Machine Learning

Google DeepMi

2026-10-26 13:28:46

DeepMind says standard Softmax and cross-entropy implementations contain numerical bugs, with Gemini credited for the initial fix

Google DeepMind researcher Jon Barron says standard implementations of Softmax and cross-entropy in PyTorch, JAX and even textbook formulations contain two subtle numerical bugs. According to Barron and a new arXiv paper, one issue can drive the correct gradient to zero when model confidence gets high enough, while another can leave a residual “ghost gradient” in distillation even when a student model exactly matches its teacher. The paper presents rewritten formulations intended to avoid adding extremely small values to 1 and to keep calculations in log space with functions such as expm1 and log1p. Barron also said in an X post that Gemini Flash 3.7 first found and fixed the bug, Gemini Flash 3.8 ran most of the experiments, and Gemini 4 Argon drafted the paper and figures. The reported results show test accuracy in a modular addition task rising from 93.3% to 100.0% in bfloat16 after the fix, while an independent engineer, Satyajit Ghana, said he reproduced the mechanism in PyTorch on CPU and confirmed the fp32 results described in the paper.

20

EVO-WAM

2026-10-19 10:21:18

EVO-WAM uses self-generated trajectories and dual verification to lift performance on unseen robotic tasks

A research team has introduced EVO-WAM, a training framework aimed at improving World Action Models, or WAMs, on unseen robotic tasks without adding new expert demonstrations or testing every candidate action in an external environment. The method asks the model to generate full video-and-action trajectories on its own, then filters those trajectories through two checks: a vision-language model evaluates whether the task shown in the generated video is actually completed, while an inverse dynamics model tests whether the generated actions are consistent with the motion depicted in the video. Only validated segments are kept and fed back into training. According to the reported results, the Cosmos3 backbone improved its average success rate on seven unseen simulation tasks in RoboTwin 2.0 from 26.9% to 68.0% after four rounds of self-training. DreamZero rose from 28.5% to 46.4% on the same benchmark. In three unseen long-horizon real-world tasks, including bowl stacking, placing two ducks into color-matched bowls, and opening an air fryer drawer before inserting bread, Cosmos3 increased from 20.0% to 76.7%, with each policy tested 10 times per task. The paper also reports that using only visual completion checks produced a 43.7% average success rate in the fourth round, while adding action-consistency verification raised that figure to 68.0%.

510

Schrödinger

2026-10-16 16:01:41

Schrödinger says machine learning simulations uncover atomic-level behavior in lithium-ion battery electrolytes

Schrödinger has published a new lithium-ion battery study in Nature, saying its team worked with researchers from Columbia University and the New York Structural Biology Center to identify how battery electrolytes behave at the atomic level. The company said the work used its physics-based AI computing platform and machine learning force field technology to move battery research beyond trial-and-error methods toward predictive design. According to the report, the study focused on electrolyte chemistry, which affects conductivity, viscosity, salt dissociation, electrochemical stability, and the formation of the solid electrolyte interphase on the anode surface. Schrödinger said earlier efforts to model complex battery chemistry often forced researchers to choose between quantum-level accuracy and large-scale simulations. Vice President Mat Halls said the company’s machine learning force field removed that trade-off. The team reported that lithium ions in electrolyte frequently form clusters coordinated by ethylene carbonate solvent molecules, and that these clusters can exist without anions. The study said as much as 40% of lithium ions cluster in standard electrolyte conditions, with clustering becoming more common in high-salt systems. Schrödinger added that its simulation results were consistent with experimental measurements including NMR spectra, electrolyte density, viscosity, ionic conductivity, and diffusion coefficients. The company said the findings could offer data support for designing faster-charging, higher-capacity, and longer-lasting lithium-ion batteries.

600

Google

2026-10-15 17:36:05

Google says AI still isn’t accelerating AI research, even as it writes most of DeepMind’s code

Google DeepMind says it found no sign that AI is materially speeding up the creation of more advanced AI systems, according to the safety report for Gemini 4 Argon, the lab’s most powerful model. The finding pushes back on fears around recursive self-improvement, the idea that AI could rapidly build smarter versions of itself until progress outruns testing, governance, or shutdown controls. Researcher Haydn Belfield said there is currently no significant end-to-end acceleration in machine learning research and development, whether AI-driven or not. DeepMind based that conclusion on a broad internal assessment that combined capability benchmarks, a macro capability index, data on AI-written code, and trajectory outcomes. The report shows that more than 75% of DeepMind’s production code is now written by AI, and the share of fully delegated pull requests rose from 7.5% to 11.9% between April and September. Even so, about 88% of pull requests still need multi-turn human steering or direct human code authorship, and all code changes remain subject to human review before submission. DeepMind also said Argon appears above trend on its internal capability index, but attributed that to increased compute and organizational investment rather than AI-led acceleration.

