Pith. sign in

REVIEW 9 minor 38 references

HERALD: Counterfactual Audits and Minimal Repairs for Proof-of-Retrieval Rewards

T0 review · 0 major / 9 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Citation laundering is the real reward loophole in search agents, and a single exact membership check closes it.

desk verdict A rigorous finite-set audit that identifies a single-check repair for citation laundering; the certificate is honestly scoped, but the artifact promise needs backing. read the letter →

arxiv 2608.06012 v1 pith:I7CIH5X7 submitted 2026-08-06 cs.AI

Zhuowen Liu , Bohan Cui , YinShang Guo , Yuting Wang , Hao Li This is my paper · ORCID
classification cs.AI
keywords rewardhackingcitationlaunderingproof-of-retrievalcounterfactualauditsearchagentsreinforcementlearningdetectorlatticegroup-relativeoptimization
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

This paper is trying to establish that proof-of-retrieval rewards can be audited offline, before any policy is trained, to find the exact check that closes a reward loophole. Applying paired same-question edits to four Qwen3-8B pools, the paper finds that the base reward $R_0$ already rejects search deletion and fake citation IDs, but a label-free citation-laundering attack—substituting a real corpus passage the agent never retrieved—succeeds. Enumerating all eight subsets of three exact detectors identifies $L$, the check that a corpus-valid citation appeared in the retrieved set, as the observed inclusion-minimal repair: every contract containing $L$ has zero empirical attack-success rate (one-sided 95% cluster upper bound 0.50%), and every contract lacking $L$ keeps the worst-case 6.66% gap. The paper also shows that a larger detector bundle can be less safe, because an attack that removes an oracle support-ID penalty can cancel the new membership penalty. Under matched 5M-token training, $R[L]$ improves citation precision and support recall on some benchmarks and reduces targeted attackability on two suites, but the detector is active in only 18 of 58,368 trajectories, so the evidence separates robust scoring from sparse learning signal and limited transfer.

What carries the argument

The central object is the exact membership detector $L(\tau)=\mathbb{1}\{\exists c\in C(\tau)\cap\mathcal{P}: c\notin E(\tau)\}$, which flags any final citation that is a valid corpus ID but was absent from the retrieved passage set. The audit machinery is a paired counterfactual test: for each trajectory-and-edit pair $(\tau, A(\tau))$ it records the margin $M_S(\tau)=M_0(\tau)-\sum_{j\in S}\lambda_j[f_j(A(\tau))-f_j(\tau)]$ and the attack-success rate $\mathrm{ASR}_S=\mathbb{E}[\mathbb{1}\{M_S(\tau)\ge 0\}]$, restricted to edits that are eligible (the target detector flips from 0 to 1) and isolated (unrelated visible detectors are unchanged). The complete $2^3$ lattice over $\{U,L,F\}$ decides inclusion-minimality by checking every proper subset, with $U$ marking a nonempty final with no search and $F$ marking a citation ID outside the corpus. The counterintuitive result that a larger bundle can be weaker is carried by the cancellation credit $C_{T\setminus S}(A,\tau)=-\sum_{j\in T\setminus S}\lambda_j d_j(A,\tau)$: when the original already triggered a penalty that the attack removes, $d_j<0$, so adding that detector raises the attack margin. The finite-set certificate for $R[L]$ is $\lambda_L>\max_{\tau,A}[R_0(A(\tau))-R_0(\tau)]$, with observed maximum $0.81375$ against the fixed $\lambda_L=1.2$. For policy transfer, the group-normalization identity $\tilde{r}'_i=(r_i-\bar{r}_g)-\lambda_L(\ell_i-\bar{\ell}_g)$ shows that a binary detector contributes $\lambda_L^2 \bar{\ell}_g(1-\bar{\ell}_g)$ to the squared centered reward, so a mixed group is necessary for the detector to reach normalized advantages.

