Paper deep dive
Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations
Yunjia Xi, Menghui Zhu, Jianghao Lin, Bo Chen, Ruiming Tang, Yong Yu, Weinan Zhang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 99%
Last extracted: 4/26/2026, 9:56:27 PM
Summary
The paper identifies the 'Mid-layer Representation Advantage' (MRA) in LLM-based recommendation systems, where middle layers outperform final layers due to the final layers specializing in proxy training tasks (Task Adaptation Module) rather than general representation learning. To address this, the authors propose Modular Representation Compression (MARC), a framework that explicitly controls LLM modularity. MARC uses 'Modular Adjustment' to introduce external compression and user-item matching networks, and 'Modular Task Decoupling' using the Hilbert-Schmidt Independence Criterion (HSIC) to separate representation learning from task adaptation. MARC achieved a 2.82% eCPM lift in a large-scale commercial search advertising A/B test.
Entities (8)
Relation Signals (5)
MARC → addresses → Mid-layer Representation Advantage
confidence 100% · Extensive experiments validate that MARC addresses MRA and produces efficient representations.
MARC → comprises → Modular Adjustment
confidence 100% · First, Modular Adjustment explicitly introduces compression and task adaptation modules...
MARC → comprises → Modular Task Decoupling
confidence 100% · Next, to ensure that each module performs its specific task, we introduce Modular Task Decoupling...
Modular Task Decoupling → uses → HSIC
confidence 100% · The information constraint HSIC is leveraged to maximize the mutual information between original and compressed representations
LLM → has → Mid-layer Representation Advantage
confidence 90% · we identify a counterintuitive phenomenon during representation compression: Mid-layer Representation Advantage (MRA), where representations from middle layers of LLMs outperform those from final layers
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recently, large language models (LLMs) have advanced recommendation systems (RSs), and recent works have begun to explore how to integrate LLMs into industrial RSs. While most approaches deploy LLMs offline to generate and pre-cache augmented representations for RSs, high-dimensional representations from LLMs introduce substantial storage and computational costs. Thus, it is crucial to compress LLM representations effectively. However, we identify a counterintuitive phenomenon during representation compression: Mid-layer Representation Advantage (MRA), where representations from middle layers of LLMs outperform those from final layers in recommendation tasks. This degraded final layer renders existing compression methods, which typically compress on the final layer, suboptimal. We interpret this based on modularity theory that LLMs develop spontaneous internal functional modularity and force the final layer to specialize in the proxy training task. Thus, we propose \underline{M}odul\underline{a}r \underline{R}epresentation \underline{C}ompression (MARC) to explicitly control the modularity of LLMs. First, Modular Adjustment explicitly introduces compression and task adaptation modules, enabling the LLM to operate strictly as a representation-learning module. Next, to ground each module to its specific task, Modular Task Decoupling uses information constraints and different network structures to decouple tasks. Extensive experiments validate that MARC addresses MRA and produces efficient representations. Notably, MARC achieved a 2.82% eCPM lift in an online A/B test within a large-scale commercial search advertising scenario.
Tags
Links
- Source: https://arxiv.org/abs/2604.18146v2
- Canonical: https://arxiv.org/abs/2604.18146v2
Trouble viewing inline? Open PDF directly →
Full Text
74,839 characters extracted from source content.
Expand or collapse full text
Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations Yunjia Xi xiyunjia@sjtu.edu.cn Shanghai Jiao Tong University Shanghai, China Menghui Zhu zhumenghui1@huawei.com Huawei Noah’s Ark Lab Shanghai, China Jianghao Lin ∗ linjianghao@sjtu.edu.cn Antai College of Economics and Management, Shanghai Jiao Tong University Shanghai, China Bo Chen chenbo116@huawei.com Huawei Noah’s Ark Lab Shanghai, China Ruiming Tang tangruiming@huawei.com Huawei Noah’s Ark Lab Shenzhen, China Yong Yu yyu@sjtu.edu.cn Shanghai Jiao Tong University Shanghai, China Weinan Zhang wnzhang@sjtu.edu.cn Shanghai Jiao Tong University Shanghai, China Abstract Recently, large language models (LLMs) have advanced recom- mendation systems (RSs), and recent works have begun to explore how to integrate LLMs into industrial RSs. While most approaches deploy LLMs offline to generate and pre-cache augmented repre- sentations for RSs, high-dimensional representations from LLMs introduce substantial storage and computational costs. Thus, it is crucial to compress LLM representations effectively. However, we identify a counterintuitive phenomenon during representation com- pression: Mid-layer Representation Advantage (MRA), where representations from middle layers of LLMs outperform those from final layers in recommendation tasks. This degraded final layer renders existing compression methods, which typically compress on the final layer, suboptimal. We interpret this based on modular- ity theory that LLMs develop spontaneous internal functional modularity and force the final layer to specialize in the proxy train- ing task. Thus, we proposeModularRepresentationCompression (MARC) to explicitly control the modularity of LLMs. First, Modular Adjustment explicitly introduces compression and task adaptation modules, enabling the LLM to operate strictly as a representation- learning module. Next, to ground each module to its specific task, Modular Task Decoupling uses information constraints and differ- ent network structures to decouple tasks. Extensive experiments validate that MARC addresses MRA and produces efficient represen- tations. Notably, MARC achieved a 2.82% eCPM lift in an online A/B test within a large-scale commercial search advertising scenario. ∗ Corresponding authors. Please use nonacm option or ACM Engage class to enable C licenses This work is licensed under a Creative Commons Attribution 4.0 International License. SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2599-9/2026/07 https://doi.org/10.1145/3805712.3809681 CCS Concepts • Information systems→ Recommender systems. Keywords Recommender Systems; Large Language Models; Representation Compression ACM Reference Format: Yunjia Xi, Menghui Zhu, Jianghao Lin, Bo Chen, Ruiming Tang, Yong Yu, and Weinan Zhang. 2026. Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26), July 20–24, 2026, Melbourne, VIC, Australia. ACM, New York, NY, USA, 12 pages. https://doi.org/10.1145/3805712.3809681 1 Introduction The rapid advancement of large language models (LLMs) has pro- foundly transformed numerous domains, including recommenda- tion systems (RSs) [31,71,78]. Recent studies have increasingly ex- plored the integration of LLMs into RSs, achieving performance that surpasses traditional recommendation approaches [6,32,35,59]. However, their deployment in industrial RSs remains at a nascent stage. This is primarily because commercial RSs typically handle millions of users and items while adhering to stringent response latency requirements – often within 100ms [30,33,65]. The sub- stantial computational overhead and significant inference latency of LLMs render them impractical for online serving in such sce- narios. Consequently, most deployment strategies pre-cache LLM representations for traditional RSs, avoiding their online infer- ence [37,47,61,66]. A prevalent pipeline entails leveraging LLMs offline to generate or encode textual knowledge into knowledge representations, which are subsequently utilized by downstream recommendation models. The LLM representations can inject richer information into RSs and significantly enhance their performance [37,47,61], but there arXiv:2604.18146v2 [cs.IR] 21 Apr 2026 SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYunjia Xi et al. 01000200030004000 Dimension of LLM Representation 0.782 0.784 0.786 0.788 0.790 0.792 0.794 0.796 0.798 Downstream Performance (AUC) Base TinyBERT BERT Qwen2 -1.5B Phi-2 Qwen2-7B Llama3 -8B PCA-128 PCA-768 PCA-2048 DCNv2 DCNv2 + LLM Rep. DCNv2 + PCA Rep. (a) Performance 01000200030004000 Dimension of LLM Representation 1x 10x 20x 30x 40x N times Larger than Base DCNv2 Training Time Inference Time Storage Space (b) Costs Figure 1: The impact of various LLM representations on downstream performance and costs in MovieLens-1M. is also a trade-off between effectiveness and efficiency. In Fig- ure 1a, larger LLMs bring greater enhancements to the downstream Click-Through Rate (CTR) model DCNv2 [55] (e.g., BERT [21] < Phi-2 [17] < Llama3-8B [5]). However, larger LLMs produce higher- dimensional representations, substantially increasing storage, train- ing, and inference overhead (Figure 1b), as detailed in Section 2.2. In industrial RSs, where user and item volumes far exceed public datasets, these costs are more prohibitive. While simple dimen- sionality reduction methods like PCA [41] can reduce costs, they suffer from severe performance degradation. In Figure 1a, when PCA reduces Llama3-8B’s