Pith. sign in

REVIEW 4 major objections 6 minor 71 references

Recommendation can run around the user, not inside one platform, by turning cross-platform behavior into a timed task state and complementary multi-source slates.

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 · grok-4.5

2026-07-30 12:44 UTC pith:O4XRIFCT

load-bearing objection Solid systems paper that names a real agent-era recsys setting and ships a coherent pipeline plus bench; headline gains are real inside the protocol but rest on an unvalidated LLM judge and synthetic trajectories. the 4 major comments →

arxiv 2607.23779 v1 pith:O4XRIFCT submitted 2026-07-26 cs.IR

ClawRec: A Claw-Native Recommender System

Chenghao Wu , Kesha Ou , Xiaolei Wang , Bowen Zheng , Bingqian Li , Enze Liu , Wayne Xin Zhao , Weitao Li
show 3 more authors
Long Zhang Sheng Chen Ji-Rong Wen
This is my paper
classification cs.IR
keywords recommender systemscross-platform recommendationpersonal agentsuser state modelingslate curationsource rolesnext-step recommendationbenchmark
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.

Most recommenders only see one site’s clicks and only rank that site’s catalog, even though real tasks unfold through searches, reading, and comparisons across many sources. This paper argues that personal agents with authorized cross-platform access make a different setup possible: recommendation centered on the user’s active task. It introduces Claw-native recommendation and builds ClawRec, which normalizes messy multi-source actions into evidence-linked events, maintains a temporally structured user state (what is active, cooling, expired, or suppressed), plans retrieval by functional source roles, and curates a non-redundant slate by marginal utility. To test the idea, it builds ClawRec-SimBench from sequences of life events realized as cross-platform browsing trajectories, then asks systems to support a held-out next step from a partial prefix. On that benchmark, ClawRec beats strong retrieval, LLM-ranking, and agent baselines on both recommendation ranking and how cleanly it recovers the current need versus stale history.

Core claim

ClawRec shows that a recommender designed for agent-accessible, cross-platform evidence—via unified events, evidence-linked temporal user state, role-aware planning, and marginal slate curation—can recover the active task and deliver complementary multi-source next-step recommendations more effectively than platform-local retrieval or generic agentic memory and search under matched budgets, reaching NDCG@20 of 0.6134 and Hit@20 of 0.6944 on ClawRec-SimBench.

What carries the argument

Evidence-linked, temporally structured user state: normalized multi-platform behaviors become provenance-backed records whose lifecycle (active → cooling → expired, plus suppression) separates the current task from durable preferences and stale noise; that state drives role-aware retrieval and marginal curation of complementary slates.

Load-bearing premise

The synthetic life-event trajectories and frozen LLM judge scores on the benchmark are faithful enough stand-ins for real human cross-platform tasks that the measured gains would still hold outside that generator–judge loop.

What would settle it

Run the same leave-one-out next-step protocol on real consented multi-site browsing logs with human relevance and state labels: if ClawRec no longer beats strong agentic and retrieval baselines on NDCG@20, Hit@20, and temporal state alignment under matched budgets, the central claim fails.

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

If this is right

  • Recommenders can target a unified complementary slate across search, guides, community experience, and products instead of one catalog’s ranking list.
  • User modeling should treat task status as first-class (active, cooling, expired, suppressed) rather than only accumulating longer history.
  • Retrieval should be planned by functional source role for the active task, not only by topical similarity to the last query.
  • Slate quality should be optimized for marginal task support so redundant near-duplicates are demoted even when each item is relevant alone.
  • Benchmarks for personal agents need latent goals inferred from behavior prefixes, not only explicit instruction-following success.

Where Pith is reading between the lines

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

  • If personal agents become the default interface, platform-local ranking metrics may understate value that only appears when complementary support is assembled across sources.
  • The same temporal-status machinery could transfer to other agent memory designs that currently blur finished tasks with lasting preferences.
  • Privacy and audit handles are not side features here: without provenance from evidence to slate, cross-platform state becomes hard to correct or trust in deployment.
  • A natural next stress test is longitudinal live use where feedback is sparse, noisy, and delayed rather than a clean held-out next click.

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

4 major / 6 minor

Summary. The paper proposes "Claw-native" recommendation: rather than ranking items within a single platform's candidate space, a recommender operating inside a personal agent (OpenClaw-style) consumes authorized cross-platform behavior traces and produces a unified, complementary slate spanning sources and content forms. The authors instantiate this with ClawRec, which (i) normalizes heterogeneous traces into provenance-linked "unified events," (ii) maintains an explicit evidence-linked user state with an active→cooling→expired/suppressed lifecycle, (iii) plans retrieval around functional source roles with behavior-adaptive source extension, and (iv) curates the final slate by marginal utility. To evaluate, they build ClawRec-SimBench: 109 simulated users × 5 life events each, where an LLM intent generator and a GUI agent synthesize cross-platform browsing trajectories, and a leave-one-out protocol asks systems to recommend content supporting a held-out final behavior. Under matched LLM backbone and budgets, ClawRec reports NDCG@20 0.6134 (+0.1126 over OpenClaw w/o ClawRec) and Hit@20 0.6944 (+0.0854 over Persona2Web-style), with ablations attributing gains to each module. All semantic metrics (relevance, SCS, diagnostics) are scored by a frozen GPT-5.5 judge.

