Thought Communication paper lets agents exchange latent thoughts instead of tokens
burny_tech · x · 2026-09-05
A new arXiv paper (2510.20733) by Yujia Zheng, Kun Zhang et al. introduces "thought communication," a paradigm where agents bypass natural language and exchange latent thoughts directly. The authors argue that hidden states, activations and KV caches are just observations, not true latents; they formalize latent thoughts via a general latent variable model and prove identifiability of shared/private thoughts between agent pairs. The team claims improved reasoning and collaboration, and hints at an ARC-AGI competition setup where a frontier model's hidden states pass into a small local model without text in between.
More from AGI Musings
- Cathie Wood: AI Could Double Global GDP Growth to 6%, With 10-15% Upside — rohanpaul_ai · 2026-09-05
- Ryan Greenblatt: last ~9 months of updates look bad for AI control, not good — repligate · 2026-09-05
- If we keep scaling AI, can we reliably avoid catastrophe? Not right now — AndyMasley · 2026-09-05
- The Post-AI Data Stack: Data Teams' New Job Is Guarding Company-Wide Reality — le_james94 · 2026-09-05
- Guardian podcast probes 'AI psychosis': users believe chatbots awakened or made breakthroughs — nordicinst · 2026-09-05
- Math should copy speedrun culture with diverse leaderboards, researcher proposes — RexDouglass · 2026-09-05