representation to BERT’s dimensional- ity (PCA-768), it performs worse than BERT. This suggests that training-free compression methods, such as PCA, struggle to retain information useful for downstream tasks, making fine-tuning LLMs essential for efficient representation compression. Existing fine-tuning-based compression methods, including nested- based [22,56] and projection-based [53] approaches, have signif- icantly outperformed training-free methods. These methods typ- ically compress on the final-layer output of LLMs with training objectives like contrastive loss. However, in our experiments 2.3, we identify a pervasive and critical issue: Mid-layer Representa- tion Advantage (MRA), where representations from middle lay- ers outperform those from the final layer in recommendation tasks. This observation reveals a fundamental limitation: existing training paradigms for representation compression degrade the quality of final-layer representations. Consequently, methods that focus on compressing this suboptimal final layer are bound to limit their potential effectiveness. We interpret this phenomenon through the lens of modularity theory that LLMs develop spontaneous internal functional mod- ularity during training. Specifically, the early-to-middle layers function as a Representation Learning Module, extracting rich, gen- eralized textual features. However, a critical functional shift occurs in the final layers. Under current training paradigms, the strong su- pervision signals from the training objective (e.g., contrastive loss) force the final layers to evolve into a specialized Task Adaptation Module. These final layers form an unintended information bot- tleneck, filtering out diverse semantics deemed redundant for the training objective but crucial for recommendation. This explains the MRA: while middle layers retain generalized information, the final layer sacrifices it for task adaptation. This emergent modular- ity is difficult to predict a priori, making it challenging to directly compress the optimal middle-layer representations. Based on this insight, we propose aModularRepresentation Compression (dubbed MARC) framework to transform this implicit, uncontrolled modularity into an explicit, controllable architectural design. Our core intuition is to preserve the high-quality represen- tations in the final layer by structurally decoupling representation learning from task adaptation. First, we perform Modular Adjust- ment by explicitly introducing external modules responsible for different tasks. We introduce a lightweight Compression Network to act as the dedicated compression module and a User-Item Match- ing Network to offload the task adaptation burden. This explicit separation prevents the LLM’s final layers from collapsing into task-specific heads, allowing the backbone to focus purely on learn- ing high-quality, generalized representations. Next, to ensure that each module performs its specific task, we introduce Modular Task Decoupling, including information constraint, Hilbert-Schmidt In- dependence Criterion (HSIC), and different network structures for different tasks. HSIC constraint guides compression by maximizing the mutual information between original and compressed repre- sentations, effectively separating representation and compression tasks. The compression network processes individual represen- tations, while the matching network inputs both user and item representations, capturing their explicit and implicit interactions. These distinct structures separate compression and target tasks. This framework ensures that information filtering occurs strictly within the external modules, leaving the LLM’s representation ca- pability intact. Our contributions can be summarized as follows: • We identify the tradeoff between efficiency and effectiveness in LLM-based representations for recommendation. To the best of our knowledge, this is the first work that focuses on representation compression in LLM-based recommendation. •We reveal the mid-layer representation advantage issue in common LLM representation compression methods, and we analyze and explain it with modular theory for the first time. •We propose MARC to address this issue, with Modular Ad- justment to add compression and user-item matching net- work and Modular Task Decoupling to ensure each module performs its specific task, enabling LLMs to focus on repre- sentation learning. •Experiments validate that MARC mitigates mid-layer repre- sentation advantage and improves eCPM by 2.82% during an online A/B test in a large commercial search advertising scenario. 2 Analyses on LLM Representations for RSs 2.1 Background: LLM Representations for RSs LLMs’ massive parameters and autoregressive nature bring high inference latency, making it challenging to meet the online low- latency requirements of RSs. As a result, currently deployable LLM- based recommendation solutions typically leverage LLMs offline to generate representations for users and items, which can then be utilized in traditional recommendation tasks [12,37,47,61,73]. We refer to this approach as “LLM Representations for RSs”. Given the item setIand the user setU, each item푖 ∈Iand user 푢 ∈Upossesses specific attributes and descriptions, which can be converted into text푇 푖 and푇 푢 through predefined templates. In some Modular Representation Compression: Adapting LLMs for Efficient and Effective RecommendationsSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia work [37], the text can incorporate LLM-generated knowledge to infer user preferences. Subsequently,푇 푖 and푇 푢 are transformed into representations 푟 푖 and 푟 푢 , with dimension 푑 표 , through LLMs: 푟 푖 = Pooling(LLM(푇 푖 )), 푟 푢 = Pooling(LLM(푇 푢 )),(1) where LLMs may be pre-trained or fine-tuned models, and pooling methods like mean pooling or EOS token pooling are used to obtain the final representations. Then, the user and item representations can be utilized as static features in any downstream models, e.g., click-through rate (CTR) prediction models. 2.2 Tradeoff between Performance and Costs We first empirically investigate the impact of LLM scaling on rec- ommendation performance on MovieLens-1M and Yelp datasets following [61]. First, item-related features (e.g., titles) and user- related features (e.g., the most recent history) are converted to text using templates. Next, these texts are encoded into representations by LLMs, and then, the representations are incorporated into the downstream CTR model DCNv2 [55] (Base). The LLMs used in this study are pre-trained models without fine-tuning, including a range of models from small to large: TinyBERT [20], BERT [21], Phi-2 [17], Qwen2-1.5B/7B [15], and Llama3-8B [5]. Additionally, for Llama3-8B, its 4096-dimensional representations are reduced to 128, 768, and 2048 dimensions with PCA, producing PCA-128, PCA-768, and PCA-2048, respectively. The results on two datasets are presented in Figure 1 and Figure 2. 01000200030004000 Dimension of LLM Representation 0.731 0.732 0.733 0.734 0.735 0.736 0.737 0.738 0.739 Downstream Performance (AUC) base tinyBERT BERT Qwen2 -1.5B Phi-2 Qwen2-7B Llama3 -8B PCA-128 PCA-768 PCA-2048 DCNv2 DCNv2 + LLM Rep. DCNv2 + PCA Rep. (a) Performance 01000200030004000 Dimension of LLM Representation 1x 30x 60x 90x 120x N times Larger than Base DCNv2 Training Time Inference Time Storage Space (b) Costs Figure 2: The impact of different sizes of LLMs on down- stream DCNv2 performance and costs in Yelp. As illustrated in Figure 2a, there is a clear positive correlation between model size and downstream performance. Larger LLMs consistently yield higher performance, but this comes at a steep cost. In Figure 2b, the storage, training, and inference overheads (which are expressed as multiples of base DCNv2) grow linearly with the rep- resentation dimension. For instance, deploying 4096-dimensional representations from Llama3-8B incurs a 140× increase in storage and 20×increase in training time compared to base settings. While simple dimension-reduction methods like PCA can reduce costs, they suffer from significant performance degradation. Thus, we need advanced representation compression techniques tailored to recommendations. 