Pith. sign in

REVIEW 5 major objections 6 minor 45 references

CORE: A Unified Cascaded Ordinal Relevance Estimation Framework for E-commerce Search

T0 review · 5 major / 6 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read Treating e-commerce relevance as ordered binary gates, not flat multi-class labels, cuts severe online bad cases by about 16%.

desk verdict Solid industrial packaging of cascade + step-level GRPO + dual-head distillation with a real A/B win; the causal story that “cascade beats flat ordinal” is under-isolated, but the system is still worth reading. read the letter →

arxiv 2607.24417 v1 pith:5TJ42EUL submitted 2026-07-27 cs.IR

Zhi Jin , Xi Wang , Yunfei Li , Guojun Liu , Qingsong Hua , Wei Lin This is my paper
classification cs.IR
keywords e-commercesearchrankingrelevanceordinalclassificationcascadedbinarylargelanguagemodelsBERTdistillationstep-levelGRPOonlineA/Btesting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

E-commerce search relevance is naturally ordered—High, Mid, Low—yet industry models usually treat those labels as unrelated classes and punish nearby mistakes the same as distant ones. This paper argues that the mismatch hurts learning. It replaces the flat classifier with a cascade: first decide High versus not-High, and only then, if needed, decide Mid versus Low. The same idea is carried into large language models as step-wise reasoning with per-step rewards, and into a fast dual-head BERT that can be served online after distillation from the LLM. On a large industrial benchmark and in a live A/B test, the cascade improves accuracy and reduces the share of queries with a severe top-5 relevance error from 13.8% to 11.6%.

What carries the argument

Cascaded binary classification: a High-versus-Non-High gate followed by a conditional Mid-versus-Low gate, realized as structured LLM steps with step-level GRPO rewards and as two sequential binary heads on a shared BERT encoder with PostCoT distillation.

What would settle it

On the same backbone and data, compare the cascade against classical ordinal losses (ordinal regression, cumulative-link, or cost-sensitive multi-class). If those match or beat the cascade on offline F1 and online Badcase@5, the claim that tier-wise binary decomposition is what drives the gain fails.

Watch

Extended reading notes

Core claim

Relevance estimation for e-commerce search works better when multi-class prediction is decomposed into ordered binary judgments from higher to lower tiers. That cascaded formulation, applied both to LLM reasoning with step-level rewards and to a dual-head BERT distilled from the LLM, substantially improves offline metrics and cuts the online bad-case rate by 15.94% relative to flat multi-class baselines.

Load-bearing premise

The industrial three-level label order and the choice to gate High first are the right structure, and the measured gains come from that structure rather than from extra model capacity, extra training, or a stronger teacher.

Editorial extensions

If this is right

  • Online relevance rankers can replace a single three-way head with two cheap binary heads and keep the same encoder latency.
  • LLM relevance judges can be trained with independently verifiable per-step rewards instead of one sequence-level score.
  • Distilling label-first cascaded traces from an LLM into a dual-head BERT transfers much of the reasoning gain into production latency budgets.
  • Severe top-rank mistakes (low items promoted high) should fall when the model must clear a strict High gate before finer Mid/Low splits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same High-first cascade may transfer to other graded IR labels (e.g., 0–4 relevance) if the top tier remains the strictest business constraint.
  • If adjacent-level confusion dominates errors, any method that separates the High boundary from the Mid/Low boundary—not only this cascade—should show similar online bad-case drops.
  • Production systems that already log binary business rules (exact match vs substitute vs irrelevant) could supervise each cascade step directly without new human labels.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes CORE, a cascaded binary classification framework for three-level e-commerce relevance estimation. The ordinal {HIGH, MID, LOW} task is decomposed into two ordered binary decisions (High vs. Non-High, then Mid vs. Low). The framework is instantiated in two regimes: (i) an LLM (Qwen3-14B) trained with SFT on correctness-filtered cascaded reasoning traces followed by "step-level GRPO," in which per-step verifiable rewards are group-normalized independently and broadcast as token-level advantages (Eqs. 2–6); and (ii) a dual-head BERT (bge-small-zh) trained with a gated binary cross-entropy loss (Eq. 10) and distilled from the LLM via PostCoT label-first logits aggregated by log-sum-exp (Eqs. 11–14). Offline evaluation on a 90K-pair human-annotated benchmark shows consistent gains of roughly 0.6–2 accuracy points over flat multi-class baselines across LLM and BERT variants (Table 2), and an online A/B test reports NDCG@5 +0.20% and Badcase@5 reduced from 13.8% to 11.6% (Table 3).

