Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new attention-free architecture that replaces Transformer attention with a stack of autoencoder-based mixing modules achieves a significant portion of attention's performance at about 1.9x fewer FLOPs when pretrained o…
A survey published on arXiv (2609.30292v1) reviews 211 studies on fake review detection from 2018 to early 2026, tracing the field from traditional machine learning and deep learning to pre-trained language model (PLM) a…
A new arXiv paper (2609.30298v1) introduces SRBench, a benchmark dataset of 45,064 labeled entries drawn from 32 curated secondary studies, for evaluating large language model performance in systematic review screening. …
A new arXiv paper (2609.31571v1) reports that self-training on strategically diverse data outperforms IID sampling, with self-training on strategically diverse but incorrect traces from Qwen3-4B beating IID distillation …
RAZOR, a training-free expert pruning method for mixture-of-experts (MoE) large language models, achieved the highest nine-task macro average among evaluated pruning methods in all eight settings tested on GLM-4.7-Flash,…
A study on arXiv (2609.30540v1) finds that the CLAP score, the standard objective metric for text-to-music faithfulness, fails to capture fine-grained musical meaning such as attribute bindings. Testing four contrastive …
LAVOIR, a method that adds value-of-information routing to single-pass decision encoders, matches a greedy oracle question policy on seen schemas (AUC 0.799 vs. 0.797) and is 14.1 points more accurate than never asking w…
A paper on arXiv (2609.30784v1) proposes a symbiotic architecture that gives frozen large language models audio-understanding capabilities by having an injector module write audio-conditioned vectors directly into the LL…
A new arXiv paper (2609.30977v1) reports that explicit time conditioning is often unnecessary in uniform discrete diffusion models (UDMs) for language, with time-agnostic predictors remaining competitive with and often o…
Tencent researchers presented KuaFu, a unified behavior-compression layer that compresses each user behavior item into 2-4 tokens of width 128-256, cutting per-item cache from 10 KB to 0.5 KB, according to the arXiv pape…
Researchers introduced JevAdvBench, described as the first adversarial benchmark for reinforcement learning for calibrated decisions (RLCD) models, with 812 typed questions across 66 scenarios and a black-box attack suit…
A study of thirteen LLM-based textual measures across seven LLMs from different providers found cross-model rank correlations averaging only 0.52, with transcript-level differences common across providers accounting for …
A new arXiv paper (2609.31603v1) introduces Belief Self-Distillation (BSD), a read-write framework that learns a compact user representation from a frozen LLM's own natural conversations and can both decode and write bel…
Researchers introduced AgentWorld, a benchmark of 100 human-annotated tasks designed to test long-horizon collaboration among multi-agent LLM systems, arguing existing multi-agent benchmarks focus on competitive settings…
Researchers introduced RFpeptides, a denoising diffusion-based deep learning pipeline for de novo design of protein-binding macrocycles, reporting medium to high affinity binders against all four diverse protein targets …
A developer building er-api, a multilingual entity resolution service for Korean, Japanese, Chinese, and English corporate data, found that Splink's full-corpus blocking, pairwise scoring, and EM calibration cannot run i…
TypeSafe AI released Jev, its first System One model, which takes unstructured state and returns typed decisions with calibrated probabilities rather than chat or code output. Jev is built on three primitives — Choice, S…
A new method called SLCA-GRPO addresses cross-segment credit misattribution in tool-calling reinforcement learning, targeting a structural failure mode where standard on-policy RL algorithms such as GRPO indiscriminately…
Researchers introduced FuseReg, a regularization method that mitigates the reconstruction-generation gap in representation autoencoders (RAEs) by regularizing layer fusion. The work addresses the open question of which p…
A developer's attempt to train a reinforcement-learning agent to complete the 1983 ZX Spectrum game Manic Miner failed, first with PPO over joystick inputs and then with move macros, according to a first-person account o…