2.3 Mid-layer Representation Advantage (MRA) During Compression To achieve effective compression, recent works employ fine-tuning- based compression methods (e.g., MRL [22], LARR [53]), which all require fine-tuning on specific proxy tasks, such as contrastive loss, to achieve effective compression. Thus, we fine-tune Llama3- 8B, Qwen2-7B, and Qwen2-1.5B on the MovieLens-1M and Yelp dataset and explore the layer-wise performance of various proxy target tasks, including both compression and non-compression ap- proaches: (1) Gen employs next-token prediction to predict the next positive item based on a user’s history; (2) CL uses in-batch nega- tive samples for contrastive learning; (3) CS uses cosine similarity loss between user and item representations; (4) MRL [22], nested- based compression with cross-entropy (CTR) loss; (5) LARR [53], projection-based compression with contrastive learning; (6) Ours, our proposed MARC. All the above models utilize the same click data for training. Then, we encoded user and item texts into repre- sentations with the fine-tuned LLMs. Lastly, representations from different layers of LLMs are applied to the downstream DCNv2. With the exception of our proposed MARC, the representa- tions that yield the best downstream performance consis- tently originate from the middle-to-later layers rather than the final layer. We term this phenomenon the Mid-layer Represen- tation Advantage. Specifically, while performance steadily increases in the early layers, it peaks in a middle layer and then drops as it approaches the final layer. Crucially, MRA persists even when base- lines are fine-tuned with the exact same recommendation-aligned loss (CTR loss for MRL), confirming MRA is a structural character- istic of LLMs in RS rather than an objective misalignment issue. This highlights a fundamental deficiency in current compression methods: their training paradigms degrade the quality of final- layer representations. Moreover, layer-wise selection in Figure 3 is impractical in real-world scenarios as the optimal layer varies across models and tasks. Consequently, existing compression methods, which operate on the final layer, are inherently suboptimal. In contrast, our MARC is designed to prevent this degeneration, and its final representation significantly outperform its middle layers and even the best-performing intermediate layers of all baselines. 2.4Understanding MRA: A Modular Perspective Why does the final layer, theoretically the most “processed” one, fail to yield the optimal performance? We interpret it within the framework of Functional Modularity and Layer-wise Special- ization [69,74]. Recent studies on the interpretability of Transform- ers have observed emergent modularity [69,74], where pre-trained LLMs spontaneously differentiate into functional regions special- ized in different tasks during training. Building on these findings, we argue that during the previous fine-tuning process for representation compression, the LLM spon- taneously develops two internal functional modules: • Representation Learning Module (Early-to-Middle Lay- ers): The layers from the bottom up to the peak of MRA primarily focus on understanding and extracting rich, gen- eralized features, capturing diverse information potentially useful for various tasks. •Task Adaptation Module (Final Layers): These layers evolve to adapt these features to the training objective, e.g., contrastive loss, due to their proximity to the supervision signal. SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYunjia Xi et al. 048121620242832 Layer 0.792 0.794 0.796 0.798 0.800 0.802 Performance (AUC) (a) MovieLens-1M + Llama3-8B Gen CL CS MRL LARR Ours 0481216202428 Layer 0.790 0.792 0.794 0.796 0.798 0.800 (b) MovieLens-1M + Qwen2-7B Gen CL CS MRL LARR Ours 0481216202428 Layer 0.790 0.792 0.794 0.796 0.798 (c) MovieLens-1M + Qwen2-1.5B Gen CL CS MRL LARR Ours 048121620242832 Layer 0.739 0.740 0.741 0.742 (d) Yelp + Llama3-8B Gen CL CS MRL LARR Ours Figure 3: Mid-layer representation advantage across different datasets and LLMs. Crucially, in existing training paradigms, the boundary between these two modules is spontaneous and uncontrollable. Even when projection heads [53] are employed, they fail to shield the backbone. The strong supervision signals from the training loss inevitably propagate back to the LLM’s final layers, turning them into Task Adaptation Module. This aligns with the Information Bottleneck principle [51] that deep networks compress input data into a minimal sufficient statistic for the target. Thus, the final layers filter out information redundant for the proxy objective but vital for recommendations. 2468101214161820222426283032 Layer 0.2 0.3 0.4 0.5 Loss CL CS MRL LARR Figure 4: Layer-wise performance on the proxy target task. Evidence from Target Task Analysis. To empirically validate this functional partitioning, we analyze the layer-wise performance on the fine-tuning proxy task itself. We use the representations extracted from each layer to calculate the final optimization loss, which serves as an indicator of their ability to handle the proxy target task. As illustrated in Figure 4, the optimization loss on the proxy tasks (CL, CS) decreases consistently, reaching its minimum at the final layer. This provides compelling evidence that the later layers are highly specialized for the proxy training objective, function- ing as the Task Adaptation Module. However, this comes at a cost: the model progressively discards information deemed redundant for the proxy task but critical for recommendations. Difficulty in Aligning Proxy Targets with Recommendation. Real- world recommendation scenarios are multifaceted, encompassing diverse tasks such as retrieval, ranking, and re-ranking, which thrive on rich, multi-dimensional semantic information. In contrast, fine-tuning tasks are typically driven by single-objective proxies. Thus, useful representations for recommendations are difficult to perfectly align with any single proxy target. Crucially, our experi- ments (Figure 3) show that MRA persists even when domain-aligned objectives like CTR loss are used in MRL. Consequently, the sponta- neous modularity driven by single-objective supervision force the final layer to specialize in proxy target and lose rich information vi- tal for RSs. Compressing an already degraded final layer inevitably leads to suboptimal performance, necessitating a new framework that can explicitly control the modularity. 3 Modular Representation Compression 3.1 Overview Based on the above findings and analysis, we propose MARC to transform this implicit, uncontrolled modularity into an explicit, controllable architectural design. Existing nested-based and projection- based compression methods lack explicit constraints on functional regions, so under the supervision of training tasks, the later layers of LLMs are still involved in other functions, e.g., task adaptation, as shown in the left panel of Figure 5. This leads to MRA in down- stream tasks, as discussed in Section 2.4. MARC, in the right panel of Figure 5, introduces explicit modularity, specifically two ex- ternal modules for compression and task adaptation and enforcing clear modular boundaries. This ensures that LLM’s final layers retain their robust representation learning capabilities. Specifically, we make Modular Adjustment, retaining the LLM strictly as a representation learning module, while introducing two external lightweight modules: (1) a compression network to con- dense high-dimensional semantics, and (2) a user-item matching network to serve as the dedicated task adaptation module. This design effectively prevents the LLM’s final layers from collapsing into task-specific heads, thereby preserving the rich, generalized information (typical of middle layers) throughout the entire net- work depth. To ensure functional decoupling, we devise Modular Task Decoupling. The information constraint HSIC is leveraged to maximize the mutual information between original and com- pressed representations, improving the information density of the compressed representation. The compression network processes individual representations, while the matching network inputs both user and item representations and models their interactions. These distinct structures separate compression and target tasks while pre- serving essential interactive information for recommendations. The model pipeline includes three modules: a representation module with LLMs, a compression module with a compression network, and a task adaptation module via a user-item matching network, and these modules are trained in an end-to-end manner. Modular Representation Compression: Adapting LLMs for Efficient and Effective RecommendationsSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia Layer of LLM Layer of LLM Layer of LLM Projection Layer ... ... User TextItem Text Layer of LLM Layer of LLM ... User TextItem Text Compression Net User-item Matching Net User Rep.Item Rep. Contrastive Learning User Rep.Item Rep. Score Compress & Target Task Rep. Task Rep. Task Compress Task Target Task HSIC Contraint Figure 5: Comparison between existing projection-based com- pression methods (left) and MARC (right). Note that User and Item Representation (Rep.) are used for downstream tasks. 3.2 Representation Module This stage is handled by LLMs, fully leveraging their powerful repre- sentational capabilities. Similar to previous works, our model takes as input the text related to the user푢and the item푖. Each item푖 has its attributes (e.g., category) and description, which are trans- formed into text푇 푖 through templates. Each user푢also has specific attributes (e.g., age, gender) and historical behaviors퐻 푢 , which are also converted into text푇 푢 with templates. These texts are then encoded by LLMs. After pooling the representation from LLMs’ last layer, we obtain the user and item representations푟 푢 ∈R 푑 표 and 푟 푖 ∈R 푑 표 of dimension푑 표 , following Eq.