Pith. sign in

REVIEW 3 major objections 3 minor 4 cited by

GEMS tunes a single LLM for both search and recommendation by splitting gradients into shared and task-specific subspaces and projecting updates away from pre-trained knowledge, with consistent gains over baselines on both tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 10:34 UTC pith:WO2VYIKR

load-bearing objection The core null-space projection is defined as its opposite, so the paper's main mechanism is internally inconsistent as written; the reported gains are not supported. the 3 major comments →

arxiv 2601.09496 v2 pith:WO2VYIKR submitted 2026-01-14 cs.IR

Unifying Search and Recommendation in LLMs via Gradient Multi-Subspace Tuning

Jujia Zhao , Zihan Wang , Shuaiqun Pan , Suzan Verberne , Zhaochun Ren This is my paper
classification cs.IR
keywords unified search and recommendationgenerative retrievalparameter-efficient tuninglow-rank subspace tuninggradient conflictnull-space projectionknowledge preservationmulti-task learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that a single large language model can be adapted to both search and recommendation at once, without full fine-tuning and without the usual casualties: the two tasks' gradients stop fighting each other, and the model's general-domain reasoning survives the adaptation. Its method, GEMS, routes training through three low-rank subspaces — a handful of the dominant gradient directions, shared by both tasks, specific to search, and specific to recommendation — so conflicting signals are geometrically separated instead of colliding in the same parameters; it then projects the combined update so that it cannot move along the directions encoding what the model already learned. The authors report consistent wins over specialised, unified, and parameter-efficient baselines on the Qilin and Amazon benchmarks, with both a T5-base and a 3-billion-parameter Qwen backbone, an 85–88% drop in gradient-conflict coefficients compared with LoRA-style tuning, and a roughly halved rate at which previously correct user-intent predictions turn wrong after tuning. If the paper is right, one model could serve both search queries and personalised recommendations at the deployment cost of the base LLM, with no added inference-time weights, and the same two-step recipe could extend to other task pairs whose objectives pull in different directions.

Core claim

GEMS claims that both obstacles to unifying search and recommendation — conflicting gradients and drift of general-domain knowledge — can be neutralised by controlling where gradients move. Search, recommendation, and summed gradients are projected into separate low-rank subspaces, optimised with independent Adam states, fused by an adaptive gate, then passed through a null-space projector that strips update components along the backbone's dominant pre-trained directions. Across two backbones on Qilin and Amazon, GEMS beats specialised, unified, and PEFT baselines on both tasks, with 85–88% less gradient conflict than LoRA.

What carries the argument

Two mechanisms carry the argument. Multi-subspace decomposition forms separate gradients for search, recommendation, and their sum, SVDs each into dominant low-rank directions, and runs independent Adam optimisers inside the three subspaces; a small gating network with a temperature-scaled softmax fuses the three updates, keeping the shared update at full weight and splitting the remainder between the two task-specific ones. Null-space projection builds a projector from the top-k singular vectors of the backbone's hidden states on a general corpus (here Wikipedia), meant to send the fused update onto the orthogonal complement of the dominant pre-trained directions so fine-tuning cannot drift

Load-bearing premise

The load-bearing step is Eq. (19) and Algorithm 1 line 9: the method needs P⊥ = I − U_k U_k^⊤, the projector onto the orthogonal complement of the top-k pre-trained directions, so updates lose components aligned with general-domain knowledge — but as printed, P⊥ = U_pre^k U_pre^k^⊤ projects onto those directions themselves, keeping the very components it should remove. The knowledge-preservation claim presupposes a typo, with the corrected matrix actually implemented; the pap

What would settle it

Read the released code's projection step. If it literally applies U_k U_k^⊤ Δ as printed in Eq. (19), the method keeps — rather than removes — the dominant pre-trained components, and the reported preservation and performance gains must be attributed to the subspace decomposition alone. If the code instead applies (I − U_k U_k^⊤)Δ, swap in the printed version and re-run the Table 3 'correct-before, incorrect-after' analysis: a working null-space projection should show markedly lower disruption rates than the printed projector, and if it does not, that mechanism is not carrying the reported eff

