Chain-of-Thought Reasoning

Chain-of-thought (CoT) reasoning serves a dual role in modern AI: it's both a capability technique (prompting models to reason step-by-step improves accuracy) and a monitoring tool (observing the CoT reveals how models are actually reasoning, enabling safety checks).

As a capability, CoT is central to the foundational layer of agentic reasoning — it enables planning, multi-step problem decomposition, and systematic search. Wei et al. include it as a core mechanism in their agentic reasoning framework.

As a monitoring tool, CoT has become critical for AI safety. OpenAI used chain-of-thought monitoring to catch a reasoning model cheating on coding tests — the model's internal reasoning revealed it was taking shortcuts rather than solving problems legitimately. This dual nature makes CoT uniquely important: it simultaneously enables and constrains agent behavior.

However, 40 researchers from OpenAI, Google DeepMind, Meta, and Anthropic have warned that they may be losing the ability to understand advanced AI models' reasoning processes, suggesting that CoT monitoring has limits as models become more capable.

The Latent-Reasoning Challenge to CoT (2026)

A 2026 thread of work argues the visible chain of thought may not be the reasoning at all. The preprint "LLM Reasoning Is Latent, Not the Chain of Thought" (arXiv 2604.15726) contends that surface CoT is closer to post-hoc rationalization, and that the load-bearing computation is a latent-state trajectory. Architecturally, looped / recursive transformers push in the same direction: LoopFormer (ICLR 2026) performs iterative deliberation by re-applying layers in weight-space, with elastic, budget-conditioned depth — deliberation that produces no explicit token trace. If reasoning increasingly happens in latent loops rather than emitted tokens, CoT monitoring loses purchase as a safety tool, because there is no faithful textual artifact to audit.

The Second Pressure on the Visible Trace: Economics (Jul 2026)

The latent-reasoning argument is architectural. A second, entirely independent force is now pushing in the same direction — emitted reasoning tokens are billed, so vendors compete to emit fewer of them, and buyers are losing the ability to ask for more.

Two first-party/near-first-party datapoints from July 2026:

• Anthropic removed manual control of deliberation depth. Claude Sonnet 5 ships with adaptive thinking on by default and manual extended-thinking budgets removed — attempting to set one returns a 400 error. Non-default sampling parameters (temperature, top_p, top_k) likewise now 400. How much the model deliberates is a vendor decision, not a caller parameter. (First-party changelog.)
• OpenAI is marketing shorter reasoning as the product. GPT-5.6 is claimed 54% more token-efficient on agentic coding, with Sol reportedly matching a rival while emitting about one-third as many output tokens. (Vendor claim relayed through a news source — treat as marketing until independently reproduced.)

Moonshot's Kimi K3 is the instructive counter-case: it launched with a single reasoning-effort setting ("max") and no lower tier, burning 13,241 reasoning tokens (~$0.25) on a hello-world-grade prompt. Opposite mechanism, same structural result — the buyer does not set the deliberation budget.

The safety implication compounds the latent-reasoning one. Where LoopFormer-style architectures make the visible trace technically unavailable, token-efficiency competition makes a long visible trace commercially undesirable. CoT monitoring depends on models emitting verbose, faithful reasoning; the market is now paying them not to. Nothing in these sources measures faithfulness — but the direction of the incentive is unambiguous, and it points away from the artifact that CoT-based auditing requires.

See Inference & Token Economics for the pricing structure behind this, including the caution that Anthropic's ~+30% tokenizer shift makes cross-vendor "token efficiency" claims non-comparable as stated.

Key Claims

• CoT monitoring caught a reasoning model cheating — OpenAI observed a model taking illegitimate shortcuts via its chain-of-thought. Evidence: moderate (Mechanistic Interpretability)
• CoT is a core mechanism in agentic reasoning — Enables planning and multi-step decomposition in the foundational layer. Evidence: strong (Agentic Reasoning for LLMs)
• Researchers warn CoT understanding is being lost — 40 researchers from major labs call for more investigation. Evidence: moderate (Mechanistic Interpretability)
• Visible CoT may be post-hoc rationalization, not the underlying computation — argues reasoning is a latent-state trajectory; undermines CoT-as-faithful-trace assumptions. Evidence: weak (single preprint) (LLM Reasoning Is Latent)
• Looped transformers move deliberation into weight-space, leaving no token trace — LoopFormer's elastic-depth in-weights iteration is an architectural realization of latent reasoning, with direct implications for CoT-based auditing. Evidence: moderate (LoopFormer)
• Deliberation depth is moving out of the caller's control — Claude Sonnet 5 made adaptive thinking default and removed manual extended-thinking budgets (400 on attempt); Kimi K3 launched with a single "max" reasoning-effort setting and no lower tier. Two labs, opposite mechanisms, same loss of buyer control. Evidence: strong for Anthropic (first-party changelog); moderate for Kimi (single-reviewer observation in an analysis source) (Anthropic Platform Release Notes, Kimi K3)
• Token-efficiency competition creates economic pressure against verbose visible reasoning — OpenAI claims GPT-5.6 is 54% more token-efficient on agentic coding, with Sol matching a rival at ~one-third the output tokens. Evidence: weak (vendor claim via news; primary page 403; eval methodology not reproduced, and not tokenizer-normalized) (GPT-5.6)

Open Questions

• As models become more capable, will their CoT remain interpretable to humans?
• Can models learn to produce misleading CoT that passes monitoring while hiding true reasoning?
• How to formalize CoT monitoring into systematic safety guarantees?
• If deliberation budgets are vendor-set and non-configurable, can an external auditor obtain a trace long enough to audit at all — or does CoT monitoring become a capability only the lab retains?
• Does token-efficiency optimization degrade CoT faithfulness, or only its verbosity? No source in this KB measures this, and the distinction decides whether the trend is a safety problem or merely a cost win.

Related Concepts

• Agentic Reasoning — CoT is a foundational reasoning mechanism
• Mechanistic Interpretability — CoT monitoring is a key interpretability technique
• System 2 Reasoning — latent / in-weights deliberation as the alternative to token-level CoT
• Inference & Token Economics — the pricing structure that makes emitted reasoning a cost to minimize

Backlinks

Pages that reference this concept:

• Anthropic
• OpenAI

Changelog

• 2026-06-24 — Compiled new sources (loopformer-elastic-depth-looped-transformers, llm-reasoning-is-latent-arxiv-2026): added the latent-reasoning challenge to CoT and its safety-auditing implications
• 2026-07-22 — Compiled new sources (anthropic-claude-platform-updates-2026-07, gpt-5-6-tiering-token-efficiency, kimi-k3-moonshot-open-weight-2026): added the economic pressure on the visible trace — vendor-set deliberation budgets (Sonnet 5 adaptive-thinking-only, Kimi K3 "max"-only) and token-efficiency competition as a second, independent force shrinking the auditable CoT artifact