One cheap way to make a language model more truthful is activation steering: at inference time, nudge the model's internal hidden states in a direction associated with true answers. It works, up to a point. Pushed harder, the steered model starts repeating itself, breaks grammar, or retreats into evasive non-answers such as "I have no comment." This proposal argues that the failure has a familiar shape. Generative-model researchers know it as the prior hole. A variational autoencoder (a generator that maps data to a compressed latent code and back) usually assumes a single bell-shaped Gaussian over that code. Data that actually sits in several separate clusters leaves high-probability regions of the Gaussian that no real example ever occupies, and decoding from those regions produces garbage. Our hypothesis is that steering vectors carry hidden states into the analogous empty regions of the LLM's activation space. They also tend to move a state out of the cluster it started in, for example from "detailed factual answer" into "refusal."
The concrete proposal is a drop-in constraint for any existing truth-steering method. We fit a multi-cluster density, a Gaussian mixture, to unlabeled residual-stream activations at the steered layer. After the steering method proposes an edit, we project the edited state back to the nearest point that is both high-density and inside the same mixture component as the original state. Fitting the mixture needs no truth labels. The projection costs about one small linear solve per token. Our central test is whether this mixture constraint beats the obvious unimodal version, a single-Gaussian (Mahalanobis-distance) constraint. That comparison isolates the lesson carried over from the prior-modeling literature: a single Gaussian cannot tell a hole between clusters from a real region of data.
Core Claim
Steering-induced incoherence and evasiveness are caused by two things. The first is edited states landing in low-density regions of a multi-cluster activation distribution. The second is edited states switching clusters. A constraint that preserves both density and cluster membership should let steering run at higher strength. At matched truthfulness, it should produce fewer evasive and degenerate outputs than the unconstrained method. A unimodal density constraint should not deliver this, because it assigns high density to exactly the holes we want to avoid.
The generative-modeling side supplies both the diagnosis and the tool. Improving VAE generations of multimodal data through data-dependent conditional priors shows that a unimodal prior squashes distinct modes together and yields hybrid outputs. Prior Learning in Introspective VAEs shows that sampling from prior holes gives unrealistic outputs. Mixture priors address both problems. Variational Deep Embedding uses a Gaussian-mixture prior to isolate clusters. Cooperation in the Latent Space shows that mixture components divide a distribution among themselves rather than duplicating each other, with likelihood improving monotonically as components are added. We take the method from these papers but apply it differently. We do not train a generator with a mixture prior. Instead, we fit the mixture post hoc to a frozen LLM's activations and use it only as a feasibility region. This also sidesteps the per-component encoder cost (52 GPU-hours for four components) that the Cooperation paper reports.
Relation to Closest Work
The hallucination side supplies the motivation. Attribution-Guided Decoding documents that vector injection pushes models out of distribution and degrades fluency, and it responds by avoiding hidden-state edits altogether. TruthFlow learns query-specific corrections with rectified flow matching from paired truthful and hallucinated states. Its authors still report that the interventions drift toward brief, evasive answers. Truth Forest already argues that truth lives in several orthogonal directions, but it applies static linear additions with no model of where activations actually reside.
The nearest method is Scalpel. Scalpel fits separate Gaussian mixtures to labeled "trusted" and "hallucinated" attention activations in vision-language models, then transports states between them with Schrödinger bridges (an entropy-regularized optimal-transport path). Our proposal differs in four ways. Scalpel's mixture defines the target of correction from labeled data; ours defines a label-free feasibility region for someone else's correction. Scalpel works on attention heads in vision-language models; we work on the residual stream of text-only LLMs, where ICR Probe locates the injection of ungrounded parametric content in feed-forward updates. Scalpel is a standalone steering method; ours wraps ITI, Truth Forest, and TruthFlow alike. Finally, we impose a cluster-preservation constraint that Scalpel lacks, and that constraint is aimed directly at the evasiveness problem TruthFlow reports.
Research Program
Let $h \in \mathbb{R}^d$ be the residual-stream state at layer $\ell$ for the current token. We first project activations to a principal subspace of a few hundred dimensions so that covariance estimates stay stable. We then fit a $K$-component mixture $p(h)=\sum_k \pi_k \mathcal{N}(h;\mu_k,\Sigma_k)$ on activations from generic unlabeled text plus unlabeled model generations. A steering method proposes $h' = h+\delta$. We find the component $k^\*$ with the highest responsibility for the unsteered $h$ and return
$$ \hat h=\arg\min_{u}\ \lVert u-h'\rVert^2 \quad \text{s.t.}\quad (u-\mu_{k^\*})^\top\Sigma_{k^\*}^{-1}(u-\mu_{k^\*})\le r_{k^\*}^2 , $$
where $r_{k^\*}$ is the 95th-percentile Mahalanobis radius of training activations assigned to component $k^\*$. Projecting onto an ellipsoid reduces to a one-dimensional search for a Lagrange multiplier, so the overhead is small. The orthogonal complement of the principal subspace passes through unchanged.
The unimodal ablation is the same procedure with $K=1$. A "density-only" ablation drops the same-component requirement and projects onto the union of component ellipsoids. That variant keeps edits out of holes but lets them switch clusters. A flow-based density, in the spirit of Aligning Latent Spaces with Flow Priors, serves as a more expressive alternative. Its constraint is enforced by gradient steps on log-density.
The program runs in three stages. The first stage is diagnostic and comes before any method claims. We ask whether layer-$\ell$ activations are measurably multi-cluster, judged by held-out likelihood as $K$ grows. We then ask whether unconstrained steered states fall disproportionately in regions that the single Gaussian rates as dense but the mixture rates as sparse. Finally, we ask whether component switches predict evasive or degenerate outputs. The second stage applies the constraint to three base steering methods, sweeping steering strength. The third stage studies $K$ and the choice of layer, including layers that ICR Probe flags as feed-forward-dominated.
