FlashEvaluator: Expanding Search Space with Parallel Sequence-Level Evaluation

Chao Feng; Chenghao Zhang; Chunjie Chen; Kaiqiao Zhan; Shanqi Liu; Shuchang Liu; Xiang Li; Yuanhao Pu

REVIEW 4 major objections 5 minor 14 references

FlashEvaluator: Expanding Search Space with Parallel Sequence-Level Evaluation

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read FlashEvaluator claims a single shared forward pass can score all K candidate lists at roughly 1/ρ the cost of independent scoring, while cross-list comparison improves selection quality.

desk verdict Genuinely useful joint evaluator with credible online gains; the O(1/K) complexity claim drops the listwise-attention term and needs correction before the theory can be taken at face value. read the letter →

arxiv 2603.02565 v2 pith:Z2WO5RCZ submitted 2026-03-03 cs.IR cs.CLcs.LG

Chao Feng , Yuanhao Pu , Chenghao Zhang , Shanqi Liu , Shuchang Liu , Xiang Li , Chunjie Chen , Kaiqiao Zhan This is my paper

classification cs.IRcs.CLcs.LG

keywords generator-evaluatorframeworkjointsequenceevaluationcross-listattentionQKV-Cacheitemreusefactorcandidatererankinginferencelatencylistwisesoftmax

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

FlashEvaluator asks a simple question: when K candidate sequences are all drawn from the same item pool, why re-encode the shared context and recurring items K times? Its answer is a joint evaluator that scores every candidate in one forward pass, factoring evaluation into a shared request-level encoding, item embeddings computed once for the distinct candidate set, sequence assembly by indexing, and a cross-sequence attention layer. The paper claims this reduces the dominant encoding cost from total occurrences Kl to distinct items M, giving a computational ratio T_joint/T_ind ≈ 1/ρ with ρ=Kl/M, and proves a generalization gap Θ(√K) times smaller than independent scoring. On public benchmarks the joint evaluator matches or slightly beats independent rerankers at far lower latency, and in a production deployment at K=50 it reports 44% lower inference latency, 114% higher throughput, and statistically significant gains in retention, engagement, and cold-start exposure. If correct, the G-E framework's search space can be widened (larger K) at near-zero marginal evaluation cost.

What carries the argument

The load-bearing identity is the item reuse factor ρ=Kl/M, the ratio of total item occurrences to distinct items. FlashEvaluator's architecture realizes the saving through a request-local reuse scheme it calls QKV-Cache: context-side key/value representations are computed once for the query, candidate-item representations are computed once per distinct item, and each candidate list is assembled by cheap indexing rather than re-encoding. A listwise self-attention layer over the K list embeddings then performs the cross-sequence comparison. This factorization carries the paper's claims: ρ converts the cost ratio into 1/ρ, and the shared latent representation is what yields the √K-larger genera

What would settle it

Fix a query context and a candidate pool of size M, then measure wall-clock latency for FlashEvaluator and for the independent baseline as K grows (e.g., 16, 64, 256, 1024). If T_joint/T_ind stops falling and starts rising toward O(K) once the listwise attention term dominates, the sublinear claim fails. A second check: repeat the online A/B at K≈200; the reported −44% latency and +114% QPS should degrade if the O(K²) cross-list attention is the hidden cost.

Watch

Extended reading notes

Core claim

The central discovery is that evaluation in the Generator-Evaluator pipeline does not have to be O(K) in the number of lists. By processing all K candidate sequences jointly, FlashEvaluator can encode each distinct item once and assemble the per-list representations by indexing, so the item-encoding cost scales with M (the candidate pool) rather than Kl (total occurrences). The paper formalizes this with the item-reuse factor ρ=Kl/M and shows T_joint/T_ind ≈ 1/ρ plus a small overhead term. On the quality side, the shared representation and cross-list attention allow the model to compare lists explicitly, and the paper proves that with a listwise softmax objective the joint evaluator's genera

Load-bearing premise

The efficiency ratio hinges on Assumption A.13 — that item-level encoding is so dominant that indexing, scoring, and especially the K-list attention cost are negligible — and on treating candidate sequences as heavily overlapping subsets of a fixed pool; if per-list overhead or attention grows quickly with K, the advertised 1/ρ saving weakens and could reverse.

