routing

This paper introduces elbow-based routing, a training-free inference-time method for MoE models that dynamically adjusts the number of active experts per token by detecting the elbow point in router probability distributions, achieving a 5.3% average latency reduction while maintaining accuracy.

This paper introduces SpecDrop, a parameter-free category-conditioned routing scheme for modular networks, showing that on vision tasks it achieves competitive accuracy while on fuzzy language partitions it reduces to no-routing baselines, suggesting granularity alignment matters more than router design.

This paper introduces MetaRoute-Bench, an open benchmark for evaluating meta-decision policies in agentic workflows, comparing routing policies on success, cost, and latency under a shared execution model.

This paper introduces an executable benchmark for learning compositional meta-routing in agentic workflows, where a budget-aware controller composes operations like retrieval, code execution, and verification. The learned policy outperforms static workflows on held-out tests but shows lower lexical generalization on challenge splits.

The paper introduces Hierarchical Copula-Gumbel-Top-K (H-CGA) routing, a method to control joint dependence among token routing choices in frozen Mixture-of-Experts models while keeping each token's routing law exactly fixed. It provides theoretical trade-offs between coherence and load dispersion and validates the mechanism with a small-scale pilot.

This paper presents ReBA (Relax Within, Balance Across), a geometry-guided load-balancing method for vision-language mixture-of-experts models. It addresses modality-specific routing imbalance by balancing image and text separately and grouping correlated visual tokens, improving robustness across resolution and composition shifts.

LlamaIndex introduces Parse Gateway, a page-level document OCR router that estimates each page's complexity and routes it to the appropriate parsing tier (LiteParse or LlamaParse), balancing cost, latency, and quality.

This paper proposes Routing-based On-Policy Distillation (ROPD), a safety realignment framework that uses two frozen teachers to preserve task performance while restoring safety, and shows it is more robust to prompt-template mismatch than existing defenses.

This paper introduces the Expert Subspace Separation Index (ESSI) to disentangle route coherence, candidate quality, and candidate-by-context interaction in sparse Mixture-of-Experts language models, revealing a pattern of coherent overlap where routing selects token-relevant experts from a shared geometric neighborhood yet multi-expert computation remains useful.

Introduces cMoLLM, a convolutionally gated mixture-of-LLMs that routes over end-to-end streams through fully differentiable dynamic convolution, improving perplexity and downstream accuracy under matched compute.

A proposed SLA-constrained carbon-aware routing policy for geo-distributed serverless clouds achieves up to 46.8% carbon reduction while maintaining zero SLA violations, evaluated on real AWS deployments.

MCPHub is an open-source centralized management and routing hub for connecting multiple MCP servers to AI clients, offering a unified dashboard, flexible routing, and access controls.

MoE2-LoRA introduces a dual-channel Routing-Conditioned Projection and a global LoRA expert pool to enable MoE-style low-rank adaptation for fine-tuning MoE models, achieving state-of-the-art accuracy while retaining general capabilities.

A guide to self-hosting servers at home using L2TP tunnels and static IPs, bypassing ISP restrictions like CGNAT.

A technical guide on setting up multicast TV distribution on a home network using ffmpeg and multicast routing protocols, with examples for local and routed scenarios.

This page describes the spacefilling curve heuristic for generating approximate solutions to the Traveling Salesman Problem, emphasizing its speed, simplicity, and practical applications in routing, logistics, and map drawing.

Reveals that MoE routing in LLMs follows a Frequency-Diversity Law analogous to Huffman coding, and proposes Subset Difference Pruning to eliminate functional redundancy and achieve optimal compression.

This paper explores Relative Positional Encoding (RPE) as an additive bias in Transformer architectures to solve the Team Orienteering Problem, demonstrating consistent improvements in collected rewards and optimality gaps over vanilla Transformer architectures.

Introduces Multi-Head Attention Residuals (MHAR), which reshapes the routing query into per-subspace heads so each feature subspace reads depth history via its own softmax. Trained from scratch on a Nemotron-based corpus, MHAR consistently improves validation loss over standard Transformers from 100M to 1B scale and also boosts mid-training downstream accuracy.

Kimi K3, an open model, rivals Fable in agentic tasks, and routing between them achieves 93% accuracy with up to 50x cost savings, establishing a new state-of-the-art approach.