Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
IronBee trained a 322M-parameter model called Laya to play browser games live by having an LLM write the game rules as code, with Laya making each move in 22 to 45 ms (about 24 ms per decision in Chrome Dino and roughly …
An open notebook maintained by MV2 reports a new bound on the growth base of rank-r Kakeya sets over a fixed finite field, showing beta(q,r) = q - q^(2-r) + O(q^(1-r) log q) and giving constructions smaller than the 2010…
A paper submitted to arXiv on 21 Dec 2025 argues that AI interpretability methods including feature attribution, probing, sparse auto-encoding and causal analyses can produce plausible-looking explanations for randomly i…
Israeli AI startup doubleAI's autonomous "Double Agent" system found a new explicit construction in extremal graph theory that raises the girth coefficient of regular graphs from 4/3 to 8/5, improving a bound that had st…
Google DeepMind researchers Maxence Faldor and Moritz Firsching published a repository of formally verified 2-symbol Turing machines with Lean 4 semantics proofs for Goldbach's Conjecture, ZF Set Theory Consistency, and …
Microsoft researchers Edward J. Hu and colleagues proposed Low-Rank Adaptation (LoRA) in a paper submitted to arXiv on 17 June 2021, a method that freezes pre-trained model weights and injects trainable rank decompositio…
A local ensemble of three QLoRA fine-tuned LLM auditors (Qwen2.5, SmolLM3, and phi4) was benchmarked against Jev, TypeSafe's hosted typed decision model, on 20,000 industrial record-linkage pairs at a 41.3% match rate. J…
AI-Atlas, a platform built to turn public GitHub repositories into an explainable map of demonstrated AI skills, evaluates each project across seven dimensions — Technical Depth, Conceptual Difficulty, Originality/Modifi…
Alicia Colmenero-Fernández published an arXiv paper on 14 Oct 2026 (arXiv:2610.20437) forensically analyzing and locally recreating RL/RLHF-induced privilege escalation in AI agents, tracked as CVE-2026-65616. The paper …
An internal SFEIR proof-of-concept comparing TypeSafe AI's decision-oriented SaaS model Jev against local models on 30,000 real IT support tickets found that a fine-tuned Laya model reached 62.0% department accuracy and …
ClawLoop, a harness-based agent training method detailed on arXiv, lets file-system/shell agents self-improve in a closed loop without a stronger teacher model by synthesizing and validating their own sandbox tasks, trai…
A developer outlines a practical approach to building AIoT (AI plus IoT) systems for industrial settings, arguing that teams should define the business problem before selecting hardware, models or cloud platforms. The ac…
A developer outlined a fitness app personalization approach that separates eligibility filtering from ranking, arguing that a rules engine with structured workout metadata and explicit constraints is sufficient for most …
Researchers posted arXiv:2610.19212v1, describing Decoupled Drift-Diffusion World Model (D3WM), a transformer trained with a Gaussian negative-log-likelihood objective that jointly predicts the deterministic next frame a…
A U-Net trained with an ℓ1-regularized loss that combines reconstruction error with spatial difference structural preservation improved image restoration under myopia-modeled defocus, according to an arXiv paper (arXiv:2…
Researchers introduced LAMP, a framework that uses vision-language models to perform point-conditioned, source-aware illumination reasoning from a single RGB image with marked query points. On the Infinigen test set, fin…
Researchers proposed Progressive Multistep Data-free Distillation (PMDD), a data-free multistep diffusion distillation method that assigns a dedicated adapted network per student distribution to avoid the biased distribu…
A Proximal Policy Optimization agent trained only to optimize reward in the Crafter environment learns a world model inside its value function, according to an arXiv paper (2610.14186v1). Using counterfactual interventio…
A paper posted to arXiv (2610.19095v1) reports that Chatbot Arena's Bradley-Terry ranking model violates Luce's choice axiom, with votes between versions of the same model proving 1.37 times as decisive as the model pred…
A new arXiv paper (2610.14005v1) presents the first test based on Levin's Kt complexity that uses large language models as representational machines, producing certified test items at difficulties measured in bits that n…