Significance. If the results hold, the paper contributes a genuinely new problem formulation (task-level next-step recommendation over agent-accessible cross-platform evidence, distinct from both cross-domain recommendation and goal-conditioned web agents) and a coherent system design with a clear mechanistic story. The experimental hygiene is above average for this area: matched backbone/decoding/budgets across baselines, a user-level dev/test split with frozen judge prompts, blinding and pooling in the judging protocol, four well-targeted ablations whose effect profiles match the claimed function of each module (e.g., removing marginal curation collapses NDCG@20 by ~0.10 while leaving Hit@20 nearly intact), and released code and data. These are real strengths. The central risk is measurement validity, not internal logic: the environment, the targets, and every semantic score are LLM-mediated artifacts of the same research stack, and the judge calibration that underwrites all headline numbers is asserted rather than quantified.

major comments (4)
  1. [§4.1.2, Table 2] §4.1.2 (Judging Protocol): every semantic measurement in the paper — rel(ri,h) feeding NDCG@20/Hit@20, the four SCS dimensions, and the three diagnostics — comes from a frozen GPT-5.5 judge, yet human–judge agreement is never quantified. The text states that paired human and LLM scores were collected on the 20-user dev split and used to refine the judge prompt, but no agreement statistic (Cohen's/weighted kappa, Spearman correlation, per-dimension mean bias) is reported anywhere. Note that §3.4.2/Table 1b validates trajectory quality, not judge fidelity, so it does not fill this gap. A systematic per-method bias of a few tenths of a relevance point would be enough to materially move the +0.1126 NDCG@20 and +0.0854 Hit@20 gaps. The revision must report the dev-split agreement statistics per metric/dimension, and ideally re-score a sample of test instances (or at least the head-to-head Cla
  2. [Appendix B (Evaluation Prompts), §4.1.2] Appendix B + §4.1.2: the scoring rubrics structurally mirror ClawRec's own architecture, which creates a correctness-risk independent of judge accuracy. The Recommendation Relevance rubric explicitly instructs the judge to reward 'source role' fit — a concept only ClawRec (and its source-role records) is designed around; the SCS dimensions Temporal Alignment and Calibration are exactly the properties the temporal-lifecycle and patch-validation modules are built to produce; and slate complementarity is what marginal curation optimizes. Baselines not architected around these concepts are thus scored against criteria embodying ClawRec's design philosophy, and an LLM judge may additionally favor well-structured, provenance-articulated cards. The pooling/blinding protocol is good and I do not suggest anything improper, but the paper should (a) acknowledge this alignment explicitly, and (b) pr
  3. [§4.2–4.3, Tables 2–3] Tables 2–3, Figures 4–6: no uncertainty estimates are reported anywhere. All numbers appear to be single runs with decoding temperature 1.0 (§4.1.3), and the Hit@20 gap of +0.0854 corresponds to roughly 38 instances over 445 test events — plausibly within run-to-run or judge-sampling noise, especially since the LLM judge itself is stochastic at temperature 1.0. Please report bootstrap confidence intervals over test instances for the headline metrics, paired significance tests against the two strongest baselines, and at least judge-score variance over repeated scoring passes. Relatedly, Hit@20 sets τ=3, which counts a 'plausible partial substitute' as a hit; given how much weight Hit@20 carries in the abstract, a sensitivity analysis at τ=4 would substantially strengthen (or appropriately temper) that claim.
  4. [§3.2, §3.4.2, §6] §3.2, §6: the benchmark's trajectories are synthesized by an LLM intent generator executed by a GUI agent, and the validity evidence (§3.4.2) is thin: 50 sampled events, 2 annotators, no inter-annotator agreement reported, and no evidence that the generated behavior distributions (Fig. 3) match real cross-platform seeking. The conclusion does acknowledge the simulated setting, which is creditable, but the abstract and §1 frame the results more strongly ('demonstrating the value of Claw-native recommendation'). Since simulator and judge come from the same research stack as the system, measured gains may partly reflect alignment with the generator–judge loop rather than transferable recommendation quality. A concrete, in-scope mitigation: a small human study (even tens of real users' retrospective next-step judgments on a subset of instances, or human-authored prefixes) showing the method
minor comments (6)
  1. [Table 2, Abstract] Table 2: the abstract's '+0.1126 / +0.0854' compares against different baselines for the two metrics (OpenClaw w/o ClawRec for NDCG@20, Persona2Web-style for Hit@20). This is legitimate but should be flagged at first mention; readers will otherwise assume a single strongest baseline. It is also worth one sentence explaining why Persona2Web attains higher Hit@20 but much lower NDCG@20 than OpenClaw (0.6090/0.4175 vs 0.5730/0.5008).
  2. [Figures 4–6] Figures 4–6: no error bars and no per-group instance counts (e.g., n per event position in Fig. 5; n for observed- vs unobserved-source in Fig. 6b). The source-shift subgroup sizes matter for interpreting the 'larger advantage on unobserved-source' claim.
  3. [§4.1.1] Baseline adaptations (§4.1.1): Persona2Web and MARS are described as 'task adaptations' and 'style' re-implementations. Please add a short paragraph on what was changed relative to the published systems and any ablation evidence that the adaptations are faithful; otherwise the comparison is against approximations of unknown fidelity.
  4. [§4.1.3] No latency/cost accounting: ClawRec's pipeline (event construction, state patching, planning, per-item assessment, marginal curation) clearly issues more LLM calls than the baselines. Even under matched retrieval/display budgets, a token/call-count comparison per decision step would make the efficiency picture honest and is easy to add from logs.
  5. [Table 2] Table 2 uses '–' for SCS on direct-LLM baselines, but the text says SCS is inapplicable only because methods 'do not produce an inspectable user state.' LLM Planner + LLMRank does reason over history; clarifying why its implicit state cannot be scored (or scoring a textual summary of it) would preempt the concern that SCS is only defined for state-exposing architectures.
  6. [Figure 2, Abstract] Figure 2 label typo: 'Query Paturalness' should be 'Query Naturalness'. Also 'introduceClawRec-SimBench' (missing space) in the abstract and 'losesubstantial' phrasing in §4.2 should be checked in a copy-edit pass.

