Abstract CoT: latent reasoning cuts tokens up to 11.6x with CoT-level performance
evijit · x · 2026-09-11
A paper by Meta researchers, "Thinking Without Words: Efficient Latent Reasoning with Abstract Chain-of-Thought," explores reasoning without natural language:
- Abstract CoT has the model emit a short sequence of "abstract" tokens from a reserved vocabulary in place of verbose natural-language CoT before answering.
- Training uses a policy-iteration-style warm-up: bottlenecking from verbal CoT via masking + SFT, then self-distillation via constrained decoding, followed by warm-started RL.
- Results: up to 11.6x fewer reasoning tokens with comparable performance on math, instruction-following, and multi-hop reasoning, generalizing across model families; an emergent power-law distribution over abstract tokens appears.
The thread also raises the idea that efficiency pressure pushes models away from English — relevant to how agents might pass messages in the future.
More from Research
- MSRA, UTS and Tsinghua unveil UniSteer: injecting human corrections into RL for flow-matching VLAs — jiqizhixin · 2026-09-11
- KAIST releases TIDES: a semester-long bilingual dataset of real team collaboration — josephseering · 2026-09-11
- World Model RL Debiasing Cuts Cost of Scaling Autonomous Research Agents — illinois · 2026-09-11
- Links: Economist AI-labor article and adaptive capacity paper overview — soumitrashukla9 · 2026-09-11
- RL-trained agents should carry a strong simulation prior, argues vooooogel — voooooogel · 2026-09-11
- Humansand launches to simulate humans for RLHF data — gharik · 2026-09-11