← Back to papers

J-CoT: Chain-of-Thought in J-Space

Junde Wu, Jiayuan Zhu, Fengling Liu, Minhao Hu, Jiazhen Pan · arXiv:2607.21981 · July 2026

Section: Mechanisms & Architecture

Bottom line: J-CoT proposes a middle ground between explicit chain-of-thought (every step verbalized) and dense latent reasoning (opaque hidden states) — using J-space vocabulary-indexed coefficients as a 'J-thought' carried across reasoning cycles, achieving the best of both worlds.

ELI5

Imagine murmuring under your breath while solving a problem: 'carry the 5… no wait… yep that's better' — vocabulary fragments that keep your place without forming full sentences. That's the J-thought: linguistically grounded but not forced into full articulation.

Key points

  • J-thought uses vocabulary-indexed J-space coefficients, not decoded language or dense hidden states
  • J-CoT-Zero matches/exceeds the strongest latent baseline without training
  • J-CoT-Train achieves 50.2% average across 8 benchmarks, +2.7 over the best latent method
  • Interface spectrum experiment shows J-CoT at the midpoint outperforms both dense and linguistic endpoints
  • Scales with model capacity — larger models benefit more from additional cycles

Why it matters

The paper uses the same J-space that Gurnee identified as a workspace — here repurposed as a computational interface, not just a diagnostic tool.

Limits

  • Requires backbone with J-lens dictionaries computed offline
  • No evaluation on creative/open-ended tasks

Connections

← Back to papers