Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2609.23019v1 Announce Type: cross Abstract: Soma instance segmentation, i.e., identifying and delineating individual cell somas as distinct instances, is crucial for cellular analysis and conne...
arXiv:2609.24576v1 Announce Type: cross Abstract: Modern Vision-Language Navigation (VLN) models rely mostly on pre-trained large Vision-Language Models (VLMs) to predict navigation actions. While th...
arXiv:2404.06135v4 Announce Type: replace Abstract: The Transformer architecture has achieved remarkable success in natural language processing and high-level vision tasks over the past few years. Ho...
arXiv:2509.22650v3 Announce Type: replace Abstract: Most existing approaches to referring segmentation achieve strong performance only through fine-tuning or by composing multiple pre-trained models,...
arXiv:2604.15281v2 Announce Type: replace Abstract: 3D policy learning promises superior generalization and cross-embodiment transfer, but progress has been hindered by training instabilities and sev...
arXiv:2605.04515v2 Announce Type: replace Abstract: Video Large Language Models (Video-LLMs) excel in general video understanding but often base physical judgments on event expectations rather than o...
arXiv:2606.07338v2 Announce Type: replace Abstract: Vision-language driving models increasingly use reasoning supervision to bridge perception, prediction, and planning, but existing driving rational...
arXiv:2608.08381v2 Announce Type: replace Abstract: Motivated by the IEEE 802.11bf effort to standardize advanced WLAN sensing, interest in Wi-Fi Channel State Information (CSI) for passive, device-f...
arXiv:2608.10706v3 Announce Type: replace Abstract: Recent vision-language models demonstrate impressive general visual understanding, yet their art interpretation remains shallow: they describe surf...
arXiv:2609.07780v2 Announce Type: replace Abstract: Automated drone surveillance has become increasingly important for public safety, critical infrastructure protection,and restricted airspace monito...
The paper introduces the Activity Chain Encoder (ACE), a self‑supervised deep learning model that transforms passively collected mobile phone location data into daily activity representations. ACE integrates pre‑trained urban embeddings, visit timing, and duration, using a Transformer to capture the sequential structure of stays, and is trained via masked activity modelling and contrastive learning without explicit activity labels. The resulting user‑level profiles are clustered and interpreted with temporal‑functional patterns and Census demographics, revealing six distinct weekday activity‑pattern groups in London that differ in daily rhythms, urban contexts, and demographic characteristics.
The paper introduces MFP, a reaction yield prediction method that uses role-aware Morgan fingerprints. It computes count-based circular fingerprints for each reaction component, aggregates them by chemical role, and combines them with transformation-sensitive difference features into a fixed-length descriptor for a feed-forward neural regressor. On the Suzuki‑Miyaura and Buchwald‑Hartwig benchmarks, MFP achieves R² scores of 0.878 and 0.969 respectively, while training an order of magnitude faster than graph or Transformer-based alternatives.
The paper introduces mutation analysis as a metric for evaluating GPU‑kernel benchmark oracles, injecting over ten thousand faults into verified CUDA implementations of 188 KernelBench problems. It shows that the current official checkers miss 16.9% of faults, with precision faults being especially problematic, and demonstrates that optimized test suites can achieve 98% detection with only two inputs per problem. The study also reveals flaws in existing patches and a fuzzing recipe that incorrectly rejects correct kernels 107 times.
SCALE (Simulation‑Calibrated Amortized Learning for Energy Materials) is a physics‑grounded learning architecture that fuses deterministic scientific operators, experimental calibration, and transformer‑scale inference to accelerate the discovery of energy materials. It converts expensive mechanistic computations and measured data into reusable models for rapid screening, ranking, inverse design, and active learning. In a case study on solid‑state metal‑hydride hydrogen‑storage, SCALE calibrated a phase‑equilibrium capacity operator against 381 experimental anchors, generated 5,000 teacher labels, and used a 2.90‑million‑parameter edge‑biased graph transformer to achieve high surrogate fidelity (MAE 0.0582 wt% H₂, RMSE 0.0833 wt% H₂, R² 0.9927, Pearson r 0.9963) while reducing per‑candidate screening cost by 10⁷–10⁸ times.
The study evaluates six deep learning architectures for predictive maintenance in Industry 4.0, focusing on Recurrent Neural Networks (RNNs) and Transformers. It finds that Transformers perform well on stable, slow-moving data but overreact to noisy, chaotic data, whereas a hybrid model combining an LSTM layer with a Transformer layer better filters noise and delivers more consistent predictions. The hybrid approach improves accuracy and reliability across varying levels of data volatility.
CTRL is a new framework for time‑series forecasting that separates semantic reasoning from quantitative prediction. It uses a frozen backbone to produce base forecasts, while LLM agents act as controllers that analyze prediction errors by decomposing them into trend, seasonal, and irregular components. The agents generate compact control signals that a lightweight residual decoder uses to correct the forecasts, and the system can adapt at test time to distribution shifts with only a few LLM calls.
The paper investigates whether the stable GAN architecture R3GAN can improve time‑series imputation when adapted to 1‑D temporal data. Using a coarse‑to‑fine refinement framework and a frequency‑domain discriminator, the authors evaluate 14 saved configurations across three datasets and find a negative result: most configurations either show negligible improvement or degrade performance compared to baseline methods. The study highlights that the usual argument—GANs optimize distributional objectives rather than point‑wise ones—does not fully explain the lack of benefit, and it poses an open problem regarding why a learned discriminator fails to provide useful refinement gradients while diffusion denoisers succeed, offering practical guidance on when adversarial refinement may be worthwhile.
The paper introduces FM4NILM, a prompt‑programmable foundation model that estimates the power usage of any requested appliance from whole‑home meter data, a natural‑language description, and optional activation examples. Trained on 645k sequences from seven public datasets, the lightweight transformer achieves competitive performance across twelve appliance requests, outperforming seven appliance‑specific baselines on key metrics such as event F1 and active‑window MAE. The model’s design allows new appliance coverage to be added via prompts and examples rather than building separate specialist networks.
The paper introduces OmicsBench, a new reasoning benchmark for multi‑omics sequences that includes 1,160 expert‑validated questions across DNA regulation, RNA processing, and protein function tasks, requiring traceable evidence chains. Evaluation of 17 large language models shows that scientific LLMs, while more accurate in classification, often lack valid evidence, suggesting shortcut learning. To address this, the authors propose tool‑augmented on‑policy distillation (TA‑OPD), a post‑training method that improves both evidence grounding and predictive performance across five Qwen3.5 models of varying sizes.
The paper investigates whether sink-aware attention head selection remains valid after 4‑bit NF4 weight‑only post‑training quantization. Using Sink Topology Consistency metrics, it finds that global rank preservation stays high across Qwen2.5 and Llama‑3.2 models, yet top‑k head overlap drops to 61–79% and layer‑specific sink‑mass shifts can be substantial. The study also shows that cross‑domain calibration degrades more than within‑domain precision and that recalibration with a small number of samples can recover most of the stability, though full‑map stability may require updating more layers.