Neural Information Causality: A New Framework for Representation Learning
Researchers Jeongho Bang and Marcin Pawłowski have introduced Neural Information Causality (Neural-IC), a novel framework integrating information causality into representation learning. Published on arXiv in May 2026, this study addresses query-separated computation where data encoding precedes query knowledge, forcing representations to function as operational messages rather than simple feature maps. The framework establishes two key principles: query-separated architectures induce random-access communication experiments obeying specific embedding inequalities, and physical capacity bounds on interfaces limit information access. This approach serves as an operational diagnostic for detecting query leakage, precision leakage, and episode-specific memory issues. The authors provide a classical one-bit Random Access Code (RAC) benchmark, demonstrating that quantum enhancement relates to fair query-conditioned access rather than total information beyond bottlenecks. Additionally, the analysis extends to CHSH-type correlation layers, showing that nested Neural-RAC protocols multiply correlation biases, with stability requirements selecting the Tsirelson threshold. Controlled simulations verify that apparent violations of these principles stem from broken query separation or undercounted capacity, offering significant insights for AI architecture design and quantum-classical boundary understanding.
Wire timeline
Neural Information Causality: A New Framework for Representation Learning
Researchers Jeongho Bang and Marcin Pawłowski have introduced Neural Information Causality (Neural-IC), a novel framework integrating information causality into representation learning. Published on arXiv in May 2026, this study addresses query-separated computation where data encoding precedes query knowledge, forcing representations to function as operational messages rather than simple feature maps. The framework establishes two key principles: query-separated architectures induce random-access communication experiments obeying specific embedding inequalities, and physical capacity bounds on interfaces limit information access. This approach serves as an operational diagnostic for detecting query leakage, precision leakage, and episode-specific memory issues. The authors provide a classical one-bit Random Access Code (RAC) benchmark, demonstrating that quantum enhancement relates to fair query-conditioned access rather than total information beyond bottlenecks. Additionally, the analysis extends to CHSH-type correlation layers, showing that nested Neural-RAC protocols multiply correlation biases, with stability requirements selecting the Tsirelson threshold. Controlled simulations verify that apparent violations of these principles stem from broken query separation or undercounted capacity, offering significant insights for AI architecture design and quantum-classical boundary understanding.
cs.AI updates on arXiv.org