Circularity Check

0 steps flagged

No definitional or fitted-input circularity; empirical gains on a self-built synthetic bench with a co-designed LLM judge are a validity risk, not a by-construction reduction.

full rationale

ClawRec is an empirical systems paper, not a first-principles derivation. Its load-bearing claims are measured NDCG@20/Hit@20 and SCS improvements of an engineered pipeline (unified events → temporal state → role-aware planning → marginal curation) over matched baselines on ClawRec-SimBench. Those metrics are not algebraically forced by fitted constants, nor is any module defined in terms of the quantity it is said to predict. There is no uniqueness theorem, ansatz, or self-citation chain that forbids alternatives and thereby forces the result. Self-citations in related work (e.g., prior agent/recsys papers from overlapping authors) are ordinary background and are not load-bearing for the reported deltas. The main methodological concern—synthetic life-event trajectories plus a frozen GPT-5.5 judge whose rubric names source-role fit, temporal alignment, and calibration, concepts ClawRec is built to optimize—is real but belongs under external validity / instrument bias, not circularity: baselines face the same bench and judge, IDCG is pooled across methods, and ablations still move the metrics when mechanisms are removed. Per the circularity criteria, that is at most a minor co-design entanglement, not Eq.X = Eq.Y by construction. Score 1 reflects only that mild entanglement; steps are empty because no quoted step reduces a claimed prediction to its inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 5 invented entities

Claims rest on engineering and evaluation assumptions rather than formal theorems: that authorized cross-platform traces are available; that synthetic life-event trajectories represent real tasks; that LLM judges track human task-level usefulness; and that functional source roles and marginal utility are the right planning/curation objectives. No physical constants; free choices are budgets, thresholds, and model picks.

free parameters (6)
  • display_budget_n = 20
    Fixed top-20 slate size for all metrics and methods; changes Hit/NDCG operating point.
  • retrieval_budget = 10
    Cap of 10 search tasks per decision; constrains coverage equally but is hand-chosen.
  • Hit_relevance_threshold_tau = 3
    Hit@20 counts success if max rel ≥ τ; τ=3 defines 'partial support' on 1–5 rubric.
  • platform_intent_score_filters = total≥9; fit/naturalness≥3
    Simulator keeps candidates with total score ≥9 and platform-fit/naturalness ≥3 on 1–5 scales; shapes trajectory distribution.
  • LLM_decoding_temperature = 1.0
    Generation and judge calls use temperature 1.0; affects variance of plans and scores.
  • backbone_and_judge_models = DeepSeek-V4-Flash / GPT-5.5
    DeepSeek-V4-Flash for systems; GPT-5.5 for semantic metrics—capacity choices that load-bear reported gaps.
axioms (6)
  • domain assumption Authorized claw-style agents can persistently observe multi-source user behaviors and return cross-source items under user permissions.
    Foundational setting in Introduction and task formulation; without it Claw-native rec is undefined.
  • domain assumption Task-level semantic usefulness (not exact URL/item ID match) is the correct ranking objective for next-step support.
    Section 3.3–3.4 and Metrics; enables graded rel but weakens classical recsys falsifiability.
  • ad hoc to paper Synthetic life events + GUI-grounded traces preserve enough structure of real cross-platform seeking for system ranking to be meaningful.
    Entire empirical case rests on ClawRec-SimBench (Section 3); authors note deployment limits in Conclusion.
  • ad hoc to paper Calibrated LLM-as-judge scores are adequate substitutes for human labels on state quality and recommendation relevance at test time.
    Section 4.1.2: human+LLM only on 20-user dev; test metrics are LLM-only after prompt freeze.
  • domain assumption Marginal/diversity-aware selection (MMR-style) improves slate utility over independent item ranking under a shared display budget.
    Section 2.5 cites Carbonell & Goldstein MMR; ablation w/o marginal curation drops NDCG sharply.
  • standard math Standard IR ranking metrics (NDCG, Hit) on graded task relevance are appropriate primary outcomes.
    Uses classical DCG/NDCG definitions with shared IDCG pool (Järvelin & Kekäläinen).
invented entities (5)
  • Claw-native recommender system (paradigm) no independent evidence
    purpose: Name the setting where rec operates on agent-accessible user environment with multi-source complementary slates.
    Framing device; not an empirical object with mass/charge, but organizes claims.
  • Unified event (provenance-linked evidence record) no independent evidence
    purpose: Normalize heterogeneous behaviors before state update.
    Internal representation; validated only via ablation performance drop when removed.
  • Evidence-linked user state (task/preference/source-role slots with active→cooling→expired/suppressed lifecycle) no independent evidence
    purpose: Separate active task from preferences and stale/negative signals with audit handles.
    Core invented schema; inspectable SCS dimensions are defined against this structure.
  • Support type / support objective / source role / retrieval task τ no independent evidence
    purpose: Plan functionally appropriate multi-source retrieval rather than query continuation.
    Planning ontology introduced in §2.4; supported by source-role match diagnostics on their bench.
  • ClawRec-SimBench (life-event, cross-platform trajectory benchmark) no independent evidence
    purpose: Make latent-task next-step recommendation evaluable.
    New dataset/protocol; quality validated by small human sample, not external standard corpus.

