Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new arXiv paper (2609.30275v1) argues that interpretability transfer under quantization is routinely certified by scale-invariant statistics reported without their noise floor, and shows on Qwen2.5-1.5B-Instruct that t…
Researchers introduced SheafDEQ, a subhomogeneous deep-equilibrium architecture with adaptive neural-sheaf propagation, in arXiv paper 2609.30277v1, and proved it admits a unique equilibrium reached globally by fixed-poi…
Researchers at UBC-NLP introduced GyroNovo, a de novo peptide sequencing framework that uses decoder errors observed during training to guide fragment imputation and rotary embeddings to encode pairwise mass differences …
Researchers introduced MatFAE, a functional neural network that learns trajectories on the Riemannian manifold of symmetric positive definite (SPD) matrices, according to a new arXiv paper (arXiv:2609.30487v1). MatFAE us…
A new arXiv paper (arXiv:2609.30503v1) introduces FedTMLE, described as the first federated targeted maximum likelihood estimation algorithm, federating targeting itself for an arbitrary target, loss, and fluctuation fam…
An audit of one commercial non-generative "System-1" model on 6,020 multiple-choice biosecurity-relevant items found pooled expected calibration error of 0.034 and pooled AUROC of 0.820 once the vendor's uncertainty fiel…
A twelve-week production-scale run of Andrej Karpathy's AutoResearch paradigm on two book-recommendation embedding systems surfaced five recurring failure modes absent from the original setting, according to an arXiv pap…
A mechanism-aware conditioning framework that injects ensemble-covariance statistics into an otherwise unchanged backbone improves rare-event emulation in data-limited chaotic systems, according to arXiv paper 2609.30746…
A new arXiv paper (2609.30957v1) establishes a theoretical analysis showing that momentum accelerates optimization in the "river-valley" loss landscape of large language model pretraining by stabilizing large learning ra…
The vLLM community added a structured-read mode to DiffusionGemma 26B-A4B via vLLM PR #57250, letting the open model return typed probabilistic decisions in a single denoising step instead of token-by-token generation. T…
Researchers identified a failure mode they call Decision-Timestamp Mismatch in on-policy self-distillation (OPSD), a technique used to add dense privileged feedback to reinforcement learning with verifiable rewards (RLVR…
MahaBodi, a Rust-based enterprise router from Laya, decides among up to 100,000 candidates in a single forward pass and beat Laya with a dense shortlist on entity linking at 100K Wikipedia pages (0.177 vs 0.121, p = 0.00…
A study of training-inference mismatch in reinforcement learning with verifiable rewards (RLVR) for large language models finds that rollouts sampled by an inference engine and gradients computed by a training engine ass…
A round-trip study of sixteen language models found that serializing tree-structured arithmetic expressions into natural language is a lossy, asymmetric channel, with swapping the generating and extracting model shifting…
AMD trained a pick-and-lift reinforcement learning policy for a UFactory XArm-6 robotic arm on AMD Instinct MI300X and MI355X GPUs using MuJoCo Warp and PPO, then deployed the same network on the real arm with a classica…
An analysis of two Mixture-of-Experts models, Qwen 3.5 35B-A3B and Qwen 3.6 35B-A3B, captured per-token router statistics across every MoE layer using first-turn MT-Bench prompts, recording which experts the router selec…
A new research approach, budget-aware LoRA merging, challenges the assumption that every layer needs the same rank budget when combining task-specific low-rank adapters, according to the paper's authors. Existing LoRA me…
An AI system built by technology firm thymia with RMIT University in Melbourne assigned elevated type 2 diabetes risk scores to 80 per cent of 7,319 UK participants who reported having the condition, using 20-second clip…
Researchers developed SASP Score, a deep learning-based biomarker built on a Guided AutoEncoder with Transformer (GAET) architecture that evaluates the combined effects of senescence-associated secretory phenotype (SASP)…
Marc Brooker built Hobson, a roughly 2-billion-parameter calibrated classifier model, by replacing the LM head of a pre-trained Qwen3.5-2B torso with a pointer head of just over 1 million parameters and fine-tuning with …