Short-Term Forecasting of Heterogeneous Wind Power Fleets: Global Temporal Fusion Transformer Leveraging Static Covariates reports that a global TFT trained on German wind farms beats site-specific models in 48-hour forecasts. The model improves accuracy for 87% of 100 synthetic farms and 77% of 11 real farms, with static covariates essential for performance.
Key points
Global TFT outperforms local models for 87% of synthetic wind farms and 77% of real German wind farms.
Altitude is the most relevant static covariate for synthetic data; hub height is decisive for the real portfolio.
Sources
Dull, Dirty, Dangerous? Working Conditions and Occupational AI Exposure study released
Dull, Dirty, Dangerous? Working Conditions and the Limits of Occupational AI Exposure presents an analysis of AI exposure across 754 US occupations using 21 O*NET items. Higher AI exposure correlates with lower dull, dirty, dangerous condition scores under both LLM and AIOE measures. Adjusted models show confidence intervals that include zero for LLM exposure and stay below zero for AIOE.
Key points
Analyzed 21 O*NET items across 754 US occupations.
Adjusted confidence intervals include zero for LLM exposure and remain below zero for AI Occupational Exposure.
Sources
Learning When to Think Aloud in Latent Reasoning shows 9.6% accuracy gain on GPT-2
Learning When to Think Aloud in Latent Reasoning introduces VITAL, a hybrid reasoning framework that selectively verbalizes tokens. VITAL achieves a 9.6% relative accuracy increase on GPT-2 and a 5.1% increase on Llama-3.2-1B compared with fully latent methods, using shorter reasoning trajectories.
Key points
VITAL uses a lightweight gating head to choose between token generation and hidden-state feedback.
VITAL is trained with curriculum-based supervised learning followed by reinforcement learning that balances correctness and conciseness.
Sources
BEAKER benchmark released with 1,000 expert-curated samples
BEAKER is a new benchmark for embodied brains in self-driving chemical laboratories, containing 1,000 expert-curated question-answer samples across 11 image and video task types. It evaluates 28 general-purpose, science-oriented, chemistry-specialized, and embodied multimodal large language models in an offline, single-turn setting. Results show models can follow coarse experimental procedures in videos but struggle with action grounding, fine-grained apparatus and state recognition, safety norms, and step ordering.
Key points
1,000 expert-curated question-answer samples across 11 image and video task types.
Evaluation of 28 multimodal large language models in an offline, single-turn setting.
Sources
Neural Stochastic Forecasting for Lateral Molecular Diffusion Simulations released
Neural Stochastic Forecasting for Lateral Molecular Diffusion Simulations introduces a neural stochastic model that predicts collective diffusion of membrane proteins. The model combines a transformer encoder, multi-scale recurrent networks, and a stochastic latent process with learned memory. It reproduces an anomalous diffusion exponent α≈1 for a seven-PIEZO system and generates trajectories up to 40,000 frames.
Key points
Model uses transformer-based protein configuration representation, recurrent networks at multiple temporal scales, and a stochastic latent process.
In the mixed system, the model captures distinct diffusion characteristics of PIEZO and TREK populations and matches time-averaged MSD, Van Hove distributions, and non-Gaussian parameters.
Sources
threejs TSL – fluid sim, marching cubes, caustics ported
threejs TSL – fluid sim, marching cubes, caustics ported from personal engine, using Eric Lengyel's Transvoxel for multi-resolution. Porting was quick and painless.
Key points
Fluid simulation, marching cubes and caustics moved to threejs TSL.
Eric Lengyel's Transvoxel employed to unlock multi-resolution marching cubes.
Sources
- Original post
- codevember-2026.felixmartinez.dev - Linked in the post
Axion 1.0 Open Agent Model released with simple training
Axion 1.0 is a fully open agent model from ByteDance that uses two SFT stages followed by RL. It achieves competitive results on HLE, SEAL-0, and FrontierScience.
Key points
Training: two SFT stages followed by RL.
Results: competitive across HLE, SEAL-0, and FrontierScience.
Sources
Action Intent as a Bridge moves video knowledge to robots
Action Intent as a Bridge moves video manipulation knowledge to robots using a two-stage intent extraction and alignment method. The paper reports extensive experiments in simulation and real-world settings that consistently outperform prior baselines.
Key points
AIT extracts explicit action intents and disentangles implicit intent from manipulation videos.
AIT aligns human and robot domains within a shared intent space and decodes the intent into executable robot actions.
Experiments show AIT transfers human knowledge more effectively than prior baselines.