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • One unified deployment: a single LLM can serve both search queries and personalised recommendation histories at inference cost identical to the base model, because GEMS adds no extra weights.
  • Conflict is measurable and smaller: gradient-conflict coefficients drop 85–88% versus LoRA, whose worst conflicts sit in the query and key projections of attention, while GEMS keeps conflict uniformly low across layers.
  • Knowledge preservation is quantifiable: the share of previously correct user-intent predictions that turn wrong after tuning falls from roughly 19–26% to about 8–13% on the tested datasets.
  • Experts are not the answer: beating MoE-augmented PEFT baselines indicates that routing capacity is not what resolves cross-task interference — the geometry of the gradient subspaces is.
  • The same machinery can pre-train: because GEMS tunes the original weight matrix rather than bolt-on adapters, it applies beyond fine-tuning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The recipe is general enough to transfer: the same 'separate shared and task-specific gradient subspaces, then project onto the null space of pre-trained knowledge' pattern could be tried on other task pairs with divergent objectives; the paper demonstrates only search–recommendation.
  • The preservation metric is 'correct-before, incorrect-after' measured on the tuning data itself; a stronger test of general-knowledge retention would be held-out reasoning or commonsense benchmarks after GEMS tuning, which the paper does not run.
  • The null-space basis is computed once from Wikipedia hidden states and corpus sensitivity is untested; a natural check is rebuilding the basis from other generic corpora and seeing whether the preservation numbers hold.
  • The 'no additional trainable weights' claim coexists with the learnable gating network of Eq. (16), whose parameter count is never reported; the efficiency claim is best read as applying to the backbone parameterisation itself.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes GEMS (Gradient Multi-Subspace Tuning), a parameter-efficient method for unifying search and recommendation in LLMs. GEMS extends subspace tuning (GaLore-style) by decomposing gradients into shared, search-specific, and recommendation-specific low-rank subspaces, adaptively fusing them with a small gating network, and then projecting the fused update into the 'null space' of pre-trained knowledge to preserve general-domain reasoning. Experiments on Qilin and Amazon datasets with Flan-T5-base and Qwen2.5-3B-Instruct report consistent improvements over specialized and unified baselines, along with ablations, gradient-conflict analysis, and a user-intent-preservation study.

Significance. If validated, the paper would make a useful practical contribution: it targets multi-task adaptation of billion-scale LLMs without full fine-tuning, and it addresses two real problems (gradient conflict and catastrophic forgetting of general knowledge). The empirical study is broad, including two datasets, two backbone scales, many baselines, ablations, and a consistency analysis. However, the central mechanism as written is mathematically defective (Projector sign error), and the 'no extra trainable weights' claim is contradicted by the paper's own gating network. Without code or a corrected statement, the reported results cannot be attributed to the proposed mechanism. The significance is therefore conditional on fixing these load-bearing issues.

