Why does textual chain-of-thought avoid the representational drift problem automatically?
This explores why reasoning expressed in words doesn't suffer the 'drift' that plagues latent (vector-space) reasoning — and the corpus suggests the answer is that text is its own anchor, though that anchoring comes with a hidden cost.
This explores why reasoning written out in words sidesteps the representational drift that breaks latent chain-of-thought — and the short version is that text can't drift, because every step has to land on an actual word. When models reason in continuous latent space, the intermediate 'thoughts' are free-floating vectors with nothing pinning them to meaning. The corpus's clearest account of the failure shows latent CoT collapsing two ways at once: gradients fade along the latent steps under outcome-only supervision, and the latent space wanders without semantic grounding — robust latent reasoning has to be propped up with extra 'space supervision' just to preserve geometric structure (Why does latent chain-of-thought fail so easily in training?). Textual CoT gets that grounding for free: each reasoning step is forced through the discrete vocabulary, so there's no continuous space to drift through in the first place.
The cost of restoring what text has by default becomes visible when researchers try to make latent reasoning behave more like text. One line of work wraps continuous thoughts in normalizing flows specifically to recover exact likelihoods, sampling, and scoring — explicitly trying to 'match the tractability of textual CoT' (Can continuous thoughts have tractable likelihoods for sampling and scoring?). That's the tell: text's anchoring isn't a feature someone engineered, it's a property you have to laboriously rebuild once you leave the token space. The flip side is that latent thought also unlocks scaling dimensions text can't reach (Can latent thought vectors scale language models beyond parameters?), which is why people keep chasing it despite the drift.
But here's the part you might not expect: the same property that prevents drift is also what limits textual CoT. Because every step is tethered to familiar token patterns from training, the model tends to reproduce the *form* of reasoning rather than perform genuine inference. Multiple notes converge on this — CoT works by constraining models to replay learned reasoning schemata, and degrades predictably under distribution shift, the signature of imitation rather than real capability (Does chain-of-thought reasoning reveal genuine inference or pattern matching?, Why does chain-of-thought reasoning fail in predictable ways?). Format matters more than logical content: invalid prompts work nearly as well as valid ones, and training format shapes strategy far more than domain does (What makes chain-of-thought reasoning actually work?, What makes chain-of-thought reasoning fail in language models?).
So the anchor that stops drift is the same leash that keeps reasoning inside the training distribution. Even within text, you can see micro-drift trying to break through: token-level memorization, dominated by 'local' dependence on the immediately preceding tokens, drives up to 67% of CoT reasoning errors — and gets worse exactly when complexity pushes the model off its familiar distribution (Where do memorization errors arise in chain-of-thought reasoning?). The deeper lesson is that grounding is never automatic; text just bundles it in. When models need grounding text can't supply, they have to reach outside themselves — interleaving reasoning with real tool calls and environment feedback to stop errors from compounding (Can interleaving reasoning with real-world feedback prevent hallucination?), because even strong textual priors can override what's actually in front of the model (Why do language models ignore information in their context?).
The thing worth taking away: 'avoiding drift automatically' isn't textual CoT being smarter — it's text trading the freedom to find genuinely new representations for the safety of staying near words it already knows. And if you don't need the chatty form, you can keep nearly all of that grounding at a fraction of the cost (Can minimal reasoning chains match full explanations?).
Sources 11 notes
Outcome supervision alone causes gradient attenuation along latent steps and lets the latent space wander without semantic grounding. Robust latent reasoning requires both dense trajectory supervision and space supervision that preserves geometric structure rather than compressing it.
NF-CoT models continuous thoughts as an autoregressive normalizing flow inside the LLM's causal stream, recovering exact likelihood, probabilistic sampling, and KV-cache compatibility. This enables policy-gradient refinement and trajectory scoring on non-verbal reasoning, matching the tractability of textual CoT.
Latent-Thought Language Models achieve superior sample and parameter efficiency by coupling fast local variational learning with slow global decoder learning. This dual-rate scheme scales few-shot reasoning across both model and latent size, creating independent scaling dimensions beyond traditional parameter scaling.
CoT works by constraining models to reproduce familiar reasoning patterns from training, not by enabling novel symbolic reasoning. Performance degrades predictably under distribution shifts—the signature of imitation rather than capability emergence.
CoT guides models to pattern-match reasoning structure rather than perform genuine inference. This explains distribution-bounded failures, why structural coherence matters more than content correctness, and why performance optimizes against interpretability.
Research shows training format shapes reasoning strategy 7.5× more than domain, demo position swings accuracy 20%, and invalid CoT prompts work as well as valid ones. CoT is pattern-guided generation, not formal logic.
Research shows CoT mirrors reasoning form without true logical abstraction. Format matters more than content, invalid prompts work as well as valid ones, and scaling reasoning creates instruction-following deficits.
STIM framework identifies local, mid-range, and long-range memorization sources in CoT reasoning. Local memorization—based on preceding tokens—accounts for up to 67% of reasoning errors, especially as complexity increases and distributional shift occurs.
ReAct demonstrates that alternating verbal reasoning with external tool queries (Wikipedia API, environment interaction) prevents error propagation by injecting real-world feedback at each step. On knowledge-intensive and interactive tasks, this approach outperforms pure chain-of-thought and reinforcement learning by 10-34% absolute accuracy.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Chain of Draft achieves equivalent accuracy to standard chain-of-thought on arithmetic, symbolic, and commonsense tasks while using only 7.6% of tokens. The 92.4% of removed tokens served style and documentation, not computation.