Hallucination Tri-Space Framework

• Hallucination Tri-Space is a framework that models hallucination as misalignment among three axes—semantic coherence, structural alignment, and knowledge grounding.
• It enables researchers to diagnose generative failures by comparing source truth, latent representations, and output responses across diverse modalities.
• Empirical studies demonstrate that tri-space analysis improves detection metrics and control interventions, paving the way for more precise model evaluations.

Hallucination Tri-Space denotes a family of three-space or three-axis formalisms that model hallucination as structured misalignment rather than as a binary error event. In the explicit text-to-image formulation, the three axes are semantic coherence, structural alignment, and knowledge grounding; in other lines of work, closely related triads appear as 3D geometric space, data or semantic space, and language or output space; as subject, relation, and object triplets; or as world truth, model knowledge, and generated behavior (Yang et al., 7 Jul 2025, Peng et al., 18 Feb 2025, Wu et al., 2024, Wang, 2024). This suggests that the term is best understood as an umbrella concept for three-component descriptions of generative failure in which the position, direction, or interaction of an error within a structured space is used to characterize hallucination type, severity, and controllability.

1. Conceptual foundations

The literature does not present a single canonical definition of Hallucination Tri-Space. One strand introduces the term explicitly for diffusion-based text-to-image generation, while several others reconstruct an equivalent three-part structure for 3D visual LLMs, large vision-LLMs, hallucination detection in embedding space, and survey-level taxonomies of AGI hallucination (Yang et al., 7 Jul 2025, Peng et al., 18 Feb 2025, Wu et al., 2024, Zavhorodnii et al., 6 Oct 2025, Wang, 2024). A common pattern is that hallucination is localized not merely as “wrong output,” but as a failure of alignment among three separable domains: the source of truth, an internal or latent representational regime, and the emitted response or decision.

These formalisms differ in what counts as a “space.” In some papers the spaces are continuous axes with explicit coordinates; in others they are discrete semantic slots or labeled regions in a learned manifold. The term therefore functions less as a fixed ontology than as a recurrent design principle: hallucination analysis becomes more informative when decomposed into three interacting components instead of a single scalar score.

Formulation	Three spaces or axes	Representative source
Tension-space	Semantic coherence, structural alignment, knowledge grounding	(Yang et al., 7 Jul 2025)
Scene-space	3D geometric space, data or semantic space, language or output space	(Peng et al., 18 Feb 2025)
Triplet-space	Subject object, relation, object object	(Wu et al., 2024)
Embedding-space	Ground-truth region, correct-LLM region, hallucination region	(Zavhorodnii et al., 6 Oct 2025)
World-model-space	Reference world, controlled view, conflict policy or output truth conditions	(Liu et al., 19 May 2026)

A recurrent implication is that tri-space formalisms separate where the evidence resides, how it is internally represented, and how it is verbalized or rendered. That separation enables finer distinctions among fabrication, relational errors, contextual override, source-conflict errors, and high-confidence false positives.

2. Canonical formalizations

The most explicit formalization appears in diffusion-based text-to-image generation, where the Hallucination Tri-Space $\mathcal{T}^3$ is defined by three principal alignment tensions: semantic coherence, structural alignment, and knowledge grounding (Yang et al., 7 Jul 2025). For prompt $p$ and diffusion step $t$, the Alignment Risk Code (ARC) is the vector

$$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$

with per-axis tension

$$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$

The magnitude $\|\vec{\tau}\|_2$ measures overall risk, $\mathrm{Var}(\vec{\tau})$ measures imbalance, and the component ratios identify the dominant failure axis. In this formulation, hallucination is modeled as trajectory drift away from an ideal alignment manifold during denoising.

