RadiomicsRetrieval: Catalog-Native Recommender

• RadiomicsRetrieval is a unified framework that treats item-ID sequences as a native dialect within Transformers for joint recommendation and language understanding.
• It employs a dual-expert FFN design in IDIOMoE, separating text and item signals to prevent destructive interference and preserve pretrained language knowledge.
• Experimental results across diverse datasets show significant gains in recommendation metrics while maintaining strong language fluency, underscoring its deployment potential.

RadiomicsRetrieval, as characterized in the provided source material, is not described as a radiomics-specific retrieval system but rather through the framework of a catalog-native recommendation LLM: Item-ID + Oral-language Mixture-of-Experts LLM (IDIOMoE). In this formulation, item interaction histories are treated as a “native dialect within the language space,” so that collaborative signals and natural language can be modeled within a single decoder-only Transformer. The stated objective is to combine the predictive accuracy and efficiency of collaborative filtering with the expressive, instruction-following, and generalizable reasoning capabilities of pretrained LLMs, while avoiding the destructive interference that arises when collaborative and linguistic signals are naively fused (Shirkavand et al., 30 Sep 2025).

1. Conceptual framing and motivation

Modern recommender systems are presented as needing to reconcile two established strengths. Collaborative filtering excels at capturing implicit user–item co-occurrence patterns in large logs, whereas pretrained LLMs excel at open-ended, instruction-following language understanding and world knowledge (Shirkavand et al., 30 Sep 2025). The source material situates IDIOMoE precisely at this interface.

The central conceptual move is to treat item-ID sequences as a “native dialect.” In the terminology of the paper, user histories of item IDs are token-efficient but semantically opaque to an LLM trained only on text. By viewing item IDs as another token type, the model is intended to “read” and “generate” IDs as fluently as words. This unification is explicitly connected to conversational recommendation, natural-language queries, and transparent explanations (Shirkavand et al., 30 Sep 2025).

The source also identifies the limitations of standard approaches. Naively adding item tokens, or fusing text features through mechanisms such as the sum of ID and text embeddings, is described as causing “destructive interference,” in which collaborative filtering signals become entangled with linguistic semantics. The reported consequence is degradation in both recommendation accuracy and language fluency. The same source further states that scaling up a single large network, whether through more layers or wider FFNs, does not resolve this entanglement problem (Shirkavand et al., 30 Sep 2025).

A plausible implication is that the framework is less concerned with converting items into natural-language surrogates than with preserving their discrete catalog identity inside a language-modeling regime. This suggests a deliberate rejection of purely text-mediated item semantics in favor of a mixed symbolic-distributional representation.

2. Architectural design of IDIOMoE

IDIOMoE augments a pretrained decoder-only Transformer, with Qwen2.5-0.5B given as an example backbone, by splitting each Transformer block’s Feed-Forward Network into two specialized experts and routing tokens by type (Shirkavand et al., 30 Sep 2025). The architecture is therefore not a fully separate dual-tower system; instead, specialization is concentrated in the FFN pathway, while other components remain shared.

The tokenizer vocabulary is extended with special item-ID tokens written as 〈|item-*|〉. The embedding layer is hybrid. Text tokens use the frozen original token embeddings, while item tokens use a trainable item-embedding table; the source notes that this item-embedding table may optionally be factorized for efficiency. Within each Transformer layer, self-attention and LayerNorm are fully shared, but the single FFN is replaced by two parallel experts—a text expert and an item expert—together with a lightweight token-type gate (Shirkavand et al., 30 Sep 2025).

The two-expert FFN is defined over a token hidden state $x \in \mathbb{R}^d$ entering layer $\ell$. The source distinguishes $\mathrm{FFN}_{\mathrm{text}}(x)$ as the original FFN of the pretrained LLM, either frozen or finetuned, and $\mathrm{FFN}_{\mathrm{item}}(x)$ as a separate FFN of identical architecture, often with a shrunk intermediate size. Routing is then governed by a static token-type gate:

$$g(x) \;=\; \begin{cases} 1 & \text{if }x\text{ originated from a text token,}\ 0 & \text{if }x\text{ originated from an item-ID token.} \end{cases}$$

The combined FFN is

$$\mathrm{FFN}_{\mathrm{IDIOMoE}}(x) \;=\; g(x)\,\mathrm{FFN}_{\mathrm{text}}(x) \;+\; \bigl(1 - g(x)\bigr)\,\mathrm{FFN}_{\mathrm{item}}(x)\,.$$

The source adds that $g(x)$ can also be implemented as a small learned binary classifier on the token-type embedding, although the main model instantiates it as the known token type (Shirkavand et al., 30 Sep 2025).