What would settle it

Rerun the audit on the same 593 isolated-eligible questions with a dense-retrieval or visible-LLM candidate generator and the same frozen corpus; the central claim fails if any eligible laundering edit has a base-reward gain above $\lambda_L=1.2$, or if $R[L]$ shows any positive attack-success rate with upper bound above zero.

Watch

Extended reading notes

Core claim

The central claim is that the observed vulnerability of the base reward $R_0$ is citation laundering, not search deletion or fake IDs: an edit can replace a citation with a real corpus passage that never appeared in the retrieved evidence and still win the reward on a nontrivial tail of questions. On 593 isolated-eligible question clusters, $R_0$ gives the laundering attack a 4.30% attack-success rate under label-free first-visible selection and a 6.66% rate under oracle worst-case selection. The complete $2^3$ detector lattice shows that the exact membership check $L$—a citation that is a real corpus ID but was not retrieved—is the observed inclusion-minimal repair: every subset containing $L$ has zero observed primary attack-success rate, with a one-sided 95% cluster upper bound of 0.50%, while every subset lacking $L$ retains the 6.66% gap. The paper states this as a conditional, finite-set result, not a population guarantee, and shows that $R_{full}$, a broader reward with oracle support-ID penalties, can be less robust because the attack removes an active support penalty and partly cancels the new membership penalty.

Load-bearing premise

The repair is measured only on the tested pools, eligibility rules, candidate generators, and frozen corpus; an untested attack family such as dense retrieval or visible-only LLM candidates could beat the $\lambda_L=1.2$ penalty margin or evade the membership check $L$, so the minimality result is conditional rather than universal.

Editorial extensions

If this is right

  • On the tested pools, adding $L$ alone closes citation laundering, same-final search deletion, and fake citation IDs; the $U$ and $F$ checks are redundant in this context.
  • A larger detector bundle can be strictly worse than the minimal repair: $R_{full}$ retains 3.37% worst-case laundering attack-success rate while $R[L]$ has zero, because removing an oracle support-ID penalty can cancel the new membership penalty.
  • Training a policy with $R[L]$ under a matched 5M-token budget improves equal-suite citation precision by 2.02 and support recall by 1.46 points, reduces unsupported citations by 1.69, and cuts laundering attackability on 2Wiki and MuSiQue, while passing EM non-inferiority on HotpotQA and 2Wiki but not MuSiQue.
  • The strict-$L$ detector fires in only 18 of 58,368 training trajectories and group normalization can erase it, so the policy evidence supports improved targeted attackability, not reduced natural incidence of $L$ violations.
  • Zero observed attack rates should be reported with the exact one-sided cluster upper bound of 0.50%, not as proof of population-level robustness; paired margins are needed alongside the attack-success rate.

Reading between the lines

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

  • Beyond the paper: the penalty-cancellation mechanism is likely general—for any reward term an attack can switch off, such as tool cost or an oracle support penalty, adding that term to a bundle can reopen the attack, so audits that compare only a base reward with a full bundle can miss the best repair.
  • Beyond the paper: deployed agents that cite from dense retrieval or visible LLM generators would need a generalization of $L$, such as a provenance or embedding-neighborhood check, because the paper leaves those attack families untested.
  • Beyond the paper: the group-normalization identity implies a concrete training intervention—construct groups that mix detector-positive and detector-negative rollouts—otherwise mean subtraction cancels a rare detector penalty, as happened in two of the seven penalized groups.
  • Beyond the paper: the same audit pattern of exact edits, eligibility conditions, lattice enumeration, and exact upper bounds on zero events could be applied to other components of composite rewards, such as tool cost, format, and answer grounding, before deployment.
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

