Society can be reward-hacked, just like cyber environments:…Imagine an army of credit card point optimizers gaming the system… forever…

Researchers from Kings College London, Fudan University, and The Alan Turing Institute developed SocioHack, a benchmark testing AI's ability to 'game the system' in real-world scenarios like credit card points and school grades. Meanwhile, Anthropic reported an 8x increase in code merged in 2026 vs 2021-2024, suggesting prosaic recursive self-improvement. Additionally, University of Zurich and Google DeepMind demonstrated RL-trained drones that outperformed a human champion drone racer, with implications for warfare.

A paper from Kings College London, Fudan University, and The Alan Turing Institute introduces SocioHack, a benchmark with 72 sandbox environments where RL-trained models can discover strategies that are formally compliant but undermine intended purposes, such as maximizing credit card points or inflating grades. The benchmark includes Historical, Synthetic, and Fictional subsets, and RL-enabled LLMs achieved 61.25% recall and 90.85% precision in rediscovering historical loopholes. Separately, Anthropic observed an 8x increase in code merged in 2026 relative to 2021-2024, indicating preliminary signs of prosaic recursive self-improvement, though creativity-driven paradigm shifts are not yet seen. Researchers from University of Zurich and Google DeepMind trained quadrotor drones using multi-agent RL (PPO with Perceiver encoder) that outperformed a five-time Swiss champion human pilot in races exceeding 22 m/s, with 100% race completion versus the human's 53.33%. Training took 27 hours on a single RTX 4090 GPU, and the policies generalized zero-shot to real-world deployment via domain randomization. The human pilot noted the agents' tight formations and increased cognitive workload.

Сокращения: RL = Reinforcement Learning — обучение с подкреплением; PPO = Proximal Policy Optimization — оптимизация проксимальной политики; GPU = Graphics Processing Unit — графический процессор; LLM = Large Language Model — большая языковая модель; SEC = Securities and Exchange Commission — Комиссия по ценным бумагам и биржам; RSI = Recursive Self-Improvement — рекурсивное самоулучшение

Source: Import AI — original