PoP: Detecting LLM Hallucinations via 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 reads internal "hesitation" from activations across transformer layers during a single forward pass. Unlike output-level metrics (logit entropy, perplexity), which fail when models are confidently wrong, and multi-sample consistency checks, which double latency and memory cost, PoP intercepts hidden states from multiple layers, fuses them into a meta-state, and applies a lightweight scoring head to assess sequence-level factual risk. On TruthfulQA, PoP achieves 75.5% AUROC for factual correctness classification at a latency overhead below 1.2%, with no extra generation passes. The key insight is that inter-layer dynamics—the trajectory of hidden states from shallow syntax to deep contextual commitment—carry more hallucination signal than any single-layer snapshot: when a model produces a factually wrong statement, its intermediate layers may "know" it is wrong even as the output layer remains confident. Limitations include reliance on white-box access to intermediate activations, limited TruthfulQA-only evaluation, and accuracy insufficient for automated rejection in high-stakes settings, making PoP best suited as a risk-flagging signal. Paper: arXiv:2608.27165, by Himal Badu.

The Model Knows It's Wrong Before It Finishes Speaking: PoP Detects Hallucinations via Inter-Layer Hesitation

Have you ever been halfway through a sentence and realized, "wait, that's not right"—but the words were already out?

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

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

This is what PoP (Prediction of Prediction) sets out to solve.

The Dilemma of Existing Hallucination Detection

Hallucination is the most troubling problem for LLMs. In high-risk scenarios like healthcare, law, and enterprise automation, a model confidently making things up is unacceptable.

Existing detection methods fall into two categories, each with serious flaws:

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

Multi-sample verification: have the model generate multiple candidate answers and check consistency. Inconsistency signals uncertainty. This is more reliable, but the cost—multiple generations, doubling latency and memory overhead—is unacceptable in real-time settings.

The dilemma: 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 upward, its representation modified at each floor. At floor 1, it holds only syntactic information; by floor 16, semantic information; by floor 32, it becomes the final prediction.

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

The intuition: if a statement is correct, the layer representations should converge relatively smoothly toward an answer. But if it's wrong, the layers may "hesitate"—shallow layers lean one direction, deep layers are forced to turn another way, or the magnitude of inter-layer change 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: while the model generates normally, capture hidden states $h_t^{(l)}$ from selected layers 2. Cross-layer fusion: fuse 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 within the forward pass the model was going to run anyway. No extra generation passes, no external verifiers, no multi-sampling.

Latency cost: <1.2%. Practically 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 probed specific layers—training a classifier on one layer's hidden state to detect hallucination. PoP finds: a single-layer snapshot is not enough; the change across layers is what matters. It's like judging whether a person is hesitating—you can't look at a single expression; you have to watch how the expression changes.

2. The shallow-to-deep commitment process

Hidden states move through a "commitment process" from shallow layers (syntactic processing) to deep layers (contextual commitment). PoP captures anomalies in this process—if the model suddenly departs at deep layers from the shallow layers' direction, or inter-layer change is abnormally large, that often signals factual error.

3. The overconfidence blind spot

PoP bypasses the failing "output-layer confidence" metric entirely. It doesn't look at how confident the model's "mouth" is, but at how coordinated the layers are inside its "head." Confident mouth, hesitant brain—that's the signature of hallucination.

A Deeper Metaphor: The Model's Micro-Expressions

PoP recalls research on "micro-expressions" in human psychology. When people lie, facial muscles leak the true emotion in fleeting (40–60 ms) micro-expressions before the main expression covers them. Trained observers can catch these.

PoP does the same:

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

The model knows it's wrong before it finishes speaking—it's just that this "knowing" hides in inter-layer dynamics, 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 and law, this accuracy is far from sufficient for automated rejection. It fits better as a "risk flag"—marking outputs for human review.
Limited TruthfulQA evaluation: the paper's main evaluation is on TruthfulQA only; cross-model, cross-domain generalization needs more validation.
White-box access requirement: PoP needs access to intermediate 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 look only at outputs; it should look at the model's internal dynamics. The model's hesitation isn't in its mouth—it's between the layers.

Resonance with "Judgment–Gate Decoupling"

This echoes an earlier finding on "judgment–gate decoupling": a model is 90% correct when judging "this is unpredictable," yet its action gate never consults the judgment module.

PoP is the other side of the same coin: the model's intermediate layers "know" the answer is wrong, but the output layer never consults them. The crack between judgment and action appears with the same structure in latent-cognition models and hallucination generation.

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

The fix is not making the output layer smarter, but giving 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