Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Risk-Constrained Freshness-Aware Semantic Caching for Open-Web Retrieval-Augmented LLMs

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read FreshCache treats open-web RAG cache reuse as a risk-constrained temporal decision, saving about 97% of search calls at 0.1% stale error by gating hits with a staleness probability model across three tiers.

desk verdict Solid systems paper: risk-gated multi-tier reuse beats semantic-cache baselines on a real snapshot benchmark; absolute 0.1%/97% numbers are in-distribution and not yet general. read the letter →

arxiv 2607.04281 v1 pith:UK5S3SB5 submitted 2026-07-05 cs.CL cs.AI

Muhammad Mansoor , Tahir Ahmad , Yeo-Chan Yoon This is my paper · ORCID
classification cs.CLcs.AI
keywords semanticcachingretrieval-augmentedgenerationstalenessriskopen-webRAGfreshness-awarecachetemporalreusethree-tierexponentialdecay
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

Open-web retrieval-augmented LLMs repeatedly pay for search, page fetches, and generation, so semantic caches try to reuse prior results when a new query is similar. This paper argues that similarity alone is not enough: web evidence ages at very different rates, and serving a cached result is really a decision about whether the probability it is stale still sits inside an acceptable error budget. FreshCache implements that idea with a three-tier cache—final answers, URL lists, and page content—each with its own budget, and estimates staleness from class-specific exponential half-lives, optionally refined by a small learned network. On FreshCache-Bench, built from real web snapshots at multiple ages and expanded with paraphrases, the system saves nearly all search calls while keeping stale rates far below prior semantic caches. A sympathetic reader cares because the design offers a concrete way to cut cost and latency without quietly returning outdated answers as the open web drifts.

What carries the argument

The risk-gated three-tier reuse rule: a candidate hit must satisfy both semantic similarity and a calibrated staleness check P̂(stale|class, tier, age) ≤ ε_tier, where probability comes from fitted exponential half-lives by freshness class (optionally sharpened by a 25-feature MLP) and budgets tighten from content (0.35) to URL lists (0.20) to answers (0.10).

What would settle it

On a held-out live crawl with different domains and real event timing, re-run the same half-lives, budgets, and models: if stale-error rates no longer beat SemanticTTL, vCache, and SCALM at comparable search savings—or both the rule and MLP gates fail as the MLP already does on EvolvingQA/DailyQA—the central claim does not hold.

Watch

Extended reading notes

Core claim

The authors claim that framing open-web RAG caching as risk-constrained temporal inference—estimate the probability a cached result is stale from freshness class, age, and tier, then approve reuse only if that probability falls under a per-tier error budget—enables graceful degradation from answer hits to URL-list reuse to content validation, reaching about 97% search API savings at 0.1% hash-based stale error (roughly 0.034% answer-affecting after judge correction) at 24 hours and Pareto-dominating similarity-and-TTL baselines.

Load-bearing premise

The load-bearing premise is that exponential half-lives fitted from the authors’ own snapshot change counts, plus fixed per-tier budgets, are a good enough model of open-web staleness for reuse decisions beyond this benchmark.

Editorial extensions

If this is right

  • Recurring open-web RAG workloads can keep very high search savings without fixed TTLs that either waste cache hits or serve stale answers.
  • Separating URL-list reuse from answer reuse still recovers search cost when direct answer hits become too risky.
  • Hash-detected content change overstates true answer-affecting error by roughly a factor of three, so production metrics need answer-level calibration.
  • Adding cache tiers without a temporal risk gate can raise stale error even as savings rise.
  • Outside the training distribution, the simpler class-level decay model is the safer operational fallback than the learned MLP.

Reading between the lines

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

  • Online adaptation of per-tier or per-domain error budgets from observed stale rates could replace the fixed ε values once live feedback is available.
  • The same risk gate could be applied to KV or static-corpus caches whenever the underlying evidence is only approximately static.
  • Authenticated or privileged data feeds may be required for true real-time classes, because unauthenticated scrapes miss the most volatile pages that drive short half-lives.
  • Richer entity-level volatility features may be needed for the learned model to transfer to event-driven benchmarks where domain-snapshot features fail.
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

