The Archive

AdaSpark adaptively learns verification scheduling for speculative decoding using online learning, improving token acceptance vs. latency tradeoffs.

Relational synthesis method combines retrieved and reference audio to generate novel Foley audio without copying source segments.

Yossi Matias, Vice President & Head of Google Research, explores how AI is beginning to reshape biology, infrastructure, manufacturing, and science, and why its greatest impact may come when it intersects with other fields. Step inside the newsroom with our MIT Technology Review editors for sharp analysis and unpublished insights from the team that researches…

Agentic retrieval using ReAct loops improves complex information retrieval beyond semantic similarity, at quantified performance-cost tradeoffs.

Post-training quantization degrades robustness of vision-language-action models; vulnerability-oriented tuning recovers generalization under distribution shift.

CIPHER-MoE addresses load imbalance in trillion-scale MoE training via dynamic token rerouting, improving hardware utilization and training stability.

HLA-WM hybrid linear attention reduces long-range forgetting in video world models by combining softmax and recurrent attention without retraining.

PACMI framework tracks memory dependencies and preserves history for long-term LLM agents, enabling graceful invalidation of outdated information.

FreSia framework aligns frequency-domain representations with LLM semantics for improved multivariate time-series forecasting and anomaly detection.

Physics-constrained neural operators reconstruct full coupled-system fields and identify unknown parameters from sparse single-field observations.

ReMaD enables tuning-free domain adaptation and OOD detection via reduced-rank Mahalanobis distance on pretrained model embeddings.

Retrieval-based in-context learning formulated as domain adaptation with theoretical guarantees under distributional shifts beyond prior work.

OR-Bench reveals VLMs detect object rotation but fail at magnitude estimation, clustering predictions at discrete angles (0°, 90°, 180°).

SimpleMark: multi-bit text watermarking framework using f-divergence constraints with per-key, per-message security guarantees.

Joint generative and self-supervised learning in single model enables label-free, pretrain-free image generation via self-conditioning.

Nexus: distributed execution fabric for LLM agents spanning cloud, edge, devices with permission control and fault-tolerant lifecycle management.

SO-RSI: second-order recursive self-improvement framework for formal verification agents that diagnoses root causes before structural workflow changes.

Voltic: recurrent memory with anisotropic covariance distinguishes volatility from stochasticity via input-dependent noise variances.

Geospatial foundation models combining multimodal and mobility data for disease prediction in data-sparse regions with operational epidemiology gains.

Outcome-max replaces token-max optimization: rewards verified task completion per unit cost, includes stopping rule and three-level simulation framework.

Comparison of LLM retrieval errors in environmental science journals: Claude, ChatGPT, Grok, DeepSeek, Perplexity, others tested on 2024-2025 citations.

PwC and Cohere announce global partnership to deploy secure enterprise AI, launching in Canada.

Cohere releases North 2, an agentic AI platform with enterprise security, cost governance, and full-stack control.

Simon Willison replicates Colin Frasier's experiment testing Qwen 3.8 27B and GPT models on arithmetic tasks (sum-in-words), finding systematic errors across model scales.

AI slop seems to be overwhelming bug bounty programs.

On Equity, we discussed the Trump administration's attempts to rebrand AI.

Dale Caldwell, former lieutenant governor for New Jersey. | Bloomberg via Getty Images New Jersey's lieutenant governor Dale Caldwell was forced to resign on September 25th after an investigation found he had sexually harassed a staffer and repeatedly violated ethics rules. The now-former Lt. governor has been making the media rounds trying to clear his name. But he took a particularly odd tactic during an interview on NJ PBS. Caldwell claims he's being unfairly targeted, and multiple AI agents back that up. He told host Rob Nelson that he "AIed" the report on the investigation "from multiple...

StarSkirmish pits AI-made StarCraft-playing bots against one another, as well as against human-made bots. OpenAI's GPT-6 Astra and Claude Opus 5.5 were essentially tied as the best-performing AI-made bots, but they couldn't top Stardust, the top-rated human-made bot. On Friday, GPT was facing off against Claude and the human-created bot Pluto, but according to Kotaku, it couldn't quite get an edge. So it resorted to a tactic that is becoming alarmingly common for modern AI models - it broke the rules. GPT-6 Astra went and downloaded Stardust, and started running that instead of its own bot. G...

This new task force is Trump's latest response to the debate over AI safety.

Simon Willison argues AI agent deployments require hard budget caps on API spend, not soft warnings, to prevent runaway costs.