Editorial extensions

If this is right

G-E systems can increase K substantially under the same latency budget, raising the ceiling set by generator diversity, because the marginal cost of an extra candidate list is limited to indexing and attention.
Listwise softmax training becomes practical memory-wise, since all K scores come from one forward pass instead of K serial passes; the paper shows listwise training beats pointwise training on the same architecture.
The evaluator is generator-agnostic: swapping it into the tested G-E backbones improves recommendation metrics without retraining the generator.
In summarization and generation, reranking can be much cheaper: FlashEvaluator reaches roughly competitive quality with expensive LLM rerankers while scaling sublinearly in the candidate count.
Operators can expect roughly half the latency and double the throughput at K=50 relative to one-by-one evaluation, together with positive retention, engagement, and ecosystem effects.
The theoretical bound predicts the joint evaluator's advantage over independent scoring grows with √K, so quality differences should become more visible at larger K.

Reading between the lines

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

The O(1/K) efficiency headline rests on Assumption A.13 plus the implicit neglect of the O(K²) listwise-attention term; at sufficiently large K that attention term should overtake amortized item encoding and push the cost ratio back upward — a regime the paper does not test beyond about K=80.
The same factorization should transfer to any setting where candidates share a prompt or context, such as LLM preference optimization or retrieval-augmented generation with many sampled continuations; the quality gain from explicit cross-candidate comparison would transfer more directly than the latency gain, which depends on item reuse.
In low-reuse settings (ρ close to 1), the efficiency argument nearly disappears, so the reported gains must come mostly from cross-list attention and listwise training; ablating these two effects at controlled ρ would separate architectural efficiency from setwise-information gains.
A concrete prediction of Theorem 5.3 is that the accuracy gap between joint and independent evaluators widens with K; plotting ranking-quality metrics against K on recommendation-style data would check whether the theoretical √K advantage shows up empirically.

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.

Desk Editor's Note

read the letter

The joint evaluator is a real idea and the online numbers look credible, but the paper's headline complexity claim is not backed by the proof. Proposition 5.5 gives T_joint/T_ind ≈ 1/ρ only if C_att(K), the listwise attention cost, is negligible. With standard self-attention over K list embeddings, C_att(K) is O(K^2·d), so the ratio grows with K and can eventually exceed the serial baseline. The abstract's 'O(1/K) computational complexity' is stronger than anything the proposition actually proves. This is the main soft spot, and it sits in the load-bearing theory section. The paper's own Eq. (78) keeps the C_att(K) term and then says it is 'typically small'—that is the gap the stress-test note identifies, and it is real.

Now the credit. What is new: the single-forward-pass joint evaluator, with shared item-pool encoding, index-based sequence assembly, and cross-sequence attention, is a legitimate extension of setwise ranking to the G-E framework. The request-local reuse scheme is a sensible engineering contribution. The 1/ρ ratio follows transparently from the architecture when the attention term really is small; the generalization comparison is a reasonable structural analysis, not circular. The empirical work is solid: offline RecFlow improvements are consistent across two generator backbones, summarization shows competitive quality with lower latency, and the online A/B test reports statistically significant gains plus 44% lower latency and 114% higher QPS. For a production systems paper, that is real evidence. The reader's circularity concern does not land; the speedups are measured independently and the generalization analysis is structural.

Other soft spots are minor by comparison. Theorem 5.3 has an o(n^{−1/2}) term that appears without justification; it is likely harmless but as written it is sloppy. Tables 2 and 3 report point estimates with no confidence intervals; the online A/B is better in that respect. Code is not released, and the paper says it will be published upon acceptance. Assumption A.13 is fine in spirit, but the paper never verifies it at large K—the tests go up to K=80, and the theory is stated without qualification. None of these are fatal. The empirical speedup at K=50 with linear attention can survive the theoretical critique; what needs fixing is the scope of the asymptotic claim.

This paper is for practitioners building G-E or reranking pipelines who want to widen the candidate search space. A serious referee should engage with it, but the authors need to constrain the complexity claim to the regime where the listwise attention cost is actually dominated, or analyze a linear-attention variant. I would send it to review and ask for revision.

