Chain-of-Thought Is a Press Release, Not the Reasoning

When an LLM writes out its reasoning step-by-step, that text may be a partial translation of a process that already happened in the hidden layers — not the process itself.

~6 min read · arXiv / South China University of Technology · April 21, 2026 · Alignment

TL;DR This position paper argues that LLM reasoning is primarily mediated by latent-state trajectories in the model's hidden activations, not by the visible chain-of-thought tokens. It formalizes three competing hypotheses, marshals existing mechanistic evidence, and runs a compute-audited empirical program showing that the winning explanatory object shifts predictably across task regimes. The upshot: the field should stop treating CoT as the default object of study for reasoning, and start treating latent dynamics as the primary target.

Most researchers studying LLM reasoning are watching the wrong screen.

When a model writes "Step 1: … Step 2: … Therefore…", it's tempting to take that text at face value — as the thing doing the reasoning. The chain-of-thought (CoT) is visible, inspectable, and satisfyingly human-like. But a growing body of mechanistic evidence suggests this transcript may be more like a press release than a board meeting: a post-hoc summary of decisions that were effectively made somewhere else, in the model's internal activations.

That is the central claim of this position paper by Wenshuo Wang. And it's worth taking seriously — because almost everything we claim about LLM faithfulness, interpretability, and inference-time reasoning interventions depends on which view is correct.

The Three-Way Debate

💡 The Core Idea Think of it this way: the chain-of-thought tokens are what gets printed in the newspaper. The actual deliberation — the hidden-state trajectory — is what happened in the room. They overlap, but they aren't the same thing. And when they diverge, it's the room that determines the outcome.

Wang structures the debate around three competing hypotheses, which the paper labels H1, H2, and H0:

• H2 (Surface CoT mediation): Reasoning is the visible natural-language trace. Hidden states are necessary to generate it, but the privileged object of reasoning is the explicit token sequence.
• H0 (Generic serial compute): Most reasoning gains are explained by extra computational budget — more decoding steps, more sampling, more search — regardless of whether that budget takes the form of CoT text or something else.
• H1 (Latent trajectory mediation): Reasoning is primarily carried by hidden-state trajectories. Surface CoT is a partial, sometimes imperfect interface to that process. Serial compute matters because it gives latent dynamics more room to iterate and stabilize.

These hypotheses make asymmetric predictions. If H2 is right, interventions on the visible trace should have the strongest causal leverage. If H0 is right, matched compute expansions (even with meaningless filler tokens) should recover most gains. If H1 is right, latent states should encode task-relevant commitments before the model fully verbalizes them, and targeted hidden-state interventions should alter reasoning even when the surface trace is partial or absent.

flowchart LR
    A[Input Prompt] --> B[Hidden-State Trajectory Z]
    B -->|partial readout| C[Surface CoT Tokens S]
    B -->|drives output| D[Final Answer]
    C -.->|sometimes influences| D
    E[Serial Compute Budget B] --> B
    E --> C

What the Evidence Actually Shows

The paper's synthesis of the existing literature is its most immediately useful contribution. Wang weights studies by diagnostic force — not by paper count — and the verdict is sobering for CoT enthusiasts.

The strongest case for H2 turns out to be narrow. Classic CoT prompting results don't isolate the surface trace as causal, because they simultaneously change visible semantics and serial compute. The genuinely strong H2 evidence comes from settings where the visible trace is made constitutive of the reasoning process — structured symbolic solvers, externalized retrieval plans, tool-use chains where intermediate actions are inspectable and binding. In those regimes, surface CoT earns its interpretive weight. But that's a special case, not a general principle.

The strongest case for H0 is real and substantial. Filler-token experiments (replacing semantically meaningful intermediate tokens with ...-style placeholders while preserving the slot count) show that some CoT gains survive without meaningful token content. Self-consistency, tree-of-thought search, and adaptive test-time compute scaling all demonstrate that how you spend inference budget matters more than whether it's expressed as coherent prose. H0 captures something important. But it can't explain why specific hidden states and features are so tightly coupled to reasoning outcomes — why probing a particular activation pattern can predict final correctness before the model has verbalized its answer.