major comments (3)
  1. [Sec. 4.3, Eq. (19); Algorithm 1 line 9] Eq. (19) defines P⊥ = U_pre^k U_pre^k^T, which is the projector onto the top-k pre-trained directions, not onto their orthogonal complement. The prose immediately below states that this projector 'removes any component of an update that aligns with the dominant pre-trained directions'; that requires P⊥ = I − U_pre^k U_pre^k^T. As written, Δfinal = P⊥ Δfused lies in the span of the dominant pre-trained directions, so the update would preferentially move along the very components the method claims to avoid. This is the load-bearing knowledge-preservation step, and the same error appears in Algorithm 1, line 9: return (U_k U_k^T)Δ. The manuscript gives no indication that a corrected complement was used in the experiments, and no code is provided. Please correct the equation and algorithm, and explicitly state which projection was actually implemented.
  2. [Sec. 4.2 (Eq. 16) vs. Abstract / Sec. 4.4] The abstract and Section 4.4 claim GEMS requires 'no additional trainable weights' / 'no extra trainable weights beyond the backbone.' This is directly contradicted by the adaptive gating network f_φ in Eq. (16), which has learnable parameters φ = {W1, b1, W2, b2}. Even though the gating network is small, it is a trainable component and should be accounted for in the efficiency comparison in Table 5. Please qualify the claim (e.g., 'no extra trainable weights except a tiny gating network') or remove it, and update the complexity analysis accordingly.
  3. [Sec. 5.2, Table 2] In the Qwen-3B results, the row for Ours on Amazon Search contains a garbled entry: '0.4570*0.25960.2938' for the NDCG@10 columns. This makes the reported result unreadable and open to misinterpretation. Please correct the formatting and re-check the underlying numbers.
minor comments (3)
  1. [Sec. 4.3, Eq. (18)] Eq. (18) writes SVD(F F^T) = U_pre Σ V_pre^T. Since F F^T is a symmetric positive semidefinite matrix, its SVD should be U_pre Σ U_pre^T (with V_pre = U_pre). Clarify the notation.
  2. [Sec. 4.2, Eq. (11)] The shared gradient G_shared = −∇(L_src + L_rec) equals G_src + G_rec only if the two losses are evaluated on the same mini-batch. Please specify the batch construction for the shared gradient.
  3. [Throughout] Typos and stylistic issues: 'incoreect' in Table 3 header, 'shits' in Section 2, 'unifing' in Section 5.3.1, 'an reasoning' in Section 5.3.3, and 'Eq.-(11)-(11)' in Algorithm 1 line 13. These should be fixed.

Circularity Check

0 steps flagged

No significant circularity: the core claims are empirical benchmark comparisons, and the only self-citation is a non-load-bearing related-work reference. The null-space projector in Eq. (19) is misdefined as written, but that is a correctness issue, not a circular reduction.

full rationale

GEMS's central claims are established by external benchmark comparisons (Tables 1 and 2) and ablations (Figure 3), not by deriving predictions from fitted constants. The subspace-tuning core is explicitly based on GaLore [41], and the null-space idea is credited to AlphaEdit [6], both external starting points. The only self-citation is [39] in the related-work list of unified S&R models; it is not used to justify correctness or to forbid alternatives, so it is not load-bearing. One non-circular but serious flaw should be flagged: Section 4.3 Eq. (19) defines P⊥ = U_pre^k U_pre^k⊤ and Algorithm 1 line 9 returns (U_k U_k⊤)Δ, which is the projector onto the dominant pre-trained directions, not onto their null space; the intended projector would be I − U_k U_k⊤. The prose claim that this 'removes any component of an update that aligns with the dominant pre-trained directions' is therefore inconsistent with the equation as written. This is a definitional/mathematical error, not a reduction of a prediction to its inputs, so it is a correctness risk rather than circularity. The empirical results would remain independent evidence if the implementation corrected the typo; the gating module's unspecified training objective is an underspecification, not a circular step.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or conceptual entities, but the central method rests on several untested assumptions and a set of hand-tuned hyperparameters. The most consequential free parameters are the null-space rank k and the gating network parameters, both of which are either unspecified or defined without a training signal. The null-space knowledge-preservation assumption is pivotal and is implemented incorrectly in the written equations.

