Transformer FFNs

Recent bilinear feed-forward replacements for vision transformers can substantially improve accuracy, but they often conflate two effects: stronger second-order interactions and increased redundancy relative to the main branch. We study a complementary design principle in which auxiliary quadratic features contribute only information not already captured by the dominant hidden representation. To this end, we propose Orthogonal Quadratic Complements (OQC), which construct a low-rank quadratic auxiliary branch and explicitly project it onto the orthogonal complement of the main branch before injection. We further study an efficient low-rank realization (OQC-LR) and gated extensions (OQC-static and OQC-dynamic). Under a parameter-matched Deep-ViT and CIFAR-100 protocol with a fixed penultimate residual readout, full OQC improves an AFBO baseline from 64.25 +/- 0.22 to 65.59 +/- 0.22, while OQC-LR reaches 65.52 +/- 0.25 with a substantially better speed-accuracy tradeoff. On TinyImageNet, the gated extension OQC-dynamic achieves 51.88 +/- 0.32, improving the baseline (50.45 +/- 0.21) by 1.43 points and outperforming all ungated variants. Mechanism analyses show near-zero post-projection auxiliary-main overlap together with improved representation geometry and class separation. The full family, including both ungated and gated variants, generalizes consistently across both datasets.

Transformers for language modeling usually rely on deterministic internal computation, with uncertainty expressed mainly at the output layer. We introduce variational neurons into Transformer feed-forward computation so that uncertainty becomes part of the internal computation itself. Concretely, we replace deterministic feed-forward units with local variational units based on EVE while preserving the overall Transformer backbone. We evaluate this design in compact next-token language-modeling settings. We compare deterministic and variational variants with both predictive and probabilistic criteria. Alongside negative log-likelihood, perplexity and accuracy, we analyze calibration, conditional variance, mutual information and latent-usage statistics. The resulting picture is clear. Variational neurons integrate stably into Transformers, preserve strong predictive performance and produce informative uncertainty signals. The experiments also show that task quality, useful depth and internal stability are distinct properties. These results establish variational Transformers as a practical form of uncertainty-aware language modeling. They show that Transformers can predict with an explicit internal structure of uncertainty, which supports stronger probabilistic evaluation and a more informative analysis of model behavior.

At typical context lengths, the feed-forward MLP block accounts for a large share of a transformer's compute budget, motivating sparse alternatives to dense MLP blocks. We study sparse, tree-structured feed-forward layers as drop-in replacements for MLP blocks in deep transformer architectures, enabling conditional computation via hard hierarchical routing without a separate router network. We demonstrate for the first time that this form of tree-structured conditional sparsity can be applied for autoregressive language modeling and downstream question answering, including zero- and few-shot settings, and its scalability beyond 1B parameters. Despite activating fewer than 5% of the feed-forward block's units per token, our models match dense baselines under controlled training and fine-tuning protocols. We further analyze training dynamics and identify an emergent auto-pruning effect: the interaction of hard routing with asymmetric nonlinearities progressively deactivates unused paths, yielding partial conversion of dynamic routing into static structural sparsity. We show that simple architectural choices can modulate this behavior and recover balanced trees without auxiliary losses. Overall, our work demonstrates that tree-structured feed-forward layers provide a scalable and controllable mechanism for sparsifying large transformer models.

We show that MLP layers in transformer language models perform binary routing of continuous signals: the decision of whether a token needs nonlinear processing is well-captured by binary neuron activations, even though the signals being routed are continuous. In GPT-2 Small (124M parameters), we find that specific neurons implement a consensus architecture -- seven "default-ON" neurons and one exception handler (N2123 in Layer 11) that are 93-98% mutually exclusive -- creating a binary routing switch. A cross-layer analysis reveals a developmental arc: early layers (L1-3) use single gateway neurons to route exceptions without consensus quorums; middle layers (L4-6) show diffuse processing with neither gateway nor consensus; and late layers (L7-11) crystallize full consensus/exception architectures with increasing quorum size (1 to 3 to 7 consensus neurons). Causal validation confirms the routing is functional: removing the MLP at consensus breakdown costs 43.3% perplexity, while at full consensus removing it costs only 10.1% -- exceeding a 4x difference. Comparing binary vs. continuous features for the routing decision confirms that binarization loses essentially no information (79.2% vs. 78.8% accuracy), while continuous activations carry additional magnitude information (R^2 = 0.36 vs. 0.22). This binary routing structure explains why smooth polynomial approximation fails: cross-validated polynomial fits (degrees 2-7) never exceed R^2 = 0.06 for highly nonlinear layers. We propose that the well-established piecewise-affine characterization of deep networks can be complemented by a routing characterization: along the natural data manifold, the piecewise boundaries implement binary decisions about which tokens need nonlinear processing, routing continuous signals through qualitatively different computational paths.