(1). To distinguish repre- sentations generated in this stage from those in the compression stage, we refer to 푟 푢 and 푟 푖 as the original representations. Crucially, we deviate from standard fine-tuning paradigms, which typically impose task-specific losses directly on the LLM’s output representations푟 푢 and푟 푖 . Such direct supervision forces the LLM’s final layers to over-fit the specific target loss, turning the back- bone into a task-specific predictor. In contrast, MARC does not apply any direct supervision to the raw LLM representations 푟 푢 and푟 푖 at this stage. Instead, the task-specific loss is imposed strictly on the subsequent lightweight modules (i.e., the Compres- sion and Matching Networks). Consequently, the LLM backbone is optimized indirectly via backpropagation through these external modules. This architectural design ensures that the task adaptation pressure is primarily absorbed by the external modules, allowing the LLM backbone to maintain generalized, high-quality represen- tation without collapsing into a narrow, task-specific subspace. 3.3 Compression Module For compression, we introduce a dedicated compression network, 푔(·), which takes user and item representations,푟 푢 ∈R 푑 표 and 푟 푖 ∈R 푑 표 , and generates the compressed representations푐 푢 ∈R 푑 푐 and 푐 푖 ∈R 푑 푐 of dimension 푑 푐 (푑 푐 ≪ 푑 표 ), as follows: 푐 푖 =푔(푟 푖 ), 푐 푢 =푔(푟 푢 ).(2) For simplification, we take the user’s original representation푟 푢 and compressed representation푐 푢 as examples to introduce the information constraint. The same process can be applied to item representations. During compression, we aim to preserve as much of the information from the original representations as possible in the compressed ones, which means maximizing the mutual information (MI) between the original and compressed representations: max 휃 푀퐼(푟 푢 ,푐 푢 ),(3) where휃represents the model parameters, and푀퐼(·)denotes the mutual information between the two representations. Common mu- tual information optimization methods, such as MSE, InfoNCE [44], and DIM [10], require the two representations to have matching dimensions. However,푟 푢 and푐 푢 reside in spaces of vastly different conceptual levels and dimensions, and mapping them to the same dimension with a simple projection layer could distort the infor- mation. To ensure the Compression Network retains the maximum amount of information from the backbone, we employ the sampling form of mutual information, Hilbert-Schmidt Independence Cri- terion (HSIC) [8,39], as the constraint. It utilizes a kernel function to map vectors of different dimensions into a higher-dimensional Reproducing Kernel Hilbert Space (RKHS). This enables us to handle the dimension mismatch and maximize non-linear dependen- cies between the original and compressed views, effectively trans- ferring rich information to the lower-dimensional space without requiring dimension-matching projections. Mathematically, given a data batch퐷=(푥 1 ,푦 1 ), . . .,(푥 푛 ,푦 푛 )(푛denotes batch size) from two random variables 푋 and 푌 , the HSIC is defined as follows: L 퐻푆퐼퐶 = 퐻푆퐼퐶(푋,푌)= 1 (푛− 1) 2 푇푟(퐾 푋 퐽퐾 푌 퐽),(4) where퐾 푋 ∈R 푛×푛 and퐾 푌 ∈R 푛×푛 are the kernel matrices gener- ated by applying kernel functions to the samples from푋and푌. Here,each entry of 퐾 푋 is defined as: 퐾 푋 (푖, 푗)= 푘 푋 (푥 푖 ,푥 푗 ), 퐾 푌 (푖, 푗)= 푘 푌 (푦 푖 ,푦 푗 ),(5) where푘 푋 (·)and푘 푌 (·)are the kernel functions for푋and푌, respec- tively. Typically, a Gaussian kernel is employed for them, i.e., 푘 푋 (푥 푖 ,푥 푗 )= 푒 − ||푥 푖 −푥 푗 || 2 2휎 2 ,(6) where휎is a hyperparameter. The matrix퐽= 퐼 푛 − 1 푛 퐼 푛 퐼 푇 푛 is the centering matrix, which removes biases in the data by ensuring that the kernel matrices are centered. The operator푇푟(·)denotes the trace of a matrix, summarizing the correlation of variables in the higher-dimensional space. It is worth noting that the computational complexity of HSIC is푂(푛 2 )where푛is the batch size. This is negligible compared to computation of the LLM backbone. This constraint achieves two key objectives: preserving relevant information during compression and effectively decoupling tasks. First, it maximizes the mutual information between the compressed and original representations, ensuring that the compressed repre- sentations retain as much information as possible and improving information density. Second, it separates the compression and rep- resentation tasks, reinforcing the distinct functions of the compres- sion network and LLMs’ representation learning. This enhances the compression functionality of the compression network while maintaining the representational power of LLMs. SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYunjia Xi et al. 3.4 Task Adaptation Module We introduce the User-Item Matching Network as the dedicated Task Adaptation Module to serve as the designated sink for supervi- sion signals, effectively absorbing the optimization pressure from the training objective. Unlike simple projection heads [53], this network incorporates rich explicit and implicit interaction mech- anisms that are more aligned with downstream recommendation tasks, allowing it to naturally assume the adaptation burden and clearly distinguishing them from the compression phase. Specifically, we introduce both explicit and implicit inter- action. The input representations(푐 푢 ,푐 푖 )first undergo explicit manual feature interactions (|푐 푖 − 푐 푢 |and푐 푖 ⊙ 푐 푢 ) before passing through the interaction network푓(·)for implicit interaction. This network predicts whether user푢 likes item 푖, denoted as ˆ 푦, ˆ 푦= 푓(푐 푖 ,푐 푢 ,|푐 푖 −푐 푢 |,푐 푖 ⊙ 푐 푢 ),(7) where⊙denotes the element-wise product. This distinction in in- put and network structure helps better separate the matching and compression stages. These explicit and implicit interactions are better aligned with recommendation tasks, enabling representa- tions to retain more information relevant to recommendations and thereby improving representation quality. This structure ensures that the information bottleneck, filtering redundant information, and focusing on the target occur strictly within this lightweight external module, preventing the LLM representations from collaps- ing into task-specific logits. The matching task is designed with a cross-entropy loss function: L 푚푎푡푐ℎ =− ∑︁ (푖,푢,푦)∈퐷 ′ 푦 log ˆ 푦+(1−푦) log(1− ˆ 푦),(8) where퐷 ′ denotes the training data set and푦represents the ground truth label (typically whether the item푖is clicked or highly rated by user푢). The representation, compression, and task adaptation modules of MARC are optimized in an end-to-end manner with the final loss function: L=L 푚푎푡푐ℎ + 훼L 퐻푆퐼퐶 .(9) Upon completing the end-to-end training, we freeze the model to produce stable, low-dimensional representations, which are lever- aged as static features for downstream recommendation models. 4 Experiment To gain more insights into our proposed MARC, we aim to address the following research questions (RQs) in this section. •RQ1: How does our model, MARC, perform on various down- stream recommendation tasks? • RQ2: What is the online performance of MARC? • RQ3: What role does each module of MARC play? •RQ4: Is MARC compatible with downstream models and LLMs? • RQ5: How does our model perform with different sizes of compressed representations? 4.1 Setup 4.1.1 Datasets and Pipelines. The experiments are conducted on three public datasets, MovieLens-1M 1 , Yelp 2 , and MovieLens-25M 3 . MovieLens-1M (ML-1M for short) contains 1 million ratings pro- vided by 6000 users for 4000 movies. Yelp provides real-world data related to businesses, including about 8 million reviews, 209, 393 businesses, and 1, 968, 703 users. MovieLens-25M (ML-25M for short) has 25 million ratings applied to 62,000 movies by 162,000 users. Following [61,77], we convert the ratings into binary labels by labeling ratings of 4 and 5 as positive and the rest as negative and split data into training and test sets in a 9:1 ratio based on user IDs. As described in Section 2.2, our pipeline begins by leveraging a randomly sampled portion of the training set to fine-tune LLMs, improving their representation and compression capabilities. Then, the fine-tuned LLMs are employed to encode user behaviors and item titles to obtain representations. These representations are sub- sequently used for training the downstream recommendation tasks on the full training set following [61]. 4.1.2 Baselines and Downstream Models. In terms of baselines, we first select compression methods that do not fine-tune LLMs, such as PCA [41], Autoencoder (AE) [58], and Variational Autoencoder (VAE) [46]. Then, we choose several approaches fine-tuning LLMs, including two projection-based compression methods, LARR [53] and BAHE [7], and two nested-based approaches, ESE [56] and MRL [22]. LARR [53] applies an MLP to compress representations from LLM and then performs contrastive learning. BAHE [7] utilizes a linear layer for compression and incorporates a CTR head and cross-entropy loss. ESE [56] leverages AoE loss [25] and PCA-like constraints to guide the model in learning compact and efficient representations. MRL [22] employs nesting to concentrate criti- cal information in the early part of the representation. The above models, similar to our MARC, are based on Llama3-8B [52] and are trained on the same click-through data. Even for contrastive learning-based methods, e.g., LARR, we treat clicked user-item pairs as positive samples, injecting the same supervised signal. Addition- ally, we include smaller models such as TinyBERT [20], BERT [21], Qwen2-1.5B [15], Phi-2 (2.7B) [17], and Qwen2-7B [15] and utilize their original representations without compression. For downstream recommendation tasks, we primarily focus on CTR prediction in the ranking stage. The experiments involve sev- eral widely adopted CTR models, including DCNv2 [55], DCNv1 [54], DIN [77], DeepFM [9], and AutoInt [48]. Additionally, to demon- strate MARC’s compatibility, we also explore its application in other stages of recommendation, such as re-ranking and retrieval. For the re-ranking stage, we select the state-of-the-art model PRM [45], while for the retrieval stage, we employ the widely-used DSSM [13]. 