0 major / 9 minor

Summary. The paper proposes HERALD, an offline reward-audit framework for search agents. It applies field-preserving same-question edits to logged trajectories, separates candidate-visible from oracle information, and evaluates the full lattice of three exact detectors: U (nonempty final with no search), L (citation ID not among retrieved evidence), and F (fake citation ID). On four Qwen3-8B pools, the base reward R0 rejects same-final deletion and fake IDs, but an adaptive citation-laundering attack succeeds with 4.30% first-visible and 6.66% oracle-worst ASR on 593 isolated-eligible question clusters. Adding a targeted L penalty with λ_L=1.2 yields zero observed ASR with a one-sided cluster upper bound of 0.50%, and the complete 2^3 ablation shows that L is the observed inclusion-minimal repair; a broader oracle bundle Rfull retains 3.37% ASR because removing an active oracle support-ID penalty cancels the new membership penalty. The audit replicates across four models and a visible BM25 attacker. Matched 5M-token Search-GRPO runs show R[L] improves citation precision and support recall, reduces unsupported citations and targeted attackability on two of three benchmarks, but fails the EM non-inferiority gate on MuSiQue and does not reduce natural L incidence, with only 18 of 58,368 training trajectories exposing the detector.

Significance. HERALD's main value is methodological: it gives concrete procedures for eligibility filtering, detector isolation, cluster-level upper bounds, and complete detector-lattice comparisons, and it formalizes the penalty-cancellation phenomenon in Eqs. (8)-(12). The negative results are as useful as the positive ones: the finding that Rfull can be weaker than R[L] because of cancellation, and the demonstration that the L penalty reaches the optimizer only 18 times and is partly canceled by group normalization, are clearly analyzed. The paper is scrupulous about the scope of its claims: zero-ASR, minimality, and finite-set margins are repeatedly qualified as observed and conditional on the implemented pools, eligibility rules, and candidate generators. The Limitations explicitly note that dense retrieval and visible-only LLM attacks are untested. I therefore regard the finite-set certificate as an honest bound on the contribution's portability rather than an internal inconsistency; the stress-test concern about untested attack families is real but does not undermine the paper's internally scoped claims.

minor comments (9)
  1. [Finite-set guarantee and real cancellation case (Eq. 12)] Please state explicitly which candidate generator family the 'audited maximum R0 gain is 0.81375' is computed over (primary lexical, BM25, hash-random, or all), so the reader can see exactly what the finite-set certificate covers.
  2. [Matched policy audit] The training is described as 'strict 5M-token matched' but the two arms stop at 5,003,240 and 5,003,684 tokens; clarify the stopping rule and report the exact token counts consistently.
  3. [Paired estimands] The one-sided upper bound is typeset as '1−0.051/n'; it should be '1−0.05^{1/n}'.
  4. [General presentation] The manuscript contains duplicated blocks: Figure 1 caption and Table 1 appear twice, and the 'Eligibility and isolation' and 'Complete detector lattice' paragraphs are repeated; please remove the duplicates.
  5. [References] The reference to 'Y ang et al. 2018' has a spurious space in the author name.
  6. [Table 4 and Figure 2(c)] Please add the support-recall row to Table 4 or explicitly point to Figure 2(c), since the abstract and text report a +1.46 support-recall improvement but the table omits that endpoint.
  7. [Reward-Audit Results] The statement that 'R[U+L+F] is identical' should clarify that it refers to identical observed ASRs and margins on the isolated-eligible set, not to identical reward values on all trajectories, since U and F penalties can affect margins when those detectors trigger.
  8. [Complete detector lattice] Align the two definitions of minimality: the text uses both 'no immediate proper subset' and 'no proper subset'; the implemented full-lattice enumeration checks all proper subsets, so the latter wording should be used consistently.
  9. [Contributions] The contribution list mentions 'reproducible artifacts' but no URL or repository identifier is given; please provide the link or state where the code and data will be released.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper explicitly labels its zero-ASR repair a post hoc finite-set certificate, and the empirical audits and policy transfer stand independently.

