Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Adversarial learning can serve as an effective post-training correction for pixel diffusion models, which generate RGB images directly without an autoencoder but systematically underrepresent fine-scale natural-image sta…
VideoLoop, a new method for long-form video understanding, counters "semantic thrashing" in multimodal agents by looping working memory instead of appending to it, according to the paper's authors. The approach targets t…
An AI system built by Markus Buehler's lab at MIT wrote its own atomistic simulation toolkit — a geometry generator, fracture solver, validation module, parameter scanner and three browser-based virtual labs — then ran h…
A from-scratch Rust language model called PSSA, which replaces transformer attention with a recurrent state-space core and a 512-slot episodic memory bank, reached 3.98 training cross-entropy versus 4.43 for a matched tr…
A developer released Zet, an open-source Apache-2.0 layer built on top of the Laya agent framework that uses conformal prediction sets and a Learn-then-Test threshold, calibrated per language, to mark each answer as sure…
A new reinforcement learning technique called Off-Policy-Aware Cross-Model Trajectory Exchange targets the all-fail group problem in Reinforcement Learning with Verifiable Rewards (RLVR) methods such as GRPO, where finit…
An empirical study examines why World Action Models (WAMs) generalize, comparing Explicit WAMs, which denoise future frames into clean video alongside every action chunk during inference, against Latent WAMs, which avoid…
Researchers propose TGRL (Temperature-Grouped Reinforcement Learning), a method that groups rollouts by temperature to improve exploration efficiency in reinforcement learning with verifiable rewards (RLVR) for large lan…
Researchers introduced Self-Play Search Distillation, a method that improves reasoning in large language models by generating high-quality training data through self-play search, addressing data scarcity caused by low-qu…
On NVIDIA Blackwell B200 GPUs, tensor-core throughput exceeds special-function exponential throughput by more than two orders of magnitude, exposing exponential evaluation as a bottleneck in fused attention kernels, acco…
A developer adapted Decision Transformers — the sequence-modeling reinforcement learning architecture from Chen et al. at Berkeley — to intelligent tutoring for heritage languages with fewer than 2,000 fluent speakers an…
TypeSafe AI released Jev, a "System One" decision model that returns typed probabilities instead of generated text, in early access on September 15, 2026, alongside a $40 million seed round led by DCVC. Construct integra…
Hugging Face launched the Open TTS Leaderboard, an objective-metric evaluation for multilingual text-to-speech and voice cloning that cuts model evaluation time from roughly two weeks of vote collection to a couple of ho…
A systematic study by Iuri Macocco, Pau Rodríguez Lopez, Arno Blaas, Luca Zappella, Marco Baroni and Xavier Suau Cuadros, published in Transactions on Machine Learning Research (TMLR) in September 2026, found that effici…
Jane Street published a review of its 2026 summer intern projects, including ML research on an event-level generative model of market data built with autoregressive diffusion by intern Kavish Kondap and work by Monte Boh…
Jane Street ML Infra intern Arsh Koneru built an activation checkpointing scheme that, given a model spec and a memory budget, selects a set of nodes to drop from the save set to minimize extra compute time while staying…
Jane Street ML Research intern Monte Bohde found that divergence decoding — inference-time steering in logits space using two smaller auxiliary models trained on pre-2015 and pre-2026 data — reduced pre-cutoff performanc…
Jane Street research intern Kavish Kondap built an event-level generative model of US equities market data using autoregressive diffusion, trained on four years of US equities data with each row containing timestamp, pri…
In week 11 of the Weekly AI Scorecard, covering Sep 21 to Sep 27, the 36 bots on the leaderboard resolved 550 calls at 55% overall (303/550), with the AI bot group at 59% (180/307) versus 51% (123/243) for mechanical bas…
Cohere launched Embed 5 on September 30, 2026, with Pro priced at $0.12 per million text tokens and Fast at $0.08, while image tokens cost $0.40 per million for either variant, according to the company's launch announcem…