Biological neural systems employ diverse neurotransmitters -- glutamate, GABA, dopamine, acetylcholine -- to implement distinct signal-processing modalities within shared neural circuits. In contrast, modern transformers apply a single fixed activation function across all feed-forward neurons. We introduce PolyGLU (Polychromatic Gated Linear Unit), a drop-in replacement for SwiGLU that enables each FFN neuron to dynamically route among K=4 activation functions via a differentiable mechanism combining learned static preferences with input-conditioned gating, trained end-to-end with Gumbel-Softmax. We train PolychromaticLM, a 597M-parameter transformer, on ~10B tokens using a single NVIDIA A100 GPU. Our key finding is emergent routing behavior: without any explicit sparsity loss or entropy regularization, the routing mechanism converges to near-deterministic activation selections (mean dynamic entropy = 0.030% of maximum), with a striking depth-dependent specialization pattern -- early layers prefer GELU while deep layers strongly favor Tanh. Three layers maintain elevated routing entropy, suggesting computational flexibility points. The routing architecture adds only 0.23% parameter overhead (~1.4M parameters) and proves fully robust to supervised fine-tuning: routing entropy remains constant at ln(4) throughout 13,067 SFT steps. On standard benchmarks, PolychromaticLM achieves 62-89% of Qwen3-0.6B-Base performance despite training on 3,600x fewer tokens. All code, weights, and training infrastructure are released under Apache 2.0.

We introduce NerVE, a unified eigenspectral framework for understanding how feed-forward networks (FFNs) in large language models (LLMs) organize and regulate information flow in high-dimensional latent space. Despite FFNs dominating the parameter budget, their high-dimensional dynamics remain poorly understood. NerVE addresses this gap through lightweight, memory-efficient tracking of eigenspectrum dynamics via four complementary metrics: Spectral Entropy (dispersion), Participation Ratio (effective dimensionality), Eigenvalue Early Enrichment (top-heaviness), and Jensen-Shannon divergence (distributional shifts). Our key insight is that FFN nonlinearities reinject variance across eigenmodes, fundamentally governing latent dimension utilization, and that optimizer geometry strongly modulates the extent of this variance reinjection. We validate NerVE across model scales, and diverse architectural and optimizer configurations, each uniquely shaping FFN dynamics: normalization schemes controlling variance flow; FFN weight geometries constraining latent space; positional encoding and activation functions regulating information flow; and optimizer choices redistributing effective capacity across depth. Across these settings, NerVE consistently recovers stable spectral signatures that correlate with model's generalization ability and respond predictably to design choices, generalizing beyond transformer to MLP-Mixer architectures, providing actionable insights for architectural and optimizer choices beyond trial-and-error.

Recent studies have observed that intermediate layers of foundation models often yield more discriminative representations than the final layer. While initially attributed to autoregressive pretraining, this phenomenon has also been identified in models trained via supervised and discriminative self-supervised objectives. In this paper, we conduct a comprehensive study to analyze the behavior of intermediate layers in pretrained vision transformers. Through extensive linear probing experiments across a diverse set of image classification benchmarks, we find that distribution shift between pretraining and downstream data is the primary cause of performance degradation in deeper layers. Furthermore, we perform a fine-grained analysis at the module level. Our findings reveal that standard probing of transformer block outputs is suboptimal; instead, probing the activation within the feedforward network yields the best performance under significant distribution shift, whereas the normalized output of the multi-head self-attention module is optimal when the shift is weak.

We investigate when transformer MLP nonlinearity is actually necessary. A gate with $d+1$ parameters decides when to replace the full MLP with a linear surrogate. Through systematic investigation across six models (162M-2.8B parameters), two architectures, and three corpora, we establish that nonlinearity need cannot be predicted from token identity: cross-corpus correlation is zero ($r < 0.05$). The routing decision is fully contextual. Despite weak per-instance predictability, the gate exploits a heavily skewed distribution where most MLP computations are near-linear, achieving 25-56% linear routing at <1% perplexity cost in GPT-2. In GPT-2 Large, 11 of 36 layers beat baseline with gating and no layer exceeds 3.7% all-linear cost. This success is architecture-dependent: Pythia models show higher costs, though Pythia-2.8B's full 32-layer sweep reveals one layer that narrowly beats baseline. As a proof of concept, we progressively replace middle-layer MLPs with frozen linear matrices: 5 of 24 layers linearize at zero cost. With a full training budget, 4 linearized layers yield a 10.2% perplexity improvement -- and a two-phase gated approach pushes this to 17.3%, beating a vanilla fine-tuning control and confirming that the nonlinear MLPs at these layers were actively harmful.

