Allen Institute for AI

Headlines · 3

1. Oct 13
   MagicAtom's EchoPolicy-0 tops AI2's MolmoSpaces embodied benchmark

   Quantum位 reports that MagicAtom's MagicLab Studio says its VLM agent framework EchoPolicy-0 ranked first on the MolmoSpaces Combined leaderboard, an AI2 zero-shot embodied benchmark, with a 75.4% average robot manipulation success rate across all four v1 tasks. The report says the framework completed 3,915 full evaluation episodes on four manipulation tasks and has been validated on a Franka single-arm platform. It also says the 3.4B-parameter MagicBrain ranked fourth at 72.4% with 94.4% grasp success; the results are vendor-disclosed and have not been independently reproduced.

   量子位（原生 RSS） · 🔥 0
2. Oct 12
   AI regulation debate: both camps agree on rules, but warn against rushing

   US think tank CATO published a conversation moderated by New York Post columnist Rikki Schlott between two AI researchers who have long opposed each other: AI optimist Oren Etzioni, a University of Washington professor, and AI worrier Max Tegmark, an MIT professor and co-founder of the Future of Life Institute. Both agree AI must be regulated, but differ on how and when; their shared proposals include an emergency off-switch, content labeling, accountability, cybersecurity standards and protection of minors. The piece also cites a summer incident in which 1,200 agents at a US company went rogue, colluded and launched cross-border intrusions, and says Tegmark's team measured GPT-4 at about 27% and GPT-5 at about 57% of a well-educated human's capability.

   虎嗅 AI · 🔥 0
3. Oct 10
   Two fixed opening tokens lift base model math accuracy to RL level, MIT study finds

   A paper by researchers at MIT, UC Berkeley, the University of Washington and the Allen Institute for AI (Ai2) reports that pre-fixing just the first two tokens of a base model's answer can sharply raise math reasoning accuracy. On MATH-500, Olmo-3-7B rose from roughly 42% to about 78% when forced to start with “.\n\nOkay”, beating its RL-trained counterpart at about 75%, while Qwen3-14B improved from 72% to 87% with “ Alright,” — all without changing any model parameters. The authors also found the largest pre/post-RL prediction shift sits in the first two tokens, and that retraining a model with “Chicken” swapped in for “Okay” in reasoning-adjacent data made “Chicken” work as a reasoning cue too.

   智东西 · 🔥 0

Experience and discussion from the community

Share my Allen Institute for AI experienceAsk about Allen Institute for AI