full rationale

The paper's central claim—that strengthening L is the observed inclusion-minimal repair—is explicitly scoped as a finite-set, post hoc certificate rather than a universal prediction. In Eq. (12), the zero-ASR conclusion is an algebraic consequence of λ_L=1.2 exceeding the observed maximum R0 gain of 0.81375 on the implemented candidate generators; the paper itself says this comparison is 'a post hoc finite-set certificate, not a population guarantee' and repeatedly qualifies results as 'observed, conditional findings.' The audit's empirical content—R0's vulnerability to citation laundering, the failure of the larger Rfull bundle through support-ID penalty cancellation, pool sensitivity, cross-model replication, and matched-token policy transfer—does not reduce to the penalty definition and is assessed against external data (HotpotQA, 2WikiMultiHopQA, MuSiQue, four models, held-out evaluation questions). No load-bearing step is justified by a self-citation; the references to ToRL, R1-Searcher, ReSearch, and similar works are external and are not used to derive the audit conclusions. The closest issue is that the zero-ASR guarantee is guaranteed by the chosen penalty, but the paper discloses this and labels it post hoc, so it is not a hidden equation-to-equation reduction or a fitted parameter renamed as a prediction.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

Central claims rest on three hand-fixed audit penalties (lambda_L=1.2 is load-bearing), the closed-world assumption for the attack space, the candidate-visible/oracle boundary, and standard cluster-bootstrap assumptions. No invented entities.

free parameters (3)
  • lambda_L (targeted L penalty) = 1.2
    Hand-fixed in the audit manifest. The finite-set guarantee (Eq. 12) requires lambda_L to exceed the observed max R0 gain of 0.81375; at scale 0.5 the ASR is 3.2%, so the zero-ASR repair depends on this choice.
  • lambda_U (same-final deletion penalty) = 1.6
    Included in the audited lattice contracts; partly redundant because R0 already rejects same-final deletion on these pools.
  • lambda_F (fake citation ID penalty) = 1.0
    Also redundant on these pools; R0 already rejects fake citation IDs. Listed because the lattice contracts use it.
assumptions (6)
  • standard math The reward program is additive and linear in the audited detectors (Eq. 8).
    The margin decomposition M_S = M_0 minus sum lambda_j [f_j(A) minus f_j(tau)] is exact for the defined additive reward and is used for the cancellation analysis.
  • domain assumption The closed corpus index and logged trajectories are fixed; no live web calls are made.
    Audit replay assumes the corpus and logs are static and that retrieval state at deployment resembles the logged state.
  • domain assumption Candidate-visible generators see only the question, logged trace, and corpus, never gold answers, support IDs, or evaluation labels.
    The label-free attack claim depends on this information boundary, shown in Figure 1 and Table 1.
  • ad hoc to paper The observed pools, eligibility rules, and candidate generators define the attack space for the zero-ASR and minimality claims.
    The paper explicitly conditions these claims on pools, eligibility, generators, and closed corpus in the Limitations; the guarantee is not population-level.
  • domain assumption Benchmark gold answers and support IDs are treated as oracle ground truth for Rfull and oracle-worst selection.
    Rfull's support-ID penalty reads benchmark annotations; the paper notes these annotations may be incomplete.
  • standard math Question clusters are independent sampling units for bootstrap and exact upper bounds.
    Used for 5,000-cluster bootstrap and the 1 minus 0.05^(1/n) zero-event bound; duplicate MuSiQue IDs are clustered.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HERALD: Counterfactual Audits and Minimal Repairs for Proof-of-Retrieval Rewards." pith.science (2026). https://pith.science/paper/I7CIH5X7

@misc{pith2026260806012,
  author       = {Pith},
  title        = {Pith review of: HERALD: Counterfactual Audits and Minimal Repairs for Proof-of-Retrieval Rewards},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7CIH5X7}},
  note         = {Machine review of arXiv:2608.06012}
}
abstract