pith-pipeline@v1.2.0-grok45-kimik3 · 31352 in / 4250 out tokens · 74152 ms · 2026-07-30T12:44:19.873499+00:00 · methodology

0 comments
read the original abstract

Recommender systems have become integral to navigating the modern digital ecosystem. Yet most deployed systems remain confined within single-platform boundaries, observing localized interaction traces and ranking items from isolated candidate spaces. This design is poorly suited to real-world tasks that unfold through searches, content consumption, and comparisons across multiple information sources. Claw-style personal agents, with persistent access to authorized cross-platform context, create an opportunity for recommendation to operate around the user rather than any single platform. In this paper, we introduce Claw-native recommender systems, a new paradigm that moves beyond platform-local ranking to produce unified, complementary recommendation slates spanning diverse sources and content forms. To instantiate this paradigm, we present ClawRec, the first recommender system designed to operate natively in this environment. ClawRec maintains an evidence-linked, temporally structured user state that connects cross-platform behaviors with cross-source recommendations. It organizes retrieval around functional source roles and selects candidates according to their marginal utility, producing non-redundant slates aligned with the user's active task. To enable rigorous evaluation, we introduce ClawRec-SimBench, a benchmark constructed from sequences of concrete life events and cross-platform behavior trajectories. Experiments show that ClawRec outperforms the strongest baselines, achieving an NDCG@20 of 0.6134 (+0.1126) and a Hit@20 of 0.6944 (+0.0854), while also improving user state quality and temporal alignment. Our code and dataset are available at https://github.com/RUCAIBox/ClawRec.

Figures

Figures reproduced from arXiv: 2607.23779 by Bingqian Li, Bowen Zheng, Chenghao Wu, Enze Liu, Ji-Rong Wen, Kesha Ou, Long Zhang, Sheng Chen, Wayne Xin Zhao, Weitao Li, Xiaolei Wang.

Figure 1
Figure 1. Figure 1: System overview of ClawRec. Cross-platform behaviors and explicit OpenClaw context [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ClawRec-SimBench construction. Latent profiles and life events define [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Benchmark distribution of simulated user behaviors and source roles. Browse dominates [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: SCS and Hit@20 across event posi￾tions. Both metrics generally improve as cross￾event evidence accumulates. especially when the appropriate source is role-compatible but not directly observed in the current behavior prefix. Marginal Curation Improves Slate Ranking. Marginal curation contributes mainly to slate-level ranking and organization. Removing it leaves Hit@20 close to the full model, decreasing onl… view at source ↗
Figure 6
Figure 6. Figure 6: Source-role-aware retrieval analysis. ClawRec achieves the strongest source-role alignment [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case study of ClawRec’s recommendation process, from cross-source behavioral evidence [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

71 extracted references · 10 linked inside Pith

  1. [1]

    BPR: Bayesian personalized ranking from implicit feedback

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. BPR: Bayesian personalized ranking from implicit feedback. InProceedings of the Twenty-Fifth 21 Conference on Uncertainty in Artificial Intelligence, UAI ’09, pages 452–461, Arlington, Virginia, USA, 2009. AUAI Press

    2009

  2. [2]

    Self-Attentive Sequential Recommendation

    Wang-Cheng Kang and Julian McAuley. Self-Attentive Sequential Recommendation. In2018 IEEE International Conference on Data Mining (ICDM), pages 197–206, 2018

    2018

  3. [3]

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer. InProceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM ’19, pages 1441–1450, New York, NY , USA, 2019. Association for Computing Machinery

    2019

  4. [4]

    Multi-behavior Recommendation with Graph Convolutional Networks

    Bowen Jin, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. Multi-behavior Recommendation with Graph Convolutional Networks. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’20, pages 659–668, New York, NY , USA, 2020. Association for Computing Machinery

    2020

  5. [5]

    A Survey on Cross- domain Recommendation: Taxonomies, Methods, and Future Directions.ACM Transactions on Information Systems, 41(2):42:1–42:39, 2022