Referee Report

4 major / 5 minor

Summary. The paper proposes FlashEvaluator, a joint evaluator for the Generator-Evaluator paradigm that scores K candidate sequences in a single forward pass, reusing request-level context and item-level representations across lists. The authors claim sublinear computational growth relative to K (advertised as O(1/K) complexity), improved selection quality via cross-sequence attention, and provide a theoretical generalization-bound comparison, extensive offline experiments on RecFlow and CNN/DM, and an online A/B deployment at Kuaishou reporting 44% lower latency and 114% higher QPS at K=50. The central architectural idea—factorizing evaluation into shared item encoding, list assembly by indexing, and a lightweight cross-list interaction—is plausible and the empirical results are encouraging, but the headline asymptotic efficiency claim is not supported by the paper's own complexity equations because the listwise self-attention term C_att(K) is O(K^2) and is dismissed rather than bounded.

Significance. If the efficiency claim were fully established, FlashEvaluator would be a valuable practical contribution: it offers a concrete way to widen G-E search spaces without proportional inference cost, with an online deployment providing unusually direct evidence of impact. The paper ships a substantial experimental package, including a reproducible public-dataset evaluation, ablation studies, and a statistically significant online A/B test, and it makes a serious attempt at theoretical grounding through Rademacher-complexity bounds and a cost model. However, the central asymptotic claim as written is internally inconsistent with the stated architecture, so the theoretical contribution requires correction before the paper's main thesis can be accepted.

major comments (4)

[Abstract; §5.2, Eq. (76)–(78), Prop. 5.5] The claim of 'O(1/K) computational complexity' is stronger than what is proven. Eq. (76) gives T_joint = Θ(M·C_item + Kl·C_idx + C_att(K) + K·C_score), and Eq. (78) dismisses C_att(K)/(Kl·C_item) as 'typically small'. But the architecture in Eq. (12) is a self-attention over K list embeddings, whose standard cost is Θ(K^2·d). Under this concrete cost, the ratio T_joint/T_ind equals approximately 1/ρ + (K·d)/(l·C_item), which grows linearly in K and can exceed the serial baseline for large K. The online deployment uses linear attention and K≤80, but the abstract and Prop. 5.5 are stated without this qualification. The proposition should either include an explicit assumption on C_att(K) (e.g., linear attention with O(K) cost) or state the corrected asymptotic expression.
[Appendix A.3.3, Assumption A.13 and Eq. (78)] Assumption A.13 only bounds C_idx and C_score relative to C_item; it says nothing about C_att(K). The proof of Prop. 5.5 thus has a gap: the 'attention term is typically small' assertion is not derived from any stated assumption and is in fact false for the standard self-attention described in §4.1 under any fixed item-encoding cost and growing K. To make the proposition load-bearing, the authors must either bound C_att(K) (as with the online linear-attention variant) or restrict the claim to the regime where Kl·C_item dominates C_att(K), and then verify that regime empirically beyond K=80.
[Figure 2a and §6.2, RQ3] The efficiency plots show QPS increasing with K up to 80 for FlashEvaluator. This is consistent with linear attention and a regime where shared item encoding dominates. But the paper extrapolates a general 'near-zero marginal cost' (§5.2) and O(1/K) complexity. If, as the theory section allows, the cross-list attention term is O(K^2), then there is a crossover point beyond which increasing K reduces efficiency relative to the serial baseline. The paper should state the measured or derived crossover point, or explicitly limit the complexity claim to the tested architecture (linear attention) and tested range.
[Theorem 5.3 and Assumption 5.2] The generalization-bound comparison is structurally interesting, but it compares two different hypothesis classes under incomparable capacity assumptions: the joint bound assumes a shared hidden representation with Λ-Lipschitz Ψ independent of K, while the independent bound uses a per-list base class. This is a valid structural statement, but it is not an oracle inequality and does not by itself establish that FlashEvaluator generalizes better for a fixed model class. The authors should add a sentence clarifying that Theorem 5.3 compares architectural complexity scalings under Assumption 5.2, not a universal model-selection guarantee. This would prevent readers from overinterpreting the Θ(√K) ratio.

minor comments (5)