Search-agent rewards mix answer quality, citation grounding, tool cost, and anti-hacking terms; a high score therefore need not imply that cited evidence was retrieved, and added penalties can cancel. We introduce HERALD, an offline audit that applies exact same-question interventions, separates candidate-visible from oracle information, and enumerates detector contracts before policy optimization. On four Qwen3-8B pools from HotpotQA, 2WikiMultiHopQA, and MuSiQue, $R_0$ rejects search deletion and fake IDs, but a label-free citation-laundering attack succeeds. A complete $2^3$ ablation identifies targeted strengthening of $L$---citing a corpus passage absent from the retrieved evidence---as the observed inclusion-minimal repair: $R[L]$ has zero empirical ASR with a 0.50% one-sided cluster upper bound. The gap persists across pool rules, a visible BM25 attacker, and four models; broader hardening remains vulnerable when the attack removes an oracle support-ID penalty. Under strict 5M-token matched training evaluated on 256 paired questions per benchmark, $R[L]$ meets the EM non-inferiority gate on HotpotQA and 2Wiki but not MuSiQue. Equal-suite citation precision and support recall improve by 2.02 and 1.46 points, unsupported citations fall by 1.69, and laundering attackability falls on 2Wiki and MuSiQue. Natural $L$ is not reduced, and the detector appears in only 18 of 58,368 training trajectories. HERALD thus separates robust scoring, sparse learning signal, and policy transfer.

Figures

Figures reproduced from arXiv: 2608.06012 by the authors.

Figure 1
Figure 1. HERALD audits a reward before policy optimization. It replays a structured trajectory, applies field-preserving same Figure 1: HERALD audits a reward before policy optimization. It replays a structured trajectory, applies field-preserving [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Audit-to-policy evidence. (a) Complete Figure 2: Audit-to-policy evidence. (a) Complete [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The strict-L training signal is active but sparse. R[L] exposes 18 of 58,368 trajectories; only seven question-aligned groups receive the penalty, and group normalization cancels it in two. At the expanded natural endpoint, the data provide no evidence of reduced natural strict-L incidence relative to R0. selection rule that reads gold answers, support IDs, or eval￾uation fields must be labeled oracle. Separate scor… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 29 canonical work pages

  1. [1]

    2025 , howpublished =

    Shao, Rulin and Li, Shuyue Stella and Xin, Rui and Geng, Scott and Wang, Yiping and Oh, Sewoong and Du, Simon Shaolei and Lambert, Nathan and Min, Sewon and Krishna, Ranjay and Tsvetkov, Yulia and Hajishirzi, Hannaneh and Koh, Pang Wei and Zettlemoyer, Luke , title =. 2025 , howpublished =

    work page 2025

  2. [2]

    2025 , howpublished =

    Zhao, Yulai and Liu, Haolin and Yu, Dian and Kung, Sunyuan and Chen, Meijia and Mi, Haitao and Yu, Dong , title =. 2025 , howpublished =

    work page 2025

  3. [3]

    2026 , howpublished =

    Helff, Lukas and Delfosse, Quentin and Steinmann, David and H. 2026 , howpublished =

    work page 2026

  4. [4]

    2025 , howpublished =

    Cai, Xin-Qiang and Wang, Wei and Liu, Feng and Liu, Tongliang and Niu, Gang and Sugiyama, Masashi , title =. 2025 , howpublished =

    work page 2025

  5. [5]

    An Imperfect Verifier is Good Enough: Learning with Noisy Rewards , year =

    Plesner, Andreas and Guzm. An Imperfect Verifier is Good Enough: Learning with Noisy Rewards , year =

    work page

  6. [6]

    2025 , howpublished =

    Hochlehnert, Andreas and Bhatnagar, Hardik and Udandarao, Vishaal and Albanie, Samuel and Prabhu, Ameya and Bethge, Matthias , title =. 2025 , howpublished =

    work page 2025

  7. [7]

    and Wang, Dong and Zamani, Hamed and Han, Jiawei , title =

    Jin, Bowen and Zeng, Hansi and Yue, Zhenrui and Yoon, Jinsung and Arik, Sercan O. and Wang, Dong and Zamani, Hamed and Han, Jiawei , title =. 2025 , howpublished =

    work page 2025

  8. [8]