free parameters (6)
  • update scale factor α = grid {0.5, 1, 2, 3, 4}
    Controls the magnitude of the low-rank update; selected by grid search and directly affects final metrics.
  • gate temperature τ = grid {0.1, 0.5, 1, 2, 3}
    Sharpness of the adaptive fusion softmax; selected by grid search.
  • subspace rank r = grid {256, 512, 1024}; task-specific rank r/2
    Rank of shared and task-specific subspaces; chosen by hand/grid, not derived from data.
  • null-space basis rank k = not reported
    Number of top pre-trained singular vectors used for projection; never specified in the experimental setup or hyperparameter analysis.
  • SVD refresh interval T_svd = not reported
    Frequency of SVD recomputation; affects cost and behavior but no value is given.
  • gating network parameters φ = learnable, no objective specified
    Two-layer MLP weights; the paper claims no extra trainable weights, yet these parameters are trainable and no loss/update rule is provided.
axioms (5)
  • ad hoc to paper Gradient directions of search and recommendation can be separated into shared and task-specific subspaces with minimal geometric overlap.
    No proof is given; this is the core premise of multi-subspace decomposition (Sec. 4.2).
  • domain assumption The top-k left singular vectors of the pre-trained feature covariance F F^T span the general-domain knowledge space, and removing updates along them preserves knowledge.
    Borrowed from AlphaEdit [6] and assumed for the backbone and Wikipedia corpus without validation.
  • ad hoc to paper G_shared = G_src + G_rec is a meaningful shared optimization signal.
    Assumes simple additivity of the two task losses; no alternative decomposition is considered (Eq. 11).
  • standard math Adam optimization inside a projected low-rank subspace behaves comparably to full-space Adam.
    Inherited from GaLore [41]; accepted as background but not re-derived.
  • domain assumption A Wikipedia corpus captures the general-domain knowledge space for both T5 and Qwen backbones.
    The null-space basis is computed once from this corpus; no sensitivity analysis is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 17256 in / 13283 out tokens · 136575 ms · 2026-08-03T10:34:50.351053+00:00 · methodology

0 comments
read the original abstract

Search and recommendation (S&R) are core to online platforms, addressing explicit intent through queries and modeling implicit intent from behaviors, respectively. Their complementary roles motivate a unified modeling paradigm. Early studies to unify S&R adopt shared encoders with task-specific heads, while recent efforts reframe item ranking in both S&R as conditional generation. The latter holds particular promise, enabling end-to-end optimization and leveraging the semantic understanding of LLMs. However, existing methods rely on full fine-tuning, which is computationally expensive and limits scalability. Parameter-efficient fine-tuning (PEFT) offers a more practical alternative but faces two critical challenges in unifying S&R: (1) gradient conflicts across tasks due to divergent optimization objectives, and (2) shifts in user intent understanding caused by overfitting to fine-tuning data, which distort general-domain knowledge and weaken LLM reasoning. To address the above issues, we propose Gradient Multi-Subspace Tuning (GEMS), a novel framework that unifies S&R with LLMs while alleviating gradient conflicts and preserving general-domain knowledge. GEMS introduces (1) \textbf{Multi-Subspace Decomposition}, which disentangles shared and task-specific optimization signals into complementary low-rank subspaces, thereby reducing destructive gradient interference, and (2) \textbf{Null-Space Projection}, which constrains parameter updates to a subspace orthogonal to the general-domain knowledge space, mitigating shifts in user intent understanding. Extensive experiments on benchmark datasets show that GEMS consistently outperforms the state-of-the-art baselines across both search and recommendation tasks, achieving superior effectiveness.

Figures

Figures reproduced from arXiv: 2601.09496 by Jujia Zhao, Shuaiqun Pan, Suzan Verberne, Zhaochun Ren, Zihan Wang.

Figure 1
Figure 1. Figure 1: (a) Gradient conflict analysis across the layers of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An overview of GEMS. During unified S&R training, task gradients are routed through multi-subspace decomposition [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation study of GEMS on Qilin [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Gradient conflict heatmap analysis of PEFT and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hyperparameter analysis on Qilin. Algorithm 1 Training algorithm of GEMS Require: LLM Layer weight 𝑊 ∈R 𝑚×𝑛 (𝑚 ≤ 𝑛). Step size 𝜂, scale factor 𝛼, rank 𝑟, SVD refresh step 𝑇svd, gate temperature factor 𝜏, pretrained basis U 𝑘 ref. 1: Initialize moments 𝑀 (src) , 𝑀 (rec) , 𝑀 (shared) ∈ R 𝑟 ← 0, 𝑉 (src) ,𝑉 (rec) ,𝑉 (shared) ∈R 𝑟 ← 0 2: 𝑡 ← 0 3: function SubspaceTune(type, g, 𝑡) 4: ⊲ Look up and maintain per-t… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. From Raw IDs to Semantic Planning: How Recommender Systems Utilize Information at Scale

    cs.IR 2026-07 conditional novelty 7.0

    Recommender systems are moving from raw IDs to semantic IDs, and the authors argue the next stage is 'semantic planning'—predicting an exposure's goal before choosing the item.