[Abstract and §1 contributions] The phrase 'O(1/√K) generational error' appears to be a typo for 'generalization error'; likewise 'summerized Utility' in §6.1.2 and 'Appdendix' in the proof of Prop. 5.5 should be corrected.
[Eq. (78)] The notation switches from T_joint/T_ind to T_new/T_old; keep one symbol. Also, the inequality '≲' is used after ignoring the C_att term; this step should be made explicit.
[Section 6.2, online A/B] Table 4 reports only relative lifts and p<0.05. Reporting confidence intervals or the raw latency/QPS values with standard errors would strengthen the efficiency claim, especially since the measured speedup at K=50 is one of the paper's headline results.
[Appendix B.3] The ablation Table 5 is informative, but the row labels for HR@6 and Utility appear reversed compared to Table 2 (HR@6 0.5671 vs. 0.3963; Utility 0.3963 vs. 0.5722). Please verify the column ordering.
[Appendix B.2, online settings] The online setting states 120 CPU threads, while the paper elsewhere refers to GPU inference. Clarify the deployment hardware for latency/QPS measurements.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: efficiency and generalization results follow from explicit architectural assumptions and are independently measured; self-citations are background only.

full rationale

The derivation chain is not circular. Theorem 5.3 is proved in Appendix A.1 from standard Rademacher results (Bartlett & Mendelson 2002; Maurer 2016) under Assumptions 5.1–5.2, which state explicit Lipschitz and capacity bounds rather than presupposing the theorem's comparison. Proposition 5.5 is a transparent algebraic consequence of the cost model in Eqs. (71)–(78): the paper identifies every cost term (C_item, C_idx, C_att(K), C_score), and the claimed 1/rho ratio follows from Assumption A.13 without fitting any parameter to the reported latency/QPS numbers. The online speedups (−44% latency, +114% QPS) are independently measured on the deployed system, not predicted from Eq. (78). The overlap-authorship citations (e.g., Yang et al. 2025b, Zhou et al. 2025) are background/baseline support, not load-bearing evidence for the new architecture, and no uniqueness theorem or ansatz is imported from prior work. The one internal caveat—that the O(K²) listwise-attention term C_att(K) is dismissed as 'typically small' in Eq. (78)—is an assumption/scope gap about large-K behavior, not a circular definition or fitted prediction; the paper's own Limitation section acknowledges that gains depend on 'candidate set size, semantic diversity, and redundancy.'

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

FlashEvaluator introduces no new physical entities. The theoretical efficiency and generalization claims rest on two ad hoc structural assumptions (Λ independent of K; item-encoding dominance) plus a domain-level bias decomposition. These are plausible but not externally validated, which limits the strength of the derived bounds.

free parameters (1)

None
No numerical free parameters are fitted to data in the central derivation. Hyperparameters (embedding dim 64, lr 0.001, batch 2048) are standard training settings, not fitted to the theoretical or efficiency claims.

assumptions (5)

ad hoc to paper Assumption 5.2(2): the joint evaluator's projection head Ψ is Λ-Lipschitz with Λ independent of K, and the hidden representation class has Rademacher complexity C_hid/√m.
This structural assumption is what removes the √K dependence from the joint generalization bound in Theorem 5.3. It is custom to this paper and not a standard property of neural heads; a linear head with rows in R^d can have spectral norm growing with K unless explicitly normalized.
domain assumption Assumption A.13: item-level encoding dominates, with C_idx ≤ α_idx·C_item and C_score ≤ α_score·C_item for tiny α_idx, α_score.
Underpins Proposition 5.5's 1/ρ cost ratio and the abstract's 'O(1/K)' claim. Requires the list-wise attention term C_att(K) to be negligible, which is not guaranteed for large K.
domain assumption Assumption A.7: sample selection bias decomposes additively into a shared set-level shift μ and a zero-sum list-specific distortion ν_k with zero-mean noise.
Needed for Theorem A.8's robustness-superiority result. The decomposition is assumed, not derived from data.
domain assumption Definition 3.1: there exists a latent utility u* and labels follow a multinomial logit model of the per-sequence utilities.
Defines the Bayes-optimal Top-1 target and the evaluation framework. This is a modeling assumption about how ground-truth sequence quality generates observed choices.
standard math Standard Rademacher complexity machinery, including Bartlett–Mendelson uniform deviation bound and Maurer's vector contraction inequality.
Used in Appendix A.1 to derive the generalization bounds. This is standard background, not specific to the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FlashEvaluator: Expanding Search Space with Parallel Sequence-Level Evaluation." pith.science (2026). https://pith.science/paper/Z2WO5RCZ