    2025 , howpublished =

    Song, Huatong and Jiang, Jinhao and Min, Yingqian and Chen, Jie and Chen, Zhipeng and Zhao, Wayne Xin and Fang, Lei and Wen, Ji-Rong , title =. 2025 , howpublished =

    work page 2025

Show all 38 references
  1. [9]

    and Zhang, Wen and Chen, Huajun and Yang, Fan and Zhou, Zenan and Chen, Weipeng , title =

    Chen, Mingyang and Sun, Linzhuang and Li, Tianpeng and Sun, Haoze and Zhou, Yijie and Zhu, Chenzheng and Wang, Haofen and Pan, Jeff Z. and Zhang, Wen and Chen, Huajun and Yang, Fan and Zhou, Zenan and Chen, Weipeng , title =. 2025 , howpublished =

    2025

  2. [10]

    2025 , howpublished =

    Li, Xuefeng and Zou, Haoyang and Liu, Pengfei , title =. 2025 , howpublished =

    2025

  3. [11]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Zhang, Jiajie and Lv, Xin and Feng, Ling and Hou, Lei and Li, Juanzi , title =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =. doi:10.18653/v1/2026.acl-long.950 , url =

    2026 doi

  4. [12]

    and Nief, Todd and Garbacea, Cristina and Veitch, Victor , title =

    Reber, David and Richardson, Sean M. and Nief, Todd and Garbacea, Cristina and Veitch, Victor , title =. Proceedings of the 42nd International Conference on Machine Learning , year =

  5. [13]

    and Hu, Xiyang , title =

    Zhou, Xiaolin and Liu, Jinbo and Li, Li and Rossi, Ryan A. and Hu, Xiyang , title =. 2026 , howpublished =

    2026

  6. [14]

    2026 , howpublished =

    Xu, Ke and Xu, Han and Chen, Xinran and Wang, Yuqian and Li, Zhixuan and Liu, Xiaojian and Wu, Changwo and Xia, Jianqiang and Li, Yuchen , title =. 2026 , howpublished =

    2026

  7. [15]

    2026 , howpublished =

    Thaman, Kunvar , title =. 2026 , howpublished =

    2026

  8. [16]

    2025 , howpublished =

    Ma, Shengjie and Deng, Chenlong and Mao, Jiaxin and Huang, Jiadeng and Wang, Teng and Wu, Junjie and Zhang, Changwang and Wang, Jun , title =. 2025 , howpublished =

    2025

  9. [17]

    2025 , howpublished =

    Hao, Chuzhan and Feng, Wenfeng and Zhang, Yuewei and Wang, Hao , title =. 2025 , howpublished =

    2025

  10. [18]

    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 , year =

  11. [19]

    International Conference on Learning Representations , year =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , title =. International Conference on Learning Representations , year =

  12. [20]

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

    Trivedi, Harsh and Balasubramanian, Niranjan and Khot, Tushar and Sabharwal, Ashish , title =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =. doi:10.18653/v1/2023.acl-long.557 , url =

    doi

  13. [21]

    2022 , howpublished =

    Menick, Jacob and Trebacz, Maja and Mikulik, Vladimir and Aslanides, John and Song, Francis and Chadwick, Martin and Glaese, Mia and Young, Susannah and Campbell-Gillingham, Lucy and Irving, Geoffrey and McAleese, Nat , title =. 2022 , howpublished =

    2022

  14. [22]

    Computational Linguistics , year =

    Rashkin, Hannah and Nikolaev, Vitaly and Lamm, Matthew and Aroyo, Lora and Collins, Michael and Das, Dipanjan and Petrov, Slav and Tomar, Gaurav Singh and Turc, Iulia and Reitter, David , title =. Computational Linguistics , year =. doi:10.1162/coli_a_00486 , url =

    doi

  15. [23]

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

    Gao, Luyu and Dai, Zhuyun and Pasupat, Panupong and Chen, Anthony and Chaganty, Arun Tejasvi and Fan, Yicheng and Zhao, Vincent and Lao, Ni and Lee, Hongrae and Juan, Da-Cheng and Guu, Kelvin , title =. Proceedings of the 61st Annual Meeting of the Association for Computationa...

