Reflection AI Interview Guide (2026): RL, Agents, and Open Frontier Models

Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both from Google DeepMind, where Antonoglou was a key architect of AlphaGo and AlphaZero and Laskin worked on reinforcement learning for Gemini. The company came out of stealth in March 2025 with $130 million (a $25 million seed led by Sequoia and CRV, then a $105 million Series A co-led by Lightspeed and CRV) at roughly a $545 million valuation, then raised about $2 billion at an $8 billion valuation in October 2025 in a round led by Nvidia, with later reports of a much larger raise. Its stated goal is open superintelligence, and it positions itself as an open frontier lab that releases model weights while selling to enterprises. The first product, Asimov, is a code-comprehension agent that reads a team’s source code, docs, and communications to answer questions about how a system works. Offices are in New York and San Francisco.

Process

There is no large public record of Reflection’s loop yet, so treat this as the likely shape based on the company’s stage and the labs its founders came from rather than a confirmed script. Expect a recruiter or hiring-manager screen first, then a split by track. On the research side (Member of Technical Staff in pre-training, post-training, or research engineering), plan for a detailed walk through your past work and papers, one or two rounds on reinforcement learning and ML fundamentals, and at least one round where you implement something from scratch. On the engineering and infrastructure side, the emphasis shifts toward coding and ML system design: agent architecture, training and serving infrastructure, and scaling. Every track carries a high general coding bar, and you should expect a behavioral conversation about ownership and working in a small, fast team. A company this size moves quickly, so the full cycle is often two to four weeks rather than the multi-month drag of a large org.

What they actually ask

• Implement a core RL algorithm from scratch, such as policy gradient, PPO, or Q-learning, then explain the update rule and reason about training instability and reward design.
• Sketch the architecture of a coding agent: the planning loop, tool use, context and memory management, and how you would keep it from going off the rails on a long task.
• Design an eval for an agent or model. What do you measure, how do you stop the metric from being gamed, and where does offline signal diverge from real production behavior?
• A strong general coding round with clean, correct code under time pressure. Data structures, complexity, and edge cases still count, even for research roles.
• For infra roles, distributed training at scale: data, model, and pipeline parallelism, GPU memory limits, throughput, and recovering from failures across thousands of accelerators.
• Serving and systems questions for applied roles: latency, context-window and cost tradeoffs, caching, and shipping an agentic system to a real customer.

Levels and comp (2026)

Reflection uses flat research-lab titles like Member of Technical Staff rather than a long ladder, and pay is built as a competitive base plus a large equity grant, with standard medical, dental, and vision benefits. Precise numbers are not published, and levels.fyi currently has no verified data points, so treat any figure as an estimate. For orientation, senior ML and research total comp at top frontier labs commonly lands in the $400K to $700K range, with base pay often between $200K and $400K and the rest in equity; exceptional or very senior hires can go well past that. Be clear-eyed about the equity: the company went from roughly a $545 million valuation in March 2025 to an $8 billion valuation eight months later, so the value of a grant swings hard on the strike or preferred price you get and on whether the valuation keeps climbing or corrects. Cross-check any specific offer against levels.fyi and recent Blind threads for the exact role, and ask directly about strike price, preferred price, vesting, and refresh policy instead of accepting a single headline number.

Prep priorities

Know reinforcement learning cold, not as trivia but as something you can code and debug: value versus policy methods, the math behind a policy-gradient update, why training goes unstable, and how reward shaping changes behavior. Read the space Reflection actually works in, which is agents, evals, and scaling, so you can talk about what makes a coding agent reliable and how you would measure it. Keep your general coding sharp, because a weak algorithmic round can sink an otherwise strong research candidate. If you are targeting infrastructure, be ready to talk concretely about large-scale distributed training and GPU efficiency. Have a tight research narrative ready, one where you can defend your choices and say what you would do differently, since the loop rewards depth over breadth. Prepare a couple of behavioral stories about high-agency work on a small team, because that is the environment you would be joining.

Frequently Asked Questions

Is Reflection AI only hiring PhD researchers? No. Alongside research roles it hires research and platform engineers, GPU infrastructure engineers, and forward-deployed engineers who take agentic systems to customers. Strong systems and coding engineers without a research background have a real path in, especially on the infra and applied side.

Do I need reinforcement learning experience to get an offer? For research roles, RL fluency is close to mandatory given the founders’ background and the company’s focus on scaling RL for coding agents. For pure infrastructure or applied engineering roles, deep RL is less of a gate, but you should still understand the training and serving pipeline you would support.

How hard is the coding bar? Expect it to be high across the board. This is a lab full of ex-DeepMind engineers, so even research interviews tend to include a round where clean, correct, efficient code matters, not just modeling ideas.

Is the equity worth it given how fast the valuation jumped? That is the real question and it has no guaranteed answer. The upside is large if the company keeps compounding, but a grant priced at an $8 billion or higher valuation needs continued growth to pay off. Ask for the strike or preferred price and model a few outcomes before you weigh the offer against a more established lab.