Dis-GEN: Integrating Discrimination in Generation

• Dis-GEN is a cross-domain research motif wherein discriminative signals are integrated directly into the generative process across various fields.
• In recommender systems, the DIG formulation unifies ranking and retrieval by learning semantic IDs through residual quantization and end-to-end training.
• In materials science and multimodal tasks, Dis-GEN models use specialized representations, such as symmetry-aware Wyckoff encoding and diffusion frameworks, to satisfy domain-specific constraints.

Dis-GEN is a label used in several distinct research settings rather than a single canonical method. In recommender systems, DIG—short for “Discrimination Is Generation”—redefines retrieval and ranking as the same optimization problem solved at different granularities, and learns Semantic IDs inside a discriminative ranker through end-to-end training. In materials science, “Dis-GEN” denotes a variational autoencoder for disordered inorganic crystal structure generation that operates on a symmetry-aware Wyckoff representation and explicitly models partial occupancies and vacancies. In adjacent multimodal work, DiffDis is explicitly described as a Dis-GEN model because it unifies image generation and image–text discrimination inside one diffusion framework (Wang et al., 14 May 2026, Petersen et al., 24 Jul 2025, Huang et al., 2023).

1. Terminological scope

In current usage, “Dis-GEN” is best understood as a family of field-specific formulations linked by a common idea: a discriminative signal is not treated as external to generation, but is incorporated into the generative representation or generation space itself.

Usage	Research area	Defining formulation
DIG / “Discrimination Is Generation”	Recommender systems	The tokenizer is embedded inside a discriminative ranking model, and the ranker naturally becomes a retrieval model
Dis-GEN	Crystal generation	A generative model for disordered inorganic crystals based on an empirical equivariant representation
DiffDis as a Dis-GEN model	Vision–language diffusion	A single diffusion framework jointly handles text-to-image generation and image–text discrimination

This distribution of usages suggests that “Dis-GEN” functions less as a standardized technical term than as a recurring research motif: generation is reformulated so that discrimination, ranking, or symmetry constraints are internal to the generative mechanism rather than appended after the fact.

2. “Discrimination Is Generation” in recommender systems

In DIG, the central problem is generative recommendation with Semantic IDs (SIDs). An item $v$ is mapped to a discrete token sequence,

$$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$

where $\mathcal{V}$ is the item set, $\mathcal{T} = \{1,\dots,K\}$ is the codebook vocabulary, and $L$ is the code length. These SIDs define the generation space: a generative model produces tokens in $\mathcal{T}$ and thereby selects items indirectly. DIG departs from earlier pipelines by arguing that the tokenizer should not be trained independently with retrieval objectives, because personalization signals are otherwise decoupled from SID construction (Wang et al., 14 May 2026).

The paper’s core claim is that ranking and retrieval are the same argmax problem at different granularities. Ranking operates in item space,

$\arg\max_{v \in \mathcal{V}} f(u,v),$

while generative retrieval operates in token space,

$$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$

If the mapping $\phi(v) = (s_1^v,\dots,s_L^v)$ is learned appropriately, then ranking over items and ranking over their SIDs become two views of the same optimization. On this view, SIDs are not merely an indexing device; they are the bridge between item space and token space.

DIG instantiates this perspective by embedding a residual quantization tokenizer inside a high-capacity discriminative ranker, specifically a DIN+DCNv2+MoE “Mixer.” The result of one training run is both a ranking model operating in item space and a retrieval model operating in token space. Items are represented by static features $\mathbf{x}_v^s$, users by $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$0, context by $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$1, and user–item cross features by $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$2. The ranker predicts

$$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$3

while retrieval at prefix depth $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$4 predicts

$$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$5

A defining design choice is the decoupling between codebook vectors $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$6, used for assignment and reconstruction, and SID embeddings $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$7, used for scoring. Residual quantization assigns

$$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$8

with $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$9, while the prefix SID representation is

$\mathcal{V}$0

This separation allows codebooks to remain fine-grained for addressing and indexing, while SID embeddings absorb the discriminative signal that ranking supplies.