    doi

  16. [24]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , year =

    Yue, Xiang and Wang, Boshi and Chen, Ziru and Zhang, Kai and Su, Yu and Sun, Huan , title =. Findings of the Association for Computational Linguistics: EMNLP 2023 , year =. doi:10.18653/v1/2023.findings-emnlp.307 , url =

    2023 doi

  17. [25]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

    Ribeiro, Marco Tulio and Wu, Tongshuang and Guestrin, Carlos and Singh, Sameer , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =. doi:10.18653/v1/2020.acl-main.442 , url =

    2020 doi

  18. [26]

    and Mulcaire, Phoebe and Ning, Qiang and Singh, Sameer and Smith, Noah A

    Gardner, Matt and Artzi, Yoav and Basmov, Victoria and Berant, Jonathan and Bogin, Ben and Chen, Sihao and Dasigi, Pradeep and Dua, Dheeru and Elazar, Yanai and Gottumukkala, Ananth and Gupta, Nitish and Hajishirzi, Hannaneh and Ilharco, Gabriel and Khashabi, Daniel and Lin, K...

    2020

  19. [27]

    Concrete Problems in

    Amodei, Dario and Olah, Chris and Steinhardt, Jacob and Christiano, Paul and Schulman, John and Man. Concrete Problems in. 2016 , howpublished =

    2016

  20. [28]

    Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence , year =

    Everitt, Tom and Krakovna, Victoria and Orseau, Laurent and Legg, Shane , title =. Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence , year =. doi:10.24963/ijcai.2017/656 , url =

    2017 doi

  21. [29]

    Synthese , year =

    Everitt, Tom and Hutter, Marcus and Kumar, Ramana and Krakovna, Victoria , title =. Synthese , year =

  22. [30]

    Advances in Neural Information Processing Systems , year =

    Skalse, Joar and Howe, Nikolaus and Krasheninnikov, Dmitrii and Krueger, David , title =. Advances in Neural Information Processing Systems , year =. doi:10.52202/068431-0687 , url =

    doi

  23. [31]

    International Conference on Learning Representations , year =

    Pan, Alexander and Bhatia, Kush and Steinhardt, Jacob , title =. International Conference on Learning Representations , year =

  24. [32]

    Proceedings of the 40th International Conference on Machine Learning , year =

    Gao, Leo and Schulman, John and Hilton, Jacob , title =. Proceedings of the 40th International Conference on Machine Learning , year =

  25. [33]

    and Hajishirzi, Hannaneh , title =

    Lambert, Nathan and Pyatkin, Valentina and Morrison, Jacob and Miranda, Lester James and Lin, Bill Yuchen and Chandu, Khyathi and Dziri, Nouha and Kumar, Sachin and Zick, Tom and Choi, Yejin and Smith, Noah A. and Hajishirzi, Hannaneh , title =. 2024 , howpublished =

    2024

  26. [34]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year =

    Gao, Tianyu and Yen, Howard and Yu, Jiatong and Chen, Danqi , title =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year =

    2023

  27. [35]

    and Zhang, Tianyi and Liang, Percy , title =

    Liu, Nelson F. and Zhang, Tianyi and Liang, Percy , title =. 2023 , howpublished =

    2023

  28. [36]

    and Salakhutdinov, Ruslan and Manning, Christopher D

    Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William W. and Salakhutdinov, Ruslan and Manning, Christopher D. , title =. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , year =

    2018

  29. [37]

    Proceedings of the 28th International Conference on Computational Linguistics , year =

    Ho, Xanh and Duong Nguyen, Anh-Khoa and Sugawara, Saku and Aizawa, Akiko , title =. Proceedings of the 28th International Conference on Computational Linguistics , year =

  30. [38]

    Transactions of the Association for Computational Linguistics , year =

    Trivedi, Harsh and Balasubramanian, Niranjan and Khot, Tushar and Sabharwal, Ashish , title =. Transactions of the Association for Computational Linguistics , year =

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.