arXiv subjects

Yurii Bilak

Publications and source records attributed to Yurii Bilak.

4 recordsLinked to original sources

arXiv2026

Mixture of Layers with Hybrid Attention

Standard Mixture-of-Experts (MoE) transformers route tokens to expert subnetworks within each layer, but the layer structure itself remains monolithic. We introduce Mixture of Layers (MoL), which replaces full-width transformer blocks (d_model) with K parallel thin blocks at reduced dimensionality (d_thin << d_model), connected via learned down/up projections and composed via top-k block routing. Scaling sparse block routing to many blocks creates an attention coverage problem, as each block sees fewer tokens. We address this by introducing hybrid attention, which pairs one shared softmax block for global context with Gated DeltaNet linear attention in routed blocks.

cs.LG↗

arXiv2026

Equifinality in Mixture of Experts: Routing Topology Does Not Determine Language Modeling Quality

Sparse Mixture-of-Experts (MoE) architectures employ increasingly sophisticated routing mechanisms -- learned routers, multi-hop trajectories, token-dependent gating. We ask: does routing topology actually determine language modeling quality? We build a geometric MoE (ST-MoE) using cosine-similarity routing against learned centroids in a low-dimensional space ($d_{space} = 64$), requiring 80% fewer routing parameters than standard linear routers. Through 62 controlled experiments on WikiText-103 at 76--84M parameters trained to convergence (50K steps, 1.64B tokens), we find that routing topology does not determine asymptotic perplexity (PPL): five cosine-routing variants are statistically equivalent within a 1-PPL margin (Two One-Sided Tests [TOST], $p < 0.05$ for all 10 pairwise comparisons; 15 runs across 3 seeds, observed range 33.93--34.72). The finding extends to hash, random-fixed, and top-1 routing (single-seed; graceful 1.1--2.2 PPL degradation) and replicates on OpenWebText (0.03 PPL gap, 6 runs, 3 seeds each). A standard linear router with 5.3$\times$ more routing parameters reaches PPL 32.76, but iso-parameter cosine routing closes 67% of this gap -- the true mechanism advantage is $\sim$1.2%. The mechanistic explanation is convergent redundancy: multi-hop updates are collinear ($\cos(\Delta h_0, \Delta h_1) = 0.805$), implementing magnitude amplification rather than compositional reasoning; a single learnable scalar replicates multi-hop performance. As a practical payoff, zero-shot relative-norm halting saves 25% of MoE FLOPs at +0.12% PPL. Expert-level specialization and causal controllability -- which coexist with topology-level equifinality -- are explored in a companion paper.

cs.AI↗

arXiv2026

Geometric Routing Enables Causal Expert Control in Mixture of Experts

Sparse Mixture-of-Experts (MoE) models scale parameters while fixing active computation per token, but the specialization of individual experts remains opaque. In a companion paper we showed that routing topology is quality-neutral: five structurally different configurations converge to statistically equivalent language modeling quality. Here we show that expert identity is nonetheless causally meaningful: individual rank-1 experts are monosemantic by construction, and cosine-similarity routing in a low-dimensional metric space makes their specialization directly inspectable. We present four lines of evidence. First, projecting expert output vectors through the unembedding matrix yields a Semantic Dictionary: 15% of experts are monosemantic specialists spanning 10 categories (temporal, geographic, cardinal, discourse, emotional, financial, military, scientific). Second, routing exhibits a frequency-to-syntax gradient: early layers separate tokens by word frequency, deeper layers by syntactic class (Zipf-confound controls, all $p < 0.001$). Third, causal interventions confirm these labels: steering toward a temporal expert's centroid increases P(temporal) by +321% (median across 44 prompts); suppressing a geographic expert drops P(geographic) by -23%; rewriting an expert's output vector halves target-category probability, and effects compose additively across layers. Fourth, the interventions are not unique to cosine routing: linear routers support comparable steering, but only cosine routing provides geometric transparency -- expert specialization is readable directly from the centroid matrix. MoE expert-level specialization is a first-class interpretability primitive: architecturally monosemantic, causally validated, and controllable at inference with zero overhead.

cs.AI↗

arXiv2025

Cosine-Similarity Routing with Semantic Anchors for Interpretable Mixture-of-Experts Language Models

Mixture-of-Experts (MoE) models improve efficiency through sparse activation, but their learned gating functions provide limited insight into routing decisions. This work introduces the Semantic Resonance Architecture (SRA), which routes tokens to experts via cosine similarity between token representations and learnable semantic anchors, making every routing decision directly traceable to anchor-token similarity scores. We evaluate SRA on WikiText-103 across 17 configurations. In a controlled multi-seed comparison (3 seeds x 4 configurations, 256 experts, $D_{ff}=256$), cosine routing achieves competitive perplexity with standard linear routing ($12.57 \pm 0.03$ vs $12.45 \pm 0.03$ for $K=1 \to 4$; $12.52 \pm 0.02$ vs $12.57 \pm 0.02$ for $K=2 \to 4$). The training recipe -- not the routing function -- drives specialization quality, while cosine routing provides inherent inspectability. We introduce a bandpass routing loss -- a floor-and-ceiling corridor on expert utilization -- that reduces dead experts from 30-45% to 0-6% and transfers to both routing types. Routing-space evaluation shows cosine routing provides significantly better word-level subtoken coherence in deeper layers ($p < 0.001$), with 44-54% of expert specialization being syntactic rather than semantic. Extended analysis reveals cosine routing maintains more stable router saturation and tighter per-expert vocabulary distributions -- structural advantages from the bounded cosine similarity range. An inference-time $k$-sweep shows that $k=5$ yields a free 0.08-0.16 perplexity gain over $k=4$. Cross-dataset validation on OpenWebText confirms generalization: cosine routing achieves comparable perplexity (44.88 vs 45.44), the bandpass loss eliminates dead experts, and specialization patterns are preserved.

cs.CL↗