Alignment Risk Code (ARC)

• Alignment Risk Code (ARC) is a dynamic 3D latent vector representation that measures hallucination risk in text-to-image diffusion models.
• It quantifies misalignment across semantic coherence, structural alignment, and knowledge grounding to diagnose hallucinations as trajectory drift.
• Paired with the TM-ARC controller, ARC enables real-time interventions that improve prompt controllability and benchmark performance.

Searching arXiv for papers related to “Alignment Risk Code (ARC)” and closely related terminology. arXiv_search(query="\"Alignment Risk Code\" OR ARC hallucination diffusion OR risk alignment agentic AI OR risk-aware stepwise alignment", max_results=10, sort_by="relevance") Alignment Risk Code (ARC) is a representation introduced for hallucination analysis and control in text-to-image diffusion models. In that formulation, ARC is a dynamic vector representation that quantifies real-time alignment tension during generation across three axes—semantic coherence, structural alignment, and knowledge grounding—and treats hallucination as trajectory drift in a latent alignment space rather than as an isolated sampling artifact. The same literature also makes clear that the acronym “ARC” is heavily overloaded across machine learning and adjacent fields; accordingly, the term “Alignment Risk Code” should be reserved for the tri-axial hallucination framework that defines ARC explicitly in those words (Yang et al., 7 Jul 2025).

1. Conceptual basis and definition

The defining claim behind ARC is that hallucinations in text-to-image generation are structured failures of alignment dynamics. Generation is modeled as a dynamic traversal in which the denoising path must remain balanced across three competing pressures: semantic coherence (SC), structural alignment (SA), and knowledge grounding (KG). Hallucination occurs when that path drifts away from the prompt-aligned manifold, so errors such as entity substitution, positional inconsistency, or commonsense implausibility are interpreted as axis-specific manifestations of alignment tension rather than as unstructured noise (Yang et al., 7 Jul 2025).

This view is formalized through the Hallucination Tri-Space, which decomposes the generation manifold into three orthogonal subspaces,

$$\mathcal{M} = \mathcal{M}_{SC} \oplus \mathcal{M}_{SA} \oplus \mathcal{M}_{KG},$$

with projection operators

$$P_i: \mathcal{M} \rightarrow \mathcal{M}_i,\quad i \in \{SC, SA, KG\}.$$

For a prompt $p$ and denoising step $t$, the instantaneous tension on axis $i$ is defined as

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

Collecting these components yields the core ARC signal,

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

The framework is explicitly cognitive in its presentation, but its operational content is geometric and control-theoretic. ARC does not merely label failures after sampling; it defines a state variable intended to diagnose and modulate alignment during the denoising trajectory itself.

2. Mathematical structure of the code

ARC is characterized by three interpretable properties. First, its magnitude measures overall misalignment:

$\|\vec{\tau}(p,t)\|_2.$

The framework also uses a prompt-dependent threshold criterion,

$|\vec{\tau}|_2 > \theta(p),$

to indicate elevated hallucination risk (Yang et al., 7 Jul 2025).

Second, the direction of the vector identifies the dominant failure axis. The component-wise distribution is operationalized through

$\mathrm{Softmax}(\vec{\tau}),$

which provides a probability-like attribution over SC, SA, and KG. In this sense, ARC is not only a scalar risk score; it is a directional code indicating whether the current instability is primarily semantic, structural, or grounding-related.

Third, imbalance is defined as anisotropy in the tension vector:

$$P_i: \mathcal{M} \rightarrow \mathcal{M}_i,\quad i \in \{SC, SA, KG\}.$$0

The paper states a compound risk condition,

$$P_i: \mathcal{M} \rightarrow \mathcal{M}_i,\quad i \in \{SC, SA, KG\}.$$1

so either high total tension or strong asymmetry can destabilize the trajectory (Yang et al., 7 Jul 2025).

