Research Notes

Tracking developments that may change how decentralized machine learning systems are built and understood.

A high-signal, system-level research log covering papers, releases, and discussions.

May 2026

Tuesday
Paper

VISTA: Decentralized Machine Learning in Adversary Dominated Environments

Akbari Nodehi et al. proposed VISTA, a decentralized ML framework using incentive-driven adaptive thresholding for robust aggregation in adversary-dominated settings, achieving convergence comparable to standard SGD.

Hanzaleh Akbari Nodehi DeML Incentive Design
Monday
Paper

Can Blockchains Reliably Train Machine Learning Models?

Li et al. proposed Proof of Training (PoT) in their paper, repurposing PoW mining power for verifiable ML training while preserving PoW incentives and security. Experiments show high throughput, strong robustness, and improved network security.

Peihao Li PoW PoT
Sunday
Paper

AgentReputation: A Decentralized Agentic AI Reputation Framework

Chishti et al. propose AgentReputation, a three-layer decentralized reputation framework for agentic AI marketplaces. It separates task execution, reputation services, and tamper-proof persistence, introducing context-conditioned reputation cards and a decision-facing policy engine to address strategic optimization, task transfer, and verification cost challenges.

Mohd Sameen Chishti Reputation
Saturday
Paper

TRUST: A Framework for Decentralized AI Service v.0.1

Huang et al. proposed TRUST, a decentralized AI auditing framework in their paper, using HDAG for reasoning decomposition, DAAN for causal attribution, and multi-tier consensus, achieving 72.4% accuracy under 30% adversarial nodes.

Yu-Chao Huang Decentralized AI

April 2026

Tuesday
Paper

Cloudless-Training: A Framework to Improve Efficiency of Geo-Distributed ML Training

Tan et al. proposed Cloudless-Training in their paper, using a two-layer architecture, elastic scheduling, and two new sync strategies (ASGD-GA and MA) to improve geo-distributed ML training efficiency, achieving up to 1.7x speedup and 24% cost reduction on Tencent Cloud.

Wenting Tan Geo-Distributed ML
Wednesday
Paper

A Tale of Two Learning Algorithms: Multiple Stream Random Walk and Asynchronous Gossip

Gholami and Seferoglu propose and analyze the Multi-Walk (MW) algorithm and Asynchronous Gossip in their paper, comparing their convergence, computation, and communication overhead under various graph topologies and data heterogeneity.

Peyman Gholami Gossip Data Heterogeneity
Monday
Paper

SpecMoE: A Fast and Efficient Mixture-of-Experts Inference via Self-Assisted Speculative Decoding

Bang et al. proposed SpecMoE, a memory-efficient MoE inference system based on self-assisted speculative decoding in their paper, which boosts throughput and reduces bandwidth without extra training.

Jehyeon Bang LLM Inference
Thursday
Paper

Tessera: Unlocking Heterogeneous GPUs through Kernel-Granularity Disaggregation

Hu et al. proposed Tessera, the first kernel-level GPU disaggregation system in their paper. It enhances throughput and cost efficiency for large model inference on heterogeneous GPU clusters via offline analysis and online adaptive scheduling.

Tiancheng Hu GPU Disaggregation
Tuesday
Paper

On Harnessing Idle Compute at the Edge for Foundation Model Training

Xue et al. proposed Cleave in their paper, a system for foundation model training using idle edge compute. Its key innovation exploits the asymmetric I/O of GEMM operations to match network bandwidth, reducing per-device communication as more devices join.

Leyang Xue Edge Computing GEMM
Saturday
Paper

MoBiE: Efficient Inference of Mixture of Binary Experts under Post-Training Quantization

Zhao et al. proposed MoBiE, the first binarization framework tailored for MoE-based LLMs, in their paper. It employs joint SVD, global gradient integration, and error constraints to significantly boost post-quantization performance and inference speed.

Zhixiong Zhao MoE Binarization
Thursday
Paper

MoE Routing Testbed: Studying Expert Specialization and Routing Behavior at Small Scale

Researchers proposed the MoE Routing Testbed in their paper to clearly evaluate expert specialization and routing dynamics at small scale, identifying balancing scope as a key factor for specialization.

Tobias Falke MoE Routing Training
Tuesday
Paper

Patterns behind Chaos: Forecasting Data Movement for Efficient Large-Scale MoE LLM Inference

Yu et al. analyze data movement bottlenecks in large-scale MoE LLM inference in their paper, proposing optimization strategies that achieve a 6.6x speedup on wafer-scale GPUs.

Zhongkai Yu MoE Profiling
Saturday
Paper

Trust-Aware Routing for Distributed Generative AI Inference at the Edge

The authors of G-TRAC propose a trust-aware coordination framework, G-TRAC, in their paper for reliable generative AI inference on decentralized edge devices, enhancing robustness via a risk-bounded shortest path algorithm and a hybrid trust architecture.