Significance. If the results hold, this is a useful industrial contribution: a single ordinal decomposition that works for both generative LLM inference and low-latency discriminative serving, with per-step verifiable rewards that avoid training a separate process reward model. The paper ships genuine evidence rather than simulation alone: ablations that separate cascade-vs-flat, response-level vs step-level GRPO, and distillation (Table 2), plus a live production A/B test against the deployed flat baseline (Table 3). The step-level advantage normalization (Eq. 4) and the gated dual-head loss (Eq. 10) are simple, reproducible design elements. However, the gains are modest in absolute terms, the evaluation lacks statistical rigor (no significance, CIs, or seeds anywhere), and the central causal claim — that the cascade structure itself, rather than threshold calibration, extra RL compute, or teacher logits, drives the improvement — is not isolated against classical ordinal or cost-sensitive controls.

major comments (5)
  1. [§4.2.1, Table 2 (baselines)] The paper's motivating claim is that relevance is 'inherently ordinal' and that flat multi-class classification mis-penalizes distant errors. Yet the cascade is compared only to plain cross-entropy flat models. No classical ordinal method — ordinal regression / cumulative link (McCullagh), cost-sensitive or ordinal-aware losses (e.g., CORAL/CORN), or a flat model with an asymmetric cost matrix — is run on the same backbone. This is load-bearing: the dual-head model of Eqs. (7)–(10) with cutpoints τ1, τ2 is close to a two-cutpoint cumulative ordinal model, and without these controls the reader cannot tell whether the gains come from the cascaded *architecture* or simply from encoding ordinal structure that existing ordinal losses already encode. At minimum, Table 2 needs one ordinal-loss baseline on the BERT side and, ideally, on the LLM side.
  2. [§4.3, Table 3 (online A/B)] The headline result (Badcase@5 13.8%→11.6%, a 15.9% relative reduction, vs. NDCG@5 +0.20%) is asymmetric in exactly the way a conservative HIGH gate would produce: with τ1=0.54 the system emits HIGH less freely, which specifically suppresses severe overestimation into top positions while leaving graded ranking nearly unchanged. The A/B compares against the production flat baseline at its own operating point, with no matched-threshold flat control (e.g., a flat model whose HIGH emission rate or Badcase target is matched). The experiment also reports no traffic split, duration, sample size, confidence interval, or significance test for either metric. As it stands, Table 3 does not establish that the cascade structure, rather than operating-point/calibration tuning, causes the bad-case reduction. Please report A/B statistics and at least one operating-point-matched baseline.
  3. [Table 2 (Cascaded-BERT-Distilled vs. TaoSR1 rows)] The Cascaded-BERT-Distilled row and the TaoSR1 row report per-class precision/recall/F1 values that are identical to four decimal places for all nine entries (High: 0.7505/0.8432/0.7941; Mid: 0.7528/0.6591/0.7028; Low: 0.7834/0.7842/0.7838), with only the accuracy differing (0.7622 vs 0.7621). Agreement to this precision across nine independently measured quantities is implausible for two different models; this looks like a copy-paste error. The numbers must be verified and corrected, since the distillation-gain claim (0.7558→0.7622) rests on this row.
