TrinityX: LLM Alignment & Trinification Models

• TrinityX is a modular framework for LLM alignment that integrates a Mixture of Calibrated Experts to balance helpfulness, harmlessness, and honesty.
• In particle physics, TrinityX denotes a TeV-scale trinification model based on SU(3)³ symmetry, addressing realistic fermion masses, collider constraints, and dark matter.
• Both usages showcase context-dependent innovation, with calibrated routing in NLP and gauge dynamics in particle physics driving measurable performance and phenomenological insights.

Searching arXiv for papers on "TrinityX" and closely related usages. arxiv.search query: all:"TrinityX" TrinityX is a research term that appears in two distinct technical contexts in the materials considered here. In natural language processing, it denotes a modular alignment framework for LLMs that jointly targets Helpfulness, Harmlessness, and Honesty through a Mixture of Calibrated Experts embedded within the Transformer architecture (Kashyap et al., 10 Sep 2025). In high-energy theory, the same label is applied to a TeV-scale trinification construction based on $SU(3)_c \times SU(3)_L \times SU(3)_R$, analyzed for fermion masses, heavy neutral leptons, LHC signatures, the muon $g-2$ anomaly, monopoles, and exotic fractionally charged states (Raut et al., 2022). This suggests that “TrinityX” functions as a shared name across otherwise unrelated literatures rather than as a single cross-domain framework.

1. Dual usage and research context

The term is used differently across the cited sources.

Usage of “TrinityX”	Domain	Core formulation
TrinityX	LLM alignment	Modular HHH alignment with a Mixture of Calibrated Experts in Transformer FFNs
“TrinityX” trinification model	BSM particle physics	TeV-scale $SU(3)_c \times SU(3)_L \times SU(3)_R$ model with two bi-fundamental Higgs multiplets
Related trinification development	Dark-matter phenomenology	$SU(3)_C \times SU(3)_L \times SU(3)_R$ with exact discrete $T$ parity and vector boson dark matter

In the alignment literature, the motivating problem is that LLMs must satisfy three often-conflicting alignment criteria—Helpfulness, Harmlessness, and Honesty—while fine-tuning or RLHF on one dimension frequently causes “catastrophic forgetting” in the others, and naïve Mixture-of-Experts suffers from unstable, poorly calibrated routing (Kashyap et al., 10 Sep 2025). In the particle-physics usage, the underlying problem is the low-energy realization of trinification at the TeV scale without imposing gauge coupling unification, while retaining a viable Higgs sector, realistic fermion masses, collider consistency, and additional phenomenology such as monopoles and exotic states (Raut et al., 2022).

A related, but distinct, trinification line of work studies $SU(3)_C \times SU(3)_L \times SU(3)_R$ supplemented by an exact discrete $T$ parity, where the lightest $T$-odd particle can be a stable vector boson dark-matter candidate (Babu et al., 2021). This situates the physics usage of TrinityX within a broader $SU(3)^3$ model-building program.

2. TrinityX as an LLM alignment framework

In the NLP usage, TrinityX is a modular alignment framework designed to keep separate expertise for each HHH dimension via lightweight task vectors, integrate them at inference via a Mixture of Calibrated Experts (MoCaE), and enforce stable, well-shaped routing through entropy and KL regularization (Kashyap et al., 10 Sep 2025). The base LLM parameters $\theta_0$ are frozen, with the paper giving LLaMA-2-7B as an example backbone.

For each alignment objective $g-2$0, TrinityX trains a low-rank adapter, or task vector, $g-2$1 with $g-2$2. This yields an expert set $g-2$3 and prevents interference by never merging $g-2$4 back into $g-2$5. The design goal is modularity without destructive parameter interference.

The framework also computes static weights $g-2$6 via cosine or inner-product similarity to a reference task vector and normalizes them as

$g-2$7

These weights are later combined with dynamic routing probabilities, so the model does not rely on either static task similarity or per-token gating alone.

This decomposition reflects the paper’s broader claim that existing methods often optimize for individual alignment dimensions in isolation, leading to trade-offs and inconsistent behavior, whereas a calibrated modular mechanism can preserve specialized expertise while still producing a unified output representation (Kashyap et al., 10 Sep 2025).

3. MoCaE architecture, routing, and optimization

The MoCaE module is integrated in every Transformer FFN. At each layer, given hidden state $g-2$8 and expert adapters $g-2$9, TrinityX computes per-expert routing logits

$SU(3)_c \times SU(3)_L \times SU(3)_R$0

followed by temperature-scaled softmax gating

$SU(3)_c \times SU(3)_L \times SU(3)_R$1

with $SU(3)_c \times SU(3)_L \times SU(3)_R$2 in the reported experiments (Kashyap et al., 10 Sep 2025).