470

Transformer

2026-10-12 04:20:34

BUPT team’s MiTA reframes efficient attention through a fast-weight FFN lens

A research team from Beijing University of Posts and Telecommunications has introduced Mixture-of-Top-k Attention, or MiTA, a method that recasts standard attention as a feed-forward network whose weights are generated by the current input and whose hidden width grows with sequence length. The paper, accepted to NeurIPS 2026, uses that fast-weight view to build a five-dimensional taxonomy for efficient attention and then proposes a design that combines reusable routed experts with a shared global expert. In the paper’s formulation, efficient attention methods fall into two broad paths: compression, which summarizes a growing fast-weight network into compact modules, and routing, which preserves raw content but only visits selected parts. MiTA tries to bridge the two. It extracts a small set of landmark queries, lets them build Top-k routed experts from the full context, and also uses the same landmarks to construct a shared global representation. Each query then reuses an expert rather than assembling its own private Top-k set. The reported results cover ImageNet-1K classification, ADE20K segmentation, Long Range Arena, and arXiv long-document classification. The paper says MiTA reaches 79.8% Top-1 on DeiT-S versus 79.9% for full attention while cutting FLOPs from 4.6G to 4.4G, reduces total LRA training time from 10.6 hours to 2.4 hours, and posts 88.0%, 92.9%, and 98.7% accuracy on arXiv classification at 16K, 32K, and 64K lengths.

800

Apple

2026-10-07 12:12:33

Apple study says lean AI agents can match or beat multi-agent systems

Apple has found in a recent study that streamlined AI agents can perform as well as, and in some cases better than, more complex multi-agent machine learning engineering systems. The research argues that simplifying AI system design may improve efficiency, while also calling into question whether multi-agent frameworks are always necessary. The item was reported by Techub News, citing Crypto Briefing. No additional technical details, benchmarks, or deployment context were provided in the source brief.

470

Numeraire

2026-10-06 11:40:10

NMR briefly hits 17.58 USDT, up 41.73% in 15 minutes

Numeraire (NMR) briefly climbed to 17.58 USDT and was last quoted at 17.19 USDT, according to Bitget market data. The token rose 41.73% over a 15-minute window. The move may be linked to Upbit’s planned listing of NMR trading pairs against the Korean won and USDT, which was cited as a possible factor behind the sharp intraday jump. Numeraire is an Ethereum-based platform that allows developers and data scientists to test and build machine learning models with higher reliability. The price update and listing-related catalyst were both included in the source report.

540

TouchScale

2026-10-08 23:43:38

TouchScale releases 500-hour vision-touch dataset, lifting robot task success from 22.5% to 57.5%

Researchers from 15 institutions, including Texas A&M University, Google DeepMind, Carnegie Mellon University, Stanford University, the University of California, Berkeley, Yale University, Microsoft, NVIDIA, Meta, and others, have introduced TouchScale, a 500-hour human vision-touch dataset collected with a single sensor stack and a unified capture and synchronization pipeline. The project is designed to test whether visual-tactile learning shows the same scaling gains seen in large video datasets when data quality and hardware consistency are held constant. The dataset includes synchronized head-mounted RGB-D, wrist RGB video, and full-hand tactile signals from gloves with 880 taxels per hand. It spans about 87,000 interaction recordings, roughly 2,000 task descriptions, more than 1,500 objects, nine environment categories, over 800 scene configurations, and around 20 data collectors. In experiments, TouchScale improved zero-shot cross-sensor tactile prediction and also produced the best visual encoder accuracy across MECCANO, Something-Something V2, and Ego-Exo4D when tactile signals were used as supervision. In real-world robot tests with an xArm6 and a BrainCo Revo 2 hand, adding TouchScale mid-training before post-training raised the average success rate across four contact-heavy tasks from 22.5% to 57.5%, without using human action labels or mapping human hand motions to robot actions.

540