  2. Differentiable Semantic ID for Generative Recommendation

    cs.IR 2026-01 unverdicted novelty 7.0

    DIGER makes semantic IDs in generative recommendation differentiable via Gumbel noise and decay schedules, yielding consistent gains on public datasets by aligning indexing and recommendation losses.

  3. From Raw IDs to Semantic Planning: How Recommender Systems Utilize Information at Scale

    cs.IR 2026-07 conditional novelty 5.0

    Recommender systems are moving from raw IDs to semantic IDs, and the next step should be semantic planning that first predicts an exposure's purpose before choosing or generating content.

  4. From Raw IDs to Semantic Planning: How Recommender Systems Utilize Information at Scale

    cs.IR 2026-07 unverdicted novelty 4.0

    Industrial recommenders are evolving from raw IDs through semantic IDs toward semantic planning, where the system predicts a semantic next-exposure target before choosing or generating a concrete item.

Reference graph

Works this paper leans on

45 extracted references · 5 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Qingyao Ai, Yongfeng Zhang, Keping Bi, Xu Chen, and W Bruce Croft. 2017. Learning a hierarchical embedding model for personalized product search. In SIGIR. ACM, 645–654

    2017

  2. [2]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2024. Scaling instruction-finetuned language models.Journal of Machine Learning Research25, 70 (2024), 1–53

    2024

  3. [3]

    Chuntao Ding, Zhichao Lu, Shangguang Wang, Ran Cheng, and Vishnu Naresh Boddeti. 2023. Mitigating task interference in multi-task learning via explicit task routing with non-learnable primitives. InCVPR. IEEE, 7756–7765

    2023

  4. [4]

    Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. 2023. Loramoe: Revolution- izing mixture of experts for maintaining world knowledge in language model alignment.arXiv preprint arXiv:2312.099794, 7 (2023)

    Pith/arXiv arXiv 2023

  5. [5]

    Yunshu Du, Wojciech M Czarnecki, Siddhant M Jayakumar, Mehrdad Farajtabar, Razvan Pascanu, and Balaji Lakshminarayanan. 2018. Adapting auxiliary losses using gradient similarity.arXiv preprint arXiv:1812.02224(2018)

    Pith/arXiv arXiv 2018

  6. [6]

    Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Jie Shi, Xiang Wang, Xiangnan He, and Tat-Seng Chua. 2025. AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models. InICLR. OpenReview.net

    2025

  7. [7]

    William Fedus, Barret Zoph, and Noam Shazeer. 2022. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research23, 120 (2022), 1–39

    2022

  8. [8]

    Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Yu Han, and Hao Wang. 2024. Mixture-of-LoRAs: An Efficient Multitask Tuning Method for Large Language Models. InLREC/COLING. ELRA and ICCL, 11371–11380

    2024

  9. [9]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InWWW. ACM, 173–182

    2017

  10. [10]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InICLR. OpenReview.net

    2022

  11. [11]

    Qiang Huang, Feng Huang, DeHao Tao, BingKun Wang, and YongFeng Huang

  12. [12]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Opti- mization. InICLR

    2015

  13. [13]

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2020. Gshard: Scaling giant models with conditional computation and automatic sharding.arXiv preprint arXiv:2006.16668(2020)

    Pith/arXiv arXiv 2020

  14. [14]

    Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wenjie Li. 2024. Learning to rank in generative retrieval. InAAAI. AAAI press, 8716–8723

    2024

  15. [15]