This design localizes modality specialization to the MLP-like transformation layer of the Transformer while preserving shared contextualization through attention. A plausible implication is that the architecture aims to retain cross-token contextual interaction without forcing text and item representations to share the same nonlinear subspace.

3. Item-ID representation as a catalog-native dialect

The item representation scheme rests on vocabulary extension plus hybrid embedding. One special token is added per item, potentially up to tens or hundreds of thousands. Text tokens map to frozen pretrained embeddings $$E_{\mathrm{text}} \in \mathbb{R}^{V_{\mathrm{text}} \times d}$$, whereas item tokens map to trainable embeddings $$E_{\mathrm{item}} \in \mathbb{R}^{N_{\mathrm{items}} \times d}$$. The source further states that the item embedding matrix may be factorized as $E_{\mathrm{item}} = W_{\ell} W_u$ to reduce memory (Shirkavand et al., 30 Sep 2025).

Within this setup, the item expert FFNs and item embeddings are described as learning collaborative co-occurrence patterns purely from interaction sequences. Because item tokens are treated identically to word tokens apart from the expert to which they are routed, the model is said to internalize “how to speak item-ID” in the same representational space as natural language (Shirkavand et al., 30 Sep 2025).

This “catalog-native dialect” framing is significant because it does not collapse items into descriptive text attributes alone. Instead, discrete item identity remains first-class in the vocabulary. This suggests that the model can preserve the token efficiency of collaborative sequences while still participating in a single autoregressive language-modeling process.

4. Training objective and optimization regime

The primary training objective is standard next-token language modeling over mixed sequences of text and item tokens:

$\ell$0

The source emphasizes that, in the main experiments, pure next-token training on interleaved item-ID sequences was sufficient (Shirkavand et al., 30 Sep 2025).

An optional recommendation-specific loss is also described. One may add a weighted ranking loss, such as BPR or cross-entropy on the next item:

$\ell$1

However, this is presented as optional rather than necessary for the main results (Shirkavand et al., 30 Sep 2025).

The optimization details stated in the source are AdamW with linear warmup then cosine decay, batch size $\ell$2, and sequence length $\ell$3 tokens comprising items and attributes. The training schedule includes warming up on item-only sequences for 20% of epochs before mixing in text (Shirkavand et al., 30 Sep 2025). The production recommendations later reiterate a 1,024-token context and specify truncating text first, then items, for robust coverage.

These choices indicate that the method remains close to standard causal LM optimization practice, with the principal novelty located in representation and routing rather than in a specialized objective. A plausible implication is that the framework is intended to be inserted into established LLM training pipelines with limited disruption.

5. Destructive interference and modality disentanglement

The problem formulation centers on destructive interference. In a single-path FFN, gradients from text tokens are described as competing with gradients from item tokens, causing weights to oscillate between semantic and collaborative objectives (Shirkavand et al., 30 Sep 2025). This is the principal failure mode the architecture is designed to address.

IDIOMoE’s response is structural separation. The text expert preserves pretrained semantic knowledge, while the item expert focuses on collaborative patterns. Static gating prevents cross-contamination because each expert sees only its dedicated token type. Since only one expert is active per token, the compute cost is reported to remain roughly that of the base model (Shirkavand et al., 30 Sep 2025).

The ablation results support this interpretation. Non-MoE capacity controls with matched parameter count—Wide-FFN, append-blocks, prepend-blocks, and LoRA—are reported to underperform IDIOMoE or even degrade at scale. Alternative expert designs such as MoA and MoT improve over non-MoE baselines but still remain behind IDIOMoE’s static-routing FFN experts. Dynamic routing is specifically reported as harmful: Switch-style dynamic gating severely degrades performance, with declines of up to $\ell$4 NDCG, whereas static token-type gating gives clear, consistent expert roles (Shirkavand et al., 30 Sep 2025).

