Software

This page collects public code repositories and software artifacts from my recent work.

Counts update automatically from Hugging Face and GitHub. Downloads are all-time totals across each model family's official Hugging Face releases, including GGUF and MLX variants.

LLM Agents and Long-Context Learning

NeoHorse-1

Open-weight 4B and 9B models for agent workflows.

• Focus: Agentic post-training for tool use, coding, and instruction following, with routing signals connecting harness execution to model learning.

• My role: Led post-training strategy, experiments, and final checkpoint selection.

• Links: Technical Report, GitHub, 4B Model, 9B Model.

NeoHorse-Jev-4B

Prefill-only structured decisions for agent workflows.

• Focus: Given application state and questions, returns typed answers and probabilities for routing requests, selecting tools, checking conditions, and rating outcomes.

• Release: Open weights and an inference runtime, with GGUF variants for local deployment.

• Links: Model and Documentation, GGUF, GitHub.

SCOPE

Self-evolving Context Optimization via Prompt Evolution.

• Problem: LLM agents receive large, dynamic execution contexts but often fail to manage useful signals from traces, errors, and tool outputs.

• Method: Treats context management as online prompt optimization, evolving agent prompts from execution traces.

• Links: Paper, Code.

• Keywords: LLM agents; prompt evolution; context optimization.

MemDLM

Memory-Enhanced DLM Training.

• Problem: Diffusion language models need stronger training signals for long-context and denoising dynamics.

• Method: Introduces memory-enhanced training with a parametric memory channel.

• Links: Paper, Code.

• Keywords: diffusion language models; parametric memory; long-context training.

Efficient LLMs and Model Compression

CMoE

Analytical FFN-to-MoE Restructuring via Activation Pattern Analysis.

• Problem: Dense LLMs are expensive at inference time, while training MoE models from scratch is costly.

• Method: Converts dense FFNs into sparse MoE layers through analytical post-training restructuring based on activation patterns.

• Links: Paper, Code.

• Keywords: FFN-to-MoE; analytical restructuring; LLM acceleration.

PreMoE

Proactive Inference for Efficient Mixture-of-Experts.

• Problem: MoE models are usually deployed with full expert capacity even when a deployment scenario may only need a subset.

• Method: Compiles sparse MoE variants before serving through expert pruning and deployment-specific sparsity.

• Links: Paper, Code.

• Keywords: MoE expert pruning; proactive inference; efficient deployment.

FuseGPT

From Pruning to Grafting: Dynamic Knowledge Redistribution via Learnable Layer Fusion.

• Problem: Structured pruning removes whole transformer blocks and can discard useful knowledge.

• Method: Reframes pruning as learnable knowledge grafting, redistributing knowledge from pruned blocks into surviving layers.

• Links: Paper, Code.

• Keywords: structured pruning; layer fusion; knowledge grafting.

LLM Evaluation

Behavioral Fingerprinting

Behavioral Fingerprinting of Large Language Models.

• Problem: Accuracy-only benchmarks miss cognitive and alignment-style differences between LLMs.

• Method: Profiles models through behavioral dimensions such as reasoning, robustness, sycophancy, and interaction style.

• Links: Paper, Code.

• Keywords: LLM evaluation; behavioral fingerprinting; alignment robustness.

Other Tools

EvolveClaw

Make your OpenClaw agent self-improving.

• Problem: Agent systems can benefit from self-improvement loops rather than static behavior.

• Method: Provides a TypeScript tool for evolving OpenClaw-style agents.

• Links: Code.

• Keywords: agent systems; self-improvement; TypeScript tooling.

ArxivFinder

arXiv Paper Finder and Notifier.

• Problem: Keeping track of relevant arXiv papers can be time-consuming.

• Method: Provides a lightweight paper-finding and notification utility.

• Links: Code.

• Keywords: research tooling; arXiv monitoring; paper discovery.