Mixing Hard and Easy Problems Boosts AI Training Efficiency 5x

Reinforcement Learning with Verifiable Rewards (RLVR) has become the go-to recipe for teaching language models to reason. Most published results assume you have tens of thousands of curated problems and a cluster of GPUs sitting idle. A new paper called Learning from Less tackles the opposite regime: what if you only have a few hundred examples and a fixed compute budget?

The short answer from the authors is that training on a mixed-complexity dataset is associated with up to 5x the sample efficiency under the same data budget. That gain comes from how you mix easy, medium, and hard problems, not from adding more data.

Why data scaling laws break at small sizes

Classical scaling work from Kaplan and Chinchilla told us how loss falls off as you increase parameters, tokens, and compute. Those laws assume abundance. In practice, teams fine-tuning a small model for a specific reasoning skill often have a few hundred verifiable questions at best and a handful of GPUs.

The authors point out that recent RLVR research has focused on using high volumes of question-answer pairs to improve LLM reasoning, citing datasets like DeepMath-103K with over 100,000 samples. Nobody has systematically characterized what happens when both data and compute are scarce, which is the situation most practitioners actually live in.

Three synthetic playgrounds with a difficulty dial

To study this in a controlled way, the team built three procedurally generated datasets where they can crank complexity up or down on demand:

• Counting Problems. Natural language prompts that ask the model to filter and count integers using operators like even count, divisible-by-N, bitwise XOR, or median. Complexity is controlled by the number of conditional filters (1 to 4) and transformations (0 to 3) applied before the final operation, creating problems with 1 to 7 total intermediate steps.