MIT

MIT report details how Pokémon Go AR data is training delivery robot navigation systems

Developed new benchmark with Anthropic revealing limitations in AI coding assistants

Hackathon team built wearable AI for physical movement guidance.

Introduced Recursive Language Models (RLMs) handling 10M+ tokens

Developed silent artificial muscle fibers using electrohydraulic actuation capable of lifting over 1kg

Collaborative paper with Oxford and CMU finds AI assistance can boost then harm human performance

collaborators:

Oxford University,Carnegie Mellon University

Published study finding GPT-5 assistance creates a 'productivity trap', boosting short-term scores but eroding independent problem-solving.

collaborators:

Oxford,Carnegie Mellon University

key finding:

AI assistance reduces persistence and hurts independent performance.

Co-authored a paper with Stanford and Google proposing a method for LLMs to self-improve their prompts.

Released a new benchmark with Anthropic revealing systematic limitations in AI coding assistants

Proposed RL training for LLMs to output multiple plausible answers instead of single guesses

research area:

RL alignment

focus:

diversity in model outputs

Reported using Pokémon Go AR data to train delivery robot navigation systems

application:

autonomous delivery robot navigation

data source:

Pokémon Go AR data

Published paper introducing Level-2 Inverse Games framework for multi-agent inference

Identified and solved critical vulnerability of 'error cascades' in LLM-based Multi-Agent Systems, developing a governance layer defense.

MIT researchers develop 'agent harness' system enabling AI agents to proactively plan and execute complex multi-step tasks autonomously

Researchers unveiled NeuroSkill, a real-time agentic system that integrates BCI with foundation models to proactively model human cognitive states.

MIT researchers publish paper on bio-integrated phone cases using algae.