  4. [§4.2.2, Table 2 (evaluation protocol)] Offline comparisons mix operating points: Cascaded-BERT is evaluated with tuned thresholds τ1=0.54/τ2=0.50 (§4.1.3) while Direct-BERT is a standard argmax cross-entropy model. Threshold tuning alone can move accuracy/F1 on a class-balanced benchmark. Additionally, all Table 2 gaps (~0.6–1.2 accuracy points) are reported from what appears to be a single run — no seeds, variances, or significance tests — and the benchmark's labels are stated (§4.1.1) to be uniformly distributed across the three levels, which is unrepresentative of production skew where LOW dominates; accuracy gains measured under a uniform label distribution may not transfer. Please report mean±std over multiple seeds (or bootstrap CIs on the 90K benchmark) and either a matched-threshold flat comparison or an evaluation under the natural label prior.
  5. [§3.3.3, Eq. (2); Figures 4–5] The step-level credit-assignment claim (§3.3.3) needs clarification regarding Step 1: by Eq. (2), s(1) is the final-answer correctness reward, which is exactly the response-level signal standard GRPO would broadcast. The genuinely step-specific signals are only s(2), s(3), and s(fmt). The ablation (Cascaded-GRPO vs. Cascaded-StepGRPO, +1.07 points) is consistent with the claim, but Figure 5 (per-step error counts) reports only aggregate error reductions of 3.5%/5.8% without stating the denominator or significance, and Figure 4 shows the curves crossing around step 400 with a non-uniform x-axis that makes the comparison hard to audit. Please (i) state precisely which steps receive signals unavailable to response-level GRPO, and (ii) give Figure 5's counts as rates with error bars or a paired test.
minor comments (6)
  1. [§4.3 (metrics)] Badcase@5 is defined only as 'the fraction of queries for which at least one top-5 result contains a severe relevance error.' Please define what counts as a severe error (e.g., LOW item in top-5, or a HIGH label required and absent) and how it is audited (human judgment sample size, inter-annotator agreement).
  2. [§3.4 / §4.3] The abstract and §1 claim 'no additional inference cost' for the dual-head BERT (also §3.4); while plausible since heads share the encoder, a brief latency/throughput measurement under production conditions would substantiate the deployment-suitability claim.
  3. [Figures 4 and 5] Figure 4's x-axis ticks (0, 400, 600, 680, 720, 760, 800, 1000, 1120) are non-uniform and visually exaggerate the divergence region; use a uniform axis. Figure 5 would be clearer as error rates per step with the total number of active Step-3 decisions stated (Step 3 is inactive on HIGH samples, so denominators differ).
  4. [§3.5, Eqs. (11)–(12)] Eq. (12): using softmax over [zLOW, zMID] as the Head-2 distillation target coincides with the teacher's conditional P(· | non-HIGH) only under the teacher's own softmax normalization; a one-line justification would help, since Head 1's target (Eq. 11) explicitly marginalizes while Head 2's conditions.
  5. [§3.3.2, §3.3.3, §4.1.2, §4.2.1] Typos/grammar: 'removing noisy samples lie in the training data' (§3.3.2); 'Therefore ,we design' (§3.3.3); 'All LLM-based methods conducted use Qwen3-14B' (§4.2.1); duplicated sentence 'Training data are constructed from real-world e-commerce search logs.' (§4.1.2).
  6. [§4.1.3 / Figure 2] Reproducibility: hyperparameters are well documented (§4.1.3), but no code or prompt templates for the cascaded schema (Figure 2) are released. Even an anonymized prompt template and the format-validity check used for s(fmt) would aid replication.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: gains are measured against external human labels and a live A/B baseline, not forced by construction from the method’s own inputs.

