Model Raising: Enhancing Capacity & Alignment

• Model raising is a set of methodological and algorithmic techniques that enhance a model’s effective capacity, performance, and alignment without solely increasing parameter count.
• In supervised learning and generative modeling, it involves strategies like meta-level feature engineering and recursive inference, leading to measurable improvements such as an increase in F1-score from 0.80 to 0.82376.
• The approach extends across fields—from AI safety with corpus-centric pretraining to dimension-raising phase transitions in physics and scale elevation in particle models—underscoring its practical and theoretical impact.

Model raising refers to a set of methodological, algorithmic, and physical principles that seek to increase the effective capacity, performance, or structural depth of models—statistical, machine learning, or physical—by means other than, or in addition to, simply enlarging model parameters. The term occurs across multiple scientific disciplines, including machine learning (meta-level feature engineering and AI alignment paradigms), condensed matter physics (dimension-raising phase transitions), and theoretical particle physics (raising supersymmetry-breaking scales). While each context adopts a field-specific interpretation, underlying all usages is the notion of achieving qualitative or quantitative gains in a target property (expressivity, alignment, prediction, or order) without incurring the traditional costs in size, dimensionality, or parameter count. The following sections survey key instantiations of model raising across fields.

1. Model Raising in Supervised Learning: Meta-level Feature Engineering

In supervised machine learning, model raising has been formalized as a meta-level feature engineering paradigm designed to augment predictive performance through the introduction of auxiliary models that articulate and exploit intra-feature relationships conditioned on target classes. Instead of only fitting a mapping $X \to Y$ (where $X$ is the set of observed features and $Y$ the target value), the model raising methodology expands the feature space by constructing, for each feature $X_j$ and each class (or slice) $k$ of $Y$, an auxiliary model $M_{(k,j)}: X_{-j} \to X_j$. These models are trained only on samples from the class $k$, and the weighted residuals of their predictions—computed across all original samples—are appended as additional columns to the feature matrix. The weights are proportional to the auxiliary models' predictive $R^2$ scores, embedding the quality of intra-feature predictability as a function of the target (Lobo-Cabrera, 2024).

In empirical proof-of-concept, this approach achieved substantial gains. For example, on the UCI Bank Marketing dataset, introducing 126 model-raising columns (for 63 features and two classes) improved the F1-score from a baseline of approximately 0.80 to 0.82376 in a Random Forest classifier. This methodology generalizes to multilabel, multiclass, and regression settings via appropriate partitioning or weighting. A principal assumption is that the conditional dependence structure among features varies meaningfully with the target; otherwise, the added columns may encode only noise or redundancy. The pipeline also supports iterative application ("second layer" raising), with safeguards against degenerate row-to-feature ratios to prevent overfitting. Key limitations include computational overhead linearly scaling with the number of features, target classes, and data samples, and the need for sufficient per-slice sample size (Lobo-Cabrera, 2024).

2. Model Raising as Corpus-Centric Pretraining for Alignment in AI Systems

Recent proposals have repurposed the term "model raising" to describe a paradigm shift in LLM training—from the standard two-stage regime (pretraining for general capability, followed by post-hoc alignment) to an integrated, corpus-centered process where value alignment is inseparable from skill acquisition (Aydin et al., 12 Nov 2025). The central premise is that a model's alignment properties should arise naturally from the very first training tokens, achieved by systematically redesigning the training corpus with alignment as an intrinsic objective.

This is operationalized via four corpus interventions:

1. First-Person Reframing: Transforming every sequence into a narrative anchored in an "I" perspective, binding knowledge and values to a persistent identity.
2. Recontextualization as Lived Experience: Rendering facts into micro-narratives, providing emotional or social context that entwines values directly in the learning signal.
3. Simulated Social Interaction: Converting monologue text into dialogic exchanges, embedding norms of reciprocity, perspective-taking, and social reasoning as foundational constructs from the outset.
4. Scaffolded Data Ordering (Curriculum Learning): Structuring the data into progressive stages of difficulty, such that ethical and epistemic primitives are mastered in lockstep.

The training objective is correspondingly expanded: rather than optimizing only the standard masked LLM loss $\mathcal{L}_{\rm MLM}$, an auxiliary alignment loss $X$0 is applied at each token during pretraining. Quantitative metrics for "early commitment" to values, such as the Value-Token Log-Likelihood Ratio (VLLR) and Knowledge-Value Mutual Information (KVMI), are used to monitor entanglement between skills and values during training.

This framework posits deeper, more stable alignment properties—reducing the risk of post-hoc misalignment and making values and capabilities intrinsically harder to disentangle—at the expense of corpus engineering and the upfront selection of targeted value systems (Aydin et al., 12 Nov 2025). Open research questions include the impact of bias in value selection, long-horizon evaluation of alignment retention, and the balance between value commitment and adaptability.

