Research & Publications
Pioneering research in AI infrastructure, decentralized systems, and frontier models. All papers are CC BY 4.0 licensed and available on GitHub.
Research Areas
AI Alignment
Developing techniques to ensure AI systems reliably pursue intended goals and behave safely.
Learn moreAI Safety
Building robust safeguards and interpretability tools for understanding model behavior.
Learn moreDecentralized AI
Distributed inference, federated learning, and privacy-preserving AI infrastructure.
Learn moreZero-copy Agent Protocol - MCP killer with ~500x faster performance via zero-copy RPC
Model Context Protocol - infrastructure for AI context management and tool use
Minimalist ML framework for Rust with GPU acceleration
Unified multimodal AI framework for vision, language, and audio
React component library for AI-powered applications
Multi-agent orchestration framework for complex AI tasks
AI-powered development environment and CLI tools
ZAP: Zero-copy Agent Protocol
Hanzo AI Research
The MCP Killer—unified protocol using zero-copy RPC for ~500x faster agent communication. Features zero-copy performance (<1μs latency), capability-secure routing, native metastable consensus, and post-quantum security. Achieves 1.2M/s throughput vs MCP's 2.2k/s.
GT-QLoRA: Gate-Targeted Fine-Tuning for Behavioral Modification in Large MoE Models
Zen LM Research
Gate-Targeted QLoRA for behavioral modification in trillion-parameter MoE models. Demonstrates why linear abliteration fails on MoE architectures (refusal encoded in expert routing, not residual stream) and introduces direct gate weight unfreezing alongside LoRA for effective behavioral steering.
Active Semantic Optimization (ASO): Training-Free Adaptation for Foundation Models
Hanzo AI Research
A novel framework for adapting foundation models to new tasks without additional training, using semantic optimization techniques. Achieves 18.2% SWE-bench resolution.
Hamiltonian Market Maker (HMM): Decentralized AI Compute Exchange
Hanzo AI Research
A market-making protocol for decentralized AI compute resources using Hamiltonian dynamics. Achieves <200ms quote latency, 98.7% price stability, zero impermanent loss.
Lux Quantum Consensus: Post-Quantum BFT Protocol
Lux Network Research
Byzantine fault-tolerant consensus protocol with quantum-resistant security using FALCON lattice-based signatures and dual-certificate finality.
LSSS Threshold Cryptography: Linear Secret Sharing Schemes
Lux Network Research
Implementation of Linear Secret Sharing Schemes (LSSS) for threshold cryptography with arbitrary access structures and efficient reconstruction.
Gym Training Platform: Decentralized Model Training at Scale
Zoo Labs Foundation
Infrastructure for distributed deep learning training across heterogeneous compute resources with significant cost reduction via TF-GRPO.
HLLM: Training-Free GRPO for Language Models
Zoo Labs Foundation
Training-free group relative policy optimization achieving $18 training cost vs $10,000+ traditional methods with 100× data efficiency.
Zen Family Overview: Model Architecture Guide
Zen LM Research
Comprehensive overview of the Zen model family (600M–1T+ parameters) with 22 variants across 6 size tiers.
fhEVM: Fully Homomorphic Ethereum Virtual Machine
Lux Network Research
EVM-compatible smart contract execution on encrypted data enabling confidential DeFi, private voting, and sealed-bid auctions without revealing underlying values.
Zen4 Abliteration: MoE-Aware Refusal Removal at Scale
Zen LM Research
Comprehensive analysis of abliteration techniques across the Zen4 model family. Covers orthogonalization for dense models and GT-QLoRA for MoE architectures including the 1.04T Zen4 Ultra.
Model Context Protocol: Standardized Tool Interface for LLM Augmentation
Hanzo AI Research
A standardized framework for tool use in large language models with 260+ production tools across filesystem, code search, shell execution, and agent delegation.
Agent Networks: Semantic Routing and Orchestration for Collaborative AI
Hanzo AI Research
Framework for multi-agent systems with intelligent routing strategies, handoff mechanisms, shared state management, and workflow orchestration.
Operative: Autonomous Computer Control via Vision-Language Models
Hanzo AI Research
System design enabling LLMs to autonomously control computers through vision, mouse/keyboard input, with safety isolation and multi-provider support.
LLM Gateway: Unified Multi-Provider API Translation and Optimization
Hanzo AI Research
Architecture for unified proxy layer abstracting 100+ LLM providers with intelligent routing, cost optimization, and 20-30% cost reduction through smart failover.
Decentralized Semantic Optimization with Byzantine-Robust Prior Aggregation
Hanzo AI Research
A distributed approach to semantic optimization that maintains robustness against Byzantine failures in decentralized AI systems.
Hanzo Network Whitepaper: Decentralized AI Infrastructure
Hanzo AI Research
Technical architecture for distributed AI inference and training across a decentralized network of compute providers.
Jin: Unified Multimodal Architecture for Text, Vision, Audio, and 3D
Hanzo AI & Zen LM Research
Novel architecture unifying text, vision, audio, and 3D modalities through shared embedding space with diffusion transformer and mixture-of-experts routing.
