Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A developer introduced OnTrack, a real-time agent monitoring system that uses streaming optimal transport to compare an agent's unfolding execution trajectory against reference runs, flagging divergence in roughly a mill…
A local evaluation of Bespoke Labs' Nimble 9B decision model against its base model Qwen3.5–9B on a nine-team consumer complaint routing task found the published benchmarks did not come from routing, the task the author …
Immich released version 3.3, adding shared people management, birthday memories, automatic stacking of edited photos, and machine learning optimizations, the open-source self-hosted photo and video management project ann…
ByteDance's Seed team identified a mechanism behind inconsistent long-context retrieval in DeepSeek models, finding that chunked KV-cache compression makes retrieval accuracy vary by up to 40 percentage points depending …
A developer has released ej, an 11 MB open model that answers typed decision questions — choice, yes/no, and ordered score — by returning one probability distribution per question in a single forward pass without generat…
Underdog Saluki 27B, a 7.89 GB 2-bit GGUF quantization of Qwen3.8-27B released under Apache 2.0, outperforms the 54 GB original model on tool calling while losing ground on competition math and reasoning, according to Ma…
A proposed research program applies homotopy type theory, contextual topos semantics, and neural tangent kernels to model AI companies as open, resource-dependent systems with feedback loops among observation, learning, …
Researchers found two bugs in FlashAttention 3 (FA3) that caused gradient norm to grow a thousandfold and loss to end 0.2 nats above FP32 attention when pretraining a 450M-parameter transformer on 50B tokens, with traini…
Qualcomm has filed a patent for storing Mixture of Experts AI models as one shared base plus small per-expert delta maps, converting models after training without retraining, according to the filing. In the patent's exam…
Researchers released a camera-native stereo VR180 dataset of 1,211 samples captured with two Blackmagic URSA Cine Immersive cameras, comprising 636 stereo video clips (2,220.8 seconds, mostly 90 fps) and 575 stereo still…
Researchers proposed SAVER, a dense-to-sparse post-training framework that trains video-language models to reason from limited temporal evidence, according to a new arXiv paper (2610.10893v1). Trained on only 1,250 rando…
Researchers proposed VICO, a co-evolutionary framework that jointly trains a vision-language model actor and an Environment-as-Rewriter that edits verifiable image-side structures such as scene graphs, chart tables, and …
A new arXiv paper (2610.10859v1) introduces a preference-driven unlearning framework that removes targeted concepts such as identity and NSFW content directly from few-step distilled (FSD) text-to-image diffusion models.…
SPLIT-RL, a staged post-training approach that trains visual reasoning and language reasoning in disjoint phases, improves average accuracy over GRPO by 1.4 to 6.1 points across Qwen3-VL models from 2B to 30B-A3B and Int…
A new arXiv paper (2610.10738v1) proposes a lossy text autoencoder that compresses web text to 2.24 bits per byte, a rate the authors report is on par with lossless text compression algorithms. The architecture uses resi…
A paper posted to arXiv (arXiv:2610.10959v1) introduces TopoGPU, a GPU streaming pipeline that computes persistence diagrams of cubical complexes induced by 2D and 3D images, achieving an average end-to-end speedup of 53…
Omni-Diffusion-Distill, a two-stage distillation framework from arXiv paper 2610.10990v1, compresses unified multimodal diffusion large language models to few-step decoding, cutting image generation from 128 to 8 steps a…
Telefónica Deutschland's o2 online shop recorded a 275% conversion-rate increase and a 17% click-through rate on personalised hero banners after deploying an event-driven AI recommendation pipeline built on Google Cloud'…
A new arXiv paper (arXiv:2610.10550v1) introduces Diffu-LoRA, a parameter-efficient method that personalizes text-to-image diffusion models by learning how to allocate low-rank adaptation capacity across layers through g…
A deployed 0.8B-parameter Qwen3.5 small language model fine-tuned on filtered, rebalanced teacher-generated supervision outperformed prompted GPT-5.4 and GPT-5.6 Sol in precision and recall on two human-curated grammar c…