# Elad Sharony — full profile

> Ph.D. student in Electrical and Computer Engineering at the Technion (advised by Prof. Shie Mannor, (RL)² Lab) and Research Intern at NVIDIA Research. Research on reinforcement learning and LLM agents: how AI systems learn from interaction and feedback, and how they plan and search efficiently, from robotics to software engineering.

This file collects everything on [esharony.me](https://esharony.me/) in plain markdown: bio, research interests, publications with abstracts, experience, education, teaching, honors, and service. The short index is [llms.txt](https://esharony.me/llms.txt). Last updated: 2026-10-15.

## Bio

I am a Ph.D. student at the Technion, advised by Prof. Shie Mannor, and a research intern at NVIDIA Research. My research is on reinforcement learning and LLM agents: how AI systems learn from interaction and feedback, and how they plan and search efficiently, from robotics to software engineering.

Previously, I was a research intern at the Bosch Center for AI, worked as a control engineer at Intel, and earned a B.Sc. in Mechanical Engineering from the Technion, *summa cum laude*.

## Research interests

My research focuses on reinforcement learning and large language model (LLM) agents. I am interested in how AI systems learn from interaction and feedback, reason and search efficiently, and generalize across domains, from robotics and control to software engineering.

- Reinforcement learning: learning from interaction and feedback, sample efficiency, exploration.
- LLM agents: efficient planning and search, from robot control to software engineering.
- Optimization: control, trajectory optimization, amortized optimization.

## Publications

### Make Every Expansion Count: Learning to Search Efficiently in LLM Code Repair (CodeTriage)

- Authors: Elad Sharony, Ido Greenberg, Jialin Song, Shie Mannor, Gal Chechik, Eli Meirom
- Status: under review
- Project page: https://esharony.me/projects/codetriage/
- arXiv: https://arxiv.org/abs/2610.15894

Abstract. LLM code-repair systems fix faulty programs by repeatedly generating revisions and running unit tests. Each revision is costly, making the choice of which repair to pursue important for search efficiency. The adaptive policies we compare against make this choice using test pass rates and search statistics, without reading the detailed test output or proposed edit descriptions. We introduce CodeTriage, a learned search policy whose small 0.6B-parameter model reads test output, failure diagnoses, and proposed edits to predict how many repair steps each edit could require to reach a fix. These predictions prioritize all pending edits across programs before generating and testing the revised code. On held-out repair tasks, CodeTriage finds fixes with 21% fewer program revisions and 12% fewer output tokens than the strongest adaptive baseline on each metric. The same model transfers without retraining to eleven unseen repair LLMs, achieving the highest success rate averaged across budgets for each one. Ablations show the value of interpreting proposed edits together with test output and diagnoses, and of learning to compare edits across different programs. Additionally, we release a corpus of 12.2 million program versions generated by four LLMs, organized into branching repair sequences with proposed edits and test results.

### VLM-Guided Experience Replay (VLM-RB)

- Authors: Elad Sharony, Tom Jurgenson, Orr Krupnik, Dotan Di Castro, Shie Mannor
- Status: under review
- Project page: https://esharony.me/projects/vlm-rb/
- arXiv: https://arxiv.org/abs/2602.01915

Abstract. Recent advances in Large Language Models (LLMs) and Vision-Language Models (VLMs) have enabled powerful semantic and multimodal reasoning capabilities, creating new opportunities to enhance sample efficiency, high-level planning, and interpretability in reinforcement learning (RL). While prior work has integrated LLMs and VLMs into various components of RL, the replay buffer, a core component for storing and reusing experiences, remains unexplored. We propose addressing this gap by leveraging VLMs to guide the prioritization of experiences in the replay buffer. Our key idea is to use a frozen, pre-trained VLM as an automated evaluator to identify and prioritize promising sub-trajectories from the agent's experiences. Across scenarios, including game-playing and robotics, spanning both discrete and continuous domains, agents trained with our proposed prioritization method achieve 15–57% higher average success rates and improve sample efficiency by 35–55% compared to previous approaches.

Key results (from the project page):
- 15–57% higher average success rates than uniform and prioritized experience replay, in discrete game-playing and simulated continuous-control manipulation.
- 35–55% better sample efficiency: fewer steps to reach the baselines' best performance.
- Matches or exceeds every baseline in success and improves sample efficiency in 23 of 24 comparisons.
- No fine-tuning required: uses frozen, off-the-shelf VLMs.
- About 12% throughput overhead via asynchronous VLM scoring.

