Leading lab adopts high-bandwidth non-text reasoning (neuralese) internally
One such breakthrough is augmenting the AI’s text-based scratchpad (chain of thought) with a higher-bandwidth thought process (neuralese recurrence and memory).
March 2027: Algorithmic Breakthroughs · Archived PDF, p. 11
At a glance
- Assessment: Emerging
- Confidence in assessment: 35%
- Predicted timing: Early 2027
- Primary source: AI 2027 archived scenario: March 2027: Algorithmic Breakthroughs, p. 11
What AI 2027 Predicted
Tracker summary: The scenario describes a major algorithmic breakthrough occurring around early 2027: augmenting text-based chain-of-thought reasoning with a higher-bandwidth “neuralese” thought process. This involves recurrence and memory mechanisms that allow models to reason in latent space rather than through token-by-token text generation. The concept implies models could process information more efficiently by thinking in a compressed, non-linguistic representation: dramatically increasing effective reasoning depth and breadth. Archived PDF, p. 11.
Scenario source: archived version and page-number reference: March 2027: Algorithmic Breakthroughs, p. 11.
How We Track This
The claim concerns internal adoption at the leading lab, not a requirement to ship a public product. The main internal model or a released model may qualify; public visibility into its reasoning architecture is the limitation.
We monitor:
- Academic research on latent-space reasoning and non-text chain-of-thought
- Recurrence architectures applied to transformers (state-space models, recurrent elements)
- Frontier lab publications on reasoning efficiency beyond text-based CoT
- Deployment of models with non-text reasoning capabilities
- Papers on “continuous thought” or “thinking in latent space”
Current Evidence
Research demonstrates latent-space reasoning mechanisms, but the reviewed sources do not establish adoption in a frontier lab’s main internal or released model.
COCONUT (Continuous Chain of Thought): A December 2024 paper from Meta demonstrated training LLMs to “reason in a continuous latent space,” where internal hidden states replace explicit text tokens as the reasoning medium. The approach enables breadth-first search-like reasoning patterns and showed promising results on logical reasoning tasks.
Later primary research: Geiping et al., February 7, 2025, demonstrate a 3.5-billion-parameter recurrent-depth model that reasons in latent space. This is proof-of-concept research, not evidence of frontier-lab production adoption. Yu et al., February 12, 2025 (RELAY), align loop iterations with supervised reasoning steps and use the resulting chains to fine-tune an autoregressive model. RELAY is relevant recurrence research, but does not establish deployment of a non-text reasoning mechanism in a frontier lab’s main model.
Interpretation: Research motivates the hypothesis, but commentary about a possible mechanism is not evidence of production use.
Public text-based reasoning traces do not establish whether a system also uses undisclosed mechanisms. Confirmation requires technical evidence of the specified non-text intermediate reasoning and its deployment.
Sources:
- Training LLMs to Reason in a Continuous Latent Space: arXiv
- Reflections on Neuralese: Alice Blair, March 12, 2025; commentary
- Thinking Without Words: Luis Cardoso, June 13, 2025; commentary
Counterevidence & Limitations
- No qualifying production deployment was identified in this review; laboratories may not disclose their internal architecture.
- Strong text-based reasoning results represent an alternative development path, not proof that latent reasoning is unnecessary.
- The academic work is promising but small-scale; scaling neuralese-style reasoning to frontier model size is unproven
- It’s possible that labs are working on this internally without public disclosure, making the “emerging” status uncertain in either direction
- Less directly readable intermediate representations create evaluation challenges, but text-based reasoning is not a complete or necessarily faithful explanation of computation.
What Would Change Our Assessment
- Upgrade to “on-track”: Verified technical evidence establishes progress toward the specified production mechanism; an unverified leak is insufficient.
- Upgrade to “confirmed”: Credible technical disclosure that a frontier lab’s main internal or released model uses non-text recurrent reasoning.
- Reassess timing after early 2027: Distinguish evidence that internal adoption is behind from insufficient public visibility into undisclosed systems; public release is not required.
- Status stays “emerging”: As long as research progresses but production deployment remains unconfirmed
Update History
| Date | Update |
|---|---|
| 2026-11-07 | Assessment reviewed. Status: Emerging; confidence: 35%. Geiping et al.’s February 2025 recurrent-depth model demonstrates latent reasoning at proof-of-concept scale. RELAY improves autoregressive reasoning using loop-aligned supervision; it does not demonstrate deployed non-text frontier reasoning. These results support a research precursor, with leading-lab internal adoption still unverified. Correction: Internal adoption can qualify without public shipment. Added dated recurrent-depth and RELAY research and clarified that RELAY improves autoregressive chains rather than establishing a deployed non-text frontier model. |
| 2026-09-06 | Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations. |
| 2026-03-13 | Active research area — COCONUT and RELAY papers demonstrate non-text reasoning pathways. No production deployment yet. Predicted timeline of early 2027 appears aggressive given current state of research. 2026-11-07 · Correction: RELAY, February 12, 2025, uses loop-aligned reasoning supervision to improve autoregressive chains; it is not evidence of a deployed non-text frontier model. The research/deployment distinction remains unresolved. |