The case for H1 comes from three converging lines. First, the residual: ordinary CoT is regularly unfaithful, incomplete, or selective relative to what the model actually "knows" at the hidden-state level. Second, positive probe evidence: propositional probes can track latent world-state representations that remain faithful even when model outputs are biased or injected. In reasoning models, hidden states encode intermediate and even future answer correctness before the relevant reasoning text is fully generated. Third, direct intervention: causal steering of small sets of latent features can improve reasoning accuracy without explicit CoT prompting, and continuous latent-space reasoning can outperform standard CoT on tasks that benefit from backtracking.

Empirical Results: The Regime Map

The paper doesn't just theorize — it runs a compute-audited empirical program across a controlled template suite and four naturalistic benchmarks (GSM8K-Platinum, HotpotQA, MATH-500, HumanEval+), using three model families (Qwen3-8B, Qwen3-32B, Llama-3.1-8B-Instruct).

The key result is what the paper calls the substrate switch:

This is the finding that makes H1 the default rather than the universal winner. Latent families dominate in ordinary settings — the kind most LLM reasoning research actually cares about. But when visible traces are forced to be constitutive (e.g., a retrieval plan that gates access to documents), surface-mediated accounts regain local force. When the problem is search-heavy and format-agnostic, compute budget becomes the dominant variable.

The mediator tests — temporal precedence, necessity under ablation, sufficiency under patching, and a direct Z-vs-S contrast — consistently support H1 in ordinary regimes. Corrupting the visible surface trace while preserving the latent trajectory is less damaging than the reverse. That asymmetry is the clearest single-study summary of what H1 predicts and observes.

Why Current Evaluations Are Broken

The methodological critique here is sharp and practically important. Most experiments in the LLM reasoning literature are non-diagnostic with respect to H1, H2, and H0 — because they move surface traces, latent states, and serial compute all at once.

Wang's proposed fix is a six-arm design that requires: (a) a baseline, (b) a targeted surface manipulation, (c) a compute-matched surface control without the task-relevant semantic content, (d) a targeted latent intervention, (e) a matched latent control or sham, and (f) a compute-only budget expansion. Only experiments with all six arms can cleanly adjudicate among the three hypotheses. Without the matched controls, any result can be claimed post-hoc by whichever hypothesis sounds most plausible.

⚠️ Watch out for

• Studies that report CoT accuracy gains without holding serial compute fixed — these can't separate H2 from H0.
• Latent probing results without paired causal follow-ups — a predictive probe isn't the same as causal evidence.
• Claims that "reasoning improved" without specifying which of S, Z, or B was actually manipulated — the ambiguity is almost never benign.
• H1 applied universally: the paper explicitly limits its claim to ordinary regimes where no external trace is made constitutive.

Why This Matters for Practitioners

If H1 is correct as a default working hypothesis, several standard practices in the field need to be revisited.

Faithfulness evaluations that compare CoT text to ground truth are measuring the quality of the press release, not the board meeting. A model can be behaviorally correct while producing an unfaithful verbal explanation, and vice versa.

Safety monitoring via CoT works best in cases where the harmful behavior requires explicit step-by-step computation to execute — the paper cites evidence that in those cases, CoT is harder to fake. But for behaviors that don't require overt step-by-step reasoning, CoT monitoring may give false assurance.

Reasoning benchmarks that measure CoT quality as a proxy for reasoning quality may be systematically measuring the wrong thing in ordinary regimes. A model that reasons well latently but verbalizes poorly would underperform; a model that generates fluent-sounding traces without reliable latent commitment would overperform.

The recommendation is concrete: treat latent-state dynamics as the default target of study, and design evaluations that explicitly separate surface traces, latent states, and serial compute before drawing conclusions about what's actually driving performance.

Source: LLM Reasoning Is Latent, Not the Chain of Thought Authors: Wenshuo Wang Published: 2026-04-21 PDF: https://arxiv.org/pdf/2604.15726