L2Rec: Dual-View LLM Adaptation for Recommendations

• The paper introduces a novel parameter-efficient dual-view framework that jointly models user behavior and item semantics for personalized recommendation.
• The methodology applies view-specific low-rank perturbations to a frozen LLM backbone, enabling adaptive cross-view fusion and significant metric improvements.
• Empirical results across diverse datasets show L2Rec outperforms baselines, achieving notable gains in offline metrics and online engagement rates.

Searching arXiv for the specified paper and closely related recommendation baselines. L2Rec is a parameter-efficient framework for adapting LLMs to personalized recommendation by jointly modeling behavioral and semantic signals within a shared Transformer parameter space. It is introduced in "L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation" (Pan et al., 26 May 2026). The central premise is that personalized recommendation requires simultaneous access to users’ interaction behavior and to item semantics, but prior approaches typically combine these signals either at the input level or at the output level, which can induce distributional mismatch or weaken end-to-end task supervision. L2Rec addresses this by applying view-specific low-rank perturbations to a frozen LLM backbone, producing complementary behavioral and semantic adaptations for each user and then integrating them through adaptive cross-view fusion (Pan et al., 26 May 2026).

1. Problem setting and motivation

L2Rec is designed for the setting in which a recommendation model must align a general-purpose LLM with user-specific preference signals. The motivating problem is twofold. First, there is a task-alignment issue: LLMs are pre-trained with next-token prediction or denoising objectives, whereas recommendation directly concerns whether a user will click an item. Second, there is a signal-integration issue: effective personalization requires both behavioral signals, such as click or view sequences, and semantic signals, such as item textual descriptions (Pan et al., 26 May 2026).

The framework is positioned against two broad classes of prior methods. One class fuses behavioral and semantic information at the input level, for example by injecting ID embeddings into token embeddings. The other class aligns separate encoders at the output level, for example with contrastive objectives. In the formulation underlying L2Rec, these alternatives are associated respectively with distribution gaps and with a lack of end-to-end task supervision. L2Rec therefore relocates the integration point from representations to parameters: both views are processed by the same frozen Transformer backbone, but with distinct low-rank, personalized perturbations layered onto that backbone (Pan et al., 26 May 2026).

This design implies that behavioral and semantic personalization are not treated as heterogeneous embedding spaces to be reconciled post hoc. Instead, they become coordinated adaptations of the same pretrained model. The paper frames this as unification in a shared parameter space, with the frozen backbone acting as an anchor that keeps both views near the LLM’s pretrained manifold (Pan et al., 26 May 2026).

2. Dual-view inputs and shared-backbone parameterization

L2Rec starts from a frozen Transformer-based LLM with weight matrices $\{W_j\}_{j=1}^J$. For a user $u$ with interaction sequence $\{(i_1,c_1),\dots,(i_L,c_L)\}$, the model constructs two distinct input views (Pan et al., 26 May 2026).

The semantic view preserves token-level text:

$$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$

This view retains the full textual content associated with the interacted items. By construction, it is suitable for exploiting the LLM’s native capability over token sequences.

The behavioral view compresses each item into a single vector after pooling and projection:

$$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$

Here, each interacted item is treated as one “ID” vector. The use of pooling and a small projector $P_U$ yields a behavior-oriented sequence representation, distinct from the token-preserving semantic stream (Pan et al., 26 May 2026).

The architectural significance of these two views is that they enter the same LLM backbone but are adapted differently at the parameter level. This is the sense in which L2Rec implements “dual-view understanding”: the semantic and behavioral streams are not handled by separate backbones, nor are they merely concatenated at the input; they are processed through shared pretrained parameters augmented by different low-rank expert combinations.

3. Dual-view Personalized Mixture-of-Experts

The core adaptation mechanism is the Dual-view Personalized Mixture-of-Experts, abbreviated DPMoE. At each layer $j$, L2Rec adapts the frozen weight matrix $W_j$ separately for each view $v\in\{\mathrm{sem},\mathrm{beh}\}$ through a low-rank mixture of experts:

$$W_{j}^{\prime\,(v)} \;=\; W_j \;+\;\Delta W_{j}^{(v)}, \quad \Delta W_{j}^{(v)} \;=\; \sum_{i\in\mathcal{E}_{sh}} B_{j,i}\,A_{j,i} \;+\; \sum_{i\in\mathcal{E}_{v}} g_{j,i}^{(v)}\; B_{j,i}\,A_{j,i}.$$

In this formulation, $u$0 is a small set of shared experts that are always active, and $u$1 is the view-specific expert pool for either the semantic or behavioral pathway. Each expert is parameterized by low-rank matrices

$u$2

Thus, DPMoE combines a shared adaptation component with a routed, view-specific component (Pan et al., 26 May 2026).