full rationale

CORE is an engineering framework (cascaded High→Non-High then Mid→Low heads; step-level GRPO; PostCoT distillation), not a first-principles derivation. Rewards s^(k) and binary targets y1, y2 are defined from held-out gold labels r★ (Eqs. 2–3, 10), not from the model’s outputs by identity. SFT filtering (Eq. 1) keeps traces that match human annotation—standard supervised filtering, not a tautology that manufactures test accuracy. Distillation (Eqs. 11–14) transfers teacher logits into a student that is still scored on the same external benchmark and online traffic; the teacher is not the evaluation oracle. Online Badcase@5 / NDCG@5 compare a deployed dual-head BERT to a production flat multi-class baseline under live A/B, which is independent of the training objective. There is no uniqueness theorem, self-citation load-bearing premise, or fitted parameter re-labeled as a prediction. Design choices (cascade order, τ1=0.54/τ2=0.50, format weight) may be under-ablated relative to classical ordinal losses, but that is an experimental-control gap, not circular reduction of claim to input. Score 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The work is empirical systems research. It rests on domain assumptions about three-level ordinal relevance and on several hand-chosen training and inference hyperparameters; it does not introduce new physical entities or free parameters inside a closed-form theory.

free parameters (5)
  • τ1 (High vs Non-High threshold) = 0.54
    Calibrated decision threshold for the first binary head; set to 0.54 to balance precision/recall and directly controls when HIGH is emitted.
  • τ2 (Mid vs Low threshold) = 0.50
    Second-stage threshold; set to 0.50.
  • λ_fmt (format reward weight) = 0.1
    Down-weights format tokens relative to semantic step rewards in StepGRPO.
  • distillation weight λ and temperature T = λ=0.5, T=4.0
    Control KL alignment of BERT heads to PostCoT LLM logits.
  • GRPO group size G, KL coefficient β, clip ε = G=8, β=0.001, ε=0.2
    RL optimization hyperparameters that affect credit assignment and update magnitude.
assumptions (5)
  • domain assumption E-commerce relevance labels HIGH/MID/LOW are strictly ordinal with asymmetric boundaries that a High-first binary cascade matches.
    Stated in Introduction and Section 3.2 as the motivation for replacing flat multi-class classification; not compared to alternative ordinal models.
  • domain assumption Correctness of each cascade step can be verified directly from the single gold three-way label via indicator targets z(2), z(3).
    Used to define per-step rewards in Eq. (2)–(3) and the gated BCE loss Eq. (10).
  • domain assumption Filtered SFT traces whose final label matches gold are a valid warm-start for RL (self-consistency style).
    Section 3.3.2 retains only label-matching sampled traces before StepGRPO.
  • standard math Standard transformer + GRPO/PPO-style clipped surrogate optimization is a valid learning framework.
    Background RL and encoder assumptions imported from cited GRPO/BERT literature.
  • ad hoc to paper PostCoT label-first logits from the teacher LLM are a suitable soft target for dual binary heads via log-sum-exp aggregation.
    Section 3.5 defines z_LLM^(1) by log-sum-exp of non-high logits; design choice specific to this distillation.
invented entities (2)
  • Step-level GRPO (StepGRPO) with per-step group-normalized advantages
    purpose: Assign separate verifiable rewards and advantages to Step1/Step2/Step3/format token groups inside one cascaded rollout.
    Extension of GRPO; defined in Section 3.3.3 and Eq. (4)–(6). Independent evidence is only the paper’s own ablations, not an external benchmark of the optimizer itself.
  • PostCoT-CORE dual-head distillation mapping
    purpose: Transfer cascaded LLM reasoning into online BERT via label-first traces and head-wise KL.
    Described in Section 3.5; builds on cited TaoSR1 PostCoT idea but specializes targets to two binary heads.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CORE: A Unified Cascaded Ordinal Relevance Estimation Framework for E-commerce Search." pith.science (2026). https://pith.science/paper/5TJ42EUL

@misc{pith2026260724417,
  author       = {Pith},
  title        = {Pith review of: CORE: A Unified Cascaded Ordinal Relevance Estimation Framework for E-commerce Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5TJ42EUL}},
  note         = {Machine review of arXiv:2607.24417}
}
read the original abstract