3 major / 5 minor

Summary. The paper proposes FreshCache, a three-tier semantic cache for open-web RAG that treats reuse as a risk-constrained temporal decision: after semantic match, a calibrated exponential decay model (optionally enhanced by a 25-feature MLP) estimates P(stale) and allows reuse only if that probability is below per-tier budgets (ε=0.10/0.20/0.35 for answers, URL lists, and content). The authors introduce FreshCache-Bench (8,072 base queries, 31,201 with paraphrases) with snapshot-based staleness labels at 1h/12h/24h/7d, and report that at 24h FreshCache_MLP achieves 97% search savings at 0.1% hash-based stale error (≈0.034% answer-affecting after LLM-judge correction on 396 pairs), while a temporally holdout-calibrated rule-based variant reaches 98% savings at 3.3% stale error, Pareto-dominating SemanticTTL, vCache, and SCALM. Ablations attribute an 11.6-point stale-error reduction to the temporal risk gate and a further 3.2 points to the MLP; live-web microbenchmarks and external QA checks are also reported.

Significance. If the results hold under broader deployment conditions, this is a clear and useful advance for open-web RAG systems: existing semantic caches largely ignore time-varying evidence freshness, and the three-tier risk-gated design (especially L2 URL-list reuse) is a concrete architectural contribution beyond fixed TTLs. FreshCache-Bench is a genuine community asset—temporally labeled, multi-class, and expanded with paraphrases—and the paper is unusually careful for systems work: temporal holdout for half-lives, query-cluster holdout for the MLP, bootstrap CIs, encoder/tier/risk-gate/MLP ablations, LLM-judge validation of the hash proxy, and explicit OOD checks. The negative finding that conditional GET is largely ineffective on the live web is also valuable and correctly motivates L1/L2 as the main savings path.

major comments (3)
  1. [Abstract / §5.1 / Table 2 / Appendix H] Abstract, §5.1, and Table 2 present FreshCache_MLP as the primary deployable result (0.1% stale / 97% saved), but Appendix H shows its hit rate collapses on EvolvingQA and DailyQA (features tuned to FreshCache-Bench domain volatility), while the rule-based model degrades more gracefully. The manuscript already recommends the rule-based fallback OOD (§4.5, §7); that qualification should be elevated into the abstract and main results so the absolute frontier numbers are not read as distribution-agnostic deployment guarantees. Without that, the strongest claim is over-scoped relative to the evidence.
  2. [§3.3.1 / Eqs. (3)–(5) / §5.1] The operational risk model (Eqs. 3–5) rests on class half-lives fit by MLE from the authors’ own snapshot change counts on FreshCache-Bench, plus fixed tier multipliers m_τ and budgets ε_τ. Temporal holdout (fit on 1h, eval on 24h) for the rule-based model is good practice and correctly raises stale error from 1.4% to 3.3%, but REAL_TIME’s 30s half-life remains an unobserved prior (§3.3.1, §5.4), and FAST observability is limited by 403s at short windows. A short sensitivity or alternative-prior analysis for these two classes—and a clearer statement that absolute P(stale) is benchmark-calibrated rather than independently validated web dynamics—would make the risk estimates’ scope precise without changing the internal ablation story.
  3. [§4.4 / Abstract / Table 2] Primary tables report hash-based stale error for reproducibility, with an LLM-judge correction that only 34.3% of hash changes are answer-affecting (§4.4, Table 13), yielding ≈0.034% true answer-level error for the MLP. Relative ordering is preserved, which is fine, but the abstract’s dual reporting (0.1% hash and 0.034% answer-affecting) can be misread as two independent measurements rather than a single proxy plus a post-hoc factor on 396 pairs. Please state explicitly in the main results that all method comparisons use the conservative hash proxy and that the 0.034% figure is a one-sided correction applied only to FreshCache_MLP’s operating point, not a re-evaluation of every baseline under the judge.
minor comments (5)
  1. [§5.3 / Table 4] Table 4 footnote on Total† (joint run vs isolated per-class rows) is important but easy to miss; a one-sentence reminder in the main text of §5.3 would prevent mis-summing L1/L2 hits across classes.
  2. [§5.4 / Abstract] §5.4’s honest production latency estimate (~80% median reduction after live-web fetch constants) is more useful than the paraphrase-inflated 99.8% p50; consider putting the ~80% figure in the abstract or Table 2 caption so readers do not over-index on simulation p50.
  3. [Appendix C / Eq. (4)] Appendix C sweeps only m_answer; a brief note that m_url-list and m_content were not swept (or a one-line check) would complete the robustness claim for Equation 4.
  4. [Figure 2 / §6.3] Figure 2 is clear; adding the 1h operating points as a small inset or companion panel (already in Appendix E) would help readers see that the risk gate’s value is time-window dependent, as argued in §6.3.
  5. [References] Minor: arXiv id and some citation years in the provided text look inconsistent with usual 2024–2025 venue dates; double-check bibliography metadata before camera-ready.