    Tianzi Zang, Yanmin Zhu, Haobing Liu, Ruohan Zhang, and Jiadi Yu. A Survey on Cross- domain Recommendation: Taxonomies, Methods, and Future Directions.ACM Transactions on Information Systems, 41(2):42:1–42:39, 2022

    2022

  6. [6]

    STAN: Stage-Adaptive Network for Multi-Task Recommendation by Learning User Lifecycle-Based Representation

    Wanda Li, Wenhao Zheng, Xuanji Xiao, and Suhang Wang. STAN: Stage-Adaptive Network for Multi-Task Recommendation by Learning User Lifecycle-Based Representation. InProceedings of the 17th ACM Conference on Recommender Systems, RecSys ’23, pages 602–612, New York, NY , USA, 2023. Association for Computing Machinery

    2023

  7. [7]

    A survey on large language models for recommendation.World Wide Web, 27(5), 2024

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. A survey on large language models for recommendation.World Wide Web, 27(5), 2024

    2024

  8. [8]

    A Survey on LLM-powered Agents for Recommender Systems

    Qiyao Peng, Hongtao Liu, Hua Huang, Jian Yang, Qing Yang, and Minglai Shao. A Survey on LLM-powered Agents for Recommender Systems. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors,Findings of the Association for Computa- tional Linguistics: EMNLP 2025, pages 11574–11583, Suzhou, China, 2025. Association for Compu...

    2025

  9. [9]

    OpenClaw: Personal ai assistant

    OpenClaw. OpenClaw: Personal ai assistant. https://github.com/openclaw/openclaw,

  10. [10]

    The use of MMR, diversity-based reranking for reordering documents and producing summaries

    Jaime Carbonell and Jade Goldstein. The use of MMR, diversity-based reranking for reordering documents and producing summaries. InProceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’98, pages 335–336, New York, NY , USA, 1998. Association for Computing Machinery

    1998

  11. [11]

    Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig

    Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. WebArena: A Realistic Web Environment for Building Autonomous Agents. InThe Twelfth International Conference on Learning Representations, 2023

    2023

  12. [12]

    MIND2WEB: Towards a generalist agent for the web

    Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, and Yu Su. MIND2WEB: Towards a generalist agent for the web. InProceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, pages 28091–28114, Red Hook, NY , USA, 2023. Curran Associates Inc

    2023

  13. [13]

    OSWORLD: Benchmarking multimodal agents for open-ended tasks in real computer environments

    Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. OSWORLD: Benchmarking multimodal agents for open-ended tasks in real computer environments. InProceedings of the 38th Internat...

    2024

  14. [14]

    Large language models empowered personalized web agents

    Hongru Cai, Yongqi Li, Wenjie Wang, Fengbin Zhu, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. Large language models empowered personalized web agents. InProceedings of the ACM on Web Conference 2025, pages 198–215, 2025

    2025

  15. [15]

    Chrome DevTools MCP

    ChromeDevTools. Chrome DevTools MCP. https://github.com/ChromeDevTools/ chrome-devtools-mcp, 2026. GitHub repository. Accessed: 2026-06-30. 22

    2026

  16. [16]

    Automatic query expansion using SMART: TREC 3

    Chris Buckley, Gerard Salton, James Allan, and Amit Singhal. Automatic query expansion using SMART: TREC 3. In Donna K. Harman, editor,Proceedings of the Third Text Retrieval Conference, TREC 1994, Gaithersburg, Maryland, USA, November 2–4, 1994, volume 500–225 ofNIST Special Publication, pages 69–80. National Institute of Standards and Technology (NIST), 1994

    1994

  17. [17]

    Robertson, S

    S. Robertson, S. Walker, S. Jones, M. Hancock-Beaulieu, and M. Gatford. Okapi at TREC-3. InText Retrieval Conference, 1994

    1994

  18. [18]

    Large Language Models are Zero-Shot Rankers for Recommender Systems

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. Large Language Models are Zero-Shot Rankers for Recommender Systems. InAdvances in Information Retrieval: 46th European Conference on Information Retrieval, ECIR 2024, Glasgow, UK, March 24–28, 2024, Proceedings, Part II, pages 364–381, Berlin, Heidelberg, 202...

    2024

  19. [19]

    Persona2Web: Benchmarking Personalized Web Agents for Contextual Reasoning with User History.arXiv preprint arXiv:2602.17003, 2026

    Serin Kim, Sangam Lee, and Dongha Lee. Persona2Web: Benchmarking Personalized Web Agents for Contextual Reasoning with User History.arXiv preprint arXiv:2602.17003, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    Agentic Recommender System with Hierarchical Belief-State Memory.arXiv preprint arXiv:2605.14401, 2026

    Xiang Shen, Yuhang Zhou, Yifan Wu, Zhuokai Zhao, Siyu Lin, Lei Huang, Qianqian Zhong, Lizhu Zhang, Benyu Zhang, Xiangjun Fan, and Hong Yan. Agentic Recommender System with Hierarchical Belief-State Memory.arXiv preprint arXiv:2605.14401, 2026

    Pith/arXiv arXiv 2026

  21. [21]

    Cumulated gain-based evaluation of IR techniques

    Kalervo Järvelin and Jaana Kekäläinen. Cumulated gain-based evaluation of IR techniques. ACM Transactions on Information Systems (TOIS), 20(4):422–446, 2002

    2002

  22. [22]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-bench and Chatbot Arena. InProceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, pages 46595–46623,...