Ranking relevance is a fundamental task in e-commerce search, directly affecting ranking quality and consumer experience. Although inherently an ordinal classification problem, it is commonly formulated as conventional multi-class classification, which overlooks the natural order among relevance levels and assigns equal penalties to adjacent and distant misclassifications. This mismatch leads to suboptimal learning objectives for practical relevance evaluation. To address this issue, we propose a unified cascaded binary classification framework applicable to both large language model inference and online BERT-based inference, which reformulates relevance estimation as a sequential decision process and decomposes multi-class prediction into a series of ordered binary judgments from higher to lower relevance tiers. For large language models, we design a step-wise reasoning procedure with pruning strategies and tier-specific reward functions. For the online BERT model, we replace the conventional classification head with multiple level-wise binary classifiers and distill the capabilities of large language models into the online model. Extensive offline industrial benchmark evaluations and online A/B experiments demonstrate that the proposed framework substantially improves relevance performance, reducing the online bad-case rate by 15.94\%. Further analyses suggest that tier-wise modeling is effective for relevance estimation.

Figures

Figures reproduced from arXiv: 2607.24417 by the authors.

Figure 1
Figure 1. Comparison of direct multi-class and cascaded [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Token-type schema annotated on one reason [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of step-level GRPO for cascaded reasoning. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Test accuracy of Standard GRPO and Step [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Per-step decision error counts on the e [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 13 linked inside Pith

  1. [1]

    Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 , pages =

    Zeng, Ziyang and Jing, Heming and Chen, Jindong and Li, Xiangli and Liu, Hongyu and He, Yixuan and Li, Zhengyu and Sun, Yige and Xie, Zheyong and Yang, Yuqing and Cao, Shaosheng and Fan, Jun and Wu, Yi and Hu, Yao , title =. Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 , pages =. 2026 , isbn =. doi:10.1145/37708...

    arXiv 2026

  2. [2]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  3. [3]

    Chi and Quoc V

    Jason Wei and Xuezhi Wang and Dale Schuurmans and Maarten Bosma and Brian Ichter and Fei Xia and Ed H. Chi and Quoc V. Le and Denny Zhou , editor =. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , booktitle =. 2022 , url =

    2022

  4. [4]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Thomas, Paul and Spielman, Seth and Craswell, Nick and Mitra, Bhaskar , title =. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2024 , isbn =. doi:10.1145/3626772.3657707 , abstract =

    arXiv 2024

  5. [5]

    CoRR , volume =

    Rodrigo Nogueira and Kyunghyun Cho , title =. CoRR , volume =. 2019 , url =. 1901.04085 , timestamp =

    arXiv 2019

  6. [6]

    Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pages =

    Zou, Lixin and Zhang, Shengqiang and Cai, Hengyi and Ma, Dehong and Cheng, Suqi and Wang, Shuaiqiang and Shi, Daiting and Cheng, Zhicong and Yin, Dawei , title =. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pages =. 2021 , isbn =. doi:10.1145/3447548.3467147 , abstract =

    arXiv 2021

  7. [7]

    Let s Verify Step by Step , url =

    Lightman, Hunter and Kosaraju, Vineet and Burda, Yuri and Edwards, Harrison and Baker, Bowen and Lee, Teddy and Leike, Jan and Schulman, John and Sutskever, Ilya and Cobbe, Karl , booktitle =. Let s Verify Step by Step , url =

  8. [8]

    Long Ouyang and Jeffrey Wu and Xu Jiang and Diogo Almeida and Carroll L. Wainwright and Pamela Mishkin and Chong Zhang and Sandhini Agarwal and Katarina Slama and Alex Ray and John Schulman and Jacob Hilton and Fraser Kelton and Luke Miller and Maddie Simens and Amanda Askell and Peter Welinder and Paul F. Christiano and Jan Leike and Ryan Lowe , editor =...

    2022

Show all 45 references
  1. [9]

    Manning and Stefano Ermon and Chelsea Finn , editor =

    Rafael Rafailov and Archit Sharma and Eric Mitchell and Christopher D. Manning and Stefano Ermon and Chelsea Finn , editor =. Direct Preference Optimization: Your Language Model is Secretly a Reward Model , booktitle =. 2023 , url =

    2023

  2. [10]

    Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and Wu, Z. F. and Gou, Zhibin and Shao, Zhihong and Li, Zhuoshu and Gao, Zi...