3. Tokenizer learning, u2i-to-u2t distillation, and unified retrieval–ranking

DIG organizes features through a three-way taxonomy. Type-I features are item-intrinsic static features $\mathcal{V}$1, encoded into SIDs and used for offline SID lookup. Type-II features are request-level features $\mathcal{V}$2, not encoded into SIDs but used as conditioning at training and inference. Type-III features are structurally lossy cross features $\mathcal{V}$3, the user–item statistics that carry much of ranking’s personalization power. These cannot be encoded into static SIDs, so DIG aggregates them into bucket-level token statistics $\mathcal{V}$4 during training and approximates them at inference with an MLP$\mathcal{V}$5.

For items assigned to the same bucket $\mathcal{V}$6 at layer $\mathcal{V}$7 in batch $\mathcal{V}$8, DIG defines

$\mathcal{V}$9

At inference time, direct aggregation is unavailable because the model is searching over token prefixes rather than known items. DIG therefore distills these statistics into

$\mathcal{T} = \{1,\dots,K\}$0

with supervision

$\mathcal{T} = \{1,\dots,K\}$1

This is the mechanism by which user–item cross features shape codebook boundaries during training and are approximately restored at token level during retrieval.

The full objective is

$\mathcal{T} = \{1,\dots,K\}$2

where $\mathcal{T} = \{1,\dots,K\}$3 is ranking BCE, $\mathcal{T} = \{1,\dots,K\}$4 is layer-wise retrieval BCE, $\mathcal{T} = \{1,\dots,K\}$5 is the commitment/alignment loss,

$\mathcal{T} = \{1,\dots,K\}$6

and $\mathcal{T} = \{1,\dots,K\}$7 is the semantic reconstruction loss,

$\mathcal{T} = \{1,\dots,K\}$8

Inference proceeds by beam search in SID space. At each layer $\mathcal{T} = \{1,\dots,K\}$9, each prefix is extended with all $L$0 candidate tokens, scored by the shared Mixer using the SID prefix and $L$1, and pruned to the top $L$2 prefixes. Completed SIDs are then mapped to items through the offline SID-to-item inverted index.

Empirically, DIG is evaluated on KuaiRec-Small, KuaiRec-Big, Taobao Ad, Meituan-Large, and Meituan-Small. On retrieval, it outperforms the best reported baseline, ETEGRec, across all five datasets. On Taobao, Recall@10 improves from $L$3 to $L$4 and NDCG@10 from $L$5 to $L$6; on Meituan-Small, Recall@10 improves from $L$7 to $L$8 and NDCG@10 from $L$9 to $\mathcal{T}$0. Ranking AUC also improves relative to the same model trained rank-only: on Taobao, $\mathcal{T}$1; on Meituan-Small, $\mathcal{T}$2. The retrieval-path AUC remains below ranking AUC, with gaps such as $\mathcal{T}$3 on Taobao and $\mathcal{T}$4 on Meituan-Large, and the paper relates larger gaps to more complex u2i features that are harder for MLP$\mathcal{T}$5 to approximate (Wang et al., 14 May 2026).

4. Dis-GEN for disordered crystal structure generation

In materials science, Dis-GEN is a generative model for crystalline inorganic materials with compositional disorder. The target objects are not ideal ordered crystals, but structures in which one or more Wyckoff sites have partial occupancy, mixed species, or vacancies. For a site $\mathcal{T}$6, the occupancy can take the form

$\mathcal{T}$7

and with vacancies included,

$\mathcal{T}$8

The model is introduced because most crystal generators assume ordered structures with one atom per site and therefore cannot directly represent mixed occupancy at the average-structure level (Petersen et al., 24 Jul 2025).