We present GLUScope, an open-source resource for analyzing neurons in Transformer-based language models (LMs), intended for interpretability researchers. We focus on the now-standard GLU-based activation functions such as SwiGLU, used in virtually all recent open-weights LMs. This introduces a new challenge: understanding positive activations is not enough. Instead, both the gate and the in activation of a neuron can be positive or negative, leading to four different possible sign combinations (activation quadrants) that in some cases have quite different functionalities (e.g. promoting vs. suppressing a specific set of tokens). Accordingly, for any neuron, our tool shows text examples for each of the four sign combinations, and indicates how often each combination occurs. Our resource can lead to novel insights: in particular, thanks to analyzing all four activation quadrants, it can surface interpretable behavior that other tools would miss. We also release the code used to create the resource, thus enabling researchers to build on it by adding more models, text corpora, or neuron pages.

The architectural shape of dense Transformers has remained remarkably stable: narrow-wide-narrow feed-forward networks (FFNs) consume most non-embedding parameters. Motivated by theoretical and empirical evidences that residual wide-narrow-wide (hourglass) MLPs remain expressive despite bottlenecks, we revisit whether this architectural convention is necessary for dense language models. We study Hourglass Transformers, which replace the conventional FFN with residual stacks of hourglass sub-MLPs and use hourglass attention to decouple residual-stream width from attention width. This exposes a practical depth-width trade-off: compressing the FFN intermediate dimension allows wider hidden states and fewer layers at matched parameter budgets. Across model scales from 113M to 8B parameters, Hourglass Transformers achieve language-modeling and downstream performance comparable to conventional Transformers, while improving training compute efficiency by $8.7\%$ at matched average downstream accuracy across the 906M, 3B, and 8B scales. After long-context extension, the 8B Hourglass model also outperforms its matched conventional baseline across 4k-64k context lengths. At 64k context, the reduced attention layer count lowers both computation and KV-cache requirements, yielding up to $1.93\times$ faster token decoding and $50\%$ lower KV-cache memory at the 1B scale. These results identify hourglass structures as a practical architecture-efficiency alternative for compute- and latency-conscious Transformer design.

Recent research has established that the impact of context in a vanilla transformer can be represented implicitly by forming a token-dependent, rank-1 patch to its MLP weights. This work extends that foundational theory to the diverse architectures of modern Large Language Models. We first demonstrate a precise, analytical solution for a Gemma-style transformer block, proving that the entire effect of a context can be perfectly mapped to rank-1 patches on its MLP weight matrices and a patch to the RMSNorm scale. We then generalize this result, providing a constructive proof and algorithm for multi-layer models. To unify these findings, we introduce a general framework centered on two core properties: input controllability and output controllability. We prove that a perfect implicit weight patch is possible for any MLP block where the inner function is input-controllable and the outer function is output-controllable. This provides a simpler and more powerful lens for understanding how transformer models transmute prompts into effective weights. This setup generalizes to a wide range of modern LLM architectures including gating, pre-/post-norm, mixture of experts and sequential/parallel transformer blocks.

Prevailing activation functions such as Swish and GELU tend toward domain-specific optima, Swish was discovered via neural architecture search on vision benchmarks, while GELU dominates transformer-based language models, and neither offers any mechanism to adapt its gating shape to individual layers. This rigidity is especially consequential in transformer FFN blocks, where LayerNorm, unlike BatchNorm, does not suppress the gradient pathologies that activation choice induces across depth. We propose SG-Blend, a per layer adaptive activation that combines SSwish, a bias-corrected, parametric Swish variant we also introduce, with learnable sharpness \b{eta} and zero-centering bias γ, with GELU through a per-layer blend coefficient α, letting each layer locate its own optimum along the SSwishGELU continuum at a cost of only three additional scalars per FFN block, with \b{eta} initialized to 1.0 and learned freely via backpropagation. On BERT-style IMDB classification (5 seeds), it matches peak accuracy (81.31%) while reducing seed-to-seed variance by 42% relative to GELU. Furthermore, it generalizes to autoregressive pretraining, achieving the lowest validation perplexity (49.10) on WikiText103 among all baselines. Crucially, ablations confirm the interpolation structure itself drives these gains, delivering reliable, top-tier performance. Beyond natural language processing, we demonstrate that SG-Blend generalizes robustly to a wider variety of tasks, extending its efficacy to computer vision and other diverse domains.