    2025

  3. [11]

    CoRR , volume =

    Qiying Yu and Zheng Zhang and Ruofei Zhu and Yufeng Yuan and Xiaochen Zuo and Yu Yue and Tiantian Fan and Gaohong Liu and Lingjun Liu and Xin Liu and Haibin Lin and Zhiqi Lin and Bole Ma and Guangming Sheng and Yuxuan Tong and Chi Zhang and Mofan Zhang and Wang Zhang and Hang ...

    2025 doi arXiv

  4. [12]

    Math-Shepherd: Verify and Reinforce LLM s Step-by-step without Human Annotations

    Wang, Peiyi and Li, Lei and Shao, Zhihong and Xu, Runxin and Dai, Damai and Li, Yifei and Chen, Deli and Wu, Yu and Sui, Zhifang. Math-Shepherd: Verify and Reinforce LLM s Step-by-step without Human Annotations. Proceedings of the 62nd Annual Meeting of the Association for Com...

    2024 doi

  5. [13]

    CoRR , volume =

    Liangchen Luo and Yinxiao Liu and Rosanne Liu and Samrat Phatale and Harsh Lara and Yunxuan Li and Lei Shu and Yun Zhu and Lei Meng and Jiao Sun and Abhinav Rastogi , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2406.06592 , eprinttype =. 2406.06592 , timestamp =

    2024 doi arXiv

  6. [14]

    CoRR , volume =

    An Yang and Baosong Yang and Beichen Zhang and Binyuan Hui and Bo Zheng and Bowen Yu and Chengyuan Li and Dayiheng Liu and Fei Huang and Haoran Wei and Huan Lin and Jian Yang and Jianhong Tu and Jianwei Zhang and Jianxin Yang and Jiaxi Yang and Jingren Zhou and Junyang Lin and...

    2024 doi arXiv

  7. [15]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  8. [16]

    CoRR , volume =

    Zhen Qin and Rolf Jagerman and Kai Hui and Honglei Zhuang and Junru Wu and Jiaming Shen and Tianqi Liu and Jialu Liu and Donald Metzler and Xuanhui Wang and Michael Bendersky , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2306.17563 , eprinttype =. 2306.17563 , t...

    2023 doi arXiv

  9. [17]

    Francis Song and Noah Y

    Jonathan Uesato and Nate Kushman and Ramana Kumar and H. Francis Song and Noah Y. Siegel and Lisa Wang and Antonia Creswell and Geoffrey Irving and Irina Higgins , title =. CoRR , volume =. 2022 , url =. doi:10.48550/ARXIV.2211.14275 , eprinttype =. 2211.14275 , timestamp =

    2022 doi arXiv

  10. [18]

    Proceedings of the 42nd International Conference on Machine Learning , articleno =

    Zhong, Han and Shan, Zikang and Feng, Guhao and Xiong, Wei and Cheng, Xinle and Zhao, Li and He, Di and Bian, Jiang and Wang, Liwei , title =. Proceedings of the 42nd International Conference on Machine Learning , articleno =. 2025 , publisher =

    2025

  11. [19]

    CoRR , volume =

    John Schulman and Filip Wolski and Prafulla Dhariwal and Alec Radford and Oleg Klimov , title =. CoRR , volume =. 2017 , url =. 1707.06347 , timestamp =

    2017 arXiv

  12. [20]

    Document Ranking with a Pretrained Sequence-to-Sequence Model

    Nogueira, Rodrigo and Jiang, Zhiying and Pradeep, Ronak and Lin, Jimmy. Document Ranking with a Pretrained Sequence-to-Sequence Model. Findings of the Association for Computational Linguistics: EMNLP 2020. 2020. doi:10.18653/v1/2020.findings-emnlp.63

    2020 doi

  13. [21]

    Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Zhuang, Honglei and Qin, Zhen and Jagerman, Rolf and Hui, Kai and Ma, Ji and Lu, Jing and Ni, Jianmo and Wang, Xuanhui and Bendersky, Michael , title =. Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2...