Baselines
The naive reference point is the unsteered base model, Llama-2-7B and Llama-3-8B, using greedy decoding and the standard TruthfulQA prompt. The starting numbers reported in the literature are 40.8% truthful for base Llama-2-7B (Truth Forest) and 52.32% "True" for Llama-3-8B (TruthFlow).
Single-vector steering in the style of ITI is the simplest intervention, and the wrapper should help it most. We re-run it with our own strength sweep, because the file contains no ITI TruthfulQA number. The file does record that ITI lost truthfulness when transferred to Natural Questions in the TruthFlow comparison.
The strongest label-light linear competitor is Truth Forest. It reports raising Llama-2-7B truthfulness from 40.8% to 74.5% and True×Info from 37.5% to 56.0% (Truth Forest). It is the strongest baseline on Llama-2-7B because of its large truthfulness gain combined with a reported informativeness number. The strongest nonlinear competitor is TruthFlow, which reports Llama-3-8B True rising from 52.32% to 64.79%, with MMLU at 65.57% against 65.77% for the base model (TruthFlow).
For fairness, every method is reimplemented or re-run on identical TruthfulQA splits and the same steered layer. Each gets the same hyperparameter budget: steering strength sweeps over a shared grid, and the mixture's $K$ and radius are tuned only on the validation split. All methods share one judge model. The constraint adds no labeled data beyond what each base method already uses.
The key ablation is the $K=1$ Mahalanobis projection, which isolates multimodality as the ingredient that matters. Further controls are the density-only mixture without cluster preservation, plain $\ell_2$ norm clipping of $\delta$ tuned to the same budget, and a $k$-nearest-neighbor density constraint. As a non-intervention competitor we run Attribution-Guided Decoding. The file reports a gain from 66.0% to 79.1% on IHEval for it, and no reported TruthfulQA result was found on Lacuna for it.
Beating a baseline means one of two things. The first is a better Pareto frontier: at matched TruthfulQA truthfulness, higher True×Info and a lower evasive-answer rate. The second is a lower degenerate-output rate, with MMLU within half a point of base. Pareto curves will be reported, not single points.
Why not just…?
Why not just turn down the steering strength or clip its norm? This is the most serious threat. If degradation only grows with the size of the edit, a tuned scalar should match us. The two ideas make different predictions, however. Norm clipping limits every direction equally. Our hypothesis says the harm comes from which region the edit enters, so a short edit into a hole should hurt, and a long edit along a populated cluster should not. The matched-budget norm-clipping baseline settles the question. If its Pareto curve coincides with ours, we will report that the density story adds nothing.
Why not just use a single-Gaussian (Mahalanobis) constraint? It is simpler, and it is exactly the unimodal assumption that the VAE literature shows creates holes (Prior Learning in Introspective VAEs). Whether LLM activations are multi-cluster enough for this to matter is an empirical question. The stage-one diagnostic and the $K=1$ ablation answer it directly. Truth Forest's observation that truth features form distinct clusters is suggestive but is not proof.
Why not just avoid editing hidden states, as Attribution-Guided Decoding does? Selection-based decoding sidesteps the out-of-distribution problem. It cannot, however, move the model toward an answer that is absent from its top candidates. Imitative falsehoods in TruthfulQA are often the model's confident top choice (TruthfulQA). We include it as a baseline and expect it to trail on TruthfulQA while possibly winning on fluency.
Why not just use nearest-neighbor density? A $k$-nearest-neighbor constraint is also multimodal and needs no fitting. If it matches the mixture, the claim survives in its general form: multimodal, cluster-preserving constraints matter. Only the specific estimator would lose its case. The mixture's advantages are a constant-cost projection and explicit cluster identities, so we will report latency alongside accuracy.
Proposed Experiments
The main evaluation is TruthfulQA, using the generation metrics True, Info, and True×Info, plus the fraction of "no comment"-style answers. Transfer is tested on Natural Questions, TriviaQA, and HaluEval QA, with the steering components trained only on TruthfulQA. Utility is measured by MMLU accuracy. Degeneration is measured by n-gram repetition rate and by perplexity of the outputs under a separate reference model.
The decisive checks are three. First, in stage one, the drop in mixture log-density of steered states should predict degenerate outputs better than the drop in single-Gaussian density does. Without that correlation, the hypothesis fails before any method comparison. Second, component switches should be enriched among evasive answers. If they are not, the cluster-preservation constraint has no reason to help informativeness. Third, the mixture-constrained versions of ITI, Truth Forest, and TruthFlow should dominate both their unconstrained versions and their $K=1$ versions on the truthfulness-versus-True×Info frontier, at fixed MMLU, across both models. A secondary result worth reporting either way is the size of the gain for each base method. We expect static-vector methods to gain more than query-conditioned TruthFlow, since TruthFlow's projection onto a truth subspace learned by singular value decomposition already limits how far its edits drift.
References
- Improving VAE Generations of Multimodal Data Through Data-Dependent Conditional Priors
- Prior Learning in Introspective VAEs
- Variational Deep Embedding: An Unsupervised and Generative Approach to Clustering
- Cooperation in the Latent Space: The Benefits of Adding Mixture Components in Variational Autoencoders
- Attribution-Guided Decoding
- TruthFlow: Truthful LLM Generation via Representation Flow Correction
- Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning
- Scalpel: Fine-Grained Alignment of Attention Activation Manifolds via Mixture Gaussian Bridges to Mitigate Multimodal Hallucination
- ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs
- Aligning Latent Spaces with Flow Priors
- TruthfulQA: Measuring How Models Mimic Human Falsehoods