The Archive

Falling token costs, cheaper models, and less spend per employee—AI adoption isn't playing out the way hyperscalers hoped.

Analysis of transformer interpretability: 53 components drive 90% of token predictions; 13 critical; 8 sufficient alone. Mass of off-predicted-token contributions exceeds predicted by 7x median.

Multi-modal beam prediction framework for V2X networks using camera and LiDAR sensing to improve vehicular communication.

Transformer-based emotion classifier applied to 1.7M Congressional speeches to correlate emotional expression with legislative outcomes.

Theoretical separation result showing randomization reduces oracle calls for online learning of thresholds on unknown orders.

Systematic audit of biases in protein-protein interaction datasets that enable ML models to exploit learning shortcuts.

Argument mining framework for political debates linking argumentative spans to debate named entities.

Study of AI agent coordination across administrative domains in network automation with fragmented authority and observability.

Contrastive framework for multimodal in-context learning that aligns MLLM reasoning paths beyond surface-level demonstration imitation.

CoGe-GCD framework enabling compositional generalization for generalized category discovery via primitive-based perception.

DoRA fine-tuning approach incorporating individual text corpora to personalize small language models for knowledge QA tasks.

Cross-family on-policy distillation method addressing degradation in knowledge transfer across different language model families.

YallaMorph benchmark evaluates Arabic morphological generation in LLMs across 600K examples with diacritized and undiacritized settings.

Kernel-managed shared memory system for multi-agent personalization with privacy controls tested on GPT-4o, Llama-3.1, Qwen-2.5.

Preventative Steering defense against malicious fine-tuning works via compensatory adaptation phase; attention projections are key residual mechanisms.

Quantum entanglement allocation sample complexity analysis for qubits and Pauli queries.

SmartWeatherAgent fuses LLMs with LightGBM for plateau weather alerts using intent recognition and hazard prediction, F1-Macro 0.605.

Formal framework for assembling context in LLM-based systems engineering design with modular units and modelling-as-code evaluation.

SA-Profile automated framework profiles sulcus angle from super-resolution MRI using implicit neural representations for trochlear dysplasia.

LLM bug-fixers fail: models claim bugs in bug-free code and damage correct programs faster than fixing actual bugs, creating pseudo-cycles.

ProbPlug lightweight confidence calibration framework for LLM binary classification using internal token features and self-attention.

Instacart is the latest app to bake a conversational AI assistant into its platform.

OpenAI's Chris Lehane argues policy framework needed now: stronger safety evidence, shared standards required before advanced AI capabilities scale.

Cymphony was valued at more than $100 million in a $25 million Series A co-led by Sequoia and SMBC Fin Atlas Beyond Fund.

Maxton Hall. | Image: Prime Video Amazon's Prime Video is launching a new AI-powered feature that lines up an actor's mouth with "human-dubbed" audio. The feature is only available with the English dub of the German series Maxton Hall for now, but Prime Video plans to expand it to "additional titles" in the future. In its announcement, Prime Video says it uses a combination of AI and visual effects technologies to make lip movements match up with translated speech. Meta and YouTube recently launched an AI-powered auto-dubbing feature for creators, while allowing them to turn on a "lip sync" o...

Data-centric post-training pipeline for financial reasoning: mines traces, distills instructions, generates knowledge-graph-guided QA pairs.

Storage-scalable knowledge-base reuse quantization for semantic communication via progressive refinement.

Federated learning framework for aging clock prediction using trust networks across distributed medical centers.

Feasibility taxonomy of 20 inference-time AI governance mechanisms as regulatory focus shifts from training to deployment.

If you want to manufacture in space—and bring the products back again—there’s a limited set of options: Wait to go to the International Space Station, or partner with a handful of start-ups launching spacecraft that spend time in orbit before they return to Earth. But the vehicle that goes to space and comes back to […]