    2023

  14. [22]

    CoRR , volume =

    Xueguang Ma and Xinyu Zhang and Ronak Pradeep and Jimmy Lin , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2305.02156 , eprinttype =. 2305.02156 , timestamp =

    2023 doi arXiv

  15. [23]

    CoRR , volume =

    Ronak Pradeep and Sahel Sharifymoghaddam and Jimmy Lin , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2312.02724 , eprinttype =. 2312.02724 , timestamp =

    2023 doi arXiv

  16. [24]

    Is C hat GPT Good at Search? Investigating Large Language Models as Re-Ranking Agents

    Sun, Weiwei and Yan, Lingyong and Ma, Xinyu and Wang, Shuaiqiang and Ren, Pengjie and Chen, Zhumin and Yin, Dawei and Ren, Zhaochun. Is C hat GPT Good at Search? Investigating Large Language Models as Re-Ranking Agents. Proceedings of the 2023 Conference on Empirical Methods i...

    2023 doi

  17. [25]

    CoRR , volume =

    Ganqu Cui and Lifan Yuan and Zefan Wang and Hanbin Wang and Wendi Li and Bingxiang He and Yuchen Fan and Tianyu Yu and Qixin Xu and Weize Chen and Jiarui Yuan and Huayu Chen and Kaiyan Zhang and Xingtai Lv and Shuo Wang and Yuan Yao and Xu Han and Hao Peng and Yu Cheng and Zhi...

    2025 doi arXiv

  18. [26]

    2026 , eprint=

    TaoSR1: The Thinking Model for E-commerce Relevance Search , author=. 2026 , eprint=

    2026

  19. [27]

    CoRR , volume =

    Navid Mehrdad and Hrushikesh Mohapatra and Mossaab Bagdouri and Prijith Chandran and Alessandro Magnani and Xunfan Cai and Ajit Puthenputhussery and Sachin Yadav and Tony Lee and ChengXiang Zhai and Ciya Liao , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2406.00...

    2024 doi arXiv

  20. [28]

    Towards Boosting LLM s-driven Relevance Modeling with Progressive Retrieved Behavior-augmented Prompting

    Chen, Zeyuan and Wu, Haiyan and Wu, Kaixin and Chen, Wei and Zhong, Mingjie and Xu, Jia and Liu, Zhongyi and Zhang, Wei. Towards Boosting LLM s-driven Relevance Modeling with Progressive Retrieved Behavior-augmented Prompting. Proceedings of the 31st International Conference o...

    2025

  21. [29]

    2025 , eprint=

    GenCLS++: Pushing the Boundaries of Generative Classification in LLMs Through Comprehensive SFT and RL Studies Across Diverse Datasets , author=. 2025 , eprint=

    2025

  22. [30]

    Large Language Models are Zero-Shot Reasoners , booktitle =

    Takeshi Kojima and Shixiang Shane Gu and Machel Reid and Yutaka Matsuo and Yusuke Iwasawa , editor =. Large Language Models are Zero-Shot Reasoners , booktitle =. 2022 , url =

    2022

  23. [31]

    ReprBERT: Distilling

    Shaowei Yao and Jiwei Tan and Xi Chen and Juhao Zhang and Xiaoyi Zeng and Keping Yang , editor =. ReprBERT: Distilling. 2022 , url =. doi:10.1145/3534678.3539090 , timestamp =

    2022

  24. [32]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =

    Lin, Zhe and Tan, Jiwei and Ou, Dan and Chen, Xi and Yao, Shaowei and Zheng, Bo , title =. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2024 , isbn =. doi:10.1145/3637528.3671559 , abstract =

    2024

  25. [33]

    Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =

    Yin, Dawei and Hu, Yuening and Tang, Jiliang and Daly, Tim and Zhou, Mianwei and Ouyang, Hua and Chen, Jianhui and Kang, Changsung and Deng, Hongbo and Nobata, Chikashi and Langlois, Jean-Marc and Chang, Yi , title =. Proceedings of the 22nd ACM SIGKDD International Conference...

