Does post-training actually create new capability, or does it just sharpen what was already there? Title: Sharpening Tax in Post-Training URL: ❓ What was the surprising finding? A pretrained base model, given nothing but a lightweight harness, can handle agentic tasks. Its single-shot accuracy is far worse, but once you let it retry, it overtakes its post-trained counterpart on solution coverage. On WebShop with gemma-4-31B, the base model passes over 85% at pass@128 while the RL version sits at 56%. 💡 Why does that happen? Post-training pushes tasks toward two extremes, always solved or never solved. In the same setting, the middle band where extra attempts still convert failures into successes collapses from 87.6% to 30.0%, while never-solved grows from 12.4% to 44.0%. There is simply no headroom left for more rollouts to recover. ❓ What is the Sharpening Tax? A single-number diagnostic for how much test-time scalability was lost relative to the base model. It comes out positive in 36 of 42 model-benchmark combinations, and it grows with model scale. It's also cheap to estimate: the value from just 8 rollouts predicts the tax at 32 rollouts with a Spearman correlation of 0.85. 💡 Can it be avoided? Not entirely, but it can be reduced. The proposed PTGS estimates each prompt's difficulty with a Bayesian posterior, then heats hard prompts to encourage exploration and cools easy ones to enforce consistency. It touches neither the RL loss nor the compute budget, and improved accuracy and coverage together under both PPO and GRPO. #ReinforcementLearning# #LLM#

This week, a human boxer fought a six-foot humanoid in a cage in San Francisco. 🥊 The machine is an EngineAI T800 1.73 m, 75 kg, the same model used in China's URKL combat league. First one imported to the US, brought in specifically for sport rather than research or industry. It is not fully autonomous. REK's fighters wear VR headsets & arm-mounted controllers and pilot the robot in real time. So the bout is a human boxer against a human operator with a metal proxy. The engineering problem being solved is teleoperation latency & balance under impact, not machine decision-making. But this is exactly why it's interesting. Getting a 75 kg bipedal machine to absorb strikes & stay upright is a genuinely hard control problem & combat is an unusually honest test of it, because you can't stage a fall. Two things that make this more than spectacle. REK says it's training models on fight footage, so the piloted data becomes training data for autonomous movement later. This same hardware appearing in Chinese league matches is now landing in the US through entertainment rather than industry, which says something about where early commercial demand for humanoids actually is. Robot theater is a fair description. But also a benchmark nobody can fake.

🤖 The T800s are officially rolling off the production line. A new batch of T800 humanoid robots has reportedly left a factory in Henan, China. The same model has been used in URKL. Humanoid robots are moving from prototypes to mass production. 🚀

🤖🔥 Steel meets sweat in Shenzhen! A 6-meter-tall T800 mech​ just landed at Futian COCO Park to launch the Ultimate Robot Knockout Legend (URKL)​ , the world’s first full-scale humanoid robot fighting championship. From July 11–16, catch free robot sparring, parades, AI avatar booths & street dances. Then on July 16, the real battle begins at Nanshan Cultural & Sports Center: 32 elite teams​ from Harvard, Stanford, Tsinghua, HKU & more fight for a ¥10 million prize pool! 🥊💰 Donnie Yen joins as special guest — because nothing says “future sports” like kung fu + giant mechs. 💥🦾 #RobotFighting# #URKL# #Shenzhen# #Mech# #AI# #Technology# #Competition# #DonnieYen# #Stanford# #Tsinghua# #FutureSports# #HardwareSiliconValley# #Event# #Innovation# #ChinaTech# 🔗 Full story: