AI Model2026
Provides low-bit quantized Hy3 (hy_v3) GGUF model weights and mixed-precision quantization recipes for running Hy3 on llama.cpp, with optional MTP self-speculative decoding and imatrix-based calibration for improved quality/speed trade-offs.
Enables RL post-training with million-token prompts under a fixed GPU budget by evaluating shared prompt state without autograd, retaining only minimal model state, and replaying short response branches; instantiated as GRPO and demonstrated on Qwen3.6-27B and GLM-5.2 up to multi-million token execution.
AI Dataset2026
Provides intermediate pretraining checkpoints for the Aether-7B-5Attn base model to enable reproducible training-dynamics research. Includes three raw checkpoints (110k, 115k, 162k steps) packaged with model.safetensors, config, and tokenizer; uses a custom aether_v2_7way architecture requiring the aether_pkg loader.
A PyTorch-native training framework for agentic reinforcement learning research that keeps researcher-facing code compact and editable. Uses an asynchronous loop to train multimodal and mixture-of-experts policies while never training on tokens the agent didn't generate; matches Megatron-style stacks under a comparable protocol and ships recipes and containers on GitHub.
Performs full-parameter post-training of trillion-parameter MoE DeepSeek-V4 models on an Ascend NPU SuperPOD, using a hierarchical optimization of model parallelism, communication orchestration, and kernel execution to increase Model FLOPs Utilization. Also builds CPT/SFT pipelines with solver-verified synthetic data for Operations Research, reporting strong zero-shot Pass@1 results.
AI Model2026
A 124B hybrid-linear Mixture-of-Experts language model optimized for instruction following, long-context reasoning and agentic workflows, activating ~5.1B parameters per token. Key features include a 256K native context (extendable to 1M), alternating KDA/MLA attention layers, and vLLM/SGLang inference support.
A pretrain-then-transfer method for streaming recommendation that decouples refreshable behavioral knowledge from task-specific geometry to enable continual model refresh without downstream interference; introduces Behavioral Multi-Token Prediction and Anchored Calibration Residual and shows 4–12% offline gains plus live Shopee A/B lifts.
Tao Feng, Fangxu Yu +10University of Illinois Urbana-Champaign, University of Maryland, College Park +3
Frames LLM routing as a sequential decision process and introduces LLMRouter plus the xRouteBench benchmark to develop, evaluate, and deploy learned routing policies across heterogeneous LLMs, optimizing response quality versus inference cost.
AI Dataset2026
Provides a machine-readable catalog of 117 AI/AX safety and deployment-readiness diagnostic criteria for assessing model intrinsic and serving/infrastructure risks. Includes MODEL-SCAN and AX-SCAN axes, bilingual source fields, per-item evidence guidance, severity/assurance metadata, and a CC BY-NC 4.0 release-candidate.
AI Agent Papers2026
Trains compact conversational agents to adapt at runtime to changing 'Harness' configurations (Skills, Hooks, prompts, tools) using Harness-Aware Training (HAT): Harness-State Augmentation, on-policy distillation, and RL to preserve generality while meeting low-latency deployment constraints.
AI Model2026
Post-trained Qwen3.8-27B variant using the COLD FUSION (GAIN+Unsloth) tuning to reduce internal reasoning-token use and improve instruction following while keeping base capabilities. Deliverables include 256k-context-compatible GGUF quants (regular and MTP, NEO IMATRIX), vision support via an mmproj, and three reasoning-effort modes (xhigh/medium/low).
Enables interactive serving of large Mixture-of-Experts (MoE) models on personal machines by adapting offload and execution to measured device bandwidth and agentic workload patterns. Key features include bandwidth-adaptive execution, semantic-aware caching of recurrent state, and an elastic GPU expert cache; supports 20+ MoE models and runs models from ~35B to 753B on consumer/workstation GPUs.