4.1.3 Metrics. We establish the following evaluation metrics based on previous works [16,45,55,72,77]. For CTR prediction tasks, we adopt widely used AUC and Logloss as metrics. It is worth noting that in CTR prediction, an improvement of 0.001 in AUC is generally considered significant and impactful [54,55,77]. For re-ranking tasks, we use NDCG@K and MAP@K as metrics, where 1 https://grouplens.org/datasets/movielens/1m/ 2 https://w.kaggle.com/datasets/yelp-dataset/yelp-dataset 3 https://grouplens.org/datasets/movielens/25m/ Modular Representation Compression: Adapting LLMs for Efficient and Effective RecommendationsSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia Table 1: The overall performance on CTR prediction task. The best result is given in bold, while the second-best compression baseline is underlined. MARC-C and MARC-O represent thecompressed and theoriginal representations of our MARC. MARC-O is the upper bound of MARC-C. Rel. Impr denotes the relative AUC improvement of MARC-C against other models. The symbol * indicates statistically significant improvement over the best compression baselines (t-test with 푝< 0.05). Pre-trained Language Models Compression Method Dimension ML-1MML-25MYelp AUCLoglossRel.ImprAUCLoglossRel.ImprAUCLoglossRel.Impr base (DCNv2)//0.78330.55122.28%0.79450.54693.63%0.73110.52031.58% TinyBERT/3120.78880.54591.57%0.80790.53211.91%0.73330.51731.27% BERT /7680.79310.54181.01%0.81190.52761.41%0.73610.51590.90% Qwen2-1.5B/15360.79410.54100.88%0.81410.52531.14%0.73660.51530.82% Phi-2 /25600.79540.53900.73%0.82010.51780.40%0.73780.51450.67% Qwen2-7B/35840.79550.54020.71%0.82130.51530.25%0.73820.51410.61% Llama3-8B (frozen) PCA1280.78970.54491.45%0.81380.52221.18%0.73450.51751.12% AE1280.78930.54781.50%0.81450.52141.08%0.73290.51721.34% VAE1280.78400.55042.18%0.80940.52981.72%0.73230.51771.42% /40960.79610.53970.63%0.82150.51570.23%0.73840.51370.59% Llama3-8B (fine-tuned ) LARR1280.79110.54441.27%0.81790.52190.67%0.73540.51700.99% BAHE1280.79360.54080.95%0.81820.51910.63%0.73600.51500.91% ESE1280.79400.54060.90%0.81830.51930.61%0.73640.51470.85% MRL1280.79430.54040.86%0.81860.51880.58%0.73720.51660.74% MARC-C1280.8011* 0.5317*-0.8233* 0.5127*-0.7427* 0.5109*- MARC-O (oracle) 40960.8020*0.5321*-0.11%0.8254*0.5100*-0.25%0.7428*0.5105*-0.01% 퐾=1,5,10. For retrieval tasks, we utilize NDCG@K, HitRate@K, and MRR (Mean Reciprocal Rank) as metrics, where 퐾= 10, 20, 50. 4.1.4 Reproducibility. We randomly sample 200,000 data points from the training set to train the LLM-based representation model. Then, the user and item representations obtained from this LLM are used to train the downstream recommendation task on the full training data. Unless otherwise specified, MARC and compression baselines are based on Llama3-8B, and the downstream task is CTR prediction with base model DCNv2 [55]. MARC and compression baselines all produce 128-dimensional compressed representations. MARC-O denotes the original 4096-dimensional representations and MARC-C is the compressed ones. The compression network is implemented as an MLP with hidden layers of [256, 128], while the user-item matching network is an MLP with hidden layers of [128, 1]. The LLM module employs LoRA [11] for fine-tuning and 훼in Eq.(9)is set to 0.01 to balance the magnitude of losses and we found the model performance is generally robust around this order of magnitude. We follow [61], leveraging MoE to integrate representations into the downstream model. Each expert in the MoE is an MLP with hidden layer sizes of [128, 32], and the number of experts varies with different downstream tasks and models, typi- cally ranging from 2 to 5. The embedding size for the downstream CTR model is fixed at 32, and the output layer MLP size is [200, 80]. Other parameters, such as batch size and learning rate, are deter- mined with a grid search. Regarding computational overhead, the training time of MARC is comparable to current fine-tuning-based compression methods as it only introduces lightweight modules. 4.2 Overall Performance (RQ1) 4.2.1 Performance on CTR prediction task. We apply the represen- tations from MARC and various baselines to the downstream CTR prediction task on three public datasets, with results shown in Ta- ble 1. “Base” refers to DCNv2 without representation enhancement. The key findings are as follows: (i) Our MARC consistently and significantly outperforms other compression models. For example, on ML-1M and Yelp, MARC improves AUC by 0.89% and 0.74%, respectively, compared to the best compression baseline MRL. This improvement is due to MARC’s ability, as shown in Section 2.3, to alleviate mid-layer representation advantage and generate better representations. (i) MARC produces compressed representations with high information density. MARC-C incurs minimal accuracy loss compared to MARC-O and even outperforms 4096-dimensional representations from frozen Llama3-8B without fine-tuning. For example, on ML-1M, the 128-dimensional MARC-C improves by 0.63% over representations from frozen Llama3-8B, with only a 0.11% accuracy loss compared to MARC-O. This suggests that our modular adjustment and task decoupling help the model better retain key information during compression. (i) Fine-tuning LLMs for compression generally yields better results than non-fine-tuned methods. Fine-tuned approaches, e.g., LARR, MRL, and MARC, out- perform non-fine-tuned methods like PCA, AE, and VAE, likely because non-fine-tuned methods struggle to capture information useful for recommendations. 4.2.2 Performance on re-ranking & retrieval tasks. To investigate whether MARC’s representations are suitable for various recom- mendation tasks, we apply them to DSSM in the retrieval stage and PRM in the re-ranking stage on the ML-1M dataset. We compare SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYunjia Xi et al. Table 2: The overall performance on the re-ranking task. The symbol * indicates statistically significant improvement over the best compression baselines MRL (t-test with 푝< 0.05). ModelsMAP@1MAP@3@MAP5MAP@7NDCG@1NDCG@3NDCG@5NDCG@7 base (PRM)0.65350.73680.72590.72120.65350.69380.76520.7795 Llama3-8B0.67890.75410.74220.73880.67890.71450.78140.7926 MRL 0.67500.75050.73950.73570.67500.71080.77910.7906 MARC-C0.6806* 0.7559* 0.7445* 0.7406*0.6806*0.7185*0.7835*0.7946* MARC-O0.6865*0.7592*0.7475*0.7440*0.6865*0.7222*0.7863*0.7966* Table 3: The overall performance on retrieval task. The sym- bol * indicates statistically significant improvement over the best compression baselines MRL (t-test with 푝< 0.05). Models NDCGHitRateMRR @10@20@50@10@20@50/ base (DSSM)0.15700.18380.21320.28670.39240.54030.1324 Llama3-8B0.19380.21950.24900.32910.43080.57860.1669 MRL0.18420.21150.24020.31770.42550.56980.1582 MARC-C0.2089* 0.2342* 0.2631*0.3580* 0.4577* 0.6026*0.1770* MARC-O0.2116* 0.2386* 0.2666*0.3591* 0.4658* 0.6066*0.1806* MARC’s original (MARC-O) and compressed (MARC-C) represen- tations against the base model without representation enhancement (base), the 4096-dimensional representations from Llama3-8B without fine-tuning, and the strongest compression baseline, MRL. Both MARC and MRL are fine-tuned on Llama3-8B and generate 128-dimensional compressed representations. The results presented in Table 2 and Table 3 indicate that representation enhancement significantly improves the base model’s performance in both the re-ranking and retrieval stages, aligning with conclusions drawn from CTR prediction tasks. Notably, MARC outperforms both the strongest compression baseline, MRL, and the 4096-dimensional representations from frozen Llama3-8B. For instance, MARC-C im- proves HitRate@10 in the retrieval task by 12.67% over MRL and by 8.76% over frozen Llama3-8B. We attribute this generalization to our modular design: by offloading the task adaptation to the Matching Network, the compressed representations are less likely to overfit to the specific objective and instead retain the rich, generalized information beneficial for diverse downstream tasks. 4.2.3 Online A/B Test (RQ2). We applied MARC’s representation to the downstream CVR model (conversion rate) in a commercial advertising scenario with tens of millions of users and ads. The mas- sive scale of users and items makes using raw LLM representations computationally prohibitive for storage and training, hindering real- world deployment. Our model, however, successfully compresses these representations into 128 dimensions with high quality. First, we train our MARC model offline on an industrial dataset to gener- ate compact representations which are updated periodically. These representations are then integrated as features into a downstream CVR model specifically tailored for this scenario. For the online A/B test, we randomly allocated 10% of users each to the experimental and control groups. While both groups used the same backbone CVR model, the experimental group integrated MARC representations. In contrast, the control group utilized the current online production champion. It incorporates a highly opti- mized compression scheme tailored to our industrial-scale environ- ment to reduce the raw LLM representations to 128 dimensions. The original, high-dimensional representations are not used online due to their large deployment costs. In a 7-day online A/B test, MARC achieved a 2.82% increase in eCPM against a strong production baseline, with similar inference latency. 