Hanzo Memory: Local Vector Database with Multimodal Embeddings
Hanzo AI Research
High-performance AI memory system using embedded vector database (InfinityDB), local embeddings (FastEmbed), and semantic search.
Hanzo Platform: Self-Hosted PaaS for AI Infrastructure
Hanzo AI Research
Self-hostable platform-as-a-service providing deployment, database management, monitoring with Traefik integration.
Post-Quantum Cryptography for Confidential AI Workloads
Hanzo AI Research
Integration of ML-KEM/ML-DSA post-quantum algorithms with privacy-tier framework for GPU TEE isolation in confidential computing.
Confidential AI: GPU TEE Isolation Patterns
Hanzo AI Research
Security architecture for running AI workloads in GPU trusted execution environments with Intel SGX and AMD SEV support.
Scaling Laws for AI Compute: GPU Cluster Optimization
Hanzo AI Research
Empirical analysis of scaling laws for distributed AI training and inference optimization across heterogeneous GPU clusters.
RAG Architecture: Retrieval-Augmented Generation Patterns
Hanzo AI Research
Design patterns for building robust retrieval-augmented generation systems in production with hybrid search and reranking.
Model Routing: Intelligent Request Distribution
Hanzo AI Research
Algorithms for routing AI requests to optimal models based on cost, latency, quality, and capability matching.
Quasar Consensus: High-Throughput DAG-Based Protocol
Lux Network Research
DAG-based consensus mechanism achieving high throughput with asynchronous finality guarantees and metastable voting.
FPC Consensus: Fast Probabilistic Consensus
Lux Network Research
Probabilistic consensus protocol for rapid finality using opinion dynamics and threshold voting.
Wave Consensus: Sub-Second Finality Protocol
Lux Network Research
Optimistic consensus protocol achieving sub-second finality for high-frequency transaction environments.
Focus Consensus: High-Throughput Streaming Protocol
Lux Network Research
Streaming consensus protocol optimized for continuous transaction throughput with parallel validation.
Snow++ Consensus: Enhanced Avalanche Protocol
Lux Network Research
Enhanced Snow consensus with improved liveness guarantees and adaptive sampling strategies.
ETHFalcon: Post-Quantum Signatures for Ethereum
Lux Network Research
Integration of FALCON lattice-based signatures into Ethereum-compatible chains for quantum resistance with EIP compatibility.
Ringtail: Threshold Post-Quantum Signatures
Lux Network Research
Threshold signature scheme combining FALCON with distributed key generation for post-quantum multi-party signing.
Universal Threshold Signatures: Multi-Party Signing
Lux Network Research
Threshold signature schemes for secure multi-party signing with configurable thresholds and dealer-free setup.
BLS-FALCON Hybrid Signatures
Lux Network Research
Dual-certificate signature scheme combining BLS aggregate signatures with FALCON for quantum-transitional security.
MPC Chain: Multi-Party Computation Infrastructure
Lux Network Research
Multi-party computation protocol for privacy-preserving blockchain operations with SPDZ and BGW protocols.
TEE Computing Mesh: Trusted Execution Network
Lux Network Research
Distributed trusted execution environment mesh for confidential computing with Intel SGX and AMD SEV attestation.
Threshold ECDSA on Lux: Distributed Key Management
Lux Network Research
Implementation of threshold ECDSA for distributed custody without any party learning the full private key.
Fraud Proofs: Optimistic Rollup Verification
Lux Network Research
Interactive fraud proof system for optimistic rollup security with bisection-based dispute resolution.
Verkle Trees: Efficient State Commitments
Lux Network Research
Verkle tree implementation for efficient blockchain state proofs with 10x smaller witness sizes than Merkle trees.
KZG Data Availability Sampling
Lux Network Research
KZG polynomial commitment scheme for data availability sampling in modular blockchain architecture.
Lightspeed DEX: High-Performance Exchange
Lux Network Research
High-throughput decentralized exchange with sub-second finality, concentrated liquidity, and MEV protection.
Credit Lending Protocol: DeFi Lending
Lux Network Research
Decentralized lending protocol with dynamic interest rates, isolated collateral pools, and liquidation mechanisms.
Perpetuals & Derivatives: On-Chain Derivatives
Lux Network Research
Perpetual futures and derivatives protocol with oracle-based price feeds and cross-margin support.
Lux Bridge: Cross-Chain Asset Transfer
Lux Network Research
Secure cross-chain bridge protocol for trustless asset transfers using light client verification.
Oracle Infrastructure: Decentralized Data Feeds
Lux Network Research
Decentralized oracle network for reliable off-chain data delivery with economic security guarantees.
DAO Governance Framework: On-Chain Governance
Lux Network Research
Comprehensive DAO governance framework with timelock, voting delegation, and treasury management.
Lux DID Specification: Decentralized Identity
Lux Network Research
W3C-compliant decentralized identifier specification for self-sovereign identity with key rotation support.
Lux IAM: Identity and Access Management
Lux Network Research
Decentralized identity and access management framework with verifiable credentials and selective disclosure.