Routing is personalized and conditioned on three signals: context $u$3, user embedding $u$4, and interaction $u$5. For each signal $u$6, routing scores are computed as

$u$7

These are fused and sparsified via Top-$u$8 selection:

$u$9

The model thereby instantiates two parallel adaptation pathways at every layer:

$\{(i_1,c_1),\dots,(i_L,c_L)\}$0

The semantic input is processed with $\{(i_1,c_1),\dots,(i_L,c_L)\}$1, and the behavioral input with $\{(i_1,c_1),\dots,(i_L,c_L)\}$2, producing two view-specific hidden outputs $\{(i_1,c_1),\dots,(i_L,c_L)\}$3 (Pan et al., 26 May 2026).

A common misconception would be to view DPMoE as a standard MoE replacement for the backbone. The framework instead keeps the main LLM frozen and adds low-rank perturbations, so the expert mechanism is an adaptation layer over the pretrained model rather than a wholesale reparameterization of the base network.

4. Cross-view fusion and optimization objectives

Once semantic and behavioral hidden states have been produced in the same parameter space, L2Rec applies a lightweight gated fusion module. Each view first passes through a residual projection:

$\{(i_1,c_1),\dots,(i_L,c_L)\}$4

A gating scalar is then computed from the concatenated view representations:

$\{(i_1,c_1),\dots,(i_L,c_L)\}$5

The fused user representation is

$\{(i_1,c_1),\dots,(i_L,c_L)\}$6

This mechanism adaptively weights behavioral and semantic evidence rather than imposing a fixed combination rule (Pan et al., 26 May 2026).

Training keeps the backbone weights $\{(i_1,c_1),\dots,(i_L,c_L)\}$7 frozen and updates only the LoRA experts $\{(i_1,c_1),\dots,(i_L,c_L)\}$8, the routers, and the fusion parameters. The total loss is

$\{(i_1,c_1),\dots,(i_L,c_L)\}$9

The recommendation loss is a contrastive objective:

$$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$0

To align the two views for the same user while separating distinct users, L2Rec introduces Bidirectional Preference Contrastive loss:

$$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$1

The third component, $$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$2, is a standard entropy-based load-balancing regularizer used to prevent a small subset of experts from dominating. Gradients flow only into the DPMoE expert parameters and fusion weights; the main LLM remains frozen (Pan et al., 26 May 2026).

This training design clarifies that L2Rec is neither full-model fine-tuning nor simple prompt conditioning. Its optimization target combines recommendation, cross-view consistency, and expert utilization regularization within a frozen-backbone PEFT regime.

5. Experimental configuration and empirical results

The experimental study covers four datasets. The industrial social platform dataset contains 42 M interactions from 1.56 M users over 4 months, with 3 M for pre-train and 1 M for fine-tune, and items have rich textual/content attributes. The Amazon Review domain-transfer setting includes Scientific with 11 k users, 5 k items, and 77 k interactions; Instruments with 27 k users, 10 k items, and 231 k interactions; and Arts with 56 k users, 23 k items, and 492 k interactions. For Amazon, the protocol pre-trains on 7 source categories and fine-tunes on target categories. Evaluation uses leave-one-out and reports Recall@10 and NDCG@10 (Pan et al., 26 May 2026).

The baselines are grouped into three categories: ID-based methods, namely SASRec and BERT4Rec; text-enhanced methods, namely S$$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$3-Rec, UniSRec, and RecFormer; and LLM-based methods, namely LLaRA (Llama2-7B) and LEARN (Baichuan2-7B). The main backbone is Qwen3-0.6B, and scaling experiments also use Llama2-7B, Baichuan2-7B, and Qwen3-8B (Pan et al., 26 May 2026).

A concise summary of the reported setup is given below.

Component	Reported configuration	Source
Backbone	Qwen3-0.6B; also scaled to Llama2-7B, Baichuan2-7B, Qwen3-8B	(Pan et al., 26 May 2026)
DPMoE	9 experts per layer (1 shared + 8 view-specific), LoRA rank $$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$4, $$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$5, Top-2 experts per view	(Pan et al., 26 May 2026)
Updated parameters	$$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$6 M $$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$7 of backbone	(Pan et al., 26 May 2026)
Optimizer	AdamW; lr$$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$8 for pre-train, $$E_u^S = \mathrm{Embed}\bigl([c_1; c_2; \dots; c_L]\bigr)\;\in\;\mathbb{R}^{T\times d}.$$9 for fine-tune	(Pan et al., 26 May 2026)
Max lengths	industrial $$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$0, Amazon $$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$1	(Pan et al., 26 May 2026)

Across all four datasets, L2Rec achieves the best N@10 and R@10. The paper highlights gains over the strongest baseline, LEARN, in N@10: Scientific improves from 0.1060 to 0.1145 (+8.02%), Instruments from 0.0878 to 0.0912 (+3.87%), Arts from 0.1252 to 0.1318 (+5.27%), and Industrial from 0.1133 to 0.1198 (+5.74%) (Pan et al., 26 May 2026).