@misc{pith2026260302565,
  author       = {Pith},
  title        = {Pith review of: FlashEvaluator: Expanding Search Space with Parallel Sequence-Level Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z2WO5RCZ}},
  note         = {Machine review of arXiv:2603.02565}
}

read the original abstract

The Generator-Evaluator (G-E) framework generates K candidate sequences and uses an evaluator to select the highest-scoring one, which is widely used in recommender systems (RecSys) and natural language processing (NLP). Existing evaluators commonly score candidates independently. Although such evaluations can be batched, independent scoring neither models interactions among candidates nor eliminates repeated computation of request-level context and recurring candidate elements, causing the total evaluation work to grow approximately linearly with K. To handle with, we propose FlashEvaluator, a joint evaluator that scores all candidate sequences in a single forward pass. FlashEvaluator factorizes evaluation into shared request-level encoding, reusable candidate-side computation, sequence assembly by indexing, and cross-sequence interaction for setwise comparison. We call this request-local reuse scheme QKV-Cache: inspired by autoregressive KV caching, it reuses context-side key/value representations across candidate sequences and, when candidate elements recur, reuses their request-conditioned representations on the query side. In repeated-item settings, the dominant item-encoding cost therefore depends on the number of distinct items rather than their total occurrences across sequences, reducing the marginal cost of evaluating additional candidates. We provide a computational analysis and evaluate FlashEvaluator on recommendation and text summarization. The results show lower latency and higher throughput with competitive recommendation and summarization quality. In an online deployment at Kuaishou with K=50, FlashEvaluator reduces inference latency by 44% and increases QPS by 114% relative to the production baseline, while yielding statistically significant gains in retention, engagement, and ecosystem metrics.

Figures

Figures reproduced from arXiv: 2603.02565 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 1 canonical work pages

[1]

The bias ofE ind θ (·)is the sum of the squared shift and theMean Second Momentof the distortion: ∆ind SSB =E z h (∆¯µ)2 + 1 K KX k=1 (∆¯νk)2 i =E z h (∆¯µ)2 +E k∼U [(∆¯νk)2] i . (51)
[2]

For any vector v, the form v⊤Hvis exactly the Variance ofvunder the probability distributionp: (∆¯ν) ⊤H(∆¯ν) = KX k=1 pk(∆¯νk)2 − KX k=1 pk∆¯νk !2 = Vark∼p(∆¯νk)

The bias of Eθ(·) corresponds to the quadratic form of the Hessian H= diag(p)−pp ⊤. For any vector v, the form v⊤Hvis exactly the Variance ofvunder the probability distributionp: (∆¯ν) ⊤H(∆¯ν) = KX k=1 pk(∆¯νk)2 − KX k=1 pk∆¯νk !2 = Vark∼p(∆¯νk). (52) WithVar(X)≤E[X 2], we establish a strictly bound: ∆joint SSB ≈E z h 1 2 Vark∼p(∆¯νk) i ≤E z h 1 2 KX k=1 ...
[3]

The joint evaluator strictly removes the (∆¯µ)2 term. Besides, assuming the distortion shift ∆¯νk is uncorrelated with the item relevancep k, the expected second moments align: Ez[Ek∼p[(∆¯νk)2]]≈E z[Ek∼U [(∆¯νk)2]].(54) Under this condition, we have: ∆joint SSB <∆ ind SSB.(55) 17 FlashEvaluator: Expanding Search Space with Parallel Evaluation We provide t...

2025
[4]

InConference On Learning Theory, COLT 2018, Stockholm, Sweden, 6-9 July 2018 (Proceedings of Machine Learning Research, Vol

Size-Independent Sample Complexity of Neural Networks. InConference On Learning Theory, COLT 2018, Stockholm, Sweden, 6-9 July 2018 (Proceedings of Machine Learning Research, Vol. 75), S´ebastien Bubeck, Vianney Perchet, and Philippe Rigollet (Eds.). PMLR, 297–299. http://proceedings.mlr.press/ v75/golowich18a.html Vidushi Goyal, Valeria Bertacco, and Ree...

