💻 computer science

Turbulence Is Not What You Need to Detect Hallucinations

Despite the theoretical appeal of viewing hallucinations as dynamical instabilities in large language models, this paper demonstrates that static representations are far more effective for detection, as adding dynamical features provides no statistically significant improvement over simple linear probes.

Original authors: Igor Itkin

Published 2026-07-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Igor Itkin

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a Large Language Model (LLM) as a giant, magical river flowing through a canyon of layers. For a long time, researchers had a very tempting idea: that when the model starts "hallucinating" (making things up), the river gets turbulent. They thought the water would start swirling, spinning, and losing its stability, creating a chaotic flow that a detector could spot.

This paper, written by independent researcher Igor Itkin, decides to test that idea with a very strict, scientific ruler. The short answer? The river might be swirling, but you can't use the swirls to catch the liars.

Here is the story of what they found, broken down into the big picture, the "myth-busting," and where the real magic actually happens.

The Big Idea: The River vs. The Snapshot

Think of the model's brain as a river. As the river flows (the model generates text), it carries a "residual stream"—a running sum of everything that has happened so far.

The Old Theory: If the river gets too choppy (turbulent), the model is hallucinating. If we measure how fast the water is spinning or how much it's drifting, we can catch the lie.
The Paper's Test: The researchers built a detector that looked at two things:
1. The Snapshot: A single, frozen picture of the water at a specific moment (the static representation).
2. The Flow: The history of how the water moved to get there (the dynamical trajectory).

They asked: Does knowing how the water moved help us spot a lie better than just looking at the water right now?

The Verdict: No.
When they compared the two, the "flow" added almost nothing.

A detector looking only at the snapshot got a score of 0.83 (a very strong performance).
Adding all the complex "turbulence" data (swirls, drift, expansion rates) changed the score by only +0.004.
Statistically, this is a null result. The confidence interval was [-0.011, +0.020], and the p-value was 0.30. In plain English: the extra data was just noise. The turbulence lens is a great way to think about how the model works, but it is not a way to detect hallucinations.

What They Ruled Out (The "Not" List)

The paper is very careful to knock down some popular ideas that looked like they were working but were actually tricks.

The "Weak Baseline" Trap:
Earlier, some tests showed that turbulence features did help. But the researchers realized those tests were comparing the turbulence against a "weak baseline" (a simple guess based on word position and probability). It's like saying a high-tech radar is amazing because it found a bird better than a person guessing where birds might be.
- The Reality Check: When they compared the turbulence against the strong baseline (the snapshot itself), the advantage vanished. The turbulence didn't add anything new; it just looked good because the competition was weak.
The "Telescoping" Illusion:
One way to measure turbulence is to look at how fast the water expands. The researchers found that a common way of calculating this (a "finite-time Lyapunov exponent") was broken. It was like a math trick where all the middle numbers canceled out, leaving only the start and end. It looked like it was measuring the whole river, but it was actually just measuring the beginning and the end. They fixed this, and the "turbulence signal" disappeared.
The "Free-Generation" Twist:
They tested if the model gets chaotic when it's actually generating text (not just reading finished text). They found that yes, the flow does amplify small changes (it's sensitive). But even though the river is physically swirling, that swirling does not give you a better signal to detect a lie than the snapshot does. The physics is real, but the detection signal is not.

Where the Signal Actually Lives

If the turbulence isn't the key, where is the "lie detector" hiding?

The researchers found that the answer is in the static snapshot, specifically in the middle of the model's layers.

The Location: The signal becomes clear in the middle-to-late layers of the network (around layers 14 to 31 in a 32-layer model).
The Shape: It's not a single, magical "truth neuron." It's a diffuse circuit. Imagine a rumor spreading through a crowd; no single person is the source, but the whole group knows the story. In the model, about 10% of the top components (neurons and attention heads) carry the signal, and they are spread out.
The Catch: This signal is task-specific. If the model is writing a story, the "lie detector" looks different than if it's answering a math question. The direction of the signal changes depending on what the model is doing.
- Example: In one model (Mistral-7B), the signal was task-specific (cosine similarity between tasks was only 0.03 to 0.42). In another model (Llama-2-7b), it was more shared (cosine similarity around 0.8).
- Conclusion: There is no single, universal "hallucination switch" that works for every model and every task.

