Tag
The article argues that relying on chain-of-thought traces for AI safety is ineffective, as they can be manipulated and do not faithfully represent model behavior, instead emphasizing the need to focus on harness control mechanisms.
This paper proposes a unified taxonomy for memory mechanisms in deep time-series models, categorizing methods from internal to external memory and identifying research gaps to address temporal dependence challenges.
This paper investigates whether Transformer and Mamba architectures exhibit similar collective dynamics in their infrared spectrum, finding that both develop near-marginal slow-mode dynamics despite different microscopic mechanisms.
This preprint introduces Neuroevolution Arena, a GPU-accelerated artificial life platform for comparing evolution and RL-based update regimes across neural architectures via a nested ecological evaluation protocol.
The paper identifies two causes why logic gate networks fail to benefit from increased depth and proposes Input-Anchored Logic Gate Networks (IALGNs) that condition each layer on original inputs, achieving consistent depth-accuracy improvements beyond 100 layers.
This paper studies transfer specificity in implicit neural representations across SIREN, ReLU MLPs, and Fourier-feature MLPs, finding that transfer magnitude and specificity depend on architecture, with ReLU being more selective and SIREN reusing weights broadly. Results suggest architecture selection should consider explicit control conditions, not just transfer magnitude.