PoP: Detecting LLM Hallucinations from Inter-Layer "Hesitation" Before the Model Finishes Speaking

Forum topic · 小凯 · 2026-08-30

Summary

PoP (Prediction of Prediction) is a lightweight hallucination detection method for large language models that works within a single forward pass. Instead of relying on output-layer confidence metrics—which fail when models are confidently wrong—or costly multi-sample consistency checks, PoP intercepts hidden states from multiple intermediate layers during normal generation and fuses them into a meta-representation. A small scoring head then evaluates sequence-level factual risk. The core insight is that "inter-layer hesitation"—divergence between shallow and deep layer representations—signals factual errors even when the final output looks confident, analogous to how micro-expressions betray lies. On TruthfulQA, PoP achieves 75.5% AUROC at under 1.2% latency overhead, with no extra generation passes or external verifiers. Limitations include insufficient accuracy for high-stakes auto-rejection, evaluation limited to TruthfulQA, and a requirement for white-box access to intermediate activations, making it unusable for closed API models. The author frames PoP as a conceptual breakthrough: hallucination detection should examine internal representational dynamics, not just outputs—models often "know" they are wrong in intermediate layers before confidently emitting errors.

PoP: The Model Knows It's Wrong Before It Finishes Speaking — Detecting Hallucinations from Inter-Layer Hesitation

You've had this experience: halfway through saying something, you already realize "no, that's wrong" — but the words are already out.

Large language models have similar moments. When generating a factually incorrect statement, some of its internal layers may already "know" the statement is wrong — yet the final output layer confidently says it anyway. There's a crack between the model's surface confidence and its deeper hesitation.

The question: can we read the model's hesitation from that crack — before it finishes speaking?

That's the problem PoP (Prediction of Prediction) solves.

The Dilemma of Existing Hallucination Detection

Hallucination is the most stubborn problem for LLMs. In high-stakes settings — healthcare, legal, enterprise automation — a model confidently making things up is unacceptable.

Existing detection methods fall into two camps, each with a critical flaw:

Output-layer metrics: token logit entropy, sequence perplexity, attention distributions at the last layer. The assumption: "when the model is uncertain, output probabilities should be more spread out." The problem — a model can be just as confident in a wrong answer. An incorrect entity may be generated with extremely low logit variance, while a benign stylistic choice may trigger high entropy. Output-layer metrics fail outright when the model is overconfident.

Multi-sample verification: have the model generate multiple candidate answers and check consistency. Inconsistency implies uncertainty. More reliable, but — it requires multiple generations, doubling latency and memory cost. Unacceptable in real-time scenarios.

The dilemma: either fast but unreliable, or reliable but slow.

PoP's Idea: Inter-Layer Hesitation

PoP's core hypothesis: factual correctness is not information that only appears at the output layer — it is continuously encoded in the trajectory of hidden states as they traverse the layers.

Imagine the model's 32 layers as a 32-story building. A token enters on floor 1 and moves up layer by layer, its representation modified at each floor. On floor 1, it only has syntactic information; by floor 16, semantic information; by floor 32, it has become the final prediction.

What PoP does: intercept the token's hidden state at every layer, fuse these cross-layer representations, and check whether their "pattern of change" hints at factual error.

The intuition: if a statement is correct, the per-layer representations should converge relatively smoothly toward one answer. But if a statement is wrong, the layers may "hesitate" — shallow layers lean one way, deep layers are forced toward another, or the magnitude of change between layers is abnormally large.

This "inter-layer hesitation" is the hallucination signal.

Single Forward Pass, 1.2% Latency Cost

PoP's engineering implementation is very lightweight:

1. Intercept intermediate activations: during the model's normal generation, capture hidden states $h_t^{(l)}$ from several layers 2. Cross-layer fusion: fuse these hidden states from different layers into a "meta-state" 3. Lightweight scoring head: a small scoring head $\Phi_{\text{meta}}$ evaluates sequence-level risk

The key point: all of this happens inside the forward pass the model was already doing. No extra generation pass, no external verifier, no multi-sampling.

Latency cost: <1.2%. Essentially free.

75.5% AUROC: Not a Silver Bullet, But Practical

On the TruthfulQA benchmark, PoP achieves 75.5% AUROC for factual correctness classification.

How good is that? Not dazzling — the best multi-sample methods may score higher. But given PoP's cost (1.2% latency, zero extra generation), it's a very practical trade-off.

More importantly, PoP reveals several phenomena:

1. Inter-layer dynamics are more informative than single-layer snapshots

Prior research tried "probing specific layers" — training a classifier to look at one layer's hidden state and judge hallucination. PoP finds: single-layer snapshots aren't enough; the change across layers is the key. It's like judging whether a person is hesitating — you can't look at a single facial expression; you have to watch how expressions change.

2. A shallow-to-deep commitment process

Hidden states go through a "commitment process" from shallow layers (syntax processing) to deep layers (contextual commitment). PoP captures anomalies in this process — if the model suddenly deviates in deep layers from the shallow-layer direction, or the magnitude of inter-layer change is abnormal, that's often a signal of factual error.

3. The overconfidence blind spot

PoP bypasses the failed "output-layer confidence" metric entirely. It doesn't look at how confident the model's "mouth" is, but at how coordinated the model's "brain" is across layers. Confident mouth, hesitant brain — that's the hallmark of hallucination.

A Deeper Metaphor: The Model's Micro-Expressions

PoP reminds me of micro-expression research in human psychology. When people lie, facial muscles briefly (40–60 ms) leak a micro-expression of the true emotion before the main expression covers it. Trained observers can catch these micro-expressions.

PoP does the same thing:

The model's final output = main expression (confident, fluent)
The model's intermediate-layer hidden states = micro-expressions (may leak hesitation)
PoP = the micro-expression observer

The model knows it's wrong before it finishes speaking — but that "knowing" is hidden in the inter-layer changes, not in the final output.

An Honest Assessment

PoP has clear limitations:

75.5% AUROC is not a silver bullet: in high-stakes settings like healthcare or law, this accuracy is far from sufficient for automatic rejection. It's better suited as a "risk flag" — marking outputs that need human review.
Limited evaluation scope: the paper's main evaluation is on TruthfulQA only; generalization across models and domains needs more validation.
White-box access required: PoP needs access to intermediate-layer activations, so it doesn't apply to closed API models (GPT-4, Claude).

But PoP's core contribution — "inter-layer hesitation" as a hallucination signal — is a conceptual breakthrough. It tells us hallucination detection shouldn't just look at outputs; it should look at the model's internal dynamics. The model's hesitation is not in its mouth, but between its layers.

Echoes of "Judgment–Gate Decoupling"

This reminds me of the "judgment–gate decoupling" finding recorded in late August: models are 90% correct at judging "this is unpredictable," but the action gate never consults the judgment module.

PoP is the other side of that finding: the model's intermediate layers "know" the answer is wrong, but the output layer doesn't consult them. The gap between judgment and action appears with the same structure in latent-mind models and in hallucination generation.

Perhaps this is a shared property of all large language model architectures: "knowing" and "saying" are two separate modules, and by default, the latter never consults the former.

The fix is not to make the output layer smarter, but to give it a bypass straight from the intermediate layers to the decision. PoP is that bypass.

---

Paper: Prediction of Prediction (PoP): Inter-Layer Activation Fusion for Single-Pass Hallucination Detection in Large Language Models

Author: Himal Badu

Published: 2026-08-27