algoverse

## AWARDS

- Winner — Smart India Hackathon (200K+ participants)
  Face recognition for Bureau of Police R&D

- Team Leader — ASEAN-India Hackathon (10+ countries)
  Marine Species Detection

- Selected Mentor — UNESCO-India-Africa Program (20+ countries)
  Voice-assisted system for farmers

## CONTACT

email:   maheepchaudhary.research@gmail.com
github:  [GitHub](https://github.com/MaheepChaudhary)
twitter: [Twitter](https://twitter.com/MaheepChaudhary)
scholar: [Google Scholar](https://scholar.google.com/citations?user=YOUR_ID)

## STRUCTURED_DATA

{
  "@type": "Person",
  "name": "Maheep Chaudhary",
  "role": "Machine Learning Engineer",
  "affiliation": "Algoverse AI Research",
  "education": ["M.Sc. AI (NTU)", "B.E. ECE (BIET)"],
  "research_focus": ["AI Safety", "Mechanistic Interpretability"],
  "publications": {
    "journal_conference": 5,
    "workshop": 7,
    "preprints": 6,
    "as_lead": 8
  },
  "venues": ["JMLR", "EMNLP", "NeurIPS", "ACM", "Springer"]
}