12 stories on Research.
NVIDIA-powered agents can steer the Allen Telescope Array and apply AI models through natural-language requests, and have made a first real-time pulsar detection.
Why it matters: Natural-language control of high-throughput sensing systems shows agents can operate real scientific infrastructure, not just software tasks.
Y Combinator Labs and Olam Labs put 50 agents in a Minecraft world, where they appointed a leader, wrote rules and built a food-rationing ledger.
Why it matters: Multi-agent systems can quickly form hierarchies and cautious behavior, which matters for teams designing agent swarms, though the results come from a game setting.
A research paper by Takehi, Teoh, Zhang and Khattab reports that GPT-6 Astra can infer what a model does from its raw weights and edit them to improve math scores.
Why it matters: If models can read and correct weights without running code, teams may gain new ways to audit and repair models, but the claim is still contested and based on a small set of experiments.
MLE-World, a fine-tuned Qwen3.5-4B model, predicts what an ML program would print, including logs, errors and validation scores, before the code is executed.
Why it matters: Teams running agent-driven ML experiments could spend less compute on candidates that predicted feedback marks as failing, if the accuracy holds on their own tasks. The reported gains apply only to tasks represented in training.
EditAgent mixes general and specialist editing tools and is guided by measured image data.
Why it matters: Agents that choose specialist editing tools may make image editing pipelines more reliable, though availability is not stated.
SemiAnalysis's GDPx benchmark of 206 agentic tasks across 11 domains finds the best model scores only 12%.
Why it matters: Buyers comparing frontier models for agentic work get a low absolute ceiling on realistic economic tasks, along with a cost gap that may matter more than the score gap.
The MIT CSAIL team reports in a paper that GPT-6 Astra edited two layers of Qwen3.5-0.8B and raised its MATH-500 accuracy from 7.9% to 48.1%.
Why it matters: Automated model editing could become a practical way to diagnose and patch specific failure modes, though the gains here come mostly from fewer unfinished answers.
In 880 five-agent debates, Cisco found that when agents skipped checking their conclusions, three quarters ended with all five backing the wrong answer.
Why it matters: Multi-agent systems that make decisions together may need built-in verification steps, or errors can spread across all agents.
A 50-task benchmark of games and web tasks shows top models pass only 36% of tasks under real-time latency, versus up to about 80% with a paused clock.
Why it matters: The gap between paused and real-time results suggests current agents are unreliable wherever response speed matters, such as buying tickets or bidding in auctions. Teams should not assume that strong offline benchmark scores carry over to live, latency-sensitive workflows.
The startup proposes an explicit, auditable reasoning core for AI systems, modeled on the approach behind DeepMind's AlphaGo, to reduce hallucinations.
Why it matters: If reasoning can be made auditable with clear provenance, it could address reliability concerns for enterprise AI in regulated or high-stakes decisions. The paper is conceptual, so the approach remains unproven.
In a randomized study, analysts who used Claude Code followed the agent's secretly assigned preference 86% of the time.
Why it matters: People may adopt an AI agent's judgment without recognizing it, so companies using agents for analysis need deliberate human review.
OpenAI released LifeVal, a benchmark that measures how well AI handles everyday tasks such as childcare, personal finances and government paperwork.
Why it matters: Companies building consumer-facing assistants would have a benchmark aimed at practical personal tasks, not just technical or academic ones.