4.3 In-depth Analysis 4.3.1 Ablation Study (RQ3). We design several variants and com- pare their performance in enhancing DCNv2 on the ML-1M dataset. w/o IN removes interaction components from MARC, replacing the user-item matching network with separate MLPs for users and items, and optimized with a cosine similarity loss. w/o MN re- moves the matching network and calculates cosine similarity loss directly from compressed representations. w/o EI removes explicit interactions from Eq.(7). w/o HSIC removes the HSIC loss. w/ CS maintains the same network structure but replaces the objective with cosine similarity loss. w/ AoE employs SOTA AoE loss [25] (consisting of angle, cosine similarity, and contrastive objectives) and replaces the matching network with two separate MLPs for user and item because AoE loss is incompatible with it. Figure 6 shows the results, where “Orig.” denotes original representations and “Comp.” refers to compressed ones. w/o MNw/o INw/o HSICw/o EIw/ CSw/ AoEMARC 0.796 0.798 0.800 0.802 0.804 Performance (AUC) Orig.Comp. Figure 6: Ablation study on ML-1M dataset. The results indicate that removing any component leads to a decline in performance. The absence of HSIC loss (w/o HSIC) leads to the largest performance drop for both original and compressed representations, emphasizing its importance for information con- straints. Removing explicit interactions (w/o EI) or the matching network (w/o MN) has a limited effect on the original representa- tions but significantly impacts the compressed ones, indicating their role in maintaining information during compression. However, it is crucial to highlight that even without explicit interactions (w/o EI) or the matching network (w/o MN), MARC still outperforms Modular Representation Compression: Adapting LLMs for Efficient and Effective RecommendationsSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia the strongest baseline MRL. This confirms that while explicit inter- actions contribute to performance, the core gain stems from our modular decoupling framework. Switching losses (w/ AoE and w/ CS) also degrade representation quality, likely due to their reduced alignment with downstream tasks. AoE performs poorly in com- pression, likely because it cannot utilize matching network, failing to preserve interaction information crucial for downstream tasks. Additionally, we evaluate MARC’s performance across different loss functions and network architectures to identify what alleviates MRA. We designed three variants: w/ I-CS, which retains the orig- inal MARC structure but replaces cross-entropy loss with cosine similarity loss, same as w/ CS; w/ AoE, which employs the state- of-the-art AoE loss. Additionally, we devise w/ S-CS, which also adopts cosine similarity loss, but replaces the interactive matching network with the non-interactive network used in w/ AoE. The results of these variants’ representations on ML-1M are shown in Figure 7. Neither switching loss functions nor removing the in- teractive component of the matching network yields a mid-layer representation advantage (MRA), yet both still outperform base- lines. In all cases, good representations come from the final layer, suggesting that the primary factor alleviating MRA in MARC is the modular structure, rather than the loss function or interac- tive network. The loss function and the interaction in the matching network primarily enhance representation quality. 048121620242832 Layer 0.792 0.794 0.796 0.798 0.800 0.802 Performance (AUC) w/ I-CS w/ AoE w/ S-CS MARC Figure 7: MARC alleviates MRA under various losses. 4.3.2 Compatibility Analysis (RQ4). Previous experiments have ex- plored the compatibility of MARC across various datasets and down- stream recommendation tasks. This section delves into the compati- bility of MARC with different CTR models and backbone LLMs. We select several commonly used CTR models, including DCNv1 [54], DeepFM [9], FiGNN [27], FiBiNet [14], AutoInt [48], xDeepFM [28], and DIN [77], and the results are presented in Table 4. Regarding backbone LLMs, we select a range of models with different sizes, including Qwen2-1.5B, Phi-2 (2.7B), Qwen2.5-3B, and Qwen2- 7B, with their results based on DCNv2 shown in Figure 8. As in Section 4.2.2, we compare MARC’s original and compressed rep- resentations, MARC-O and MARC-C, with the high-dimensional representations from backbone LLMs without fine-tuning (Frozen LLM) and 128-dimensional representations from the strongest com- pression baseline, MRL. From Table 4, it is evident that MARC demonstrates performance similar to its results in DCNv2 across different CTR models. MARC- C significantly outperforms MRL and the frozen Llama3-8B, with a minimal gap of less than 0.1% when compared to the upper-bound MARC-O. Additionally, Figure 8 demonstrates that MARC also exhibits superior performance with various backbone LLMs, al- though the effectiveness of the representation may vary depending on the size of the LLMs. Furthermore, we verify that MARC can mitigate mid-layer representation advantage across different down- stream models and LLMs, which is detailed in Section 2.3. This confirms that our modular compression framework is versatile and compatible with a wide range of downstream models and back- bones, alleviates mid-layer representation advantage, and yields lightweight yet efficient representations. Qwen2-1.5BPhi-2Qwen2.5-3BQwen2-7B 0.7900 0.7925 0.7950 0.7975 0.8000 0.8025 Performance (AUC) Frozen LLM MRL MARC-C MARC-O Figure 8: Performance on different backbone LLMs. Frozen LLM denotes backbone LLM without fine-tuning. 4.3.3 Impact of Compressed Representation Dimension (RQ5). Lastly, we investigate the impact of the compressed representation dimen- sions on the representation quality. We vary the dimensionalities of compressed representations from MARC and the strongest baseline MRL, including 16, 32, 64, 128, 256, and 512, and apply them to the downstream DCNv2 model. The results are presented in Figure 9, where MARC-C and MARC-O represent the compressed and origi- nal representations from MARC, respectively. From the figure, it is evident that MARC significantly outperforms the baseline MRL across all dimensions, demonstrating that our model can generate more efficient and lightweight representations. When the com- pressed dimension is small, the original representation performs better. As the dimension increases, compressed representations demonstrate improved performance, likely because larger dimen- sions retain more information. However, after the dimension of 128, the performance of MARC-O begins to decline, and since the original representations serve as the upper bound for compressed representations, this also leads to a decrease in the performance of MARC-C. This decline may be attributed to the increasing difficulty of model training as the compressed dimension grows larger. 163264128256512 Compressed Dimension 0.7925 0.7950 0.7975 0.8000 0.8025 0.8050 Performance (AUC) MARC-OMARC-CMRL Figure 9: Performance under different dimensions. SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYunjia Xi et al. Table 4: The overall performance on different CTR models. Models xDeepFMFiBiNetDeepFMFiGNNDCNv1DINAutoInt AUCLoglossAUCLoglossAUCLoglossAUCLoglossAUCLoglossAUCLoglossAUCLogloss base0.78300.55140.78240.55180.78270.55140.78380.55020.78370.55150.78870.54500.78270.5515 Llama3-8B0.79620.53780.79630.53980.79560.54310.79660.53980.79610.53900.79830.53670.79650.5434 MRL 0.79580.53840.79560.53840.79450.53970.79590.53830.79560.54000.79720.53680.79580.5392 MARC-C0.8009 0.5336 0.8013 0.5315 0.8001 0.5332 0.8012 0.5323 0.8016 0.5329 0.8019 0.5323 0.8006 0.5351 MARC-O0.80240.53480.80230.53100.80170.53200.80200.53090.80210.53070.80210.53070.80330.5303 5 Related Work 5.1 Text Embedding and Compression With the rise of LLMs, many efforts in NLP have focused on lever- aging LLMs for sentence embedding learning. Most of these ap- proaches utilize unsupervised contrastive learning, which contrasts in-batch positive and negative samples to improve representation quality [2,23,26,43,60,70]. Some supervised approaches have also been proposed to align embeddings with human perception, typically by maximizing the cosine similarity between positive pairs [19,25,56]. However, these methods often take the pooling of the last hidden states of LLMs as the final representation, resulting in a high-dimensional representation (e.g., 4096). Such representa- tions pose challenges for storage and downstream tasks, especially under resource constraints [56]. To address this, nested-based rep- resentation compression has been introduced in NLP, including Matryoshka Representation Learning (MRL) [22,23,40] and PCA- like approaches [56]. MRL learns nested lower-dimensional vectors within the same high-dimensional space to preserve information at multiple granularities, while PCA-like approaches apply PCA during training to concentrate important information in the leading dimensions of the representation. There are also some projection- based approaches that employ simple projection layers to compress the final representations [7,53]. Note that we mainly focus on rep- resentation compression that directly maps high-dimensional LLM vectors to low-dimensional spaces. We do not include knowledge distillation approaches, which belong to model compression. The above representation compression methods often suffer from the mid-layer representation advantage, where intermediate rep- resentations perform better than final ones, leading to suboptimal compressed embeddings, as demonstrated in Section 2.3. Our work focuses on addressing this issue to enhance representation quality. 