The Causal Proof (The Steering Experiment)

To prove this wasn't just a coincidence, they tried to "steer" the model. They pushed the model's internal state in the direction of the "hallucination" signal.

The Result: When they pushed the model toward the "hallucination" direction, it actually made the model produce less faithful text.
The Strength: The effect was real but modest. When they used a strong external judge to check the text, the improvement in detecting lies was small. The model didn't turn into a chaotic liar instantly; it just got slightly worse at being honest.
The Mechanism: They used a mathematical "decomposition theorem" to show that the effect wasn't just a static change; it was a Jacobian-propagated effect. This means the change happened because the signal traveled through the non-linear layers of the model, flipping signs along the way (like a ripple that turns into a wave).

The Final Takeaway

The paper concludes that the "turbulence" idea is a beautiful metaphor for understanding how LLMs work, but it is a failed tool for detection.

Did they prove turbulence doesn't exist? No, the flow does get sensitive.
Did they prove turbulence helps detect lies? No, the paper explicitly rules this out with high confidence. The trajectory adds 0.004 to the detection score, which is statistically indistinguishable from zero.
What works? A simple probe on the static representation (the snapshot) works best, reaching an AUC of 0.83.

The authors warn us not to get seduced by fancy metaphors. Just because a model's internal river looks like a storm doesn't mean the storm is the alarm bell. The alarm bell is already ringing in the quiet, static snapshot.

The Bottom Line: If you want to catch a hallucination, don't watch the river swirl. Just take a good, clear photo of the water right now. The photo tells you everything you need to know.

Technical Summary: Turbulence Is Not What You Need to Detect Hallucinations

Problem Statement

Large Language Models (LLMs) are inherently nonlinear systems, leading to a hypothesis that hallucinations (ungrounded or fabricated outputs) arise from a loss of dynamical stability within the model's computation flow. Specifically, the "turbulence lens" suggests that hallucinated generations correspond to a regime of "edge of chaos" or unstable flow, where small input perturbations amplify into large output deviations, distinct from stable "laminar" regimes.

While statistical accounts explain why hallucinations are inevitable in expectation, they do not identify whether a specific generation is hallucinated at the moment of production. The central research question is: Do the dynamical signatures of the residual stream trajectory (e.g., finite-time expansion rates, drift, spectral entropy) provide detectable signals of hallucination that are not already captured by the static hidden-state representation?

Methodology

Experimental Setup

The study employs a "faithful" setup where models are evaluated on their own generated text, ensuring the residual stream reflects the actual generation process rather than an encoder pass over another model's output.

Datasets: Primary benchmark is RAGTruth (grounded hallucination spans in data-to-text, QA, and summarization tasks). A secondary dataset, PsiloQA, is used for cross-dataset checks.
Models: Experiments span multiple architectures and scales, including Mistral-7B, Llama-2-7b/13b, and a grid of five 2026-era models (Phi-4, Mistral-Nemo, Gemma-4, Qwen3, Qwen3.6).
Labels: Hallucination spans are annotated based on source support. For modern models without human labels, a large language model judge (Gemma-3-12B-it) is used, calibrated against human annotations.

Feature Extraction and Comparison

The paper compares two feature sets for binary hallucination detection:

STATIC: A probe on the static representation (pooled hidden states reduced via PCA) plus a weak unsupervised control (log-probabilities, signal levels, token position).
STATIC + DYNAMICS: The static features augmented with dynamical signatures extracted from the residual stream trajectory:
- Finite-time expansion proxy: A least-squares slope of log-norms of trajectory increments (a proxy for Lyapunov exponents).
- Drift: Cumulative drift off the "laminar manifold."
- Spectral entropy and effective rank: Measures of trajectory shape complexity.
- Early-expansion asymmetry: Difference between early and late expansion rates.

Evaluation Protocol

Metric: Area Under the ROC Curve (AUC).
Validation: Response-grouped cross-validation with a 2000-resample bootstrap to ensure no data leakage between training and test folds.
Decisive Test: The core experiment measures the marginal gain ( $\Delta$ AUC) of adding dynamical features to the static probe.
Controls:
- Weak Baseline: Unsupervised scores (log-prob, etc.) to test if dynamics help over a naive baseline.
- Strong Baseline: The static representation itself (PCA of hidden states).
- Dimensionality-Matched Control: Ensuring the static probe uses the same number of components as the dynamical features to rule out budget artifacts.
- Null Hypothesis: A "matched-influence" null (ablating top-k components of a random direction) is used for causal tests to avoid inflated significance from low-variance random component sets.

