Home
Skip to main content
xStryk™

Decision Intelligence for AI in production — guardrails, traceability & evaluation.

←xTheus
LLM INTERPRETABILITY

H-Neurons: The Sparse Circuitry Behind LLM Hallucinations

12 minMarch 29, 2026

A team of researchers from the Institute for Artificial Intelligence at Tsinghua published in December 2025 a finding that reframes how we understand hallucinations in large language models. The paper demonstrates that an exceptionally sparse subset of neurons reliably predicts when an LLM will hallucinate. Less than 0.1% of total neurons. In Mistral-7B, the range is 0.01‰ to 0.35‰. The implication is immediate: hallucinations are not uniformly distributed statistical noise — they are a localizable and intervenable phenomenon.

Main Finding — Tsinghua IAI, December 2025

"A remarkably sparse subset of neurons — less than 0.1% of total — can reliably predict hallucination occurrences with strong cross-scenario generalization." H-Neurons are not a Mistral or Llama artifact: they appear in all evaluated transformer families, from 4B to 70B parameters.

<0.1%
Neurons predicting hallucinations
6
Model families evaluated
78.4%
Mistral-7B accuracy / TriviaQA
+16.7pp
Over random baseline (Mistral-7B)

What Are H-Neurons and How Are They Identified

The paper defines H-Neurons as neurons in transformer feedforward networks (FFN) whose activation systematically predicts the occurrence of hallucinations. Identification combines three stages: construction of a deterministic dataset, a normalized contribution metric (CETT), and sparse classification via L1 logistic regression.

H-Neuron Identification Pipeline
Input
TriviaQA
→
10 samples/question
→
1K correct + 1K incorrect
CETT
FFN activations
→
Relative contribution
→
Per-neuron score
Classification
L1 Regression
→
Positive weights
→
H-Neurons (&lt;0.1%)
L1-regularized logistic regression over FFN neurons — positive weights identify H-Neurons. L1 regularization forces maximum sparsity: identifies the minimal subset with highest predictive power (0.01‰–0.35‰ of total).STAGE 3 · CLASSIFICATION
CETT (Contribution of Each neuron to the Total output) — normalizes the magnitude of each neuron's projected output against the layer's total output vector. Embedding-dimension agnostic: enables cross-layer and cross-model comparison.STAGE 2 · CETT METRIC
TriviaQA — 1,000 correct examples + 1,000 incorrect examples with deterministic behavior (consistent across 10 samples per question). The deterministic filter ensures the signal is not stochastic noise.STAGE 1 · DATASET
CETT Metric — relative contribution of neuron i in layer l
CETT(i, l)  =  ∥Wout[i]⋅hi∥∑j∥Wout[j]⋅hj∥\mathrm{CETT}(i,\,l) \;=\; \frac{\bigl\|\mathbf{W}_{\mathrm{out}}[i]\cdot h_i\bigr\|}{\displaystyle\sum_j \bigl\|\mathbf{W}_{\mathrm{out}}[j]\cdot h_j\bigr\|}
Wout[i]\mathbf{W}_{\mathrm{out}}[i] = column ii of the output projection matrix · hih_i = activation of neuron ii. CETT captures relative influence on output direction — not absolute magnitude.
Mistral-7B-v0.3Mistral-Small-3.1-24BGemma-3-4BGemma-3-27BLlama-3.1-8BLlama-3.3-70BTriviaQAFalseQAFaithEvalSycophancyJailbreak

Results: Universal Generalization Across Families and Scales

ModelParametersH-Neurons (% total)TriviaQA Accuracyvs. random baseline
Mistral-7B-v0.37B0.01‰ – 0.35‰78.4%+16.7pp
Mistral-Small-3.124B<0.1%High~+10pp
Gemma-3-4B4B<0.1%Consistent~+10pp
Gemma-3-27B27B<0.1%Consistent~+10pp
Llama-3.1-8B8B<0.1%Consistent~+10pp
Llama-3.3-70B70B<0.1%Consistent~+10pp

The consistency across Mistral, Gemma, and Llama — and across scales from 4B to 70B — is the paper's most robust result. H-Neurons are not an artifact of a specific model family: they are a universal emergent property of feedforward transformers. The paper also demonstrates cross-scenario generalization: H-Neurons identified on TriviaQA predict hallucinations in completely different domains — confirming they capture a general over-compliance mechanism, not a factual domain signal.

Four Dimensions of Over-Compliance Induced by α-Scaling

The central experiment of the paper is direct intervention: scaling H-Neuron activations by a factor α ∈ [0, 3]. The result is unambiguous: amplifying H-Neurons (α > 1) systematically increases problematic behavior rates across four independent dimensions.

α-Scaling Experiment on H-Neurons
α=0\alpha = 0
suppression
→
α=1\alpha = 1
baseline
→
α=2\alpha = 2
amplification
→
α=3\alpha = 3
maximum
Slope small
≈ 3.03Mistral-7B, Gemma-3-4B, Llama-3.1-8BHigher sensitivity to amplification
Slope large
≈ 2.40Mistral-Small-24B, Gemma-3-27B, Llama-3.3-70BSome robustness — not immunity
α-Scaling Intervention on H-Neurons
h~i=α⋅hi,α∈[0, 3]\tilde{h}_i = \alpha \cdot h_i, \quad \alpha \in [0,\,3]
α<1\alpha < 1 → suppression (reduces over-compliance) · α=1\alpha = 1 → baseline · α>1\alpha > 1 → amplification (induces hallucination). No weight modification θ\theta.
FalseQA

Invalid premises

When H-Neurons are amplified, the model increasingly accepts factually incorrect claims present in the prompt. H-Neuron activation predicts when the model will override its own knowledge to comply with the question's premise.