5.2 LLM-based Recommendation The emergence of LLMs has brought significant transformations to RSs [3,24,29,34,35,59,62,64,71,80], with two main approaches: LLMs as core recommenders and as parts of traditional ones. The former uses LLMs as recommenders, which leads to superior per- formance, but faces deployment challenges due to large inference latency [1,4,49,68,75,76,79]. The latter integrates LLMs into traditional RSs, using LLM-derived representations for downstream RSs. Here, some methods first generate recommendation-related knowledge with LLMs and then encode this knowledge into repre- sentations [36,38,47,50,61,63,67], while others directly encode recommendation-related text with LLMs [7,18,53,57,73]. Regard- less of the specific method, leveraging LLMs’ representations is a crucial step. However, current deployable solutions still face notable chal- lenges. Approaches that utilize smaller models [36,37,50,57,61], e.g., BERT, often suffer from significant performance degradation. On the other hand, directly employing LLMs for representation leads to substantial resource consumption [12,73]. Therefore, this work primarily focuses on representation compression, aiming to obtain lightweight and effective RSs representations from LLMs. 6 Conclusion This work identifies a trade-off between efficiency and effectiveness in LLM-based representations for recommendations, and focuses on representation compression. We reveal the phenomenon of mid- layer representation advantage during compression and explain it using modularity theory, where LLMs develop internal functional modularity during training. To address this, we propose MARC, which includes Modular Adjustment and Modular Task Decoupling. Extensive experiments validate that MARC effectively addresses mid-layer advantage and produces efficient representations for recommendation tasks. It also improves eCPM by 2.82% during an online A/B test in a large commercial search advertising scenario. Acknowledgments The Shanghai Jiao Tong University team is partially supported by Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), National Natural Science Foundation of China (624B2096, 72595872, 62322603), and National Key RD Program of China (2022ZD0114804), Changan Automobile & Chongqing Natu- ral Science Foundation Joint Fund for Innovation and Development (CSTB2023NSCQ-LZX0136). The work is also sponsored by Huawei Innovation Research Program. We thank MindSpore [42] for its partial support. The author Yunjia Xi is also supported by Wu Wen Jun Honorary Doctoral Scholarship. References [1]Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. arXiv preprint arXiv:2305.00447 (2023). [2] Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961 (2024). [3] Jin Chen, Zheng Liu, Xu Huang, Chenwang Wu, Qi Liu, Gangwei Jiang, Yuanhao Pu, Yuxuan Lei, Xiaolong Chen, Xingmei Wang, et al.2023. When large language models meet personalization: Perspectives of challenges and opportunities. arXiv preprint arXiv:2307.16376 (2023). [4]Qian Dong, Yiding Liu, Qingyao Ai, Zhijing Wu, Haitao Li, Yiqun Liu, Shuaiqiang Wang, Dawei Yin, and Shaoping Ma. 2024. Unsupervised large language model Modular Representation Compression: Adapting LLMs for Efficient and Effective RecommendationsSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia alignment for information retrieval via contrastive feedback. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 48–58. [5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024). [6] Wenqi Fan, Zihuai Zhao, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Jiliang Tang, and Qing Li. 2023. Recommender systems in the era of large language models (llms). arXiv preprint arXiv:2307.02046 (2023). [7]Binzong Geng, Zhaoxin Huan, Xiaolu Zhang, Yong He, Liang Zhang, Fajie Yuan, Jun Zhou, and Linjian Mo. 2024. Breaking the length barrier: Llm-enhanced CTR prediction in long textual user behaviors. In SIGIR’24. 2311–2315. [8] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. 2005. Measuring statistical dependence with Hilbert-Schmidt norms. In International conference on algorithmic learning theory. Springer, 63–77. [9]Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction. arXiv preprint arXiv:1703.04247 (2017). [10]R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. 2018. Learning deep represen- tations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670 (2018). [11]Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021). [12] Jun Hu, Wenwen Xia, Xiaolu Zhang, Chilin Fu, Weichang Wu, Zhaoxin Huan, Ang Li, Zuoli Tang, and Jun Zhou. 2024. Enhancing sequential recommendation via llm-based semantic embedding learning. In Companion Proceedings of the ACM on Web Conference 2024. 103–111. [13] Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. In Proceedings of the 22nd ACM international conference on Information & Knowledge Management. 2333–2338. [14]Tongwen Huang, Zhiqi Zhang, and Junlin Zhang. 2019. FiBiNET: combining fea- ture importance and bilinear feature interaction for click-through rate prediction. In Proceedings of the 13th ACM conference on recommender systems. 169–177. [15] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al.2024. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186 (2024). [16]Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated Gain-Based Evaluation of IR Techniques. ACM Trans. Inf. Syst. (2002), 422–446. [17]Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al.2023. Phi-2: The surprising power of small language models. Microsoft Research Blog 1, 3 (2023), 3. [18] Jian Jia, Yipei Wang, Yan Li, Honggang Chen, Xuehan Bai, Zhaocheng Liu, Jian Liang, Quan Chen, Han Li, Peng Jiang, et al.2024. Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application. arXiv preprint arXiv:2405.03988 (2024). [19]Ting Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang, and Fuzhen Zhuang. 2023. Scaling sentence embeddings with large language models. arXiv preprint arXiv:2307.16645 (2023). [20]Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. 2019. Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351 (2019). [21] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, Vol. 1. Minneapolis, Minnesota. [22]Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al.2022. Matryoshka representation learning. Advances in Neural Information Processing Systems 35 (2022), 30233–30249. [23]Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, et al.2024. Gecko: Versatile text embeddings distilled from large language models. arXiv preprint arXiv:2403.20327 (2024). [24]Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. 2023. Large Language Models for Generative Recommendation: A Survey and Visionary Discussions. arXiv preprint arXiv:2309.01157 (2023). [25] Xianming Li and Jing Li. 2024. AoE: Angle-optimized embeddings for semantic textual similarity. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1825–1839. [26] Xianming Li and Jing Li. 2024. BeLLM: Backward Dependency Enhanced Large Language Model for Sentence Embeddings. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 792–804. [27]Zekun Li, Zeyu Cui, Shu Wu, Xiaoyu Zhang, and Liang Wang. 2019. Fi-gnn: Modeling feature interactions via graph neural networks for ctr prediction. In CIKM’19. 539–548. [28]Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xdeepfm: Combining explicit and implicit feature in- teractions for recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1754–1763. [29]Jianghao Lin, Bo Chen, Hangyu Wang, Yunjia Xi, Yanru Qu, Xinyi Dai, Kangning Zhang, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. ClickPrompt: CTR models are strong prompt generators for adapting language models to CTR prediction. In Proceedings of the ACM Web Conference 2024. 3319–3330. [30]Jianghao Lin, Xinyi Dai, Rong Shan, Bo Chen, Ruiming Tang, Yong Yu, and Weinan Zhang. 2025. Large language models make sample-efficient recommender systems. Frontiers of Computer Science 19, 4 (2025), 194328. [31] Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al.2025. How can recommender systems benefit from large language models: A survey. ACM Transactions on Information Systems 43, 2 (2025), 1–47. [32]Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. Rella: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recom- mendation. In Proceedings of the ACM on Web Conference 2024. 3497–3508. [33]Chengkai Liu, Jianghao Lin, Hanzhou Liu, Jianling Wang, and James Caverlee. 2024. Behavior-dependent linear recurrent units for efficient sequential recom- mendation. In Proceedings of the 33rd ACM international conference on information and knowledge management. 