A second canonical form is geometric rather than dynamical. In embedding-space hallucination classification, answer texts are embedded with all-MiniLM-L6-v2, reduced with UMAP, and treated as three clusters by construction: ground truth, correct LLM outputs, and hallucinated outputs (Zavhorodnii et al., 6 Oct 2025). If $U_{GT}, U_C, U_H$ are the projected sets, their centroids $\mu_{GT}, \mu_C, \mu_H$ define a tri-space whose geometry is analyzed via Euclidean distances

$d(\mu_i,\mu_j)=\|\mu_i-\mu_j\|_2.$

For 500 samples in 3D UMAP, the reported averages remain stable across seeds: $p$0 is about $p$1–$p$2, $p$3 about $p$4–$p$5, and $p$6 about $p$7–$p$8 (Zavhorodnii et al., 6 Oct 2025). In that setting, distance from $p$9 is interpreted as a proxy for severity of informational distortion.

A third canonical form is reference-world based. HalluWorld defines a world

$t$0

with a controlled view function $t$1 and a truth function $t$2 under a source-conflict policy $t$3 (Liu et al., 19 May 2026). Hallucination is then falsehood with respect to the reference world, not with respect to free-form human judgment. The tri-space interpretation here is operational: evidence or observability alignment, temporal or planning depth, and source or reference alignment define three major axes along which hallucination profiles differ across gridworlds, chess, and terminal tasks.

3. Multimodal and task-specific instantiations

In 3D visual LLMs, hallucination is defined as generated text that does not align with the 3D scene or point cloud, including objects that do not exist in the scene and incorrect relationships between objects (Peng et al., 18 Feb 2025). The paper distinguishes “scene conflict” from knowledge conflict and text-image conflict, and its analysis naturally separates three interacting spaces: 3D geometric space, data or semantic space, and language or output space. Object hallucinations are formalized by attribute mismatch,

$t$4

while relation hallucinations are predicate mismatches,

$t$5

The underlying causes identified in ScanNet-derived data are uneven frequency distribution of objects, strong correlations between objects, and limited diversity in object attributes (Peng et al., 18 Feb 2025).

For large vision-LLMs, Tri-HE operationalizes a discrete triplet-level tri-space using semantic triplets $t$6 extracted from model outputs (Wu et al., 2024). Hallucination is judged against an image scene graph $t$7: object hallucination occurs when $t$8 or $t$9; relation hallucination occurs when $$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$0 but $$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$1; and prediction error occurs when the components exist but the specific triplet is not true in $$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$2. The resulting question-level hallucination rate is

$$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$3

This triplet-space makes relation hallucination observable alongside object hallucination rather than collapsing both into object presence.

At survey level, AGI hallucination is organized around three conflict types: conflict in intrinsic knowledge of models, factual conflict in information forgetting and updating, and conflict in multimodal fusion (Wang, 2024). A plausible implication is that this taxonomy reconstructs a broad tri-space of world or truth space, model knowledge or latent representation space, and output or behavior space. The survey’s multimodal cases—language, vision-language, video-language, audio-language, 3D, and agents—repeatedly describe hallucination as misalignment among perception, latent fusion, and emitted language or action.

A methodologically distinct but structurally similar tri-space appears in fine-grained domain generalization through Hyperbolic State Space Hallucination (Bi et al., 10 Apr 2025). There, images are mapped into a Euclidean VMamba state space, style-diverse hallucinated states are generated in feature space, and both original and hallucinated states are projected into a hyperbolic manifold for consistency regularization. The paper describes this as image space $$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$4 Euclidean state space $$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$5 hyperbolic manifold, with hallucination occurring in the intermediate state space and invariance enforced in the third.

4. Geometric, representational, and information-theoretic reinterpretations

Several papers generalize the tri-space idea from modality structure to latent geometry. One asks what geometric hallucination detection metrics actually measure by varying five properties—incorrectness, confidence, irrelevance, incoherence, and incompleteness—and analyzing Hidden Score, Matrix Entropy, and Attention Score across domains (Yeats et al., 9 Feb 2026). Its synthesis proposes a multidimensional hallucination space in which correctness, relevance, and coherence behave as distinct axes, and shows that many raw geometric statistics are domain-sensitive. A simple perturbation normalization yields AUROC gains of $$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$6 points in multi-domain settings, which suggests that the geometry of hallucination must be disentangled from domain geometry (Yeats et al., 9 Feb 2026).