Dis-GEN addresses this with an empirical equivariant representation derived from crystallography. Each crystal is encoded as a pair $\mathcal{T}$9. The matrix $\arg\max_{v \in \mathcal{V}} f(u,v),$0 has one column per Wyckoff site and rows encoding partial occupancies, Wyckoff multiplicity, a binary disordered-site indicator, fractional coordinates, and Wyckoff letter. Because structures contain variable numbers of sites, $\arg\max_{v \in \mathcal{V}} f(u,v),$1 is zero-padded to a fixed maximum number of columns, with padding indicators included so that real and padded sites are distinguishable. The vector $\arg\max_{v \in \mathcal{V}} f(u,v),$2 contains the six lattice parameters $\arg\max_{v \in \mathcal{V}} f(u,v),$3 followed by a 230-dimensional one-hot encoding of the space group.

The model is a VAE. A CNN branch encodes $\arg\max_{v \in \mathcal{V}} f(u,v),$4, an MLP branch encodes $\arg\max_{v \in \mathcal{V}} f(u,v),$5, and the concatenated representation parameterizes

$\arg\max_{v \in \mathcal{V}} f(u,v),$6

with reparameterized sampling

$\arg\max_{v \in \mathcal{V}} f(u,v),$7

A symmetric decoder reconstructs $\arg\max_{v \in \mathcal{V}} f(u,v),$8 and $\arg\max_{v \in \mathcal{V}} f(u,v),$9. The training objective is

$$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$0

using cross-entropy for categorical outputs, MSE for fractional coordinates and partial occupancies, and MAE in practice for lattice parameters. The reported coefficients are $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$1, $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$2, $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$3, $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$4, $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$5, $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$6, $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$7, and $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$8.

Training data are drawn from the ICSD. The raw database contains about $$\arg\max_{(s_1,\ldots,s_L) \in \mathcal{T}^L} g(u, s_1,\ldots,s_L).$$9 entries, of which $\phi(v) = (s_1^v,\dots,s_L^v)$0 are reported as disordered, $\phi(v) = (s_1^v,\dots,s_L^v)$1 as ordered, and $\phi(v) = (s_1^v,\dots,s_L^v)$2 as infeasible to parse. After filtering out P1, structures with atomic number $\phi(v) = (s_1^v,\dots,s_L^v)$3, more than 9 Wyckoff sites, multiplicity $\phi(v) = (s_1^v,\dots,s_L^v)$4, more than 6 disordered Wyckoff sites, more than 6 atom types on a Wyckoff site, and “atomic charge state as partial occupancy,” the final dataset contains $\phi(v) = (s_1^v,\dots,s_L^v)$5 structures, of which $\phi(v) = (s_1^v,\dots,s_L^v)$6 have at least one partially occupied Wyckoff site and $\phi(v) = (s_1^v,\dots,s_L^v)$7 do not.

5. Symmetry preservation, generation workflow, and empirical performance

A central requirement of Dis-GEN is symmetry consistency. Since the decoder predicts space group, Wyckoff letters, and multiplicities, the reconstructed structure can be checked against crystallographic tables. The resulting symmetry-matching accuracy (SMA) is defined as the fraction of reconstructed sites for which the predicted multiplicity and Wyckoff letter are consistent with the predicted space group. On the test set, the model reports lattice-parameter MAEs of $\phi(v) = (s_1^v,\dots,s_L^v)$8 Å for $\phi(v) = (s_1^v,\dots,s_L^v)$9 and $\mathbf{x}_v^s$0 degrees for $\mathbf{x}_v^s$1, space-group accuracy of $\mathbf{x}_v^s$2, disordered-site-indicator accuracy of $\mathbf{x}_v^s$3, and SMA of $\mathbf{x}_v^s$4. For disordered Wyckoff sites, partial occupancy MAE is about $\mathbf{x}_v^s$5, multiplicity accuracy is $\mathbf{x}_v^s$6, Wyckoff-letter accuracy is $\mathbf{x}_v^s$7, and fractional-coordinate MAE is $\mathbf{x}_v^s$8. For ordered sites, species occupancy accuracy is $\mathbf{x}_v^s$9, multiplicity accuracy $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$00, Wyckoff-letter accuracy $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$01, and fractional-coordinate MAE $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$02 (Petersen et al., 24 Jul 2025).