    2023

  23. [23]

    G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment. In Houda Bouamor, Juan Pino, and Kalika Bali, editors,Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 2511–2522, Singapore, 2023. Association for Computational Linguistics

    2023

  24. [24]

    DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence.arXiv preprint arXiv:2606.19348, 2026

    DeepSeek-AI, Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chengyu Hou, Chenhao Xu, Chenze Shao, Chong Ruan, Conner Sun, Damai Dai, et al. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence.arXiv preprint arXiv:2606...

    arXiv 2026

  25. [25]

    GPT-5.5 system card

    OpenAI. GPT-5.5 system card. https://openai.com/index/gpt-5-5-system-card/ ,

  26. [26]

    Can movies and books collaborate? cross-domain collaborative filtering for sparsity reduction

    Bin Li, Qiang Yang, and Xiangyang Xue. Can movies and books collaborate? cross-domain collaborative filtering for sparsity reduction. InProceedings of the 21st International Joint Conference on Artificial Intelligence, IJCAI’09, pages 2052–2057, San Francisco, CA, USA,

    2052

  27. [27]

    Accessed: 2026-07-01

    2026

  28. [28]

    Xiang, Nathan N

    Weike Pan, Evan W. Xiang, Nathan N. Liu, and Qiang Yang. Transfer learning in collaborative filtering for sparsity reduction. InProceedings of the Twenty-Fourth AAAI Conference on Artificial Intelligence, AAAI’10, pages 230–235, Atlanta, Georgia, 2010. AAAI Press

    2010

  29. [29]

    A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems

    Ali Mamdouh Elkahky, Yang Song, and Xiaodong He. A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems. InProceedings of the 24th International Conference on World Wide Web, WWW ’15, pages 278–288, Republic and Canton of Geneva, CHE, 2015. International World Wide Web Conferences Steering Committee

    2015

  30. [30]

    Transfer learning for collaborative filtering via a rating-matrix generative model

    Bin Li, Qiang Yang, and Xiangyang Xue. Transfer learning for collaborative filtering via a rating-matrix generative model. InProceedings of the 26th Annual International Conference on Machine Learning, ICML ’09, pages 617–624, New York, NY , USA, 2009. Association for Computing Machinery

    2009

  31. [31]

    A deep framework for cross-domain and cross-system recommendations

    Feng Zhu, Yan Wang, Chaochao Chen, Guanfeng Liu, Mehmet Orgun, and Jia Wu. A deep framework for cross-domain and cross-system recommendations. InProceedings of the 27th In- ternational Joint Conference on Artificial Intelligence, IJCAI’18, pages 3711–3717, Stockholm, Sweden, 2018. AAAI Press

    2018

  32. [32]

    CoNet: Collaborative Cross Networks for Cross- Domain Recommendation

    Guangneng Hu, Yu Zhang, and Qiang Yang. CoNet: Collaborative Cross Networks for Cross- Domain Recommendation. InProceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM ’18, pages 667–676, New York, NY , USA,

  33. [33]

    Cross-domain recommendation: An embedding and mapping approach

    Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. Cross-domain recommendation: An embedding and mapping approach. InProceedings of the 26th International Joint Conference on Artificial Intelligence, IJCAI’17, pages 2464–2470, Melbourne, Australia, 2017. AAAI Press. 23

    2017

  34. [34]

    CATN: Cross-Domain Recommendation for Cold-Start Users via Aspect Transfer Network

    Cheng Zhao, Chenliang Li, Rong Xiao, Hongbo Deng, and Aixin Sun. CATN: Cross-Domain Recommendation for Cold-Start Users via Aspect Transfer Network. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’20, pages 229–238, New York, NY , USA, 2020. Association for Computing Machinery

    2020

  35. [35]

    LSTM networks for online cross-network recommen- dations

    Dilruk Perera and Roger Zimmermann. LSTM networks for online cross-network recommen- dations. InProceedings of the 27th International Joint Conference on Artificial Intelligence, IJCAI’18, pages 3825–3833, Stockholm, Sweden, 2018. AAAI Press

    2018

  36. [36]

    One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction

    Xiang-Rong Sheng, Liqin Zhao, Guorui Zhou, Xinyao Ding, Binding Dai, Qiang Luo, Siran Yang, Jingshan Lv, Chi Zhang, Hongbo Deng, and Xiaoqiang Zhu. One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction. InProceedings of the 30th ACM International Conference on Information & Knowledge Management, CIKM ’21, pages 4104–41...