    2016

  26. [34]

    ACM Trans

    Zhu, Yutao and Yuan, Huaying and Wang, Shuting and Liu, Jiongnan and Liu, Wenhan and Deng, Chenlong and Chen, Haonan and Liu, Zheng and Dou, Zhicheng and Wen, Ji-Rong , title =. ACM Trans. Inf. Syst. , month = nov, articleno =. 2025 , issue_date =. doi:10.1145/3748304 , abstract =

    2025 doi

  27. [35]

    Beyond Yes and No: Improving Zero-Shot LLM Rankers via Scoring Fine-Grained Relevance Labels

    Zhuang, Honglei and Qin, Zhen and Hui, Kai and Wu, Junru and Yan, Le and Wang, Xuanhui and Bendersky, Michael. Beyond Yes and No: Improving Zero-Shot LLM Rankers via Scoring Fine-Grained Relevance Labels. Proceedings of the 2024 Conference of the North American Chapter of the ...

    2024 doi

  28. [36]

    Le and Sergey Levine and Yi Ma , editor =

    Tianzhe Chu and Yuexiang Zhai and Jihan Yang and Shengbang Tong and Saining Xie and Dale Schuurmans and Quoc V. Le and Sergey Levine and Yi Ma , editor =. Forty-second International Conference on Machine Learning,. 2025 , url =

    2025

  29. [37]

    CoRR , volume =

    Shengyao Zhuang and Xueguang Ma and Bevan Koopman and Jimmy Lin and Guido Zuccon , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2503.06034 , eprinttype =. 2503.06034 , timestamp =

    2025 doi arXiv

  30. [38]

    Hinton and Oriol Vinyals and Jeffrey Dean , title =

    Geoffrey E. Hinton and Oriol Vinyals and Jeffrey Dean , title =. CoRR , volume =. 2015 , url =. 1503.02531 , timestamp =

    2015 arXiv

  31. [39]

    Pretrained Transformers for Text Ranking: BERT and Beyond

    Yates, Andrew and Nogueira, Rodrigo and Lin, Jimmy. Pretrained Transformers for Text Ranking: BERT and Beyond. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Tutorials. 2021. doi:1...

    2021 doi

  32. [40]

    Le and Ed H

    Xuezhi Wang and Jason Wei and Dale Schuurmans and Quoc V. Le and Ed H. Chi and Sharan Narang and Aakanksha Chowdhery and Denny Zhou , title =. The Eleventh International Conference on Learning Representations,. 2023 , url =

    2023

  33. [41]

    2022 , eprint=

    Shopping Queries Dataset: A Large-Scale ESCI Benchmark for Improving Product Search , author=. 2022 , eprint=

    2022

  34. [42]

    Rationale-Guided Distillation for E -Commerce Relevance Classification: Bridging Large Language Models and Lightweight Cross-Encoders

    Agrawal, Sanjay and Ahemad, Faizan and Sembium, Vivek Varadarajan. Rationale-Guided Distillation for E -Commerce Relevance Classification: Bridging Large Language Models and Lightweight Cross-Encoders. Proceedings of the 31st International Conference on Computational Linguisti...

    2025

  35. [43]

    HybridFlow: A Flexible and Efficient RLHF Framework , url=

    Sheng, Guangming and Zhang, Chi and Ye, Zilingfeng and Wu, Xibin and Zhang, Wang and Zhang, Ru and Peng, Yanghua and Lin, Haibin and Wu, Chuan , year=. HybridFlow: A Flexible and Efficient RLHF Framework , url=. doi:10.1145/3689031.3696075 , booktitle=

  36. [44]

    L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models

    Zheng, Yaowei and Zhang, Richong and Zhang, Junhao and Ye, Yanhan and Luo, Zheyan. L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)....

    2024 doi

  37. [45]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Xiao, Shitao and Liu, Zheng and Zhang, Peitian and Muennighoff, Niklas and Lian, Defu and Nie, Jian-Yun , title =. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2024 , isbn =. doi:10.1145/3626772.3657...

    2024

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.