arXiv:2608. 06631v1 Announce Type: cross Abstract: Cryptanalytic extraction has been demonstrated for ReLU networks, for networks using componentwise activations such as GELU or SiLU, and for a Transformer's final projection matrix.
arXiv:2608. 02064v1 Announce Type: new Abstract: Feed-forward networks (FFNs) account for a large fraction of Transformer parameters, yet their hidden width is usually constant across depth.
arXiv:2606. 26587v1 Announce Type: cross Abstract: Low-bit floating-point formats and semi-structured sparsity are increasingly supported by modern accelerators, yet combining them for LLM activation compression remains challenging: activations contain input-dependent outliers that dominate block scales in FP4 quantization, and directly applying N:M sparsity masks discards moderate values, coupling sparsification loss with quantization error.
arXiv:2606. 19379v1 Announce Type: cross Abstract: Transformer feed-forward networks (FFNs) are often treated as nonlinear stores of computation, yet how nonlinear a trained FFN block actually is has rarely been measured.
Robust CurveMoE is a mixture‑of‑experts framework that protects neural networks against perturbations defined by multiple norm constraints. It connects norm‑specialized models through a low‑loss path, selectively expertises only influential layers, and shares the rest of the parameters across routing paths. The method introduces contribution‑guided partial updating to reduce curve‑construction cost and provides a theoretical bound on the objective gap between partial and full optimization, achieving consistent improvements in clean, norm‑specific, and Union accuracy on CIFAR‑100 and ImageNet‑100.
arXiv:2606. 26538v1 Announce Type: cross Abstract: Deep Transformers are composed of uniformly stacked residual blocks, yet their deepest layers often add little value.
arXiv:2411. 09816v5 Announce Type: replace Abstract: Large neural networks achieve state-of-the-art performance on many tasks, yet their sheer size hinders deployment on resource-constrained devices.
arXiv:2607. 11990v1 Announce Type: cross Abstract: Feedforward network (FFN) blocks account for a large fraction of the parameters and computation in Transformer architectures, yet their internal structure remains difficult to interpret due to the additive superposition induced by the residual stream.
The paper introduces GaugeLasso, a method that applies symmetric group‑lasso penalties to transformer channels during training, enabling entire tensor slices to be zeroed out while maintaining dense tensors for GPU efficiency. By calibrating channel penalties based on inference utility per compute, the network self‑organizes into depth‑dependent structural profiles that can be dramatically smaller than the original architecture, achieving up to 255‑fold compression on a polynomial division task and outperforming hand‑designed baselines on language modeling and autoencoding benchmarks. The approach also accelerates training and reveals over‑provisioned axes that guide subsequent design iterations.
The paper introduces Mixture of Activations (MoA), a token‑adaptive feedforward network design that mixes multiple activation functions using lightweight gates while sharing linear projections. It also presents learnable activations (LA) as an input‑independent variant. The authors theoretically prove that MoA strictly surpasses both fixed‑activation FFNs and LA in expressive power, and empirically demonstrate that MoA achieves lower loss and better scaling on dense and MoE language models from 0.12 B to 2 B parameters with minimal overhead.
arXiv:2606. 16768v1 Announce Type: new Abstract: Training billion-parameter Transformers is often brittle, with transient loss spikes and divergence that waste compute.
arXiv:2609.37717v1 Announce Type: new Abstract: Decoder-only transformers are trained only through a terminal next-token prediction loss, yet this loss constrains every intermediate hidden state thro...