    2021

  37. [37]

    Personalized Transfer of User Preferences for Cross-domain Recommendation

    Yongchun Zhu, Zhenwei Tang, Yudan Liu, Fuzhen Zhuang, Ruobing Xie, Xu Zhang, Leyu Lin, and Qing He. Personalized Transfer of User Preferences for Cross-domain Recommendation. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, WSDM ’22, pages 1507–1515, New York, NY , USA, 2022. Association for Computing Machinery

    2022

  38. [38]

    Large language model enhanced recommender systems: A survey.arXiv preprint arXiv:2412.13432, 2024

    Qidong Liu, Xiangyu Zhao, Yuhao Wang, Yejing Wang, Zijian Zhang, Yuqi Sun, Xiang Li, Maolin Wang, Pengyue Jia, Chong Chen, et al. Large language model enhanced recommender systems: A survey.arXiv preprint arXiv:2412.13432, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5)

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). InProceedings of the 16th ACM Conference on Recommender Systems, pages 299–315, 2022

    2022

  40. [40]

    Chat-rec: Towards interactive and explainable llms-augmented recommender system.arXiv preprint arXiv:2303.14524, 2023

    Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. Chat-rec: Towards interactive and explainable llms-augmented recommender system.arXiv preprint arXiv:2303.14524, 2023

    Pith/arXiv arXiv 2023

  41. [41]

    MAMDR: A Model Agnostic Learning Framework for Multi-Domain Recommendation

    Linhao Luo, Yumeng Li, Buyu Gao, Shuai Tang, Sinan Wang, Jiancheng Li, Tanchao Zhu, Jiancai Liu, Zhao Li, and Shirui Pan. MAMDR: A Model Agnostic Learning Framework for Multi-Domain Recommendation. In2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 3079–3092, 2023

    2023

  42. [42]

    Towards open-world recommendation with knowledge augmentation from large language models

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. Towards open-world recommendation with knowledge augmentation from large language models. InProceedings of the 18th ACM Conference on Recommender Systems, pages 12–22, 2024

    2024

  43. [43]

    Adapting large language models by integrating collaborative semantics for recommen- dation

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. Adapting large language models by integrating collaborative semantics for recommen- dation. In2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 1435–1448. IEEE, 2024

    2024

  44. [44]

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems, pages 1007–1014, 2023

    2023

  45. [45]

    Recommendation as instruction following: A large language model empowered recommendation approach.ACM Transactions on Information Systems, 43(5):1–37, 2026

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. Recommendation as instruction following: A large language model empowered recommendation approach.ACM Transactions on Information Systems, 43(5):1–37, 2026

    2026

  46. [46]

    Recommender systems meet large language model agents: A survey.Foundations and Trends® in Privacy and Security, 7(4):247–396, 2025

    Xi Zhu, Yu Wang, Hang Gao, Wujiang Xu, Chen Wang, Zhiwei Liu, Kun Wang, Mingyu Jin, Linsey Pang, Qingsong Weng, et al. Recommender systems meet large language model agents: A survey.Foundations and Trends® in Privacy and Security, 7(4):247–396, 2025

    2025

  47. [47]

    A survey of large language model empowered agents for recommendation and search: Towards next-generation information retrieval.arXiv preprint arXiv:2503.05659, 2025

    Yu Zhang, Shutong Qiao, Jiaqi Zhang, Tzu-Heng Lin, Chen Gao, and Yong Li. A survey of large language model empowered agents for recommendation and search: Towards next-generation information retrieval.arXiv preprint arXiv:2503.05659, 2025

    Pith/arXiv arXiv 2025

  48. [48]

    Starec: An efficient agent framework for recommender systems via autonomous deliberate reasoning

    Chenghao Wu, Ruiyang Ren, Junjie Zhang, Ruirui Wang, Zhongrui Ma, Qi Ye, and Wayne Xin Zhao. Starec: An efficient agent framework for recommender systems via autonomous deliberate reasoning. InProceedings of the 34th ACM International Conference on Information and Knowledge Management, pages 3355–3365, 2025

    2025

  49. [49]

    On softmax direct preference optimization for recommendation.Advances in Neural Information Processing Systems, 37:27463–27489, 2024

    Yuxin Chen, Junfei Tan, An Zhang, Zhengyi Yang, Leheng Sheng, Enzhi Zhang, Xiang Wang, and Tat-Seng Chua. On softmax direct preference optimization for recommendation.Advances in Neural Information Processing Systems, 37:27463–27489, 2024. 24

    2024

  50. [50]

    Mgfrec: Towards reinforced reasoning recommendation with multiple groundings and feedback

    Shihao Cai, Chongming Gao, Haoyan Liu, Wentao Shi, Jianshan Sun, Ruiming Tang, and Fuli Feng. Mgfrec: Towards reinforced reasoning recommendation with multiple groundings and feedback. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, pages 49–58, 2026

    2026

  51. [51]

    Deep research for recommender systems.arXiv preprint arXiv:2603.07605, 2026

    Kesha Ou, Chenghao Wu, Xiaolei Wang, Bowen Zheng, Wayne Xin Zhao, Weitao Li, Long Zhang, Sheng Chen, and Ji-Rong Wen. Deep research for recommender systems.arXiv preprint arXiv:2603.07605, 2026

    arXiv 2026

  52. [52]

    Agentcf: Collaborative learning with autonomous language agents for recommender systems

    Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. Agentcf: Collaborative learning with autonomous language agents for recommender systems. InProceedings of the ACM Web Conference 2024, pages 3679–3689, 2024

    2024

  53. [53]

    Jiani Huang, Xingchen Zou, Lianghao Xia, and Qing Li. Mr. rec: Synergizing memory and rea- soning for personalized recommendation assistant with llms.arXiv preprint arXiv:2510.14629, 2025

    arXiv 2025

  54. [54]