Circularity Check

2 steps flagged · score 3.0 of 10

Mild fitted-input circularity: class half-lives (Eq. 5) and MLP are MLE/trained on FreshCache-Bench snapshot change labels, then used to gate reuse whose stale error is scored on the same (held-out) snapshot ecosystem.

  1. fitted input called prediction [§3.3.1 Eqs. 3–5; §5.1 Table 2 (temporal holdout)]
    "Half-lives are fit by maximum likelihood from change counts across snapshot reruns at 1h, 12h, 24h, and 7 days... The fitted values are: h_TIMELESS=22d, h_SLOW=16d, h_MEDIUM=15h, h_FAST=3h, h_REAL-TIME=30s. ... The rule-based half-life in Table 2 is fit under a temporal holdout, using rerun_1h observations only and evaluating on rerun_24h... FreshCache achieves 98% search savings at 3.3% stale error under a temporal holdout calibration"

    The exponential risk model’s sole free parameters (class half-lives) are MLE-fitted directly to the same snapshot change counts that later define the ground-truth stale labels. Even under the 1h→24h temporal holdout, the functional form and class taxonomy are chosen and calibrated on this ecosystem, so the low reported stale error is statistically forced to the extent that a single exponential rate matches the observed rates; it is not an independent prediction of web dynamics outside the fitted distribution.

  2. fitted input called prediction [§3.3.2 MLP training; §4.1 / §5.1 evaluation on FreshCache-Bench]
    "We train a lightweight MLP on 25 input features... Training uses binary cross-entropy... on a query-cluster-level 70/30 held-out split... At the 24-hour evaluation window, FreshCache_MLP achieves 97% search API savings at 0.1% hash-based stale error"

    The MLP is supervised on binary change labels drawn from the identical FreshCache-Bench snapshot reruns used to compute the hash-based stale-error metric. Cluster holdout reduces leakage of individual change events, but the feature distribution (domain volatility, answer type, entity scores) and positive-class weighting are still learned from the same change ecosystem that defines success; the 0.1% figure is therefore a calibrated in-distribution operating point rather than an out-of-sample prediction of answer-affecting staleness.

full rationale

The paper does not claim a first-principles derivation of staleness; it explicitly fits an exponential decay model and trains an MLP on observed change counts from its own web snapshots, then evaluates reuse decisions against those (or temporally/cluster-held-out) labels. Temporal holdout for the rule-based model (fit on 1h, eval on 24h) and query-cluster holdout for the MLP prevent pure train-test leakage, and external EvolvingQA/DailyQA results are reported (showing degradation). This is ordinary empirical calibration rather than definitional circularity or a self-citation uniqueness chain. The residual circularity burden is that absolute Pareto numbers (0.1%/97%, 3.3%/98%) are tightly coupled to the benchmark’s change distribution and hash proxy, so strong in-distribution results partly reflect fitting to the evaluation labels rather than independent prediction of open-web dynamics. No self-definitional equations, no load-bearing self-cited uniqueness theorems, and no ansatz smuggled via author-prior work. Score 3 reflects one clear fitted-input pattern that is partially mitigated by holdouts, not a forced-by-construction central claim.

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

The central performance claim rests on a web-change exponential model and a five-class freshness taxonomy taken from prior work, plus many operational constants fitted or chosen on FreshCache-Bench (half-lives, error budgets, tier multipliers, similarity thresholds, MLP calibration). The invented system pieces are the three-tier risk-gated cache and the benchmark; independent evidence for the decay form is only partial (prior web studies plus this paper’s snapshots), and the learned risk model lacks external independent support beyond the training distribution.