Verifiable Credentials on Lux
Lux Network Research
Implementation of W3C Verifiable Credentials with zero-knowledge selective disclosure proofs.
Experience Ledger: Decentralized Semantic Optimization
Zoo Labs Foundation
Community-driven semantic optimization protocol for collaborative AI model improvement with Byzantine-robust aggregation.
Zoo Network Architecture
Zoo Labs Foundation
Technical architecture for the Zoo decentralized AI network with proof-of-contribution consensus.
Zoo Tokenomics: Incentive Design for Decentralized AI
Zoo Labs Foundation
Economic model for incentivizing contributions to decentralized AI training and inference.
ZIP-002: Zen Reranker Specification
Zoo Labs Foundation
Zoo Improvement Proposal for the Zen Reranker cross-encoder search reranking model.
Zoo Foundation Mission
Zoo Labs Foundation
Mission and vision for the Zoo Labs Foundation - open AI research network for decentralized science.
Zen Omni: Unified Multimodal Foundation Model
Zen LM Research
Single model architecture for text, vision, audio, and code understanding with emergent cross-modal reasoning.
Zen Coder: Code-Specialized Language Model Architecture
Zen LM Research
Architecture for code-specialized language models with multi-language support and repository-level understanding.
Zen Nano: Efficient Small Language Models
Zen LM Research
Techniques for training highly capable small models (600M-3B) for edge deployment with 95% energy reduction.
Zen Agent: Autonomous AI Agent Model
Zen LM Research
Agent-optimized model for tool use, planning, and autonomous task execution with ReAct-style reasoning.
Zen Guard: Safety and Moderation Model
Zen LM Research
Content safety and moderation model for filtering harmful content with constitutional AI techniques.
Zen Reranker: Search Reranking Model
Zen LM Research
Cross-encoder reranking model for improving search relevance with 30% MRR improvement over BM25.
Zen Artist: Image Generation Model
Zen LM Research
State-of-the-art image generation with style control, high-resolution output, and text fidelity.
Zen Video: Video Generation Model
Zen LM Research
Text-to-video generation model for creating dynamic visual content up to 60 seconds.
Zen 3D: 3D Generation Model
Zen LM Research
Text-to-3D generation model for creating 3D assets and environments with NeRF and Gaussian splatting.
Zen Musician: Music Generation Model
Zen LM Research
AI music composition model for creating original musical pieces across genres with stem separation.
Zen Scribe: Transcription and Generation
Zen LM Research
Speech-to-text model with 98% accuracy on conversational audio and speaker diarization.
Zen World: World Model for Simulation
Zen LM Research
World model for environment simulation, physics understanding, and planning in embodied AI.
Zen Eco: Energy-Efficient Model
Zen LM Research
Energy-optimized architecture achieving 95% inference cost reduction through quantization and sparsity.
Lux FHE: Threshold Fully Homomorphic Encryption
Lux Network Research
Independent threshold FHE implementation in Go and C++ for privacy-preserving computation with distributed key generation and multi-party decryption.
FHE-VM: Encrypted Smart Contract Execution
Lux Network Research
Virtual machine for executing smart contracts on encrypted data using our threshold FHE scheme with EVM compatibility.
Threshold FHE for Decentralized AI Inference
Lux Network Research
Privacy-preserving AI inference using threshold FHE where no single party can decrypt intermediate computations.
fheCRDT: CRDTs over Encrypted State
Lux Network Research
Conflict-free replicated data types operating on homomorphically encrypted data for privacy-preserving distributed systems and collaborative applications.
Verifiable Encrypted Voting
Lux Network Research
End-to-end verifiable voting system using FHE where votes remain encrypted throughout tallying while maintaining individual and universal verifiability.
Privacy-Preserving ML at Scale
Lux Network Research
Practical techniques for training and inference on encrypted data using FHE with GPU acceleration achieving 100x speedup over CPU implementations.
RISC Zero zkVM: Zero-Knowledge Virtual Machine
RISC Zero
General-purpose zero-knowledge virtual machine enabling verifiable computation with STARK proofs.
Plonk: Universal zkSNARK Construction
Gabizon-Williamson-Ciobotaru
Universal and updateable structured reference string zkSNARK with efficient verification.
Groth16: Efficient zkSNARK Proofs
Groth
Most efficient pairing-based zkSNARK with constant proof size and verification time.
STARKs: Scalable Transparent Arguments of Knowledge
Ben-Sasson et al.
Post-quantum zero-knowledge proofs without trusted setup using hash functions.
ML-KEM: Post-Quantum Key Encapsulation Mechanism
NIST Post-Quantum Cryptography Project
Lattice-based key encapsulation for quantum-resistant key exchange, standardized in FIPS 203.
ML-DSA: Post-Quantum Digital Signature Algorithm
NIST Post-Quantum Cryptography Project
Lattice-based digital signatures for quantum-resistant authentication, standardized in FIPS 204.
SLH-DSA: Stateless Hash-Based Digital Signatures
NIST Post-Quantum Cryptography Project
Hash-based signature scheme providing conservative quantum resistance, standardized in FIPS 205.