3. Recursive Scaling and Effective Capacity in Diffusion Models

Model raising in generative modeling appears as a strategy for scaling Masked Diffusion Models (MDMs) via recursive depth rather than or in addition to parameter count or denoising step increases (Carballo-Castro et al., 16 Jun 2026). In Recursive Masked Diffusion Models (R-MDMs), the standard single application $X$1 of the denoiser at each diffusion step $X$2 is replaced by an $X$3-fold recursive application of the same $X$4:

$X$5

Because all $X$6 passes use shared parameters, the model achieves an effective depth of $X$7 (for Transformer depth $X$8) without increasing parameter count—trading inference compute for expressive depth. On structured sequence-generation tasks such as Sudoku completion and Countdown, R-MDMs with $X$9-fold recursion routinely match the quality of non-recursive MDMs with $Y$0-times more parameters. For instance, a $Y$1 R-MDM ($Y$210.6M params) reaches performance equivalent to a single-pass $Y$3 MDM ($Y$453.1M params) at the same step count.

Moreover, recursive depth can partially substitute for the number of diffusion steps $Y$5 at fixed total compute budget $Y$6. On 9x9 Sudoku, a baseline single-pass model at $Y$7 (so $Y$8) is matched by an $Y$9, $X_j$0 R-MDM (only $X_j$1 passes) at the same 95% Valid Puzzle Rate, reflecting a 2.7-fold reduction in forward passes. These empirical results establish recursive depth as a third scaling axis for model raising in diffusion-based generative architectures, enabling flexible compute allocation and remarkable parameter efficiency without underlying architectural changes (Carballo-Castro et al., 16 Jun 2026).

4. Dimension-Raising Phase Transitions in Statistical Physics

In condensed matter physics, model raising arises in the context of driven many-body systems that exhibit emergent macroscopic dimensionality exceeding that of their microscopic instantaneous dynamics. In "dimension-raising" phase transitions, periodically driven 1D Hamiltonians are shown to yield effective 2D or higher-dimensional behavior above a critical driving frequency. Concretely, an array of classical rotors with couplings alternated between orthogonal directions realizes at high-frequency driving ($X_j$2) a Floquet effective Hamiltonian that is isotropic in both dimensions:

$X_j$3

This emergent 2D phase, in contrast to the 1D instantaneous configuration, can undergo true phase transitions (e.g., a Berezinskii–Kosterlitz–Thouless (BKT) transition or, for intermediate frequencies, a first-order transition with hysteresis). Generalization to higher dimensions is possible by appropriate scheduling of coupling along additional spatial axes. Experimental realization is feasible in cold-atom optical lattices and programmable Rydberg tweezer arrays (Chen et al., 12 Mar 2025).

5. Raising Scales in Theoretical Particle Physics

In high-energy physics, the term "raising" refers to the increase of critical scales through model extensions. One notable instance is in supersymmetric (SUSY) models of electroweak symmetry breaking. By merging the elementary-Goldstone-Higgs scenario with supersymmetry, the spontaneous symmetry breaking scale $X_j$4 (derived from an enhanced global $X_j$5 symmetry in the super-Higgs sector) is identified with the SUSY-breaking scale, both lying in the multi-TeV regime. The model delivers a pseudo–Nambu–Goldstone boson Higgs whose vacuum expectation value is radiatively induced and suppressed relative to $X_j$6, yielding $X_j$7 with $X_j$8. Consequently, all SUSY particles (superpartners) obtain masses at several TeV, aligning with LHC bounds, while the light Higgs mass and vacuum alignment dynamics remain consistent with Standard Model phenomenology. Additional features, such as $X_j$9 and a requisite singlet superfield, are direct consequences of the symmetry structure (Alanne et al., 2016).

6. Conceptual Synthesis and Theoretical Implications

Across machine learning, AI safety, condensed matter, and particle physics, model raising functions as a lever for enhancing systems' effective capacity, alignment, or dimensionality—either by engineering the learning environment, modifying inference dynamics, or extending symmetry structure. A recurring theme is the substitution of traditionally expensive scaling (more parameters, longer diffusion or simulation times, higher effective dimension) by structured overloading: recursion, meta-level modeling, emergent connectivity, or symmetry-induced scale elevation.

A plausible implication is that future systems in multiple disciplines will increasingly rely on model raising mechanisms to overcome trade-offs between resource constraints and functional requirements, whether in low-sample, compute-limited, or physically realizable scenarios. For each context, however, the design and efficacy of model raising is contingent on the underlying structure of interactions (data, features, symmetries, or coupling), and the presence of strong target-conditional dependencies or emergent properties is essential for substantive benefit.

7. Limitations, Assumptions, and Open Questions

The effectiveness of model raising is conditional on several factors:

• In machine learning, if intra-feature dependencies are not target-modulated, or if the auxiliary models are weak, performance may degrade.
• Recursive generative models require careful allocation of compute budget, balancing recursion against pipeline depth for optimal quality (Carballo-Castro et al., 16 Jun 2026).
• Corpus-centric alignment is sensitive to the choice of foundational values and susceptible to bias propagation in the corpus design phase (Aydin et al., 12 Nov 2025).
• In physical systems, dimension-raising transitions depend on precise control of driving frequency and noise, and exhibit novel critical behavior distinct from equilibrium analogs (Chen et al., 12 Mar 2025).
• Theoretical frameworks involving scale raising in high-energy models impose specific constraints on vacuum alignment and radiative correction mechanisms (Alanne et al., 2016).

Open questions remain regarding the limits of benefit, the scaling behaviors in ultra-high-dimensional or compositional tasks, and the long-term entanglement of targets (such as skills and values) under iterative model raising. The empirical and theoretical characterization of model raising across domains continues to be a dynamic area of interdisciplinary research.