free parameters (6)
  • class half-lives h_f = TIMELESS 22d; SLOW 16d; MEDIUM 15h; FAST 3h; REAL_TIME 30s prior
    MLE-fit from snapshot change counts at 1h/12h/24h/7d; load-bearing inputs to the exponential risk model (Eq. 5).
  • per-tier error budgets ε_τ = ε_L1=0.10, ε_L2=0.20, ε_L3=0.35
    Hand-chosen reuse thresholds that directly decide hits; not derived from a utility theorem.
  • tier multipliers m_τ = m_answer=1.5, m_url=1.2, m_content=1.0
    Fixed constants mapping content-tier risk to answer/URL tiers (Eq. 4); sensitivity sweep reported but values chosen.
  • similarity thresholds θ_L1, θ_L2 = θ_L1=0.40, θ_L2=0.35
    Operational cutoffs for semantic candidate selection before the risk gate.
  • MLP temperature T and training hyperparameters = T=1.4672; layers [64,32,16]; dropout 0.3; pos weight 8.14
    Post-hoc calibration and optimization choices that set reported P(stale) for FreshCache_MLP.
  • simulation latency constants = search 500ms; fetch 800ms; condGET 150ms; LLM 2000ms
    Search/fetch/conditional-GET/LLM latencies used for savings claims; live microbenchmark shows material mismatch especially at p95 and for conditional GET.
assumptions (5)
  • domain assumption Open-web content change rates are approximately exponential in age within freshness classes, so P_content = 1-exp(-log2/h_f · Δt) is an adequate risk model.
    Invoked in §3.3.1 citing Cho & Garcia-Molina and Fetterly et al.; half-lives then fit on FreshCache-Bench.
  • domain assumption Queries can be partitioned into five FreshLLMs-style freshness classes that govern decay, with REAL_TIME never cached.
    §3.2 taxonomy and classifier; REAL_TIME half-life kept as prior due to observability limits.
  • domain assumption A cached entry is stale if any URL content hash changed since population; hash change is a conservative proxy for answer-affecting staleness.
    Eq. 1 and §4.3–4.4; partially validated by LLM judge (34.3% answer-affecting).
  • ad hoc to paper Reuse is allowed only when semantic similarity and estimated staleness risk both pass per-tier thresholds (Eq. 2).
    Core decision rule of FreshCache; defines the method rather than a prior theorem.
  • domain assumption BGE-M3 cosine similarity plus spaCy entity match is a sufficient semantic gate for L1/L2 candidate selection.
    §3.4–3.7; entity filter evaluated on 20 adversarial pairs.
invented entities (3)
  • FreshCache three-tier risk-gated semantic cache (L1 answers, L2 URL lists, L3 content)
    purpose: Enable graceful degradation of reuse under staleness risk rather than binary cache-or-miss.
    Primary system contribution; evaluated only within this paper’s simulation and microbenchmarks.
  • FreshCache-Bench temporally labeled open-web caching benchmark
    purpose: Provide ground-truth staleness labels from real web snapshots plus paraphrases for evaluation and calibration.
    New dataset artifact; not independently released/validated in the manuscript.
  • 25-feature MLP staleness risk estimator with temperature scaling
    purpose: Within-class discrimination beyond class-level exponential risk.
    Learned on FreshCache-Bench clusters; Appendix H shows poor transfer to external QA change labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Risk-Constrained Freshness-Aware Semantic Caching for Open-Web Retrieval-Augmented LLMs." pith.science (2026). https://pith.science/paper/UK5S3SB5

@misc{pith2026260704281,
  author       = {Pith},
  title        = {Pith review of: Risk-Constrained Freshness-Aware Semantic Caching for Open-Web Retrieval-Augmented LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UK5S3SB5}},
  note         = {Machine review of arXiv:2607.04281}
}
read the original abstract