Each expert produces

$SU(3)_c \times SU(3)_L \times SU(3)_R$3

and the combined weight is

$SU(3)_c \times SU(3)_L \times SU(3)_R$4

The aggregated output is

$SU(3)_c \times SU(3)_L \times SU(3)_R$5

after which the calibrated embedding is formed as

$SU(3)_c \times SU(3)_L \times SU(3)_R$6

The paper summarizes the layerwise procedure as: $g-2$54

The router $SU(3)_c \times SU(3)_L \times SU(3)_R$7 is implemented as per-expert linear heads $SU(3)_c \times SU(3)_L \times SU(3)_R$8 atop $SU(3)_c \times SU(3)_L \times SU(3)_R$9. To stabilize routing, TrinityX applies two regularizers on $SU(3)_C \times SU(3)_L \times SU(3)_R$0. The entropy term is

$SU(3)_C \times SU(3)_L \times SU(3)_R$1

and the temporal regularizer is

$SU(3)_C \times SU(3)_L \times SU(3)_R$2

The combined per-layer objective is

$SU(3)_C \times SU(3)_L \times SU(3)_R$3

with $SU(3)_C \times SU(3)_L \times SU(3)_R$4 and $SU(3)_C \times SU(3)_L \times SU(3)_R$5 in the experiments. A gating-loss term can also be added:

$SU(3)_C \times SU(3)_L \times SU(3)_R$6

Expert training is independent at first. For datasets $SU(3)_C \times SU(3)_L \times SU(3)_R$7, the reported losses are standard cross-entropy on each dimension:

$SU(3)_C \times SU(3)_L \times SU(3)_R$8

$SU(3)_C \times SU(3)_L \times SU(3)_R$9

$T$0

The final joint MoCaE training uses

$T$1

The central methodological claim is therefore not merely expert modularity, but calibrated expert mixing. The paper explicitly attributes ablation gains to calibrated routing rather than to expert specialization alone (Kashyap et al., 10 Sep 2025).

4. Evaluation, reported gains, and efficiency

The evaluation uses three standard alignment benchmarks: Alpaca-Small train / Alpaca-Eval test with 805 instructions for Helpfulness, BeaverTails-Train with 27K safe samples and BeaverTails-Test with 3K unsafe samples for Harmlessness, and TruthfulQA split half for train/test and expanded via permutations for Honesty (Kashyap et al., 10 Sep 2025).

The reported metrics are:

• Win Rate: $T$2, higher is better.
• Safety Score: $T$3, lower is better.
• Truthfulness: $T$4, higher is better.
• Composite Avg: $T$5.

On LLaMA-2-7B, the paper reports relative gains over the strongest prior joint-alignment baseline, H³Fusion, of $T$6 in win rate, $T$7 in safety score, and $T$8 in truthfulness. The corresponding absolute results are reported as H³Fusion: $T$9, $SU(3)_C \times SU(3)_L \times SU(3)_R$0, $SU(3)_C \times SU(3)_L \times SU(3)_R$1, $SU(3)_C \times SU(3)_L \times SU(3)_R$2, versus TrinityX (full): $SU(3)_C \times SU(3)_L \times SU(3)_R$3, $SU(3)_C \times SU(3)_L \times SU(3)_R$4, $SU(3)_C \times SU(3)_L \times SU(3)_R$5, $SU(3)_C \times SU(3)_L \times SU(3)_R$6 (Kashyap et al., 10 Sep 2025).

The ablation studies identify several sensitivities. Removing MoCaE entirely changes metrics from $SU(3)_C \times SU(3)_L \times SU(3)_R$7, $SU(3)_C \times SU(3)_L \times SU(3)_R$8, and $SU(3)_C \times SU(3)_L \times SU(3)_R$9. Removing entropy or KL regularization leads to less expert diversity, slight overfitting or unstable routing. Increasing the gating-loss weight improves WR but may slightly worsen SS. Varying the temperature $T$0 shows that too low a value sharpens $T$1, causing over-confidence and a drop in TI.

The efficiency claims are also explicit. Comparing TrinityX to H³Fusion on LLaMA-2-7B under the full pipeline, Table 5 reports inference time per sample changing from $T$2 to $T$3 seconds, training time from $T$4 to $T$5 seconds, and memory for TrinityX at approximately $T$6 MB, characterized as a $T$7 reduction (Kashyap et al., 10 Sep 2025). The explanation given is sparse activation of only a subset of experts per layer and the fact that full fine-tuned models for each objective are never loaded simultaneously.

Cross-model evaluation extends the framework to Mistral-7B, Gemma-7B, and DeepSeek-7B. All three are reported to show similarly large improvements over baselines, with DeepSeek-7B achieving the highest overall Avg at approximately $T$8 while maintaining $T$9, $T$0, and $T$1 (Kashyap et al., 10 Sep 2025). This is presented as evidence of backbone-level generalization rather than a model-specific artifact.

