Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A research direction called Orthrus achieves parallel token generation in large language models without altering the output distribution, generating up to 32 tokens per forward pass by inserting a trainable diffusion att…
A team of engineers building recommendation systems for financial services clients found that collaborative filtering fails structurally in fintech due to sparse interaction data, cold-start problems, and regulatory comp…
Researchers have developed Architecture-driven Shift (ADS), a lightweight metric that captures logit shift trends in continual learning models using few data samples, addressing the high computational cost of traditional…
Researchers at arXiv have identified that Batch Normalization causes gradient skew in dynamic sparse training (DST) methods, leading to slower convergence compared to dense neural network training. The team proposed Spar…
Large language models fundamentally cannot perform causal discovery from observational data due to a mathematical limitation proven in a new study, which shows that fine-tuning and other learning methods produce predicto…
Researchers have developed representation-conditioned diffusion models that generate synthetic training images by conditioning on learned representations from DINOv2, DINOv3, and CLIP, achieving a 10.76 percentage point …
Researchers introduced DynaSchedBench, a diagnostic framework for the Dynamic Flexible Job Shop Scheduling Problem that uses a Sequential Event-Space Calibrator to control instance difficulty through a novel Schedule Str…
A new benchmarking study from arXiv:2605.27467v1 finds that Liquid Neural Networks (LNNs) significantly outperform traditional LSTMs in parameter efficiency and robustness across four sequential data modalities, includin…
Researchers introduced Sequential Bayesian Belief Tracking (SBBT), a framework that estimates the likelihood of a correct final answer from partial reasoning traces by calibrating observation likelihoods and updating a t…
Researchers have developed RAG-Coding, a method that uses four large language model agents to automate ICD-10-CM medical coding by grounding decisions in official coding tabular lists and guidelines. On the MDACE dataset…
Researchers in Japan fine-tuned the LLaVA-1.5-7B vision-language model on up to 4,000 bridge damage images to automate damage assessment and repair priority scoring, addressing significant inter-rater variability in mand…
Researchers have developed ICG, a framework that uses multimodal large language models and personalized preference alignment to generate cover images tailored to individual users. The system extracts semantic features fr…
Researchers developed a translation model that converts soft prompts—parameter-efficient but uninterpretable LLM adaptations—into natural language text. The translated prompts outperformed existing interpretability metho…
The 10th Affective & Behavior Analysis in-the-Wild (ABAW) Workshop and Competition, held at CVPR 2026, introduced new challenges for multimodal human-centered AI, including emotional mimicry intensity estimation, ambival…
Researchers introduced MTM-Bench, a controlled benchmark that isolates three distinct language roles—instruction, content, and response—across English, Spanish, and Chinese to evaluate multilingual LLM task execution. Te…
Researchers at Simorgh have developed a region-aware hybrid retrieval method combining BM25 lexical matching and dense semantic similarity with regional weighting heuristics to improve culturally grounded question answer…
A new jailbreak framework called PAST2HARM has achieved up to 100% attack success rates against state-of-the-art multimodal AI systems, including Gemini Nano, GPT Image 2, and SD XL, by exploiting past tense reformulatio…
Researchers at Poolside released two new Mixture-of-Experts AI models, Laguna M.1 and Laguna XS.2, designed for long-horizon software engineering tasks. The 225.8-billion-parameter M.1 and 33.4-billion-parameter XS.2, tr…
Researchers have released GenSBI, an open-source library for simulation-based inference built entirely in JAX, implementing flow matching, score matching, and denoising diffusion methods. The library offers three transfo…
Researchers have introduced LaneRoPE, a novel positional encoding method that enables multiple language model sequences to collaborate during parallel reasoning and generation tasks. The approach uses inter-sequence atte…