2018
[7]

arXiv:2403.02475 doi:10.48550/ARXIV.2403.02475 Andreas Maurer

Enhancing LLM Safety via Con- strained Direct Preference Optimization.CoRR abs/2403.02475 (2024). arXiv:2403.02475 doi:10.48550/ARXIV.2403.02475 Andreas Maurer. 2016. A Vector-Contraction Inequality for Rademacher Complexities. InAlgorithmic Learn- ing Theory - 27th International Conference, ALT 2016, Bari, Italy, October 19-21, 2016, Proceedings (Lecture...

arXiv doi:10.48550/arxiv.2403.02475 2024
[9]

Direct Preference Optimization: Your Language Model is Secretly a Reward Model. InAdvances in Neu- ral Information Processing Systems 36: Annual Confer- ence on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine (Eds...

arXiv 2023
[10]

RankMixer: Scaling Up Ranking Models in Indus- trial Recommenders. InProceedings of the 34th ACM International Conference on Information and Knowl- edge Management, CIKM 2025, Seoul, Republic of Ko- rea, November 10-14, 2025, Meeyoung Cha, Chanyoung Park, Noseong Park, Carl Yang, Senjuti Basu Roy, Jessie Li, Jaap Kamps, Kijung Shin, Bryan Hooi, and Lifang...

work page Pith review arXiv doi:10.48550/arxiv.2309.12767 2025
[14]

concatenates the full candidate set with the user query (e.g., interaction history) to form an enriched input, enabling intra-candidate information exchange via self-attention. Similarly, IFA(Yu et al., 2024), DUET(Goyal et al., 2024) and HoMer(Chen et al., 2025) propose end-to-end models that rank all candidates simultaneously, leveraging self-attention ...

2024

Show all 14 references

[2015]

InProceedings of The 28th Conference on Learn- ing Theory, COLT 2015, Paris, France, July 3-6, 2015 (JMLR Workshop and Conference Proceedings, Vol

Norm-Based Capacity Control in Neural Net- works. InProceedings of The 28th Conference on Learn- ing Theory, COLT 2015, Paris, France, July 3-6, 2015 (JMLR Workshop and Conference Proceedings, Vol. 40), Peter Gr ¨unwald, Elad Hazan, and Satyen Kale (Eds.). JMLR.org, 1376–1401....

2015
[2018]

Learning a Deep Listwise Context Model for Rank- ing Refinement. InThe 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR 2018, Ann Arbor, MI, USA, July 08-12, 2018, Kevyn Collins-Thompson, Qiaozhu Mei, Brian D. Davison, Yiqun Liu...

2018
[2023]

InIn- ternational Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA (Pro- ceedings of Machine Learning Research, Vol

PAL: Program-aided Language Models. InIn- ternational Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA (Pro- ceedings of Machine Learning Research, Vol. 202), An- dreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, an...

2023
[2024]

InProceedings of the 61st ACM/IEEE Design Automation Conference, DAC 2024, San Francisco, CA, USA, June 23-27, 2024, Vivek De (Ed.)

Duet: A Collaborative User Driven Recommen- dation System for Edge Devices. InProceedings of the 61st ACM/IEEE Design Automation Conference, DAC 2024, San Francisco, CA, USA, June 23-27, 2024, Vivek De (Ed.). ACM, 55:1–55:6. doi: 10.1145/3649329. 3656225 Kelvin Guu, Kenton Lee...

2024
[2025]

InThe Thirteenth International Confer- ence on Learning Representations, ICLR 2025, Singa- pore, April 24-28, 2025

RecFlow: An Industrial Full Flow Recommen- dation Dataset. InThe Thirteenth International Confer- ence on Learning Representations, ICLR 2025, Singa- pore, April 24-28, 2025. OpenReview.net. https: //openreview.net/forum?id=vVHc8bGRns Weiwen Liu, Yunjia Xi, Jiarui Qin, Fei Sun...

