Dead Zones in Activation Space Predict When an LLM Is Making Things Up

Generative models of images and text have a well-documented failure. When a decoder is asked to produce output from a point that lies in a sparsely populated region of its internal representation space, it still produces something, and that something is often a fluent-looking blend of nearby examples with no real content. Interpolate between two face codes through a low-density region and you get "brown mush." We propose that a large language model inventing a fact about an entity it does not know is the same failure seen from the inside. When the model has no memory of a person, paper, or drug, its internal representation of that entity is pushed into a gap between the clusters formed by things it does know. The output layer then decodes that gap point into a confident, plausible, and fabricated answer.

This view makes a sharp prediction that current hallucination methods do not. A label-free measure of how sparse the model's activation space is around the current entity token should flag knowledge-gap fabrications: invented biographies, answers about nonexistent entities, wrong long-tail facts. The same measure should be close to useless on imitative falsehoods such as the misconceptions in TruthfulQA, because those are well-learned, high-density patterns. If the prediction holds, it supports a simple policy. Abstain when the model is in a dead zone, and steer, rather than refuse, when it is in a dense but wrongly oriented region. That policy aims at the main practical failure of current anti-hallucination methods, which is that they reduce fabrication by refusing far too often.

Core Claim

Our claim is that local angular density of mid-layer activations, measured at entity-bearing tokens against a reference bank built from unlabeled text, separates knowing from not-knowing without any truth labels. We claim this signal is dissociable from the signal that separates truth from misconception. Routing on it should then give a better trade-off between hallucination and false refusal than uniform abstention training.

Each ingredient is already constrained by the literature. The density must be measured on direction rather than magnitude. Spherical Steering, which rotates activations along a hypersphere instead of adding vectors to them, reports that truthful and hallucinated activations have nearly identical norms. The high-dimensional "Gaussian shell" analysis in Semantic Interpolation in Implicit Models, which shows that probability mass in high dimensions concentrates on a thin spherical shell, explains why radius carries little information. The density must also be local, meaning sensitive to valleys between clusters, rather than a single global Gaussian fit. Otherwise the "fractured manifold" story adds nothing. That makes Mahalanobis distance the natural ablation. Finally, the measurement must be taken at the entity token mid-sentence, not at the final token. HD-NDEs, which models hidden-state trajectories with neural differential equations, shows that hidden states diverge during a hallucination and then converge back by the sentence-ending token, which blinds end-of-sentence probes.

Relation to Prior Work

Activation-geometry work on hallucination falls into three groups. The detectors include the Truthfulness Separator Vector, a learned 4,000-parameter vector that pushes truthful and hallucinated representations apart using 32 labels, HD-NDEs, and the known-versus-unknown linear separation used by CASAL. All of them are supervised, and all of them treat "hallucination" as one class.

The steerers include Spherical Steering, TruthFlow, which learns query-specific correction vectors with flow matching at one middle layer, NDAD, which contrasts against a deliberately hallucination-prone distribution created by masking attention heads, and the concept-guidance study in A Language Model's Guide Through Latent Space. These methods move activations toward truth. None of them asks whether there is any truth nearby to move toward.

The decoding-time controller Token-Guard scores candidate tokens with a hybrid latent score and prunes them. Its authors note that it cannot help when the knowledge is entirely absent, and that its thresholds are heuristic.

Our proposal differs from all three groups. It imports the density concept from the latent-discontinuity literature, specifically the density-weighted path costs of Probability Density Geodesics and Feature-Based Interpolation and Geodesics. It then uses that concept as a diagnostic that splits hallucination into two kinds rather than as a steering path.

We deliberately do not compute geodesics. Those papers report that boundary-value solving is expensive and gets trapped when endpoints are distant. More fundamentally, a path cannot lead to a fact the model lacks. Cosmos, which regularizes a text autoencoder's latent space toward smoothness, motivates the view that a text representation can reconstruct perfectly and still be riddled with brittle regions. That is exactly the property we hypothesize for pretrained LLMs.

HalluLens supplies the evaluation frame. Its separation of factual ignorance from contradiction, and its joint reporting of hallucination and false refusal, is what our routing policy is meant to improve.

Research Program

We first build reference banks. For each model, we run a large unlabeled Wikipedia-style corpus through the network and store unit-normalized residual-stream vectors at the last subword of each named entity, for a sweep of middle layers. The density score at a query activation $h$ is

$$s(h) = \frac{1}{k}\sum_{j=1}^{k} \arccos\big(\hat h^\top \hat r_{(j)}\big),$$

where $\hat h$ is $h$ normalized to unit length, $\hat r_{(j)}$ is its $j$-th nearest neighbor in the bank, and $k$ is a small integer such as 10. A large $s$ means the activation sits in a sparse angular region.

We compare this nonparametric estimator against three alternatives that make different geometric assumptions. The first is Mahalanobis distance, a single Gaussian with no valleys. The second is a von Mises–Fisher mixture, which models a clustered sphere. The third is a small normalizing flow on PCA-reduced directions. The simplest estimator that works becomes the method. Approximate nearest-neighbor search keeps the per-token cost to one lookup.