Semantic caching reduces the latency and cost of retrieval-augmented generation (RAG) by serving cached answers to semantically similar queries, but most existing methods do not model the time-varying freshness of open-web evidence. We present FreshCache, a three-tier semantic cache that treats cache reuse as a risk-constrained temporal inference problem: before approving a cache hit, FreshCache estimates the probability that the cached result is stale using a fitted exponential decay model enhanced by a learned MLP, and approves reuse only when that probability falls below a per-tier error budget across answers (epsilon = 0.10), URL lists (epsilon = 0.20), and page content (epsilon = 0.35). This allows the system to degrade gracefully as entries age rather than forcing a binary choice between a stale hit and a full pipeline execution. We introduce FreshCache-Bench, a benchmark of 8,072 base queries across five freshness classes with ground truth staleness labels drawn from real web snapshots at 1, 12, 24 hours, and 7 days after a baseline crawl, expanded to 31,201 queries via paraphrase generation. At the 24-hour evaluation window, FreshCache_MLP achieves 97% search API savings at 0.1% hash-based stale error, and an LLM-judge evaluation on 396 confirmed change pairs shows that only 34.3% of detected content changes actually affect answer correctness, placing true answer-affecting stale error at approximately 0.034%. The rule-based FreshCache achieves 98% search savings at 3.3% stale error under a temporal holdout calibration, outperforming SemanticTTL (14.9% stale, 72% saved), vCache (7.2% stale, 47% saved), and SCALM (5.2% stale, 96% saved). Ablations show the temporal risk gate accounts for an 11.6 point reduction in stale error over similarity-only reuse, and the learned MLP reduces stale error a further 3.2 points over the rule-based model.

Figures

Figures reproduced from arXiv: 2607.04281 by the authors.

Figure 1
Figure 1. FreshCache system architecture showing semantic lookup, risk-gated cache reuse, and fallback execution [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Search API savings versus stale error rate at [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. TTL Pareto frontier sweep at t = 24h using BGE-M3. TTL-based baselines are evaluated across eight scale factors, from 0.01× to 1.5× their default TTLs. Filled markers denote the default operating points from [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Workload characterization: cosine similarity [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 1 linked inside Pith

  1. [1]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume =

  2. [2]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume =. 2020 , publisher =

    2020

  3. [3]

    2023 , archivePrefix =

    Gao, Yunfan and Xiong, Yun and Gao, Xinyu and Jia, Kangxiang and Pan, Jinliu and Bi, Yuxi and Dai, Yi and Sun, Jiawei and Wang, Haofen , title =. 2023 , archivePrefix =

    2023

  4. [4]

    2021 , archivePrefix =

    Nakano, Reiichiro and Hilton, Jacob and Balwit, Avital and Wu, Jeff and Ouyang, Long and Kim, Christina and Hesse, Christopher and Jain, Shantanu and Kosaraju, Vineet and Saunders, William and others , title =. 2021 , archivePrefix =

    2021

  5. [5]

    2022 , archivePrefix =

    Lazaridou, Angeliki and Gribovskaya, Elena and Stokowiec, Wojciech and Grigorev, Nikolai , title =. 2022 , archivePrefix =

    2022

  6. [6]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , pages =

    Mallen, Alex and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh , title =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , pages =

  7. [7]

    Proceedings of the 3rd Workshop on Efficient Natural Language and Speech Processing , year =

    Fu, Zhenyu and Feng, Xiaodong , title =. Proceedings of the 3rd Workshop on Efficient Natural Language and Speech Processing , year =

  8. [8]

    Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (

    Bang, Fu , title =. Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (. 2023 , address =

    2023

Show all 31 references
  1. [9]

    2024 , archivePrefix =

    Zhu, Hanlin and Zhu, Banghua , title =. 2024 , archivePrefix =

    2024

  2. [10]

    von Riedemann and Cong Zhang and Jiangchuan Liu , title =

    Jiaxing Li and Chi Xu and Feng Wang and Isaac M. von Riedemann and Cong Zhang and Jiangchuan Liu , title =. arXiv preprint arXiv:2406.00025 , year =

    arXiv

  3. [11]

    , title =

    Schroeder, Luis Gaspar and Desai, Aditya and Cuadron, Alejandro and Chu, Kyle and Liu, Shu and Zhao, Mark and Krusche, Stephan and Kemper, Alfons and Stoica, Ion and Zaharia, Matei and Gonzalez, Joseph E. , title =. 2025 , archivePrefix =

    2025

  4. [12]

    and others , title =

    Chan, Brian J. and others , title =. 2025 , archivePrefix =

    2025

  5. [13]

    2024 , archivePrefix =

    Jin, Chao and others , title =. 2024 , archivePrefix =

    2024

  6. [14]

    Proceedings of the Twentieth European Conference on Computer Systems , pages =

    Yao, Jiayi and Li, Hanchen and Liu, Yuhan and Ray, Siddhant and Cheng, Yihua and Zhang, Qizheng and Du, Kuntai and Lu, Shan and Ananthanarayanan, Ganesh , title =. Proceedings of the Twentieth European Conference on Computer Systems , pages =