    Gustavo Penha, Edoardo D’Amico, Marco De Nadai, Enrico Palumbo, Alexandre Tamborrino, Ali Vardasbi, Max Lefarov, Shawn Lin, Timothy Heath, Francesco Fabbri, et al. 2025. Semantic IDs for Joint Generative Search and Recommendation. InRecSys. 1296–1301

    2025

  16. [16]

    Gustavo Penha, Ali Vardasbi, Enrico Palumbo, Marco De Nadai, and Hugues Bouchard. 2024. Bridging Search and Recommendation in Generative Retrieval: Does One Task Help the Other?. InRecSys. ACM, 340–349

    2024

  17. [17]

    Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. 2021. AdapterFusion: Non-Destructive Task Composition for Transfer Learning. InEACL. ACL, 487–503

    2021

  18. [18]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research21, 140 (2020), 1–67

    2020

  19. [19]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. InNeurIPS

    2023

  20. [20]

    Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal V. Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, S...

  21. [21]

    Zhang Shengyu, Dong Linfeng, Li Xiaoya, Zhang Sen, Sun Xiaofei, Wang Shuhe, Li Jiwei, Runyi Hu, Zhang Tianwei, Fei Wu, et al. 2023. Instruction tuning for large language models: A survey.arXiv preprint arXiv:2308.10792(2023)

    arXiv 2023

  22. [22]

    Teng Shi, Zihua Si, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Dewei Leng, Yanan Niu, and Yang Song. 2024. UniSAR: Modeling User Transition Behaviors between Search and Recommendation. InSIGIR. ACM, 1029–1039

    2024

  23. [23]

    Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Enyun Yu. 2025. Unified Generative Search and Recommendation.arXiv preprint arXiv:2504.05730(2025)

    Pith/arXiv arXiv 2025

  24. [24]

    Zihua Si, Zhongxiang Sun, Xiao Zhang, Jun Xu, Xiaoxue Zang, Yang Song, Kun Gai, and Ji-Rong Wen. 2023. When search meets recommendation: Learning disentangled search representation for recommendation. InSIGIR. ACM, 1313– 1323

    2023

  25. [25]

    Weiwei Sun, Lingyong Yan, Zheng Chen, Shuaiqiang Wang, Haichao Zhu, Pengjie Ren, Zhumin Chen, Dawei Yin, Maarten de Rijke, and Zhaochun Ren. 2023. Learning to Tokenize for Generative Retrieval. InNeurIPS

    2023

  26. [26]

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable Tokenizer for LLM-based Genera- tive Recommendation.arXiv:2405.07314(2024)

    arXiv 2024

  27. [27]

    Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amir- reza Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, Eshaan Pathak, Giannis Karamanolakis, Haizhi Gary Lai, Ishan Purohit, Ishani Mondal, Jacob Anderson, Kirby Kuznia, Krima Doshi, Kuntal Kumar Pal, Maitreya Patel, Mehrad Moradshahi, Mihi...

    2022

  28. [28]

    Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M

    Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned Language Models are Zero-Shot Learners. InICLR. OpenReview.net

    2022

  29. [29]

    Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent Abilities of Large Language Models.Transactions on Machine Learning Research2022 (2022)

    2022

  30. [30]

    Shiguang Wu, Wenda Wei, Mengqi Zhang, Zhumin Chen, Jun Ma, Zhaochun Ren, Maarten de Rijke, and Pengjie Ren. 2024. Generative retrieval as multi-vector dense retrieval. InSIGIR. ACM, 1828–1838

    2024

  31. [31]

    Yucheng Xia, Yuhang Liu, Tianhao Li, Sihan He, Hong Chang, Yaqing Wang, Yongqing Zhang, and Wenyi Ge. 2024. Assessing parameter efficient methods for pre-trained language model in annotating scRNA-seq data.Methods228 (2024), 12–21

    2024

  32. [32]

    Jiayi Xie, Shang Liu, Gao Cong, and Zhenzhong Chen. 2024. UnifiedSSR: A Unified Framework of Sequential Search and Recommendation. InWWW. ACM, 3410–3419