Dynamic Contextual Orthogonalization treats hallucinations as orthogonal noise relative to a context-defined semantic manifold in the residual stream (Zhao et al., 2 Jun 2026). For residual anchor

$$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$7

each attention-head output $$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$8 is decomposed into

$$\vec{\tau}(p, t) = \left[ \tau_{SC}(p, t),\; \tau_{SA}(p, t),\; \tau_{KG}(p, t) \right]^\mathsf{T} \in \mathbb{R}^3,$$9

with orthogonality metric

$$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$0

The paper itself is a two-part decomposition, but it explicitly supports a tri-space reinterpretation consisting of context-aligned semantic subspace, parametric or prior semantic space, and hallucinatory outlier space (Zhao et al., 2 Jun 2026).

HARP similarly starts from a two-subspace decomposition,

$$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$1

derived from the singular vectors of the unembedding matrix $$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$2 (Hu et al., 15 Sep 2025). The reasoning subspace is used for hallucination detection, reducing dimensionality to approximately $$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$3 of the original while improving AUROC; on TriviaQA the reported AUROC is $$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$4, outperforming the previous best method by $$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$5 (Hu et al., 15 Sep 2025). The paper also suggests a more general multi-space extension in which retrieval or factuality directions would be separated from reasoning or decision directions, bringing it close to an explicit tri-space.

A more abstract formulation comes from rate-distortion theory for membership testing (Guo et al., 31 Jan 2026). The three spaces are fact or data space $$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$6 with sparse true set $$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$7, representation or memory space $$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$8, and decision or score space $$\tau_i(p, t) = \left\| \nabla_{z_t} \mathcal{A}_i(z_t, p) \right\|.$$9. Memory is measured by

$\|\vec{\tau}\|_2$0

and in the sparse regime $\|\vec{\tau}\|_2$1, the optimal rate-distortion function converges to

$\|\vec{\tau}\|_2$2

where $\|\vec{\tau}\|_2$3 and $\|\vec{\tau}\|_2$4 are the score distributions on facts and non-facts (Guo et al., 31 Jan 2026). In the log-loss case, the optimal non-fact distribution places some mass at the same high-confidence point as the fact distribution, yielding a formal explanation for high-confidence hallucination as a consequence of lossy compression rather than of defective training.

5. Dynamic behavior, control, and empirical signatures

The explicit Hallucination Tri-Space in text-to-image generation also serves as a control space. ARC-based analysis shows that a full 3D representation outperforms lower-dimensional reductions: classification accuracy rises from $\|\vec{\tau}\|_2$5 in 1D and $\|\vec{\tau}\|_2$6 in 2D to $\|\vec{\tau}\|_2$7 in 3D, while the silhouette score rises from $\|\vec{\tau}\|_2$8 and $\|\vec{\tau}\|_2$9 to $\mathrm{Var}(\vec{\tau})$0 (Yang et al., 7 Jul 2025). The associated TM-ARC controller applies axis-specific interventions in latent space, and on DrawBench with SDXL the reported metrics improve from CLIP $\mathrm{Var}(\vec{\tau})$1, PickScore $\mathrm{Var}(\vec{\tau})$2, ImageReward $\mathrm{Var}(\vec{\tau})$3, FID $\mathrm{Var}(\vec{\tau})$4 for vanilla sampling to CLIP $\mathrm{Var}(\vec{\tau})$5, PickScore $\mathrm{Var}(\vec{\tau})$6, ImageReward $\mathrm{Var}(\vec{\tau})$7, FID $\mathrm{Var}(\vec{\tau})$8 for ARC (Yang et al., 7 Jul 2025).

A dynamical tri-space also emerges in context titration studies of LLM hallucination. Using a tri-perspective detector built from semantic deviation, factual extension, and logical inference, one study measures external hallucination together with internal drift in hidden states and attention maps (Wei et al., 22 May 2025). Across six open-source LLMs, hallucination rates and representation drift increase monotonically and plateau after $\mathrm{Var}(\vec{\tau})$9–$U_{GT}, U_C, U_H$0 rounds; convergence of JS-Drift to about $U_{GT}, U_C, U_H$1 and Spearman-Drift to about $U_{GT}, U_C, U_H$2 is interpreted as an “attention-locking” threshold beyond which hallucinations become resistant to correction (Wei et al., 22 May 2025). The same study distinguishes relevant-context assimilation errors from irrelevant-context topic-drift errors, suggesting a third axis spanning assimilation versus diffusion modes.

