Ulam builds ultra-hard RL tasks and benchmarks in mathematics, STEM, and cybersecurity for frontier AI labs. We focus on research-level problems that challenge leading models.
Our RL tasks pair difficult problems with verifiable rewards, while our benchmarks measure meaningful progress on those problems. We use expert review and task-specific checks to distinguish sound reasoning from plausible mistakes and identify failure modes that further training can address.
Ultra-hard RL tasks and benchmarks that help frontier AI labs train stronger reasoning and measure progress in mathematics, STEM, and cybersecurity.
Difficult mathematical problems with executable verifiers and precise reward signals. Models must produce constructions, witnesses, or certificates that satisfy explicit conditions.
Scientific reasoning tasks that demand more than a plausible explanation: quantitative predictions, computational results, and testable solutions checked against task-specific requirements.
Challenging security environments for training and evaluating agents on investigation, defensive analysis, and tool use, with defined objectives and checks on task outcomes.
ErdosBench, STEMBench, and CyberBench test models on difficult problems beyond standard benchmarks. We assess correctness, meaningful partial progress, and failure modes—and build private evaluations around the capabilities your team needs to strengthen.
Evaluate mathematical reasoning, scientific research, and cybersecurity capabilities. Use the results to identify recurring errors and prioritize further training.
Start with difficult math, STEM, and cybersecurity problems packaged with verifiers and reward signals. Commission new tasks around the capabilities your model needs to strengthen.
Train agents on tasks that require tool use, multiple steps, and decisions based on intermediate results. We help teams hillclimb Terminal-Bench Science, GDP.pdf, and other popular benchmarks with targeted RL environments.
We usually start by providing sample RL tasks or exploring together which reasoning capabilities need to be strengthened. From there, we can provide off-the-shelf datasets or build RL environments matched to those needs.
Book a call with us.
We are a team of PhD-level mathematicians and engineers tackling research-level open problems. We use AI to push frontier models to their limits - and beyond - by solving real mathematical problems, not just measuring performance on established benchmarks. We care about the evidence, environments, and training signals that can move LLM reasoning to its next level.
Introducing our correctness-first evaluation of AI research across six STEM domains, with ten audited model runs and the open STEMBench-small companion set.
Read full article →