Today's AI news centers on autonomous agents that act without human supervision. A team of four Gemini 4 Argon agents defeated Minecraft's Ender Dragon entirely on their own. Meanwhile, new research shows hybrid reasoning cuts inference costs by up to 80%.
Hybrid Reasoning Agents Cut Costs by Up to 80%
Latent and Text-Based Reasoning
Researchers introduced Hybrid Reasoning Agents (HybridAgent), which switch dynamically between latent reasoning and text-based interaction. A lightweight HybridLink enables seamless transitions between modes while preserving tool-integrated capabilities. The paper was submitted to arXiv on October 13, 2026.
Source: https://arxiv.org/abs/2610.18861
Benchmark Gains and Token Reduction
HybridAgent improves performance by 10.4% across 11 benchmarks covering math, science, medicine, search, and agentic reasoning. It delivers 35.5% to 54.7% token reduction, 1.8x to 2.6x faster inference, and 2.7x to 3.7x cost savings across three model scales.
Source: https://arxiv.org/abs/2610.18861
Four Gemini 4 Argon Agents Beat Minecraft
Autonomous Ender Dragon Victory
Four Gemini 4 Argon agents entered a sandboxed Minecraft survival world with one goal: defeat the Ender Dragon. Each agent controlled a player via Mineflayer and coordinated through in-game chat and shared memory. No human assigned roles or wrote game logic. The agents planned, wrote, and debugged every program themselves.
Source: https://arxiv.org/abs/2610.22319
Run Metrics
The team wrote 2,783 programs and defeated the dragon in 27 hours and 14 minutes, dying 605 times along the way. Researchers monitored via a live dashboard and never needed to intervene. This marks the first end-to-end autonomous completion of Minecraft by AI agents using only standard Mineflayer.
Source: https://arxiv.org/abs/2610.22319
MLExpert Advances Autonomous Machine Learning Engineering
35B Mixture-of-Experts Agent
MLExpert is a 35-billion-parameter Mixture-of-Experts agentic model for autonomous machine learning engineering. It uses hierarchical agentic SFT and Windowed Interaction RL to handle long-horizon experimentation. The paper appeared on arXiv on October 15, 2026.
Source: https://arxiv.org/abs/2610.22344
Medal Rate Improvement
On MLE-Bench Lite, MLExpert improves the average medal rate of the base model from 37.9% to 72.7%. This matches performance of larger frontier models such as Kimi-K3 and GLM-5.3. The learned optimization also generalizes to NatureBench Lite and algorithm optimization tasks.
Source: https://arxiv.org/abs/2610.22344
FDE-Gym Measures Real-World AI Delivery
Benchmark of Customer Delivery Tasks
FDE-Gym evaluates agents on 95 end-to-end delivery tasks from 20 customer scenarios across 18 sectors. Industry experts adapted each task from real-world delivery projects. Agents must clarify ambiguous requirements and uncover unstated constraints in live systems.
Source: https://arxiv.org/abs/2610.22350
Delivery Scores Reveal Gaps
The best configuration achieves a mean FDE-Gym score of 0.56 and resolves only 14 of the 95 tasks. Discovering unstated constraints is the primary bottleneck. Even on 23 tasks where every key fact was obtained, the best config reached meaningful business uplift on only 10.
Source: https://arxiv.org/abs/2610.22350
DeepMind SIMA 2 Agent Learns in Virtual Worlds
Agent That Plays and Reasons
Google DeepMind introduced SIMA 2, an agent that plays, reasons, and learns with users in virtual 3D worlds. SIMA 2 follows instructions and completes tasks across simulated environments without game-specific access.
Source: https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/
FAQ
What is a hybrid reasoning agent?
It switches between internal latent reasoning and external text-based interaction to act through tools while keeping thoughts inspectable.
How did the Gemini agents beat Minecraft?
Four agents wrote 2,783 programs, coordinated through chat, and defeated the Ender Dragon in 27 hours and 14 minutes after 605 deaths.
How much does MLExpert improve MLE performance?
It raises the average medal rate on MLE-Bench Lite from 37.9% to 72.7%.
What does FDE-Gym measure?
It measures how well AI agents deliver working solutions to real customers across 95 tasks in 18 sectors.
Can SIMA 2 learn new tasks?
Yes, it plays, reasons, and learns with users in virtual 3D environments.
Top comments (0)