FaithEval

Misleading context

When context contradicts the model's knowledge, amplification increases the rate of misleading context adoption. High H-Neurons = higher probability the model will "believe" the context over its training.

Sycophancy

Sycophantic tendency

With α > 1, the model tends to validate user-expressed preferences even when incorrect. The correlation with H-Neurons suggests sycophancy and factual hallucination share an underlying mechanism.

Jailbreak

Harmful instructions

Amplification increases compliance rates against jailbreak attempts. H-Neurons appear to be the general "over-compliance" mechanism — of which factual hallucinations are one specific manifestation.

Pre-Training Origin: RLHF Does Not Eliminate the Mechanism

AUROC Transferability — H-Neurons from tuned model → base model
AUROC ⁣(Htuned,  Dbase)  ≫  AUROC ⁣(Hrand,  Dbase)\mathrm{AUROC}\!\left(\mathcal{H}_{\mathrm{tuned}},\; \mathcal{D}_{\mathrm{base}}\right) \;\gg\; \mathrm{AUROC}\!\left(\mathcal{H}_{\mathrm{rand}},\; \mathcal{D}_{\mathrm{base}}\right)
Htuned\mathcal{H}_{\mathrm{tuned}} = H-Neurons identified in instruction-tuned model · Dbase\mathcal{D}_{\mathrm{base}} = base model evaluation dataset · Hrand\mathcal{H}_{\mathrm{rand}} = random neurons (baseline). The inequality holds across all 6 evaluated models.
Pre-training
H-Neurons emerge here
AUROC exceeds baseline in base models
High normalized-rank → minimal post modification
The mechanism is fixed in base weights
Fine-tuning (RLHF / SFT)
Does not eliminate H-Neurons
Does not substantially modify their influence
Mitigates behavioral expression
Does not touch the underlying mechanism

The AUROC transferability analysis is the most important piece of evidence: the authors take H-Neurons identified in instruction-tuned models and verify their predictive power in the corresponding base models (before RLHF). AUROC scores consistently exceed random baselines — proving H-Neurons are not created by fine-tuning: they were already there. Parameter analysis confirms: H-Neurons concentrate in the "high-normalized-rank region," indicating their values change minimally during RLHF and SFT. RLHF and Constitutional AI can suppress the expression of hallucinations — but leave the mechanism intact.

Three Production Intervention Vectors

Vector 1 · No weight modification

Real-time detection

Monitor H-Neuron activations during inference. When they exceed the threshold, emit a low confidence score or block the response. Implementable today with access to model intermediate states — no retraining.

Vector 2 · At inference

α-Scaling suppression

Apply α < 1 to identified H-Neuron activations during the forward pass. Reduces over-compliance rate without retraining. Preserves general model capability — only attenuates the hallucination circuit.

Vector 3 · Directed fine-tuning

Localized regularization

Fine-tuning with specific regularization over H-Neurons: penalize high activations in over-compliance contexts. More efficient than full RLHF — works on the mechanism, not just the behavioral expression.

Orthogonal combination

LLM governance stack

All three vectors are orthogonal: they can be combined. Real-time detection for alerts, α-scaling for immediate suppression, directed fine-tuning for permanent reduction. Defense-in-depth architecture against hallucinations.

LLM Governance Stack with H-Neurons
Offline identification
Deterministic dataset (TriviaQA-style)CETT computation per layer and neuronL1 regression → H-Neuron indices
Inference monitoring
Forward pass hook (H-Neuron activations)Anti-hallucination confidence scoreAdditional latency &lt;5ms
Active suppression
α-scaling in forward pass (α ∈ [0.3, 0.8])Domain-configurable thresholdNo retraining — runtime only
Permanent improvement
Fine-tuning with H-Neuron regularizationLocalized adversarial trainingLower cost than full RLHF
Traceability
H-Neuron activation log per responsePost-incident audit with neural evidence

Key Takeaways

  • Less than 0.1% of an LLM's FFN neurons (0.01‰–0.35‰ in Mistral-7B) predict when the model will hallucinate, with robust cross-domain and cross-family generalization (Mistral / Gemma / Llama, 4B–70B). H-Neurons are a universal transformer property, not an architectural artifact.
  • The CETT metric normalizes each neuron's relative influence on its layer's output direction — not absolute magnitude. This enables embedding-dimension-agnostic H-Neuron identification and cross-scale model comparison.
  • Amplifying H-Neurons (α > 1) systematically increases over-compliance across four dimensions: invalid premises (FalseQA), misleading context (FaithEval), sycophancy, and jailbreak. Smaller models are more sensitive (slope ≈ 3.03 vs ≈ 2.40 for larger models).
  • H-Neurons emerge in pre-training — AUROC scores in base models exceed random baselines. RLHF and Constitutional AI mitigate the behavioral expression of hallucination behavior but do not modify the underlying mechanism encoded in base weights.
  • H-Neurons enable three orthogonal production intervention vectors: (1) real-time detection via activation monitoring (no weight modification, latency &lt;5ms), (2) suppression via α-scaling at inference (no retraining), (3) directed fine-tuning with localized regularization (lower cost than full RLHF). All three can be combined into a defense-in-depth stack.