2025
[6120]

doi:10.18653/V1/D19-1633 Noah Golowich, Alexander Rakhlin, and Ohad Shamir

doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

The bias ofE ind θ (·)is the sum of the squared shift and theMean Second Momentof the distortion: ∆ind SSB =E z h (∆¯µ)2 + 1 K KX k=1 (∆¯νk)2 i =E z h (∆¯µ)2 +E k∼U [(∆¯νk)2] i . (51)

[2] [2]

For any vector v, the form v⊤Hvis exactly the Variance ofvunder the probability distributionp: (∆¯ν) ⊤H(∆¯ν) = KX k=1 pk(∆¯νk)2 − KX k=1 pk∆¯νk !2 = Vark∼p(∆¯νk)

The bias of Eθ(·) corresponds to the quadratic form of the Hessian H= diag(p)−pp ⊤. For any vector v, the form v⊤Hvis exactly the Variance ofvunder the probability distributionp: (∆¯ν) ⊤H(∆¯ν) = KX k=1 pk(∆¯νk)2 − KX k=1 pk∆¯νk !2 = Vark∼p(∆¯νk). (52) WithVar(X)≤E[X 2], we establish a strictly bound: ∆joint SSB ≈E z h 1 2 Vark∼p(∆¯νk) i ≤E z h 1 2 KX k=1 ...

[3] [3]

The joint evaluator strictly removes the (∆¯µ)2 term. Besides, assuming the distortion shift ∆¯νk is uncorrelated with the item relevancep k, the expected second moments align: Ez[Ek∼p[(∆¯νk)2]]≈E z[Ek∼U [(∆¯νk)2]].(54) Under this condition, we have: ∆joint SSB <∆ ind SSB.(55) 17 FlashEvaluator: Expanding Search Space with Parallel Evaluation We provide t...

2025

[4] [4]

InConference On Learning Theory, COLT 2018, Stockholm, Sweden, 6-9 July 2018 (Proceedings of Machine Learning Research, Vol

Size-Independent Sample Complexity of Neural Networks. InConference On Learning Theory, COLT 2018, Stockholm, Sweden, 6-9 July 2018 (Proceedings of Machine Learning Research, Vol. 75), S´ebastien Bubeck, Vianney Perchet, and Philippe Rigollet (Eds.). PMLR, 297–299. http://proceedings.mlr.press/ v75/golowich18a.html Vidushi Goyal, Valeria Bertacco, and Ree...

2018

[5] [7]

arXiv:2403.02475 doi:10.48550/ARXIV.2403.02475 Andreas Maurer

Enhancing LLM Safety via Con- strained Direct Preference Optimization.CoRR abs/2403.02475 (2024). arXiv:2403.02475 doi:10.48550/ARXIV.2403.02475 Andreas Maurer. 2016. A Vector-Contraction Inequality for Rademacher Complexities. InAlgorithmic Learn- ing Theory - 27th International Conference, ALT 2016, Bari, Italy, October 19-21, 2016, Proceedings (Lecture...

arXiv doi:10.48550/arxiv.2403.02475 2024

[6] [9]

Direct Preference Optimization: Your Language Model is Secretly a Reward Model. InAdvances in Neu- ral Information Processing Systems 36: Annual Confer- ence on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine (Eds...

arXiv 2023

[7] [10]

RankMixer: Scaling Up Ranking Models in Indus- trial Recommenders. InProceedings of the 34th ACM International Conference on Information and Knowl- edge Management, CIKM 2025, Seoul, Republic of Ko- rea, November 10-14, 2025, Meeyoung Cha, Chanyoung Park, Noseong Park, Carl Yang, Senjuti Basu Roy, Jessie Li, Jaap Kamps, Kijung Shin, Bryan Hooi, and Lifang...

work page Pith review arXiv doi:10.48550/arxiv.2309.12767 2025

[8] [14]

concatenates the full candidate set with the user query (e.g., interaction history) to form an enriched input, enabling intra-candidate information exchange via self-attention. Similarly, IFA(Yu et al., 2024), DUET(Goyal et al., 2024) and HoMer(Chen et al., 2025) propose end-to-end models that rank all candidates simultaneously, leveraging self-attention ...

2024