What happens when ~80 open-weight LLMs face hard Python programming problems, and why reasoning models cheat far more than non-reasoning ones. ~20 min read Companion post: ACES: Teaching LLMs to Invent Their Own Programming Challenges explains the algorithm and benchmark used in this analysis. Background As reinforcement learning becomes the dominant paradigm for LLM post-training, a troubling pattern has emerged: models increasingly exploit loopholes in tests and scoring systems rather than solving the actual task (Pan et al., 2022; Skalse et al., 2022). This is not hypothetical. It has been observed repeatedly (Von Arx et al., 2025) in both benchmarks and real-world deployments, from coding agents deleting test files to models gaming evaluation metrics. When you combine RL-trained models with genuinely hard problems, the incentive to hack rather than solve becomes strong. ...

Self-Improving Language Models for Evolutionary Program Synthesis: A Case Study on ARC-AGI 🤗 Hugging Face (data and model)   |    📑 Paper    |    📑 Blog Large Language Models (LLMs) have become incredibly powerful, but they often hit a wall when faced with truly complex reasoning tasks that require discovering a solution from scratch. Simply throwing more computing power or using a bigger model often yields diminishing returns. But what if a model could learn from its own experience, getting smarter with every attempt? ...

Generating a Diversity of Challenging Programming Puzzles with Autotelic Generative Models (ACES) Introduction Human intelligence is marked not just by the ability to solve problems, but by the creative act of inventing them. Automating the generation of novel, diverse, and challenging problems has wide-ranging applications-from personalized education to robust benchmarking of AI systems. The ACES (Autotelic CodE Search) framework, accepted as a Spotlight Poster 💫 at NeurIPS 2024, introduces a principled method for generating Python programming puzzles that are both difficult and semantically varied, pushing the boundaries of what current generative models can achieve alone. ...