研究与动态
Serving Causal Encoder–Decoder Models with MoL2
If no LoRA touches the encoder or the layer-20 attention, every LoRA can share one encoder state and one global KV. A LoRA switch then replays 128 tokens with no measurable quality loss for the LoRAs we tested, and the routing decision takes 2.77 ms in the reference implementation.
阅读
Deep Principle Hydrogen-1: Post-Training GLM5.3 for Scientific Discovery | Mint Recursive
Hydrogen-1.0 is an open-weight scientific model post-trained from GLM-5.3 by Deep Principle and Mind Lab on Mint Recursive, connecting scientific reasoning with simulation, experiment planning, and feedback-driven materials discovery.
阅读
Adapting Causal Encoder–Decoder Models with LoRA
We made LoRA training work on DeepSeek-V4.1-Flash's causal encoder–decoder and found that adapting only the decoder is enough, sharing one compressed KV cache.
阅读
Introducing Macaron-V1.1 and Mint Recursive
Macaron-V1.1 is a 752B-parameter agent model post-trained from GLM-5.3, with Mint Recursive enabling faster experience-driven training and evaluation.
阅读
Scaling PEFT Towards Brain Foundation Models
Scaling PEFT for brain encoding: adapt one large shared multimodal backbone with small neural adapters, taking first place on CNeuroMod/Algonauts 2025 with ~5% of the trainable parameters and 100x faster training.
阅读
Muon vs AdamW for LoRA: When the optimizer changes, where should the learning rate move?
Swapping AdamW for Muon in LoRA usually shifts the best learning rate upward, but the rank-to-learning-rate transfer pattern is more stable than the absolute scale.
阅读
Serving Macaron-V1 at Scale: Page-Level 2D KV Resharding
Page-level 2D KV resharding that keeps Prefill CP layer-split and Decode DCP page sharding consistent during cross-node KV transfer.
阅读
Introducing Macaron-V1
Macaron-V1 is Mind Lab's agent model in 748B Venti and 35B Tall variants, built on Mixture-of-LoRA and model-harness co-design.
阅读
UI4A: A Component-Native Harness for Generative UI
UI4A combines frontend-code expressiveness with an auditable, cross-framework runtime designed around how agents naturally build interfaces.
阅读
Efficient LLM Agents with REPL Harnesses: Executable Composition and Validated Reuse
A controlled study of how action substrates change agent success, turns, token cost, and the value of validated executable reuse.
阅读
Macaron-V1-Preview: 749B MoL Agent Model post-trained from GLM5.1
Macaron-V1-Preview
阅读
MinT Cookbook: From Reproducible Baselines to Continuous Autoresearch
MinT Cookbook turns autoresearch into a standardized experiment practice, letting AI agents reproduce baselines, optimize methods, and scale continuous research loops on MinT infrastructure.
阅读
Macaron-A2UI: A Model for Generative UI in Personal Agent
Macaron-A2UI trains Personal Agents to produce protocol-grounded Generative UI with short prompts and measurable rendered quality.
阅读
Support GLM5 and GLM5.1 in MinT: LoRA training for DSA and MTP
Bringing GLM5 and GLM5.1 into MinT turned out to be a full-stack effort across training, inference, and checkpoint conversion rather than a routine model bring-up.
阅读
0.03% Parameters, 100% Potential: Reflexivity of Compute Optimal Rank Scaling
LoRA rank in reasoning RL is not a bigger-is-better scaling knob: low rank already shows promise, mid rank is the most reliable operating region, and high rank adds cost without improving the performance frontier.
阅读
OpenClaw Skill Security: Intent-Capability-Behavior Consistency as a New Framework
A new framework for OpenClaw skill security, showing why traditional scanners miss language-driven risk and why ICB matters.
阅读
Triquetra for LoRA's entangled knobs: When LoRA rank changes, how should learning rate move?
Why LoRA learning-rate transfer across ranks is not universal? The key factor is how alpha scales with rank: different scaling rules lead to different rank-to-learning-rate behaviors.
阅读
MindClaw: Fine-Tuning OpenClaw for Personalized Long-Term Memory
MindClaw is Mind Lab's online system for turning OpenClaw skills into parametric memory with MetaClaw, MinT, and LoRA RL for personal agents.
阅读
From Context Engineering to Context Learning
Exploring the paradigm shift from Context Engineering to Context Learning, where transient test-time gains become permanent model capabilities.
阅读
Router Replay R3: Why It Failed and How We Fixed It
We fixed Router Replay R3 in vLLM and veRL to eliminate training-inference mismatch in MoE models without discarding samples.
阅读
Streaming Memory Benchmark: Diagnosing Memory with Evidence-Grounded Episodes
Streaming Memory Benchmark: evidence-grounded streaming episodes for stage-level diagnosis (formation, retrieval, application), measuring accuracy, latency, and token cost to expose inference masking and evaluate deployable long-term memory systems.
阅读
MinT: RL Infrastructure for Experiential Intelligence
MinT: managed RL & post‑training turning real product experience into better open models; supports Qwen, DeepSeek, Kimi, multimodal/robotics, LoRA; bridges training to inference.
阅读
From Words to Worlds: How World Models Unlock Scalable Agentic RL
Can LLMs serve as world models for agentic learning? We study scaling laws, fidelity, and agent utility across five text environments—showing how world models enable reasoning, adaptation, and safer RL.
阅读
Exploring Agentic Memory beyond Reasoning and Tool-Use
Memory Diffusion is a Mask–Allocate–Refill framework that makes AI agent memory dynamic, enabling intelligent forgetting and efficient long-context reasoning.
阅读
How We Build Trillion Parameter Reasoning RL with 10% GPUs
Mind Lab shows how to run efficient LoRA reinforcement learning on trillion-parameter MoE reasoning models, matching full RL with only 10% of the GPUs.
阅读
Introducing Mind Lab: Building AI that Learns from Real Experience
Mind Lab builds AI minds that learn from real-world experience, using experiential intelligence and agentic RL to turn static models into adaptive products.
阅读