    Multi-agent collaborative filtering: Orchestrating users and items for agentic recommendations

    Yu Xia, Sungchul Kim, Tong Yu, Ryan A Rossi, and Julian McAuley. Multi-agent collaborative filtering: Orchestrating users and items for agentic recommendations. InProceedings of the ACM Web Conference 2026, pages 8649–8652, 2026

    2026

  55. [55]

    Webshop: Towards scalable real-world web interaction with grounded language agents.Advances in Neural Information Processing Systems, 35:20744–20757, 2022

    Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents.Advances in Neural Information Processing Systems, 35:20744–20757, 2022

    2022

  56. [56]

    Mind2web 2: Evaluating agentic search with agent-as-a-judge.Advances in Neural Information Processing Systems, 38, 2026

    Boyu Gou, Zanming Huang, Yuting Ning, Yu Gu, Michael Lin, Weijian Qi, Andrei Kopanev, Botao Yu, Bernal Jimenez Gutierrez, Yiheng Shu, et al. Mind2web 2: Evaluating agentic search with agent-as-a-judge.Advances in Neural Information Processing Systems, 38, 2026

    2026

  57. [57]

    AgentCF++: Memory-enhanced LLM-based agents for popularity-aware cross-domain recommendations

    Jiahao Liu, Shengkang Gu, Dongsheng Li, Guangping Zhang, Mingzhe Han, Hansu Gu, Peng Zhang, Tun Lu, Li Shang, and Ning Gu. AgentCF++: Memory-enhanced LLM-based agents for popularity-aware cross-domain recommendations. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2566–2571, 2025

    2025

  58. [58]

    Webvoyager: Building an end-to-end web agent with large multimodal models

    Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. Webvoyager: Building an end-to-end web agent with large multimodal models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6864–6890, 2024

    2024

  59. [59]

    Memorybank: Enhancing large language models with long-term memory

    Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19724–19731, 2024

    2024

  60. [60]

    Long- memeval: Benchmarking chat assistants on long-term interactive memory.arXiv preprint arXiv:2410.10813, 2024

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Long- memeval: Benchmarking chat assistants on long-term interactive memory.arXiv preprint arXiv:2410.10813, 2024

    Pith/arXiv arXiv 2024

  61. [61]

    Weblinx: Real-world website navigation with multi-turn dialogue.arXiv preprint arXiv:2402.05930, 2024

    Xing Han Lù, Zdenˇek Kasner, and Siva Reddy. Weblinx: Real-world website navigation with multi-turn dialogue.arXiv preprint arXiv:2402.05930, 2024

    arXiv 2024

  62. [62]

    Mem0: Building production-ready ai agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production-ready ai agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    Pith/arXiv arXiv 2025

  63. [63]

    Lamp: When large language models meet personalization

    Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. Lamp: When large language models meet personalization. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7370–7392, 2024

    2024

  64. [64]

    Longlamp: A benchmark for personalized long-form text generation.arXiv preprint arXiv:2407.11016, 2024

    Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A Rossi, Franck Der- noncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, et al. Longlamp: A benchmark for personalized long-form text generation.arXiv preprint arXiv:2407.11016, 2024

    Pith/arXiv arXiv 2024

  65. [65]

    A-mem: Agentic memory for llm agents.Advances in Neural Information Processing Systems, 38:17577– 17604, 2026

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-mem: Agentic memory for llm agents.Advances in Neural Information Processing Systems, 38:17577– 17604, 2026. 25

    2026

  66. [66]

    Do llms recognize your preferences? evaluating personalized preference following in llms.arXiv preprint arXiv:2502.09597, 2025

    Siyan Zhao, Mingyi Hong, Yang Liu, Devamanyu Hazarika, and Kaixiang Lin. Do llms recognize your preferences? evaluating personalized preference following in llms.arXiv preprint arXiv:2502.09597, 2025

    Pith/arXiv arXiv 2025

  67. [67]

    Learning personalized agents from human feedback.arXiv preprint arXiv:2602.16173, 2026

    Kaiqu Liang, Julia Kruk, Shengyi Qian, Xianjun Yang, Shengjie Bi, Yuanshun Yao, Shaoliang Nie, Mingyang Zhang, Lijuan Liu, Jaime Fernández Fisac, et al. Learning personalized agents from human feedback.arXiv preprint arXiv:2602.16173, 2026. 26 A Privacy and Audit Controls Privacy Controls.ClawRec uses local-first processing for behavior evidence. Raw brow...

    arXiv 2026

  68. [69]

    Apollonion: Profile-centric dialog agent.arXiv preprint arXiv:2404.08692, 2024

    Shangyu Chen, Zibo Zhao, Yuanyuan Zhao, and Xiang Li. Apollonion: Profile-centric dialog agent.arXiv preprint arXiv:2404.08692, 2024

    Pith/arXiv arXiv 2024

  69. [2009]

    Morgan Kaufmann Publishers Inc

  70. [2018]

    Association for Computing Machinery

  71. [2026]

    Accessed: 2026-04-29

    GitHub repository. Accessed: 2026-04-29

    2026