Research Scientist, Reinforcement Learning

Pramaana Labs; Palo Alto (CA); On-site; USD 150,000 - 230,000; Full time

Job summary

Pramaana Labs, a Palo Alto AI lab, is seeking an RL Post-Training Researcher to scale reasoning capabilities of foundation models. You will apply RLVR to new domains, use Lean-derived rewards, and push autoformalization and proving through novel RL algorithms and data evals.

You will own the full loop from rollout to evaluation. Ideal candidates will have deep RL research experience at scale, expertise with deterministic reward signals, and a strong systems mindset to build robust evals with

Qualifications

• Deep, hands-on research experience in reinforcement learning applied to reasoning models at scale.
• Experience working with RLVR, test-time RL, or exact deterministic reward signals.
• Strong algorithmic and systems intuition — comfortable writing custom RL loops, managing data pipelines, and building robust evals from scratch.
• High autonomy: the ability to take a fuzzy problem area and independently drive it to state-of-the-art results without day-to-day direction.

Responsibilities

• Scale RLVR to new, real-world domains using the Lean kernel as a deterministic reward signal.
• Design and implement novel RL algorithms and test-time compute optimizations tailored for formal environments and proof search.
• Push the frontier of autoformalization, training models to map highly technical natural language into strict formal specifications.
• Curate data and build evals that tightly correlate with verifiable downstream reasoning capabilities.
• Own the whole loop: rollout sampling, reward design, the update, the eval.

Skills

Reinforcement learning

RLVR

Custom RL loops

Tools

Lean

Coq

Isabelle

Job description

Company

Pramaana Labs is a frontier AI lab building the verification layer for AI. Founded in 2025 and headquartered in Palo Alto, we turn complex human knowledge, including tax codes, legal rules, clinical guidelines and security protocols, into a formal representation, so every AI answer can be traced, challenged, and proved.

Pramaana's architecture pairs foundation models trained to formalize and reason with a symbolic world model encoded in Lean 4. Pramaana was founded by a team out of Google, DeepMind, and Glean, combining frontier AI researchers and formal methods experts.

The role

We're training foundation models to natively interact with symbolic world models. As an RL Post-Training Researcher, you'll take foundation models and scale their reasoning capabilities: applying RLVR to new domains using verified rewards from the Lean kernel, pushing the frontier of autoformalization and proving, and innovating on RL algorithms, data, and evals. Your work will also define how our models leverage test-time compute to solve long-horizon logical tasks.

What you'll actually work on

• Scale RLVR to new, real-world domains using the Lean kernel as a deterministic reward signal.

• Design and implement novel RL algorithms and test-time compute optimizations tailored for formal environments and proof search.

• Push the frontier of autoformalization, training models to map highly technical natural language into strict formal specifications.

• Curate data and build evals that tightly correlate with verifiable downstream reasoning capabilities.

• Own the whole loop: rollout sampling, reward design, the update, the eval.

You Will Thrive in This Role If You Have

• Deep, hands-on research experience in reinforcement learning applied to reasoning models at scale.

• Experience working with RLVR, test-time RL, or exact deterministic reward signals.

• Strong algorithmic and systems intuition — comfortable writing custom RL loops, managing data pipelines, and building robust evals from scratch.

• High autonomy: the ability to take a fuzzy problem area and independently drive it to state-of-the-art results without day-to-day direction.

Especially Strong Candidates May Also Have

• A background in formal verification (Lean, Coq, or Isabelle)

• A strong publication record in top-tier venues (NeurIPS, ICML, ICLR).