Mechanistic Analysis

Layer-wise Probing: Identifying where hallucination becomes linearly decodable across network depth.
Causal Steering: Intervening on the residual stream at the peak layer to observe downstream effects on hallucination signals and behavioral faithfulness.
Circuit Attribution: Decomposing the hallucination direction into contributions from individual attention heads and MLP neurons.
Cross-Task/Model Transfer: Testing if the hallucination direction and circuit generalize across tasks, datasets, and model families.
Decomposition Theorem: A formal decomposition of linear readouts into a direct term and a Jacobian-propagated indirect term to explain sign flips observed in ablation studies.

Key Results

1. The Decisive Negative Result

The central finding is that dynamical features add no significant information beyond the static representation.

Static Performance: A probe on the static representation alone achieves an AUC of 0.831 on faithful Mistral-7B.
Dynamical Increment: Adding all dynamical features increases the AUC by only +0.004 (95% CI: $[-0.011, +0.020]$ , $p=0.30$ ).
Robustness: This null result holds across:
- Three random seeds and various PCA budgets.
- Dimensionality-matched controls (static budget = 9 components).
- Multiple models (Llama-2, Phi-4, Qwen, Gemma) and scales (up to 35B).
- Different hallucination labels (human, self-judge, external judges, consensus).
- Per-token analysis (where dynamics actually degrade performance slightly, suggesting overfitting to noise).
- Free-generation flows (where the autoregressive loop genuinely amplifies perturbations, yet this amplification still yields no detection signal).

2. The Static Signal is Real and Causal

While dynamics add nothing, the static signal is robust:

Decodability: Hallucination is linearly decodable from the static hidden state, emerging through depth to a broad plateau (Layers 14–31 in Mistral) with peak AUC $\approx 0.71$ .
Causality: Steering the residual stream along the learned hallucination direction at the peak layer causally shifts the downstream signal (dose-response correlation $\approx 0.99$ ) and modulates behavioral faithfulness (though the behavioral effect is modest and asymmetric).
Circuit Structure: The circuit producing the signal is diffuse (top-20 components carry only $\sim 10\%$ of the signal in Mistral) and task-specific (cross-task cosine similarity is low, 0.03–0.42).
Transfer: The mechanism transfers partially across datasets (RAGTruth $\to$ PsiloQA) but is highly model-idiosyncratic (e.g., Llama-2-7b shows a concentrated, task-shared circuit, unlike Mistral).

3. Methodological Illusions Identified

The paper identifies three specific "traps" that previously made turbulence theories appear validated:

Degenerate Estimator: A naive finite-time Lyapunov exponent calculation telescopes to depend only on the first and last increments, masquerading as a complex physics quantity.
Position Confound: In some datasets, hallucinations occur in contiguous spans, making token position a strong predictor; dynamical features correlated with prefix length then appeared to help.
Weak Baseline: Dynamical features show significant gains only when compared to weak unsupervised baselines (log-probs). When compared to the strong static representation baseline, the gains vanish.

4. Cross-Detector Commensurability

Distinct mechanistic detectors (supervised difference-in-means, drift, density, ICR) do not share a single robust circuit. Apparent sharing dissolves when tested against a matched-influence null and across multiple seeds.

Significance and Claims

The paper claims three primary contributions:

A Rigorous Negative Test: It provides a well-powered, replicated refutation of the hypothesis that residual stream dynamics offer a superior or independent signal for hallucination detection compared to static representations. The "turbulence" framing is a useful interpretive lens for understanding model behavior but is not a viable detection mechanism.
Mechanistic Characterization of Grounded Hallucination: It establishes that grounded hallucination is linearly decodable, causally active, and localized to mid-to-late network layers, but that its specific implementation (direction, circuit concentration, task-sharing) is highly dependent on the model architecture and task, challenging the notion of a universal "truth direction."
Methodological Correction: It highlights the necessity of using strong baselines (static representations), matched-influence nulls, and seed-stability checks to avoid false positives in mechanistic interpretability. The paper argues that "clean negatives" are valuable results that prevent the deployment of false safety signals.

The authors conclude that the "turbulence" lens is a good way to think about hallucination as a loss of stability, but the static representation already encodes the necessary information to detect it, rendering the dynamical trajectory redundant for detection purposes.