Henrik Abgaryan

Bio

PhD Student, LAMSADE PSL

Expected Graduation: 2027

AI Scientist and PhD Candidate specialized in LLM Reasoning for Combinatorial Optimization. I build rigorous benchmarking frameworks (ACCORD, STARJOB) and scalable generative models. Previously architected AI features for 150M+ users at Picsart, combining theoretical depth (Diffusion, GANs) with production-grade engineering.

Publication and Preprints

• [1] ACCORD: A large-scale open dataset and framework for constraint-aware LLMs
  NeurIPS 2025 Differentiable Learning of Combinatorial Algorithms Workshop
  Authors: Henrik Abgaryan, Tristan Cazenave, Ararat Harutyunyan
  Released ACCORD, a 90K instance dataset and evaluation framework spanning multiple NP-hard problems (Knapsack, Bin Packing, TSP, VRP, JSSP). ACCORD fine-tuning (LLAMA 8B) reduces optimality gaps from > 80% to as low as 0-12% across problem sizes, outperforming strong prompting baselines.
• [2] LLMs can Schedule
  ECML PKDD 2025 DSO Workshop (Oral Presentation)
  Authors: Henrik Abgaryan, Tristan Cazenave, Ararat Harutyunyan
  Demonstrated end-to-end LLM-based solving for NP-hard Job Shop Scheduling, with systematic evaluation across instance sizes and baselines.
• [3] STARJOB: An LLM-native dataset for Job Shop Scheduling
  AAAI 2025 Workshop on LLMs for Planning (Oral Presentation)
  Authors: Henrik Abgaryan, Tristan Cazenave, Ararat Harutyunyan
  Released STARJOB, a 120K instance dataset formatted for LLM fine-tuning on Job Shop Scheduling. To our knowledge, the first large-scale LLM-native JSSP dataset designed specifically for supervised fine-tuning.
• [4] SchedulExpert: Graph Attention Meets Mixture-of-Experts for JSSP
  Learning and Intelligent Optimization (LION 19), 2025 (Oral Presentation)
  Authors: Henrik Abgaryan, Tristan Cazenave, Ararat Harutyunyan
• [5] Randomized Greedy Sampling for the Job Shop Scheduling Problem
  Learning and Intelligent Optimization (LION 18), 2024 (Oral Presentation)
  Authors: Henrik Abgaryan, Tristan Cazenave, Ararat Harutyunyan

Summer Schools & Talks

• Data Science Meets Combinatorial Optimisation Euro PhD School
  Technical University of Eindhoven, The Netherlands. (2025)
  Participated in advanced doctoral training at the intersection of data science and combinatorial optimisation.
• Talk: "GNNs for Job Shop Scheduling Problem"
  Lamsade Doctoral Spring School, Ecole du Luat, Ecouen-Ezanville, France. (2024)
  Presented research on applying Graph Neural Networks to the Job Shop Scheduling Problem.

Industry Experience

• Machine Learning Scientist, Picsart AI
  2021-2023
  • Generative AI: Engineered personalized text-to-image systems (LoRA) and geometry-preserving GANs to ensure subject consistency and facial fidelity.
  • Mobile Optimization: Designed Teacher-Student distillation frameworks to compress high-capacity GANs for real-time, on-device edge deployment.
  • Classical CV: Optimized performance on mobile CPUs by engineering gradient mapping pipelines for material transfer, eliminating neural overhead.
  • Scale: Architected 4-bit quantization pipelines, reducing end-to-end latency by 12% for 150M+ monthly active users.
  • Leadership: Transitioned 35+ research prototypes into production environments using large-scale distributed training systems.
• Software Engineer Intern, Synergy International Systems
  Software Engineering & Database Systems (2019)
  • Engineered core software components and designed a comprehensive database system to optimize business operations.

Education

• PhD, Artificial Intelligence
  LAMSADE, Université Paris-Dauphine PSL
  2023-Present
  • Focus: Machine Learning for Combinatorial Optimization
  • Advisors: Tristan Cazenave, Ararat Harutyunyan
• MSc, Computer Science (AI Systems and Data)
  Université Paris-Dauphine PSL
  2022
  • Graduated with Distinction
• B.S. in Computer Science
  American University of Armenia
  2017
  • Relevant coursework: CS104 Linear Algebra, CS111 Discrete Mathematics, CS107 Probability, CS108 Statistics, CS213 Optimization, CS392 Special Topics in CS: Distributed Algorithms

Technical Expertise

• Core AI: LLM Reasoning (CoT/ToT), GRPO, Reinforcement Learning, Graph Neural Networks (GNNs), Combinatorial Optimization.
• Software & Tools: PyTorch, Tensorflow, JAX, Docker, Multi-GPU Training.
• Programming Languages: Python, C++
• Languages: English, French, Armenian (Native), Russian

Teaching and Leadership

• Université Paris-Dauphine PSL (2023-2026)
  Teaching Associate & Assistant
  
  • Graduate (M1): Artificial Intelligence, Introduction to ML (MIAGE Apprentissage).
  • Undergraduate (L2): Data Analytics 2, Algorithms and Programming 2, Computer Tools (HTML/Excel).
• American University of Armenia (2022)
  Teaching Associate & Assistant
  
  • Undergraduate: Causal Inference, Data Visualization.

Honors and Awards

• Dauphine Doctoral Scholarship (2023-2026)
• 1st Place, IASD Master's DL Competition: Designed a parameter-efficient neural network for Go that outperformed peer models in a round-robin tournament. (2022)
• PSL Excellence Scholarship (awarded to top students based on GPA) (2022)
• National Academic Scholarship (2017)
• Oral selections: AAAI 2025 LLM4Plan; ECML PKDD 2025 DSO; LION 18 & 19
• Reviewer: NeurIPS 2024, NeurIPS 2025

Interests

GYM, Painting, Swimming, Running, Healthy Lifestyle.

AI News

Latest AI news from around the web