Empirical task-specific findings repeatedly support tri-space decompositions. In 3D-LLMs, Random Point Cloud Pair Evaluation and Opposite Question Evaluation are designed to test whether outputs follow point-cloud evidence or language priors; on ScanQA-SR-Opposite, the reported $U_{GT}, U_C, U_H$3 values are $U_{GT}, U_C, U_H$4 for LL3DA and $U_{GT}, U_C, U_H$5 for 3D-LLM, showing that opposite spatial questions often receive almost identical answers (Peng et al., 18 Feb 2025). In Tri-HE, relation hallucination is frequently more severe than object hallucination; for LLaVA-1.5, the image-level overall hallucination rate is $U_{GT}, U_C, U_H$6, with object hallucination $U_{GT}, U_C, U_H$7 and relation hallucination $U_{GT}, U_C, U_H$8 (Wu et al., 2024). In HalluWorld, perceptual hallucination on directly observed information is reported as near-solved for frontier models, while multi-step state tracking, causal forward simulation, and abstention remain difficult; for GPT-5.5 in the terminal setting, the uncertainty category still shows $U_{GT}, U_C, U_H$9 hallucination (Liu et al., 19 May 2026).

These results indicate that tri-space frameworks are not only taxonomic. They expose separable failure profiles that standard aggregate accuracy metrics can hide: a model may be strong in direct perception but weak in causal rollout, strong in object grounding but weak in relational grounding, or strong in prompt semantics but weak in source-conflict arbitration.

6. Limitations, misconceptions, and directions

A common misconception is that Hallucination Tri-Space refers to one fixed mathematical object. The literature instead presents several non-equivalent but structurally analogous formalisms. Some are explicit and axis-based, as in ARC; some are cluster-based, as in truth/correct/hallucination manifolds; some are discrete symbolic spaces, as in subject-relation-object triplets; and some are reconstructed from broader taxonomies or geometric theories (Yang et al., 7 Jul 2025, Zavhorodnii et al., 6 Oct 2025, Wu et al., 2024, Wang, 2024). This suggests that Hallucination Tri-Space is a family-resemblance concept rather than a single standardized framework.

The same papers also delimit current weaknesses. Embedding-space classification experiments focus mainly on fabrication even though the taxonomy is broader (Zavhorodnii et al., 6 Oct 2025). Geometric detection metrics depend strongly on embedding choice, UMAP parameters, and domain normalization (Yeats et al., 9 Feb 2026). The 3D-LLM analysis notes that detailed hallucination typing is focused on short QA, that spatial-relation failure is not fully explained at the architectural level, and that GPT-4-generated QA augmentations may contain minor annotation noise (Peng et al., 18 Feb 2025). Triplet-level LVLM evaluation depends on GPT-4 judging and finite scene-graph coverage (Wu et al., 2024). ARC-style control depends on external evaluators, approximate orthogonality of axes, and per-backbone calibration of thresholds and controller gains (Yang et al., 7 Jul 2025). HalluWorld, while controlled and reproducible, operates in synthetic and semi-synthetic reference worlds rather than open real-world knowledge environments (Liu et al., 19 May 2026).

Future directions in the literature converge on finer-grained typing and better causal attribution. Proposed extensions include adding more hallucination subtypes as separate centroids in embedding space, learning dedicated classifiers on distance-based features, extending tri-axial control to more axes, relating geometric directions to mechanistic failure modes, and using controlled reference worlds to isolate source-conflict, temporal-depth, and observability effects (Zavhorodnii et al., 6 Oct 2025, Yang et al., 7 Jul 2025, Liu et al., 19 May 2026). A plausible implication is that the long-term value of Hallucination Tri-Space lies less in any particular triad than in the methodological claim shared across these works: hallucination becomes more measurable, more diagnosable, and more controllable when decomposed into three interacting dimensions instead of being treated as a unitary error class.