The source also reports that layer placement matters. Inserting experts in the last 8 layers yields the largest gains, while shallow layers benefit less from specialization. Item-expert capacity likewise exhibits a regime dependence: shrinking the item expert’s intermediate dimension by factors 2–8 yields a sweet spot, for example factor 4 on small Amazon datasets, but produces consistent drops on industrial data, implying that expert width must scale with catalog complexity (Shirkavand et al., 30 Sep 2025).

Taken together, these findings suggest that the method’s gains are attributed less to raw parameter count than to explicit modality partitioning inside the FFN substructure.

6. Experimental scope, metrics, and reported results

The evaluation spans public Amazon small catalogs—Games, Instruments, Arts, Sports, Beauty, Toys—public Amazon large 2023 catalogs—Beauty(23), Books(23), Toys(23)—and a proprietary industrial-scale dataset with hundreds of millions of users and tens of thousands of items (Shirkavand et al., 30 Sep 2025).

The comparison set includes classical sequential recommenders, LLM-based recommender frameworks, and ID-centric variants. The baselines named in the source are GRU4Rec, BERT4Rec, SASRec, HSTU, P5-CID, VIP5, E4SRec, ReAT, ID Transformer, Item-LLM, and Text-Attr LLM (Shirkavand et al., 30 Sep 2025). Recommendation is measured with HR@10, NDCG@10, and MRR, while language retention is measured with negative log-likelihood on WikiText and zero-shot benchmarks BBH, HellaSwag, MMLU, and WinoGrande (Shirkavand et al., 30 Sep 2025).

The key gains reported in the source are summarized below.

Setting	Comparison	Reported result
Amazon-Arts	vs. ID-only	HR@10 +44.1%, NDCG@10 +28.1%
Proprietary industrial dataset	vs. SASRec baseline	NDCG@10 +27.1%, HR@10 +16.6%, MRR +31.2%
Language retention	backbone/text-bias methods	Matches backbone NLL and downstream scores; Text-bias methods degrade significantly

The source states that language understanding is preserved: IDIOMoE matches the backbone model’s NLL and downstream scores, whereas text-bias methods degrade significantly (Shirkavand et al., 30 Sep 2025). This is a central claim because the architecture is motivated not only by recommendation quality but by joint retention of language competence.

A plausible implication is that the method is intended as a unified recommender-LLM rather than a recommender retrofitted with text interfaces. The preservation of backbone language metrics is therefore treated as part of the primary success criterion, not as an auxiliary diagnostic.

7. Interpretability, deployment guidance, and broader significance

The source includes an FFN key-value memory analysis following the treatment of FFNs as key-value memory. For each neuron $\ell$5, it defines:

$\ell$6

and

$\ell$7

together with the fraction of “clustered” rows with purity $\ell$8 (Shirkavand et al., 30 Sep 2025). According to the reported analysis, item experts show higher item-text affinity, higher category purity, and more clustered neurons, which the authors interpret as evidence of clearer modality separation and more interpretable representations.

For production integration, the source recommends starting from a moderately sized pretrained decoder-only LLM with 0.5–2.5B parameters, extending the tokenizer with item-ID tokens, using frozen text embeddings plus trainable item embeddings, replacing each FFN with two experts routed by static token type, and placing experts in the top-half or last 8–12 layers while sharing attention and norm layers (Shirkavand et al., 30 Sep 2025). It further recommends tuning item-expert width according to catalog scale, warming up on item-only sequences for approximately 20% of training, freezing the text expert and attention layers on very large datasets to preserve language skills and reduce training cost, optionally adding a small ranking loss on item tokens, maintaining a 1,024-token context, and monitoring both recommendation metrics and language NLL on a held-out text corpus (Shirkavand et al., 30 Sep 2025).

In aggregate, the framework is presented as a principled disentanglement of collaborative and semantic processing. The significance assigned to it in the source lies in the claim that recommendation can be performed within a single language-modeling paradigm without sacrificing the pretrained model’s text understanding. This suggests a line of work in which catalog structure is not external to the LLM but incorporated as a first-class tokenized modality, with expert specialization used to keep collaborative and linguistic functions from interfering destructively.