Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A proposed zeroth-order paradigm for LLM preference alignment addresses likelihood displacement, a problem that motivates alternative ways to extract information from preference pairs with small likelihood margin. The ap…
A method called SpectralShift extends the effective context window of Gated DeltaNet, a linear attention architecture used in place of softmax attention for long-context modeling, by reparameterizing the model's spectral…
An open-source developer published an arXiv paper (2503.23303), a Hugging Face model (DeepMostInnovations/sales-conversion-model-reinf-learning), and a training dataset (DeepMostInnovations/saas-sales-conversations) in M…
A new arXiv paper (2609.17804v1) reports that large language models solve grade-school math word problems through a four-stage sequential pipeline — Schema Abstraction, Operation Planning, Operand Binding, and Computatio…
Researchers introduced GVD (Governed Versioning and Deduplication), a framework that unifies cross-document version linking with rule-level conflict resolution under an auditable update policy, according to an arXiv pape…
A new arXiv paper (2609.17847v1) introduces "individuated utility" functions that condition reward models on both the individual and their decision context, arguing that annotator disagreement in subjective tasks reflect…
A study measuring 54 configurations of Qwen2.5-7B-Instruct on vLLM 0.12 across L4, A100, and H100 GPUs found that 18 of 36 configurations reach the cost, quality, and latency Pareto frontier, with combined optimization m…
Researchers proposed OBC-Prune, an outcome-based calibration method for pruning large reasoning models that weights reasoning tokens by their causal contribution to correct answers rather than by uniform statistical sali…
A study posted to arXiv (2609.17532v1) uses large language model-derived features from free-text respiratory therapy notes plus a logistic regression pipeline to improve prediction of extubation failure (EF) in invasivel…
Researchers released DantinoX, an open-source JAX/Flax library that runs three language-generation paradigms — autoregressive decoding, discrete masked diffusion, and continuous flow-matching — on a single modular Transf…
A systematic benchmark of open-source instruction-tuned LLMs for key-value pair extraction found that models including Gemma, Mistral, Qwen2.5, LLaMA 3, and DeepSeek approach supervised layout-aware systems on clean text…
A new arXiv paper (2609.17571v1) reports that grokking in Transformers is a spectral recoding of an existing distributed circuit rather than a switch between modules, based on behavior-aligned exact activation games the …
Researchers introduced Fathom, a per-query read-depth key scan for sparse decoding over offloaded KV caches, which at one million tokens on Qwen3-8B runs a decode step 1.67x faster in GPU time than the 136-bit scans of D…
The University of Bozen/Bolzano has released lecture notes from its 2025/2026 PhD course on Physics Informed Neural Networks, published as arXiv:2609.17638v1. The notes cover Physics Informed Deep Neural Networks (PINN) …
Researchers introduced Diffusion Skill Discovery (DSD), a skill discovery method that uses a diffusion model to approximate the entropy gradient of a policy-induced state distribution through score matching, according to…
A paper posted to arXiv (2609.17654v1) proposes a regularized least squares approach for training quadratic neural networks that yields a lower bound on the training optimization problem's solution when the regularizatio…
A new arXiv paper (2609.17691v1) introduces "draft trees," a speculative sampling method that accelerates diffusion model generation by up to 8.3% over the reflection maximal coupling baseline in practical settings. The …
Researchers introduced REVERSAL-BENCH, a benchmark that controls environmental reversibility via a continuous parameter ρ ∈ [0, 1] and provides a reset oracle to test state recoverability across eight manipulation settin…
A paper published on arXiv (2609.17697v1) proposes composite-gradient learning (CGL), a method that integrates a model predictive control (MPC) controller into deep reinforcement learning (DRL) training by treating the D…
Researchers proposed Sharp Mode Connectivity (SMC), a sharpness-aware reformulation of mode connectivity that enforces low loss across the entire neighborhood of a weight-space interpolation curve rather than only along …