The emergence of hallucination is then written as trajectory drift,

$$P_i: \mathcal{M} \rightarrow \mathcal{M}_i,\quad i \in \{SC, SA, KG\}.$$2

with

$$P_i: \mathcal{M} \rightarrow \mathcal{M}_i,\quad i \in \{SC, SA, KG\}.$$3

This formulation makes total tension and tension imbalance jointly responsible for drift intensity. A plausible implication is that ARC is intended to encode not only risk level but also failure geometry: magnitude captures stress, direction localizes the active axis, and variance encodes whether the trajectory is being dominated by a single unresolved pressure.

3. TM-ARC and online intervention during sampling

ARC is paired with a controller called TensionModulator, or TM-ARC, which uses the code as an online feedback signal during diffusion sampling. At each denoising step, TM-ARC estimates the total tension, the axis-wise tensions, and the imbalance structure, then applies targeted latent-space interventions to steer the trajectory back toward the ideal prompt-aligned manifold. The update rule is

$$P_i: \mathcal{M} \rightarrow \mathcal{M}_i,\quad i \in \{SC, SA, KG\}.$$4

with a gating function

$$P_i: \mathcal{M} \rightarrow \mathcal{M}_i,\quad i \in \{SC, SA, KG\}.$$5

so correction strength increases with overall tension (Yang et al., 7 Jul 2025).

The intervention is axis-specific. SC-Gate mitigates semantic drift by reactivating attention to prompt-critical entities. SA-Tuner mitigates structural errors by adjusting spatial or positional encodings. KG-Augment mitigates grounding failures by injecting or emphasizing factual or commonsense priors. KG-Augment has two modes: static injection, which prepends KG-related prompt embeddings to the text encoder input, and dynamic modulation, which reweights cross-attention based on $$P_i: \mathcal{M} \rightarrow \mathcal{M}_i,\quad i \in \{SC, SA, KG\}.$$6.

The controller is described as latent-space only, training-free, modular, and compatible with pretrained backbones. Those properties are central to its intended role. ARC therefore functions simultaneously as diagnosis and control variable: the same vector that identifies the failure axis is also used to determine where and how corrective effort should be applied.

4. Empirical characterization

The empirical program supporting ARC has two parts: representational validation of the code itself and end-to-end evaluation of TM-ARC as a mitigation mechanism. In a synthetic controlled experiment with 1000 deterministic prompt-image pairs containing unambiguous geometry, color, and position, a standard DDPM still produced 27.8% misaligned samples. Clustering these failures yielded three stable groups corresponding to the SC, SA, and KG axes, which is presented as evidence that hallucination modes are tri-axial rather than undifferentiated (Yang et al., 7 Jul 2025).

The representational quality of ARC was evaluated against lower-information baselines. For clustering quality, the reported values were: PCA with Silhouette 0.41, Calinski-Harabasz 752.3, and Davies-Bouldin 1.37; random 3D mapping with Silhouette 0.21, Calinski-Harabasz 389.1, and Davies-Bouldin 2.10; and alignment features with Silhouette 0.63, Calinski-Harabasz 1480.5, and Davies-Bouldin 0.84. A dimensionality ablation further reported 78.2% accuracy and silhouette 0.42 for 1D ARC, 86.7% accuracy and silhouette 0.52 for 2D, and 91.6% accuracy and silhouette 0.63 for 3D, supporting the claim that hallucination structure is inherently multi-axial (Yang et al., 7 Jul 2025).

Prompt-level controllability experiments reported faithfulness improvements of +14.3 for “Flying Elephant” and +9.7 for “Red Triangle,” with the dominant ARC component shrinking after intervention. On standard benchmarks, the framework was evaluated on DrawBench and Pick-a-Pic across Stable Diffusion XL, Stable Diffusion 1.5, PixArt-sigma, and Hunyuan-DiT. Reported DrawBench results for ARC included, for SDXL, CLIPScore 29.3, PickScore 21.6, ImageReward 0.85, and FID 20.0; for SD1.5, CLIPScore 28.7, PickScore 21.1, ImageReward 0.89, and FID 20.0; for PixArt-sigma, CLIPScore 29.3, PickScore 21.5, ImageReward 0.79, and FID 19.8; and for Hunyuan-DiT, CLIPScore 29.4, PickScore 22.6, ImageReward 0.81, and FID 18.2. The aggregate claims were that ARC achieved the best PickScore in 6 of 8 settings and the lowest FID in 7 of 8 settings, while reducing hallucination without compromising image quality or diversity (Yang et al., 7 Jul 2025).

