The Illusion of Superposition / Latent CoT Superposition Analysis
Summary
The Illusion of Superposition is a paper-level handle for a latent-CoT interpretability study. It tests whether continuous latent thoughts actually maintain multiple candidate reasoning paths in superposition, rather than collapsing to a discrete token interpretation or shortcutting directly to the answer.
Interface
- Systems studied: Soft Thinking-style soft token mixtures and Coconut-style latent reasoning.
- State inspected: hidden states at latent reasoning positions and intermediate Transformer layers.
- Probes: Logit Lens, token-level soft-versus-argmax intervention, and entity-level belief tracking.
- Positive result boundary: shallow from-scratch models on a symbolic ProsQA variant can show superposition-like belief evolution.
- Negative result boundary: off-the-shelf and fine-tuned pretrained models tend to collapse or shortcut.
Role In The Wiki
Use this entity when discussing whether latent reasoning, recurrent depth, hidden thoughts, or soft tokens are actually used for multi-path reasoning. Its role is mainly cautionary: extra hidden computation is not evidence of useful latent deliberation unless ablations and probes show that the hidden state carries the intended uncertainty or intermediate state.
For foundation time-series work, this maps to dynamic compute and latent-state evaluation. A TSFM that spends extra hidden steps on hard windows, regimes, or candidate futures needs the same kind of mechanism audit: no-loop/no-latent ablations, state probes, and checks that the model has not simply learned an easier shortcut.