Generation proceeds by sampling the latent space after VAE training. The paper compares three latent samplers on $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$03 generated samples: a multivariate Gaussian, a Gaussian mixture model, and a kernel density estimator (KDE) fitted to training latents. KDE performs best. Its generation error pass rate is $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$04, SMA is $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$05, validity under a $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$06 Å minimum interatomic distance criterion is $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$07, and SMACT-based charge neutrality is $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$08. By contrast, the Gaussian and GMM samplers show severe degradation in SMA, at $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$09 and $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$10 respectively. This indicates that staying close to the empirical latent manifold is important for preserving the learned symmetry structure.

The reconstruction pipeline includes explicit crystallographic post-processing. Generated structures are decoded into lattice parameters, space group, and Wyckoff-site descriptions; symmetry-inconsistent or unphysical outputs are discarded; fractional coordinates are refined with PyXtal to the nearest valid Wyckoff position; and final structures are filtered by validity and charge-neutrality heuristics. The paper also describes conditional generation by latent optimization against target composition and symmetry constraints. In a Zn–V–O case study, this procedure fills composition regions underrepresented in Materials Project data, including disordered candidates. Subsequent evaluation with a MACE force field and supercell sampling places most generated phases above the convex hull. The paper attributes this to the absence of property-conditioned training and to the broader difficulty of evaluating disordered phases, which require large supercells and many sampled configurations. Limitations listed explicitly include restriction to compositional disorder on Wyckoff sites, exclusion of very complex structures, dependence on filtered ICSD coverage, modest charge-neutrality rates, and the absence of finite-temperature disorder modeling.

6. Related uses and broader research landscape

A broader “Dis-GEN” concept appears in vision–language diffusion. DiffDis asks whether a powerful text-to-image diffusion generator can be turned into a cross-modal Dis-GEN model that both generates high-quality images from text and discriminates image–text semantics within a single diffusion framework and single architecture. It keeps a latent diffusion generator for text-conditional image synthesis, reformulates image–text discrimination as a diffusion process over text embeddings conditioned on the image, and shares the same image encoder and latent UNet backbone across the two tasks. Under full pretraining from Stable Diffusion, the paper reports both better generation and better discrimination than separate baselines, including an average $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$11 improvement in ImageNet zero-shot accuracy over discriminative-only training and a $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$12 FID improvement over generative-only training in the from-scratch comparison (Huang et al., 2023).

The name also appears in graph representation learning as DISGEN, “a model-agnostic framework designed to disentangle size factors from graph representations.” In that usage, the problem is size generalization in graph neural networks: training on relatively small graphs and testing on significantly larger ones. DISGEN introduces size- and task-invariant augmentations together with a decoupling loss based on the residual of the optimal linear map between hidden size-related and task-related representations, and the paper reports gains of up to $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$13 on real-world datasets (Huang et al., 2024).

In biomedical informatics, INDIGENA uses “Dis-GEN” to denote the disease–gene association prediction problem: for a disease described by phenotypes, rank genes by their likelihood of association. That paper’s method is inductive rather than transductive, using graph-projected phenotype ontologies and phenotype-embedding aggregation rather than disease-node embeddings, and reports that ConvKB-D with Graph 4 and BMA reaches approximately MR $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$14, H@100 $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$15, and AUC $$\phi : \mathcal{V} \rightarrow \mathcal{T}^L,\quad \phi(v) = (s_1^v,\ldots,s_L^v),$$16 in the inductive setting (Zhapa-Camacho et al., 1 Feb 2026).

Taken together, these usages indicate that “Dis-GEN” is presently a cross-domain naming pattern rather than a single method family. Its recurring theme is the relocation of a discriminative, structural, or relational constraint into the representational core of a generative or ranking process, but the concrete formalism depends entirely on the field: SIDs and beam search in recommendation, Wyckoff-site occupancy tensors in crystallography, diffusion over embeddings in multimodal modeling, disentangled nuisance factors in graph learning, and ontology-driven ranking in disease–gene prediction.