hillclimb

RL environments for post-training math LLMs

Member of Technical Staff - Environments & Post-Training

Natural Language Processing & LLMsConfirmed live in the last 24 hours

Full-timeSan Francisco, CA, USA$300k /yrIn-person

About the Job

Responsibilities

Turn machine learning research problems into environments that reward a model for real progress and good research taste.
Test whether each environment's reward is repeatable and tracks the quality of the work it scores.
Build systems that create environments at scale and catch broken ones before they reach training.
Run post-training experiments to find which environments are trainable.

Desired Qualifications

Prior experience with large language models, reinforcement learning, reinforcement learning from human feedback or AI feedback, reinforcement learning with verifiable rewards, post-training, evaluations, graders, synthetic data, model training, or coding agents at a frontier or data lab.
Understanding what makes a reward evaluation trustworthy and how models exploit an untrustworthy evaluation.
Background in designing automated systems for building, validating, and quality-checking environments.

About Hillclimb

Hillclimb's mission is to improve agent capabilities towards recursive self-improvement. To do so, we work on creating environments and evals that test and push frontier models' ability to ideate, experiment, and produce impactful research.

We're a founding team of former DeepMind & quant researchers, backed by Tier 1 VCs, Jeff Dean, Paul Graham, Amjad Masad, and angel investors from OpenAI, Anthropic, DeepMind, xAI & Meta Superintelligence Labs.

About this role:

Your work will consist of the following and many more:

Turn ML research problems into environments that reward a model for real progress and good research taste
Test whether each environment's reward is repeatable and tracks the quality of the work it scores
Build systems that create environments at scale and catch broken ones before they reach training
Run post-training experiments to find which environments are trainable

Preferred Experience:

Prior experience with LLMs, RL, RLHF/RLAIF, RLVR, post-training, evals, graders, synthetic data, model training, or coding agents at a frontier/data lab
Understanding of what makes a reward evaluation trustworthy, and how models exploit one that is not
Background in designing automated systems for building, validating, and quality checking environments.

Benefits:

Health, dental, vision insurance
Competitive 401k match
Free lunch & dinner

About the Company

hillclimb

RL environments for post-training math LLMs

Hillclimb is a San Francisco-based technology startup that creates reinforcement learning (RL) environments designed for post-training large language models (LLMs). Its core offering is specialized RL environments created in collaboration with mathematicians to improve LLM performance on mathematics tasks, aiming for state-of-the-art results in this domain. The product works by providing tailored RL environments that generate feedback signals to guide post-training fine-tuning of LLMs, helping developers refine and evaluate math-focused capabilities. Hillclimb differentiates itself from competitors by its explicit mathematical focus and close collaboration with mathematicians to design the RL environments, focusing on the math reasoning and problem-solving abilities of LLMs rather than generic model training. The company's goal is to push the math performance of LLMs and related AI systems, enabling higher-quality mathematical reasoning and problem-solving in AI applications.

Company size: 1-10
Company stage: Seed
Total funding: $130K
Headquarters: San Francisco, CA
Founded: 2025

Recently Posted Jobs

Explore more opportunities at hillclimb

See more jobs