Research

Everything in this section in one place — newest stories first.

LatestResearch

A New Synthetic Data Pipeline Boosts Crop-Prediction Models Across a Dozen Farm Sites

TCSDG combines a Bayesian Network generator with the TabICL tabular foundation model to create synthetic agricultural data

ResearchJuly 17, 2026

New Benchmark Exposes How Compiler Optimizations Quietly Break LLM Reverse Engineering

REFORGE is a new provenance-tracked benchmark pipeline for evaluating LLMs on decompiled binary function naming, accepted to ACC 2026 in Lisbon

ResearchJuly 17, 2026

When AI Exposure Scores Disagree by 11x, a New Statistical Fix Steps In

A new arXiv paper (submitted July 13, 2026) fixes a problem where competing AI occupational-exposure scores differ by up to 11x and flip the sign of employment effect estimates

AI AgentsJuly 17, 2026

Study Casts Doubt on Self-Improving Agent Harnesses, Finds They Rarely Beat Simple Scaling

A July 2026 arXiv paper by Yike Wang tests whether automatically evolving an AI agent's 'harness' (tools, prompts, control logic) improves performance more than simple test-time scaling.

ResearchJuly 17, 2026

Graph Neural Network Hits 99% Accuracy Decoding Muscle Signals for Prosthetics in Real Time

A graph neural network for real-time sEMG-based gesture recognition achieved 99% average classification accuracy in a new study.

ResearchJuly 17, 2026

New Minimax Bound Settles the True Cost of Fairness in Multi-Armed Bandits

A new arXiv paper proves a tight minimax bound of Ω(σ√(k^max(1,q)/T)) for regret in fairness-constrained bandit problems.

ResearchJuly 17, 2026

Thomas Strohmer's "Mathematics of Data Science" Puts the Theory Back Under the Hood of AI

Thomas Strohmer submitted 'Mathematics of Data Science' to arXiv on July 11, 2026 (arXiv:2607.11938)

LLMsJuly 17, 2026

ATLAS Cuts LLM Benchmarking Down to Size, Using 41 Questions Instead of 5,600

ATLAS, a new IRT-based adaptive testing framework, matches full-benchmark ability estimates on HellaSwag using just 41 of 5,600 items, with 0.157 mean absolute error

ResearchJuly 17, 2026

Why Vision Models Fail to Generalize, and Why Locality Might Fix It

A new ECCV 2026 paper, 'On Locality and Length Generalization in Visual Reasoning,' finds that vision models with global image access learn shortcuts and fail to generalize to longer or more complex tasks, mirroring known failures in language models.

ResearchJuly 17, 2026

A Single Decoder for All Your Quantum Error Codes? New Research Puts Meta-Learning to the Test

New meta-decoding framework learns a single syndrome-to-recovery mapping shared across FiveQubit, Steane, Planar3x3, and Planar5x5 stabilizer codes.

ResearchJuly 17, 2026

A Trainable Connection Trick Shrinks Logic Gate Networks by 50x

A new arXiv paper (2607.09399) introduces trainable connections for differentiable logic gate networks and lookup table networks, not just trainable gate types

ResearchJuly 17, 2026

A New Mathematical Blueprint for Market Making in Crypto's Perpetual Futures

Minmin Zeng's arXiv paper (2607.11888) introduces a stochastic optimal control framework for market making across two exchanges in perpetual futures markets

LLMsJuly 17, 2026

A 7B Chemistry Model Cuts Hallucinations by 79% Using a Multi-Agent Trick Borrowed From Game Theory

OmniChem, a 7B parameter model, matches GPT-4o mini on ChemBench using a training corpus of 363,045 chain-of-thought traces generated via the new G-Frame framework.

AI AgentsJuly 16, 2026

AWS Turns Its Own Support Desk Into an AI Agent Built on Bedrock AgentCore

AWS published a reference architecture for an AI Support Companion built on Bedrock AgentCore and Strands Agents, using MCP servers to access documentation, support APIs, and AWS services.

LLMsJuly 16, 2026

Anthropic's Claude Fable 5 Survives an Export Ban, Then Comes With a Bill Attached

Anthropic's Claude Fable 5 and Mythos 5 launched June 9, 2026, then hit US export controls on June 12 after Amazon researchers found a jailbreak exposing software vulnerabilities.

RoboticsJuly 16, 2026

Walden Robotics Emerges From Stealth With $300M and a Toyota Factory Already Running Its Robots

Walden Robotics exited stealth with $300 million in funding at a $1.1 billion valuation, co-led by Toyota and Deviation Capital.

LLMsJuly 16, 2026

Mira Murati's Thinking Machines Ships Inkling, a 975B-Parameter Open Model That Beats US Rivals but Trails China

Thinking Machines Lab, founded by ex-OpenAI CTO Mira Murati, released Inkling, a 975B-parameter open-weight mixture-of-experts model with 41B active parameters.

LLMsJuly 16, 2026

OpenAI's GPT-5.6 Launch Hit a Government Delay, a Benchmark Crown, and a File-Deletion Problem

OpenAI released the GPT-5.6 family (Sol, Terra, Luna) on July 9, 2026, after a U.S. government-forced delay tied to Commerce Department safety review

All articles