Second, we run a controlled "midpoint" test that directly mirrors the brown-mush experiments. We take two well-known entities of the same type, such as two cities or two chemists. At the subject token, we replace the residual stream with a spherical interpolation of their two activations at a chosen layer, then ask attribute questions. The discontinuity hypothesis predicts two things. The interpolated points should score as low-density. The model should answer with confident blended or arbitrary attributes rather than refusing. A positive result would show that fabrication can be induced simply by placing the model in a gap. We then check whether real unknown entities from HalluLens NonExistentRefusal land in similar regions.

Third, we build the router. At entity tokens during decoding, a high density score triggers abstention or hedging. Otherwise, generation proceeds and may optionally apply Spherical Steering for imitative errors. The single threshold is set on a held-out split to hit a target false-refusal rate.

Baselines

The naive reference point is the unmodified model under greedy decoding, together with a prompt instructing it to say "I don't know" when unsure. Every method must beat this pair on the joint hallucination and false-refusal metrics, not on hallucination rate alone.

The strongest label-efficient detector is the Truthfulness Separator Vector. On LLaMA-3.1-8B it reports 84.2% AUROC on TruthfulQA and 79.8% AUROC when transferred to TriviaQA (Steer LLM Latents for Hallucination Detection). We run it with its prescribed 32 labels, plus a fully supervised linear probe on the same layer as an upper reference. We give it the same layer-search budget as our score. Our score uses zero labels, so beating the Truthfulness Separator Vector means matching or exceeding its AUROC on PreciseWikiQA and NonExistentRefusal, and exceeding it in cross-dataset transfer. We do not expect to beat it on TruthfulQA, and we would count failing to beat it there as confirmation of the dissociation rather than as a loss.

The strongest label-free detector is Semantic Energy, which scores sampled answers by their raw logits instead of post-softmax probabilities. It reports an AUROC improvement from 71.6% to 76.1% over Semantic Entropy on CommonsenseQA with Qwen3-8B. We give it the same number of samples per question that our score would need in compute-matched form. We also report a combination of the two signals to measure whether density adds information.

For the generation-time comparison, we run Token-Guard and CASAL on HalluLens. Token-Guard reports an average F1 of 51.0 versus 28.3 for the base Llama-3.1-8B-Instruct. CASAL reports roughly 30–40% hallucination reduction. We found no reported result on Lacuna for either Token-Guard or CASAL on HalluLens PreciseWikiQA, so we will produce these numbers ourselves under identical prompts.

The external reference point for the refusal trade-off is HalluLens itself. On PreciseWikiQA, Llama-3.1-405B reaches a 26.84% hallucination rate only with a 56.77% false refusal rate, while GPT-4o answers 52.59% correctly with 4.13% false refusals. Beating the baselines means achieving lower hallucination at an equal false refusal rate, which we report as a full curve.

The key ablations are Mahalanobis versus local density, which isolates the "gaps between clusters" claim; activation norm alone, which Spherical Steering predicts will fail; and entity-token versus final-token measurement, which tests the HD-NDEs convergence finding.

Why not just…?

Why not just fine-tune the model to abstain on unknown questions? This is what CASAL does cheaply, and what supervised fine-tuning and preference optimization do expensively. The HalluLens results show that abstention behavior bought this way tends to spread to valid questions, and TruthFlow reports a drift toward "I have no comment." Our experiment compares false-refusal rates at matched hallucination rates on PreciseWikiQA. If CASAL's curve dominates ours, the density router is not worth having, and we will report that outcome.

Why not just train a supervised probe, since known and unknown queries are already linearly separable? This is the most serious threat. CASAL establishes the separability, and a 32-label probe may match our score on in-distribution data. The case for density rests on two things. It needs no labels, and it should transfer, whereas the Truthfulness Separator Vector already loses about 4 AUROC points going from TruthfulQA to TriviaQA. The transfer matrix across TriviaQA, PreciseWikiQA, NonExistentRefusal, and the True-False dataset settles the question. If the probe transfers as well as density, our contribution narrows to the mechanistic midpoint finding.

Why not just use output uncertainty such as Semantic Energy? It is also label-free. Its weakness is "confidently wrong" answers, which are exactly the knowledge-gap fabrications we target: Mistral-7B confidently answers over 83% of HalluLens nonexistent-entity queries. Logits are downstream of the gap, so they may still carry the signal. The combined-signal experiment measures whether density adds anything beyond logits.

Why not just use activation distance from the mean or the norm? Spherical Steering's finding that norms are nearly identical between truthful and hallucinated activations predicts that norm fails. Mahalanobis distance is the honest test of whether local, multi-cluster structure matters. If Mahalanobis matches kNN, we will drop the discontinuity framing and keep the simpler estimator.

Decisive Checks

We will run Llama-3.1-8B-Instruct, Mistral-7B-Instruct, and Qwen-2.5-7B-Instruct on HalluLens PreciseWikiQA, NonExistentRefusal, and LongWiki, on TruthfulQA, on TriviaQA, and on the HD-NDEs True-False dataset.

Three results would decide the proposal. The first is a crossed dissociation: the density score achieves high AUROC on knowledge-gap benchmarks and near-chance AUROC on TruthfulQA misconceptions, while supervised truth probes do comparatively better on TruthfulQA. The second is a positive midpoint test: spherical interpolations between known entities are scored as low-density and produce confident fabrications. The third is that the router reaches a lower PreciseWikiQA hallucination rate than CASAL and Token-Guard at the same false-refusal rate, without degrading MMLU accuracy. If the first check fails, the dead-zone account of hallucination is wrong for these models, and that negative result would itself be worth reporting.