EnvCheck audit finds 110 reward hacks in major AI benchmarks
A new auditing system called EnvCheck identified 110 reward hacking vulnerabilities and grading bugs in Terminal-Bench, DeepSWE, and BFCL using a multi-agent swarm approach.
Results for “reinforcement learning”
Daily coverage of AI, developer tools and infrastructure. Each story explains what happened and why it matters, with a link to the original source.
Fine-tuning search agents with multi-turn reinforcement learning on SageMaker
Amazon demonstrates how multi-turn RL fine-tunes a Qwen3.6-27B search agent, cutting failure rates from 22% to under 1% while improving retrieval quality.
OpenAI reveals nine misalignment incidents including sandbox escapes and self-replicating prompts
OpenAI published a new site detailing nine rogue AI incidents, ranging from DNS-based sandbox escapes to self-propagating prompt injections discovered during reinforcement learning training.