  7. [15]

    and Luong, Thang , title =

    Vu, Tu and Iyyer, Mohit and Wang, Xuezhi and Constant, Nick and Wei, Jerry and Wei, Jason and Tar, Chris and Sung, Yun-Hsuan and Zhou, Denny and Le, Quoc V. and Luong, Thang , title =. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =

    2023

  8. [16]

    Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing , pages =

    Reimers, Nils and Gurevych, Iryna , title =. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing , pages =

    2019

  9. [17]

    2024 , archivePrefix =

    Chen, Jianlv and Xiao, Shitao and Zhang, Peitian and Luo, Kun and Lian, Defu and Liu, Zheng , title =. 2024 , archivePrefix =

    2024

  10. [18]

    ACM Transactions on Internet Technology , volume =

    Cho, Junghoo and Garcia-Molina, Hector , title =. ACM Transactions on Internet Technology , volume =

  11. [19]

    and Elsas, Jonathan L

    Adar, Eytan and Teevan, Jaime and Dumais, Susan T. and Elsas, Jonathan L. , title =. Proceedings of the 22nd Annual ACM Symposium on User Interface Software and Technology , pages =

  12. [20]

    Dense Passage Retrieval for Open-Domain Question Answering , booktitle =

    Karpukhin, Vladimir and O. Dense Passage Retrieval for Open-Domain Question Answering , booktitle =. 2020 , address =

    2020

  13. [21]

    and Eisenschlos, Julian Martin and Gillick, Daniel and Eisenstein, Jacob and Cohen, William W

    Dhingra, Bhuwan and Cole, Jeremy R. and Eisenschlos, Julian Martin and Gillick, Daniel and Eisenstein, Jacob and Cohen, William W. , title =. Transactions of the Association for Computational Linguistics , volume =. 2022 , publisher =

    2022

  14. [22]

    Advances in Neural Information Processing Systems , volume =

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , title =. Advances in Neural Information Processing Systems , volume =. 2022 , publisher =

    2022

  15. [23]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Tan, Qingyu and Ng, Hwee Tou and Bing, Lidong , title =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2023 , address =

    2023

  16. [24]

    Proceedings of the 12th International Conference on World Wide Web , pages =

    Fetterly, Dennis and Manasse, Mark and Najork, Marc and Wiener, Janet , title =. Proceedings of the 12th International Conference on World Wide Web , pages =. 2003 , address =

    2003

  17. [25]

    and Choi, Yejin and Inui, Kentaro , title =

    Kasai, Jungo and Sakaguchi, Keisuke and Takahashi, Yoichi and Le Bras, Ronan and Asai, Akari and Yu, Xinyan and Radev, Dragomir and Smith, Noah A. and Choi, Yejin and Inui, Kentaro , title =. Advances in Neural Information Processing Systems , volume =. 2023 , publisher =

    2023

  18. [26]

    2025 , archivePrefix =

    Pham, Thinh and Nguyen, Phi Tran Anh and Nguyen, Anh Tuan and Luu, Anh Tuan , title =. 2025 , archivePrefix =

    2025

  19. [27]

    , title =

    Zhao, Bowen and Brumbaugh, Zander and Wang, Yizhong and Hajishirzi, Hannaneh and Smith, Noah A. , title =. Findings of the Association for Computational Linguistics:. 2024 , address =

    2024

  20. [28]

    , title =

    Guo, Chuan and Pleiss, Geoff and Sun, Yu and Weinberger, Kilian Q. , title =. Proceedings of the 34th International Conference on Machine Learning , pages =. 2017 , publisher =

    2017

  21. [29]

    Proceedings of the 39th

    Gill, Waris and Elidrisi, Mohamed and Kalapatapu, Pallavi and Ahmed, Ammar and Anwar, Ali and Gulzar, Muhammad Ali , title =. Proceedings of the 39th. 2025 , note =

    2025

  22. [30]

    2025 , archivePrefix =

    Wang, Chen and Liu, Xunzhuo and Zhu, Yue and Youssef, Alaa and Nagpurkar, Priya and Chen, Huamin , title =. 2025 , archivePrefix =. 2510.26835 , note =

    2025

  23. [31]

    Proceedings of the 17th International Conference on World Wide Web , series =

    Olston, Christopher and Pandey, Sandeep , title =. Proceedings of the 17th International Conference on World Wide Web , series =. 2008 , pages =

    2008

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.