cosinegate-dynamic-routing

name	cosinegate-dynamic-routing
title	CosineGate: Semantic Dynamic Routing via Cosine Incompatibility
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.22206
keywords	["dynamic-routing","efficient-inference","residual-networks","gating"]
description	Achieve efficient neural networks via self-supervised dynamic routing using Cosine Incompatibility Ratio (CIR). Ground gating decisions in geometric novelty rather than learned heuristics, enable per-sample/per-block binary routing via Gumbel-softmax, constrain with progressive FLOPs regularization—maintaining accuracy while reducing computation 28.5% on CIFAR-10.

Overview

Dynamic routing based on semantic novelty between identity and residual paths.

Core Technique

Cosine Incompatibility Ratio:

# CIR = 1 - cos(x, F(x))
# Low CIR: identity sufficient (skip block)
# High CIR: block adds novelty (execute)

cir_score = 1.0 - cosine_similarity(input_x, residual_output)
should_skip = cir_score < threshold

Differentiable Gating: Gumbel-softmax for training, deterministic for inference.

Performance

• 91.3% accuracy with 28.5% FLOPs saving
• No auxiliary supervision needed

References

• Cosine incompatibility ratio for routing
• Geometric grounding of gate decisions
• FLOPs-constrained gating