### Learning Multiple Initial Solutions to Optimization Problems (MISO)

- Authors: Elad Sharony, Heng Yang, Tong Che, Marco Pavone, Shie Mannor, Peter Karkus
- Venue: ICRA 2026 (IEEE International Conference on Robotics and Automation), Vienna
- Project page: https://esharony.me/projects/miso/
- arXiv: https://arxiv.org/abs/2411.02158
- Code: https://github.com/EladSharony/miso

Summary (from the project page). Many real-world applications, from autonomous driving to robotic manipulation, require solving similar optimization problems sequentially under strict runtime constraints. The performance of local optimizers is highly sensitive to the initial solution: poor initialization can lead to slow convergence or convergence to suboptimal local minima. MISO (Learning Multiple Initial SOlutions) is a framework that trains a single neural network to predict multiple diverse initial solutions for optimization problems. Unlike ensemble methods that require training and running multiple models, MISO uses a single efficient network with diversity-promoting training objectives.
- Single-optimizer setting: select the most promising initial solution using a selection function, then run a single optimizer.
- Multiple-optimizers setting: initialize multiple parallel optimizers, each with a different predicted solution, and select the best final output.
- Guaranteed improvement: by including the default initialization among predictions, MISO is guaranteed to perform equal or better than the baseline.
- Efficient inference: unlike ensembles, MISO's inference time remains nearly constant as the number of predicted solutions K increases.

## Research and industry experience

- NVIDIA Research — Research Intern (May 2026 – present).
- Bosch Center for Artificial Intelligence (BCAI) — Research Intern (Apr 2024 – Oct 2024). Developed VLM-based trajectory embeddings for diversity-aware experience replay in robotic manipulation RL.
- Intel Corporation — Control Engineer (Apr 2021 – Jan 2023). Designed and integrated a data-driven autotune pipeline for thermal controllers, and developed and deployed a pneumatic control system to ensure stable and reliable thermal contact.

## Education

- Ph.D., Electrical and Computer Engineering (Direct Track), Technion — Israel Institute of Technology (Mar 2022 – present, expected 2027). Research: integrating fast (System 1) and slow (System 2) thinking into RL and optimization for robotics and autonomy. Advisor: Prof. Shie Mannor. GPA 92.0.
- B.Sc., Mechanical Engineering, Technion — Israel Institute of Technology (Oct 2018 – Mar 2022, Early Bird). Thesis: non-contact topographic sensing using vibration and machine learning to predict the distance between surfaces. Advisors: Prof. Itzhak Bucher and Dr. Yoav Vered. GPA 94.9 (Summa Cum Laude).

## Teaching and mentoring

- Teaching assistant, Technion: Control Theory; Control System Design; Control Systems 1.
- Project supervision, Technion: Developing Robotic Knot-Tying Skills; Reinforcement Learning in Isaac Simulator.

## Honors and awards

- Reamim — selective program supporting outstanding research-oriented students (Oct 2018 – Mar 2022).
- President's List for Academic Excellence (Fall 2021/22, Spring 2020/21, Fall 2020/21).
- Dean's List for Academic Excellence (Spring 2019/20, Fall 2019/20, Spring 2018/19).

## Service

- Reviewer: CoRL 2025, ICLR 2026, ICML 2026, NeurIPS 2026.
- Technical editor: *Reinforcement Learning: Foundations* by Shie Mannor, Yishay Mansour, and Aviv Tamar (Cambridge University Press, 2026).

## News

- May 2026: Started a research internship at NVIDIA Research.
- Jan 2026: MISO was accepted to ICRA 2026 in Vienna.
- Apr 2024: Started a research internship at the Bosch Center for AI.
- Mar 2022: Started my Ph.D. at the Technion with Prof. Shie Mannor.

## Links

- Homepage: https://esharony.me/
- CV: https://esharony.me/cv.html
- Google Scholar: https://scholar.google.com/citations?user=OmuTq-MAAAAJ&hl=en
- GitHub: https://github.com/EladSharony
- LinkedIn: https://linkedin.com/in/EladSharony
- X: https://twitter.com/elad_sharony
- Email: eladsharony@campus.technion.ac.il