The online A/B test is conducted on a setting with 1.5 M DAU and 6% traffic, comparing treatment against a DLRM baseline. The reported gains are +9.24% click-through rate ($$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$2) and +3.15% reply rate ($$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$3) (Pan et al., 26 May 2026).

These findings support the paper’s claim that the gains derive from the adaptation design rather than merely from model scale. The backbone-scaling analysis reports that L2Rec’s improvements persist for larger backbones, including Llama2-7B, Baichuan2-7B, and Qwen3-8B (Pan et al., 26 May 2026).

6. Ablations, sensitivity, and low-resource behavior

The ablation study isolates the contributions of adaptation, dual-view modeling, bidirectional alignment, and personalization in routing. Removing adaptation and using the backbone only causes a 39% drop in N@10 on the Industrial dataset. Removing the semantic view yields drops of 11.7% on Scientific, 8.6% on Arts, and 17.9% on Industrial. Removing the behavioral view yields drops of 7.2%, 6.1%, and 11.9% on the corresponding datasets. Removing the BPC loss yields degradations ranging from 5.6% to 10.4%, and removing personalization in the router reduces performance by 2.0% to 2.8% (Pan et al., 26 May 2026).

These ablations indicate that the framework’s effectiveness is not attributable to a single component. Both views matter, explicit cross-view alignment matters, and personalized routing matters. The particularly large degradation without adaptation underscores that the frozen backbone alone is insufficient for the recommendation task under the reported setup.

The low-resource analysis examines training ratios of 1%, 5%, 10%, 50%, and 100%. L2Rec outperforms LEARN by large margins throughout this range; a highlighted example is that at 10% data on Scientific, L2Rec already beats LEARN at 100% (Pan et al., 26 May 2026). This suggests that the parameter-level dual-view design is especially useful when supervision is limited, although the paper reports this as an empirical observation rather than as a formal sample-efficiency theorem.

Hyperparameter sensitivity analysis shows that gains with LoRA rank $$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$4 saturate beyond approximately $$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$5, while gains from increasing Top-$$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$6 experts flatten beyond approximately $$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$7. The paper therefore recommends moderate settings such as $$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$8–16 and $$E_u^B = P_U\bigl([\mathrm{pool}(\mathrm{Embed}(c_i))]_{i=1}^L\bigr)\;\in\;\mathbb{R}^{L\times d}.$$9–4 for favorable cost-benefit trade-offs (Pan et al., 26 May 2026).

7. Deployment characteristics and interpretive significance

L2Rec is explicitly presented as a practical PEFT-style recommendation architecture. The parameter overhead is approximately

$P_U$0

extra parameters, which is about 5% of a 600 M backbone. The compute overhead consists of routing networks per layer plus two expert matrix multiplications of rank $P_U$1, and is described as still far cheaper than full fine-tuning (Pan et al., 26 May 2026).

The deployment guidelines are correspondingly modular. The LLM is frozen and served together with small adapter weights. User embeddings $P_U$2 are precomputed offline at periodic intervals. Per-layer LoRA experts $P_U$3 and router weights are stored in a key-value store. Adapter rank $P_U$4 and Top-$P_U$5 are chosen to maintain acceptable latency, with $P_U$6 and $P_U$7 given as an example. The fused user and item embeddings, $P_U$8 and $P_U$9, are integrated into downstream matching and ranking pipelines as plug-and-play embeddings (Pan et al., 26 May 2026).

In interpretive terms, L2Rec occupies a specific point in the design space of LLM-based recommendation. It does not simply append recommendation heads to an LLM, and it does not rely solely on textual prompting or standalone sequential encoders. Its defining claim is that behavioral and semantic personalization should be unified at the parameter level of a shared pretrained Transformer. A plausible implication is that this parameter-level coupling reduces representational misalignment relative to approaches that reconcile views only at the input or output stage. That implication is consistent with the framework’s use of shared experts, view-specific experts, and bidirectional preference contrastive alignment, all of which are directed toward coordinated dual-view adaptation (Pan et al., 26 May 2026).

Within the scope of the reported results, L2Rec is characterized by three properties: a frozen-backbone adaptation regime, dual-view personalized low-rank expert perturbations, and adaptive fusion of aligned behavioral and semantic representations. The combination yields reported improvements in offline recommendation metrics and in online engagement metrics while maintaining modest additional parameter and compute cost (Pan et al., 26 May 2026).