Chanh Nguyen Edge Generative AI
Friday
Paper

Real-Time AI Service Economy: A Framework for Agentic Computing Across the Continuum

This paper analyzes resource allocation for real-time AI services across the device-edge-cloud continuum. It finds that the structure of service-dependency graphs is key to whether decentralized, price-based allocation can scale reliably.

Lauri Lovén Edge Computing Mechanism Design
Thursday
Paper

MoEless: Efficient MoE LLM Serving via Serverless Computing

The MoEless team proposed the first serverless computing-based MoE model serving framework in their paper, significantly reducing inference latency and cost through expert load prediction and optimized resource scheduling.

Hanfei Yu Serverless Inference
Wednesday
Paper

MoE-Sieve: Routing-Guided LoRA for Efficient MoE Fine-Tuning

Researchers propose MoE-Sieve in the paper, which applies LoRA adapters only to frequently activated experts based on routing analysis, significantly reducing fine-tuning compute and storage costs.

Andrea Manzoni MoE LoRA

March 2026

Tuesday
Paper

Route Experts by Sequence, not by Token

Wen et al. proposed SeqTopK in their paper, shifting the expert routing budget from token to sequence level for end-to-end dynamic allocation, improving model performance while maintaining the same overall budget.

Tiansheng Wen Routing Sparsity

February 2026

Monday
Paper

SeedFlood: A Step Toward Scalable Decentralized Training of LLMs

Kim et al. proposes SeedFlood, a novel decentralized training method designed to scale large models across complex networks. By leveraging the seed-reconstructible nature of zeroth-order updates, it reduces communication overhead to near-zero and independent of model size, significantly improving training efficiency.

Jihun Kim Gossip Zero-th Order Optimization
Saturday
Paper

Privacy-Preserving Mechanisms Enable Cheap Verifiable Inference of LLMs

Pal et al. proposes a novel insight that privacy-preserving LLM inference methods can enable verifiable inference at marginal extra cost. It designs two new protocols that add minimal computational tokens to provide guarantees, significantly reducing verification overhead.

Arka Pal Verifiable Inference
Friday
Paper

Node Learning: A Framework for Adaptive, Decentralised and Collaborative Network Edge AI

The paper proposes 'Node Learning', a decentralized paradigm where intelligence resides at individual edge nodes and expands via selective peer interaction, addressing scaling bottlenecks of centralized AI in heterogeneous, mobile, and resource-constrained environments.

Eiman Kanjo Edge AI
Thursday
Paper

Agentic AI Reasoning for Mobile Edge General Intelligence: Fundamentals, Approaches, and Directions

Luo et al. propose a joint optimization framework for Mobile Edge General Intelligence (MEGI) in their paper to address the challenges of deploying LLM-based agentic AI reasoning on resource-constrained edge devices.

Mingyi Luo Edge Computing MEGI MoE
Thursday
Paper

The Rise of Sparse Mixture-of-Experts: A Survey from Algorithmic Foundations to Decentralized Architectures and Vertical Domain Applications

This paper provides a comprehensive survey of the Sparse Mixture-of-Experts (MoE) architecture, covering recent advancements from algorithmic foundations to decentralized architectures and vertical domain applications, serving as a resource for researchers and practitioners.

Dong Pan MoE Survey
Wednesday
Paper

Distributed Hybrid Parallelism for Large Language Models: Comparative Study and System Design Guide

This paper provides a comprehensive review of parallel strategies for distributed LLM training and inference, offering methodological guidance for optimal system design through mathematical models and case studies.

Hossam Amer Parallelism
Tuesday
Paper

InfiCoEvalChain: A Blockchain-Based Decentralized Framework for Collaborative LLM Evaluation

The authors propose a blockchain-based decentralized LLM evaluation framework in their paper. It incentivizes global contributors as independent validators for large-scale benchmarking across heterogeneous compute nodes, addressing issues like opacity and hardware variance in centralized evaluations.

Yifan Yang LLM Evaluation
Sunday
Paper

Mugi: Value Level Parallelism For Efficient LLMs

In the paper, Price et al. study VLP for LLMs, extending it to nonlinear ops and small-batch GEMMs, and propose the Mugi architecture to improve throughput, energy efficiency, and sustainability.

Daniel Price VLP GEMM
Saturday
Paper

Privacy-Preserving LLM Inference in Practice: A Comparative Survey of Techniques, Trade-Offs, and Deployability

In the paper, Andreoletti et al. survey privacy-preserving LLM inference techniques, comparing TEE, crypto-augmented designs, and FHE in terms of trust, scalability, and deployment readiness under a strong end-to-end privacy model.

Davide Andreoletti Confidential Inference
Thursday
Paper

Mosaic Learning: A Framework for Decentralized Learning with Model Fragmentation

In the paper, Biswas et al. introduced Mosaic Learning, a decentralized learning framework that uses model fragmentation across multiple gossip graphs to improve node-level performance, especially under non-IID data.

Sayan Biswas Fragmentation Non-IID Gossip