AI Briefing#lora
The latest AI and developer news about #lora, with the original source and a short summary.
The latest AI and developer news about #lora, with the original source and a short summary.
A new script and container allow users to fine-tune TRELLIS.2 with LoRA on custom 3D assets, significantly improving geometric accuracy for specific domains like footwear.
The feature allows users to train LoRA adapters for specific models on specific tasks using designated harnesses like hermes, pi, or opencode.
The new adapters reduce image generation from 40 to 6 steps and remove the separate prompt-upsampling phase, saving 41 to 80 seconds per image.
Kokon 0.10 introduces a 'Prompt+' feature that uses the existing Qwen3-VL 8B encoder to expand prompts, adding roughly 2 seconds of processing time on an RTX 5070 Ti, alongside Windows-only support for custom Qwen-Image 2.1 LoRAs.
The new LoRA improves coordinate accuracy from 41% to 74% compared to the base model by returning page numbers and bounding boxes for every extracted field.
A rank-16 LoRA distilled from representation analysis cuts the slop score by 83% without system prompts or significant quality loss.
The open-weight model reduces thinking tokens to roughly half of the base model's usage while maintaining valid tool calls in agent loops.
The ML Intern agent trained, evaluated, and published six specialized models, including a 0.8B prompt rewriter and a 4-step text-to-image distillation, on Hugging Face hardware for about USD 103.
Both studies achieved an F1 score of 90.14% on the ADReSSo benchmark, surpassing the previous best record
Jared Palmer released Kev, a lightweight decision-making model based on Qwen3.5, offering simultaneous processing of Yes/No, multiple-choice, and rating tasks with Apple Silicon support.
Researchers at the University of Waterloo have released PAW, an open-source framework that compiles English text descriptions into small Neural Programs executable locally.
Video Rebirth has open-sourced 'HyperFlow,' a data-free self-distillation LoRA that reduces MiniMax-H3 inference steps from 49 to 8.
Fyxer achieved a 90% 90-day user retention rate and a 53% draft acceptance rate with an AI assistant trained on 500,000 hours of human EA data.
AgileRL has released AgileRL Arena v1.0, featuring a manifest-based reinforcement learning SDK and LoRA LLM fine-tuning capabilities.
TRL v1.14 introduces Async GRPO functionality that synchronizes LoRA adapters using storage buckets and proxies instead of NCCL.
The MultiMatte model, which fine-tunes SAM 3 with LoRA to isolate specific objects via text prompts, has been released, showing significant performance improvements on major benchmarks.
H3-World is a technology that enables precise text-based control of character and camera movements with minimal data.
Mind Lab released the LoRA-based Macaron-V1, demonstrating the potential of continual learning.
Ramp Labs has open-sourced PorTAL, a framework that supports cross-model LoRA adaptation.
A public leaderboard has launched to compete on LoRA fine-tuning speed under identical hardware and dataset conditions.
This introduces the latest research methodologies of Foundation models for tackling time series data, classified according to how they utilize LLMs.
LG AI Research introduces a case where it combined EXAONE's Image Understanding and Generation technologies to realize the Hwadamsup media art.
A new LoRA methodology that reduces LLM hallucination by leveraging Sub-Riemannian geometry has been unveiled.
Research has been published on preventing fine-tuning poisoning by restricting the training scope to a subspace of trusted LoRA adapters.