These results establish the intended status of ARC as more than a descriptive visualization. In the cited formulation, it is a compact state representation with measurable clustering behavior, prompt-level controllability, and benchmark-level mitigation effects.

5. Relation to broader risk-aware alignment research

ARC belongs to a wider landscape of methods that make alignment explicitly sensitive to risk structure, but adjacent frameworks use different objects, objectives, and deployment loci. In agentic AI, “risk alignment” refers to calibrating an AI system’s decision-making under uncertainty to the risk attitudes of users, developers, or society, rather than to a tri-axial latent code for generative failure (Clatterbuck et al., 2024). In language-model fine-tuning, Risk-aware Stepwise Alignment (RSA) treats alignment as a token-level risk-aware constrained policy optimization problem with nested risk measures, targeting tail risks and model drift rather than hallucination trajectories in diffusion sampling (Zhang et al., 30 Dec 2025). At inference time, DARC—Disagreement-Aware Alignment via Risk-Constrained Decoding—reranks candidates by maximizing a KL-robust entropic satisfaction objective and introduces explicit risk budgets for heterogeneous human preferences, again without defining any object called Alignment Risk Code (Zou et al., 9 Mar 2026).

Other nearby systems emphasize interpretable reward decomposition or test-time configurability. ARCANE frames alignment as multi-agent collaboration through rubrics represented as weighted sets of verifiable criteria, and uses those rubrics to steer long-horizon agents without retraining (Masters et al., 5 Dec 2025). ARCO—Adaptive Rubric CO-evolution—generates per-step natural-language criteria and predicts rubric-conditioned step-level rewards for multi-step LLM agents; the paper is explicit that ARCO is not an “Alignment Risk Code” and should not be conflated with ARC in that sense (Tian et al., 19 Jun 2026).

Taken together, these works suggest a broad family resemblance: risk-sensitive alignment increasingly favors decomposed, auditable, or distribution-aware control signals over single opaque reward scalars. ARC’s specific contribution within that landscape is narrower and more concrete: a 3D latent tension code for semantic, structural, and grounding instability during text-to-image generation.

6. Scope, ambiguity of the acronym, and limitations

The acronym “ARC” is used in multiple unrelated literatures. It denotes Adaptive Risk Control in online calibration (Zecchin et al., 2024), Alignment-based Redirection Controller in redirected walking for virtual reality (Williams et al., 2021), ARC-class fusion power plants in fusion proliferation analysis (Ball et al., 2024), and the Abstraction and Reasoning Corpus benchmark in program-synthesis work on abstract reasoning (Lei et al., 23 May 2025). That ambiguity matters because several papers relevant to alignment and risk use ARC-adjacent terminology without introducing Alignment Risk Code as such. In the cited literature, the explicit definition of ARC as “Alignment Risk Code” belongs to the text-to-image hallucination framework described above (Yang et al., 7 Jul 2025).

The same source also makes the main limitations visible. ARC is defined over three specific axes—SC, SA, and KG—so other failure modes may not fit neatly into the Hallucination Tri-Space. The framework assumes that hallucination risk can be captured by latent alignment tensions, which may not fully cover all generative artifacts. Validation is concentrated on text-to-image diffusion models, and the grounding component may depend on the availability of useful auxiliary priors. These are not peripheral caveats: they delimit the current meaning of ARC. It is a specific representational and control framework for tri-axial hallucination dynamics, not a general-purpose alignment taxonomy, not a universal risk metric, and not a synonym for the broader study of risk alignment in agentic AI (Yang et al., 7 Jul 2025).