The Archive

Fine-tuned SHuBERT-ByT5 model for real-time sentence-level sign language translation achieving BLEU 15.9 on test set.

Nilay Patel argues AR glasses require continuous camera streaming to cloud due to computational and power constraints; no edge-only solution currently viable.

Stratechery weekly digest covering Xbox discontinuation and Zuckerberg commentary; not AI-focused.

Agora: auction-based task allocation framework for dynamically routing LLM agent requests to heterogeneous expert models.

Tokenizer vocabulary transplantation fixes Bengali ASR failure in Moonshine by replacing English-centric byte tokenizer with BanglaBERT.

PAC-ACT reinforcement-learning post-training for Action Chunking Transformers improves industrial robot contact manipulation policy.

There are many ways to optimize code for GPUs. In this post, you’ll learn how kernel fusion can improve memory bandwidth and reduce kernel launch overhead,... There are many ways to optimize code for GPUs. In this post, you’ll learn how kernel fusion can improve memory bandwidth and reduce kernel launch overhead, along with multiple ways to apply it in NVIDIA CUDA code. A common bottleneck when writing GPU code is that GPU compute is so fast that even high-bandwidth device memory doesn’t use the GPU kernel fully. Kernel fusion addresses this by… Source

TrustX ARC: 12-dimension risk classification framework for governance of enterprise agentic AI systems across seven deployment types.

AI performance comes down to three dimensions: Accuracy: How well the model reasons and produces outputs Throughput: How many tokens per second a... AI performance comes down to three dimensions: Deployments must balance all three: High accuracy is wasted if responses are slow, and raw throughput means little if each user’s experience is laggy. Practical systems therefore optimize accuracy, throughput, and interactivity together. This post focuses on throughput and interactivity, and how model-design choices shape both without… Source

Physics-constrained ML surrogate with entropy generation constraints for chemical kinetics in turbulent flow simulation.

Knowledge graphs and explainable AI techniques integrated for defensible pre-demolition urban mining assessment decisions.

Network-based graph framework representing 160 idiomatic expressions across 8 languages using cognitive-linguistic feature annotations.

Evolutionary algorithms with cardinality constraints for large-scale portfolio optimization under real-world conditions.

TCLA: training-free adaptation method for medical vision-language models to handle domain shift and class bias in few-shot settings.

Research argues frontier AI systems need dynamic vocabulary and solution-space adaptation beyond fixed representational frames for open-ended reasoning tasks.

GR-RTRMC: graph-regularized low-rank matrix completion method using Grassmann manifold optimization.

VLMs encode correct counts internally but fail to output them; probes detect misalignment between internal representations and verbalized answers.

CoCoT-EEG: contrastive self-supervised foundation model with multiscale convolution for EEG signal decoding.

Digital Services Act may force Meta to make big changes on its platforms.

GatedLinear: adaptive routing mechanism for time series forecasting that routes between complementary linear bases for diverse temporal patterns.

Researchers demonstrate cryptographically-undetectable backdoors in neural networks that enable adversarial examples while evading white-box detection.

Freya-TTS: 183M-parameter Turkish text-to-speech model using flow-matching diffusion with frozen AudioVAE2 latents.

TSAI-MetaFraud: multimodal benchmark dataset for fraud and risk detection in metaverse virtual economies.

ALICE: unified pathology foundation model via multi-stage distillation from eight expert models, trained on 25M+ pathology images.

SAGEAgent learns to selectively acquire multimodal clinical data for cancer survival prediction, reasoning about cost-benefit tradeoffs in diagnostic workflows.

Energy profiling of edge VLM inference reveals language generation dominates energy costs, not vision processing, across five models and two hardware platforms.

Large-scale study of LLM CLI coding-agent failure trajectories reveals onset patterns and recovery mechanisms, treating failure as temporal process not final outcome.

VGGT geometric foundation model encodes co-visibility of image pairs as emergent 3D-aware representation, useful for reconstruction and robotic localization.

Rashomon Explanation paradigm couples prediction with explainability as complementary objectives, showing self-explanation can improve rather than degrade model accuracy.

Square Root and Hapax Correction techniques derive well-calibrated likelihood ratios for forensic authorship verification without auxiliary calibration data.