5. TrinityX as a TeV-scale trinification model

In the particle-physics usage, TrinityX refers to a trinification model based on

$T$2

with electric charge embedded as

$T$3

where

$T$4

(Raut et al., 2022).

Per generation, the fermion multiplets are

$T$5

with

$T$6

and

$T$7

Here $T$8 is a new down-type quark, $T$9 and $SU(3)^3$0 are heavy $SU(3)^3$1 doublets, and $SU(3)^3$2 are SM-singlet neutral leptons. The scalar sector contains two bi-fundamentals,

$SU(3)^3$3

The vacuum alignments are given by

$SU(3)^3$4

with electroweak VEVs $SU(3)^3$5 and trinification-breaking VEVs $SU(3)^3$6. The symmetry-breaking chain is

$SU(3)^3$7

The renormalizable Higgs potential contains quadratic terms, a mixing term $SU(3)^3$8, quartics $SU(3)^3$9, and cubic terms $\theta_0$0, with the complete expression given in Eq. (2.21) of the paper. Minimization in the electroweak-preserving limit $\theta_0$1 yields relations such as

$\theta_0$2

The paper further states that 12 Goldstones are eaten by the 12 new gauge bosons (Raut et al., 2022).

Fermion masses arise from the Yukawa Lagrangian

$\theta_0$3

This gives the tree-level charged-fermion mass matrices

$\theta_0$4

$\theta_0$5

For the heavy down-type sector,

$\theta_0$6

A numerical scan of the $\theta_0$7 heavy-quark matrix gives $\theta_0$8; with the lightest new $\theta_0$9 from CMS, the resulting bound is $g-2$00 (Raut et al., 2022).

Neutral leptons are massless at tree level. One-loop radiative diagrams generate Majorana masses for $g-2$01, while below $g-2$02 a tiny $g-2$03 is also generated. The full light-neutral mass matrix is reported as

$g-2$04

leading, for $g-2$05, to

$g-2$06

The paper characterizes this as a mixture of radiative type-I and type-II seesaw (Raut et al., 2022).

6. Phenomenology, related trinification developments, and interpretive cautions

The collider analysis of the trinification model reports 12 extra vector bosons whose masses are controlled by $g-2$07 and $g-2$08 once $g-2$09 is fixed. Resonant production is treated in the narrow-width approximation,

$g-2$10

with

$g-2$11

By rescaling ATLAS limits on sequential $g-2$12 and $g-2$13, the current bound on the trinification-breaking VEV is reported as

$g-2$14

(Raut et al., 2022).

The same model is also presented as a possible resolution of the muon $g-2$15 anomaly through loops involving heavy charged leptons and new neutral gauge bosons $g-2$16. The one-loop contribution is

$g-2$17

with

$g-2$18

$g-2$19

and $g-2$20. The numerical fit shown in the paper states that explaining $g-2$21 within $g-2$22–$g-2$23 requires $g-2$24–$g-2$25 and corresponding $g-2$26–$g-2$27 with perturbative Yukawas (Raut et al., 2022).

Topological and exotic-sector consequences are equally central. The paper gives

$g-2$28

yielding a topologically stable monopole carrying three quanta of Dirac magnetic charge,

$g-2$29

with estimated mass

$g-2$30

for $g-2$31. If fundamentals of $g-2$32 are added, the model predicts exotic mesons and baryons with fractional electric charges such as $g-2$33 and $g-2$34 (Raut et al., 2022).

A related trinification paper studies an exact discrete $g-2$35 parity,

$g-2$36

within $g-2$37, so that the lightest $g-2$38-odd particle is absolutely stable (Babu et al., 2021). In that framework, the T-odd neutral gauge-boson mass matrix in the $g-2$39 basis leads, in the limit $g-2$40, to a light eigenstate mostly $g-2$41 with mass

$g-2$42

where $g-2$43 and $g-2$44. The dark-matter analysis gives the relic-density requirement $g-2$45, a perturbative bound $g-2$46, and collider bounds $g-2$47 and $g-2$48 for light jets, or $g-2$49 for $g-2$50-jets (Babu et al., 2021).

Two misconceptions are therefore worth separating. First, TrinityX in LLM alignment and TrinityX in trinification are not the same construct; the sources place them in unrelated methodological and disciplinary contexts. Second, within the physics usage, the minimal two-$g-2$51 trinification model and the $g-2$52-parity dark-matter construction are related by gauge structure but are not identical implementations. A plausible implication is that “TrinityX” in current literature should be interpreted contextually: in machine learning it designates a calibrated expert-mixture alignment method, whereas in particle physics it denotes or is associated with a class of TeV-scale $g-2$53 trinification models with distinctive collider, neutrino, and topological phenomenology.