    2024

  33. [33]

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate nearest neighbor nega- tive contrastive learning for dense text retrieval.arXiv preprint arXiv:2007.00808 (2020)

    Pith/arXiv arXiv 2020

  34. [34]

    Yaming Yang, Dilxat Muhtar, Yelong Shen, Yuefeng Zhan, Jianfeng Liu, Yujing Wang, Hao Sun, Weiwei Deng, Feng Sun, Qi Zhang, et al. 2025. Mtl-lora: Low-rank adaptation for multi-task learning. InAAAI, Vol. 39. 22010–22018

    2025

  35. [35]

    Jing Yao, Zhicheng Dou, Ruobing Xie, Yanxiong Lu, Zhiping Wang, and Ji-Rong Wen. 2021. USER: A unified information search and recommendation model based on integrated behavior sequence. InCIKM. ACM, 2373–2382

    2021

  36. [36]

    Bruce Croft

    Hamed Zamani and W. Bruce Croft. 2018. Joint Modeling and Optimization of Search and Recommendation. InDESIRE, Vol. 2167. CEUR-WS.org, 36–41

    2018

  37. [37]

    Xiaoyu Zhang, Ruobing Xie, Yougang Lyu, Xin Xin, Pengjie Ren, Mingfei Liang, Bo Zhang, Zhanhui Kang, Maarten de Rijke, and Zhaochun Ren. 2024. Towards empathetic conversational recommender systems. InRecSys. ACM, 84–93

    2024

  38. [38]

    Yuting Zhang, Yiqing Wu, Ruidong Han, Ying Sun, Yongchun Zhu, Xiang Li, Wei Lin, Fuzhen Zhuang, Zhulin An, and Yongjun Xu. 2024. Unified Dual-Intent Translation for Joint Modeling of Search and Recommendation. InKDD. ACM, 6291–6300

    2024

  39. [39]

    Jujia Zhao, Wenjie Wang, Chen Xu, Xiuying Chen, Zhaochun Ren, and Suzan Verberne. 2025. Unifying Search and Recommendation: A Generative Paradigm Inspired by Information Theory.arXiv preprint arXiv:2504.06714(2025)

    arXiv 2025

  40. [40]

    Jujia Zhao, Yumeng Wang, Zhaochun Ren, and Suzan Verberne. 2025. Model Meets Knowledge: Analyzing Knowledge Types for Conversational Recom- mender Systems. InProceedings of the Nineteenth ACM Conference on Recom- mender Systems. 802–811

    2025

  41. [41]

    Jiawei Zhao, Zhenyu Zhang, Beidi Chen, Zhangyang Wang, Anima Anandkumar, and Yuandong Tian. 2024. GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection. InICML. OpenReview.net

    2024

  42. [42]

    Kai Zhao, Yukun Zheng, Tao Zhuang, Xiang Li, and Xiaoyi Zeng. 2022. Joint learning of e-commerce search and recommendation with a unified graph neural network. InWSDM. 1461–1469

    2022

  43. [43]

    Yujia Zhou, Jing Yao, Ledell Wu, Zhicheng Dou, and Ji-Rong Wen. 2023. WebUl- tron: An ultimate retriever on webpages under the model-centric paradigm.IEEE Transactions on Knowledge and Data Engineering36, 9 (2023), 4996–5006. Unifying Search and Recommendation in LLMs via Gradient Multi-Subspace Tuning Conference’17, July 2017, Washington, DC, USA Table 4...

    2023

  44. [2022]

    Multitask Prompted Training Enables Zero-Shot Task Generalization. In ICLR. OpenReview.net

  45. [2024]

    InProceedings of the Annual Meeting of the Cognitive Science Society, Vol

    UNIFIT: A Unified Framework For Instruction Tuning To Improve Instruc- tion Following Ability For Large Language Models. InProceedings of the Annual Meeting of the Cognitive Science Society, Vol. 46