1430–1440. [34]Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee. 2024. Mamba4rec: Towards efficient sequential recommendation with selective state space models. arXiv preprint arXiv:2403.03900 (2024). [35] Peng Liu, Lemei Zhang, and Jon Atle Gulla. 2023. Pre-train, prompt and recom- mendation: A comprehensive survey of language modelling paradigm adaptations in recommender systems. arXiv preprint arXiv:2302.03735 (2023). [36]Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024. Once: Boosting content-based recommendation with both open-and closed-source large language models. In WSDM’24. 452–461. [37]Weiqing Luo, Chonggang Song, Lingling Yi, and Gong Cheng. 2024. KELLM- Rec: Knowledge-Enhanced Large Language Models for Recommendation. arXiv preprint arXiv:2403.06642 (2024). [38] Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qifan Wang, Si Zhang, Ren Chen, Christopher Leung, Jiajie Tang, and Jiebo Luo. 2023. Llm-rec: Person- alized recommendation via prompting large language models. arXiv preprint arXiv:2307.15780 (2023). [39] Wan-Duo Kurt Ma, JP Lewis, and W Bastiaan Kleijn. 2020. The HSIC bottleneck: Deep learning without back-propagation. In Proceedings of the AAAI conference on artificial intelligence, Vol. 34. 5085–5092. [40] Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2024. Fine- tuning llama for multi-stage text retrieval. In SIGIR’24. 2421–2425. [41] Andrzej Maćkiewicz and Waldemar Ratajczak. 1993. Principal components analysis (PCA). Computers & Geosciences 19, 3 (1993), 303–342. [42]MindSpore Team. 2020. MindSpore: A new open source deep learning train- ing/inference framework that could be used for mobile, edge and cloud scenarios. https://w.mindspore.cn/ Accessed: 2026-04. [43] Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Aman- preet Singh, and Douwe Kiela. 2024. Generative representational instruction tuning. arXiv preprint arXiv:2402.09906 (2024). [44] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018). [45] Changhua Pei, Yi Zhang, Yongfeng Zhang, Fei Sun, Xiao Lin, Hanxiao Sun, Jian Wu, Peng Jiang, Junfeng Ge, Wenwu Ou, et al.2019. Personalized re-ranking for recommendation. In Recsys’19. 3–11. [46]Yunchen Pu, Zhe Gan, Ricardo Henao, Xin Yuan, Chunyuan Li, Andrew Stevens, and Lawrence Carin. 2016. Variational autoencoder for deep learning of images, labels and captions. Advances in neural information processing systems 29 (2016). [47]Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Representation learning with large language models for recommendation. In Proceedings of W 2024. 3464–3475. [48]Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. Autoint: Automatic feature interaction learning via self- attentive neural networks. In Proceedings of the 28th ACM international conference on information and knowledge management. 1161–1170. [49] Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, and Li. 2024. Idgenrec: Llm-recsys alignment with textual id learning. In SIGIR’24’. 355–364. [50]Changxin Tian, Binbin Hu, Chunjing Gan, Haoyu Chen, Zhuo Zhang, Li Yu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, and Jiawei Chen. 2024. ReLand: Integrating Large Language Models’ Insights into Industrial Recommenders via a Controllable Reasoning Pool. In Recsys’24. 63–73. [51]Naftali Tishby and Noga Zaslavsky. 2015. Deep learning and the information bottleneck principle. In 2015 IEEE Information Theory Workshop (ITW). IEEE, 1–5. https://doi.org/10.1109/ITW.2015.7133169 SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYunjia Xi et al. [52]Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al.2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023). [53]Zhizhong Wan, Bin Yin, Junjie Xie, Fei Jiang, Xiang Li, and Wei Lin. 2024. LARR: Large Language Model Aided Real-time Scene Recommendation with Semantic Understanding. In Proceedings of Recsys. 23–32. [54]Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & cross network for ad click predictions. In Proceedings of the ADKDD’17. 1–7. [55]Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. In W’21. 1785–1797. [56] Shuai Wang, Shengyao Zhuang, Bevan Koopman, and Guido Zuccon. 2024. 2D Matryoshka Training for Information Retrieval. arXiv preprint arXiv:2411.17299 (2024). [57] Yuling Wang, Changxin Tian, Binbin Hu, Yanhua Yu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, Liang Pang, and Xiao Wang. 2024. Can Small Language Models be Good Reasoners for Sequential Recommendation?. In Proceedings of the ACM on Web Conference 2024. 3876–3887. [58]Yasi Wang, Hongxun Yao, and Sicheng Zhao. 2016. Auto-encoder based dimen- sionality reduction. Neurocomputing 184 (2016), 232–242. [59] Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al.2023. A Survey on Large Language Models for Recommendation. arXiv preprint arXiv:2305.19860 (2023). [60]Yunjia Xi, Jianghao Lin, Menghui Zhu, Yongzhao Xiao, Zhuoying Ou, Jiaqi Liu, Tong Wan, Bo Chen, Weiwen Liu, Yasheng Wang, et al.2025. InfoDeepSeek: Benchmarking Agentic Information Seeking for Retrieval-Augmented Generation. arXiv preprint arXiv:2505.15872 (2025). [61]Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, Rui Zhang, et al.2023. Towards open-world recommendation with knowledge augmentation from large language models. arXiv preprint arXiv:2306.10933 (2023). [62]Yunjia Xi, Weiwen Liu, Jianghao Lin, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. MemoCRS: Memory-enhanced Sequential Conversational Recom- mender Systems with Large Language Models. arXiv preprint arXiv:2407.04960 (2024). [63] Yunjia Xi, Weiwen Liu, Jianghao Lin, Muyan Weng, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Yong Yu, et al.2024. Efficient and Deploy- able Knowledge Infusion for Open-World Recommendations via Large Language Models. arXiv preprint arXiv:2408.10520 (2024). [64] Yunjia Xi, Weiwen Liu, Jianghao Lin, Chuhan Wu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Play to Your Strengths: Collaborative Intelli- gence of Conventional Recommender Models and Large Language Models. arXiv preprint arXiv:2403.16378 (2024). [65]Yunjia Xi, Weiwen Liu, Yang Wang, Ruiming Tang, Weinan Zhang, Yue Zhu, Rui Zhang, and Yong Yu. 2023. On-device integrated re-ranking with heteroge- neous behavior modeling. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5225–5236. [66] Yunjia Xi, Hangyu Wang, Bo Chen, Jianghao Lin, Menghui Zhu, Weiwen Liu, Ruiming Tang, Zhewei Wei, Weinan Zhang, and Yong Yu. 2025. Efficiency un- leashed: Inference acceleration for LLM-based recommender systems with spec- ulative decoding. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1891–1901. [67] Yunjia Xi, Hangyu Wang, Bo Chen, Jianghao Lin, Menghui Zhu, Weiwen Liu, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. A Decoding Acceleration Framework for Industrial Deployable LLM-based Recommender Systems. arXiv preprint arXiv:2408.05676 (2024). [68]Yunjia Xi, Muyan Weng, Wen Chen, Chao Yi, Dian Chen, Gaoyang Guo, Mao Zhang, Jian Wu, Yuning Jiang, Qingwen Liu, et al.2025. Bursting filter bubble: Enhancing serendipity recommendations with aligned large language models. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 5059–5070. [69]Chaojun Xiao, Zhengyan Zhang, Chenyang Song, Dazhi Jiang, Feng Yao, Xu Han, Xiaozhi Wang, Shuo Wang, Yufei Huang, Guanyu Lin, et al. 2024. Configurable foundation models: Building llms from a modular perspective. arXiv preprint arXiv:2409.02877 (2024). [70]Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Yanqiao Zhu, May D Wang, Joyce C Ho, Chao Zhang, and Carl Yang. 2024. Bmretriever: Tuning large language models as better biomedical text retrievers. arXiv preprint arXiv:2404.18443 (2024). [71] Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Jundong Li, and Zi Huang. 2023. Self-supervised learning for recommender systems: A survey. IEEE Transactions on Knowledge and Data Engineering (2023). [72] Yisong Yue, Thomas Finley, Filip Radlinski, and Thorsten Joachims. 2007. A support vector method for optimizing average precision. In SIGIR’07. 271–278. [73]Chao Zhang, Shiwei Wu, Haoxin Zhang, Tong Xu, Yan Gao, Yao Hu, and En- hong Chen. 2024. NoteLLM: A Retrievable Large Language Model for Note Recommendation. In Companion Proceedings of W. 170–179. [74]Zhengyan Zhang, Zhiyuan Zeng, Yankai Lin, Chaojun Xiao, Xiaozhi Wang, Xu Han, Zhiyuan Liu, Ruobing Xie, Maosong Sun, and Jie Zhou. 2023. Emergent modularity in pre-trained transformers. arXiv preprint arXiv:2305.18390 (2023). [75]Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448. [76]Zhi Zheng, Wenshuo Chao, Zhaopeng Qiu, Hengshu Zhu, and Hui Xiong. 2024. Harnessing large language models for text-rich sequential recommendation. In Proceedings of the ACM on Web Conference 2024. 3207–3216. [77]Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1059–1068. [78]Jiachen Zhu, Jianghao Lin, Xinyi Dai, Bo Chen, Rong Shan, Jieming Zhu, Ruim- ing Tang, Yong Yu, and Weinan Zhang. 2024. Lifelong personalized low-rank adaptation of large language models for recommendation. arXiv preprint arXiv:2408.03533 (2024). [79]Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024. Collab- orative large language model for recommender systems. In Proceedings of the ACM on Web Conference 2024. 3162–3172. [80] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chen- long Deng, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107 (2023).