Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Combining AI and human fact-checkers by AI confidence beats either alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 07:08 UTC pith:D7DBCOFM

load-bearing objection Core claim—confidence-based hybridization beats AI alone—is credible and replicates; the 'assistance further improves' claim is confounded and fails in the paper's own T3 replication, so it needs to be scoped or retracted. the 3 major comments →

arxiv 2510.26518 v2 pith:D7DBCOFM submitted 2025-10-30 cs.AI cs.HC

Human-AI Complementarity: A Goal for Amplified Oversight

Rishub Jain , Sophie Bridgers , Lili Janzer , Rory Greig , Tian Huey Teh , Vladimir Mikulik This is my paper
classification cs.AI cs.HC
keywords Human-AI ComplementarityAmplified OversightScalable OversightLLM RatingAI AssistanceHybridizationFact-verificationOver-reliance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Human oversight of AI often fails at hard verification tasks, but this paper argues that an AI's own uncertainty can be used to keep humans in the loop where they still add value. Its central experiments show that routing fact-checking examples to human raters whenever the AI rater's confidence is low lifts whole-set accuracy above the AI alone (89.3% vs. 87.7%), because humans are more accurate than the AI exactly on those low-confidence examples. The paper then shows that the human side can be improved further by giving raters the AI assistant's search results and selected verbatim evidence—evidence without predicted labels, explanations, or confidence—rather than the assistant's conclusions. When those assisted human ratings are majority-voted and hybridized, accuracy reaches 91.3%. If this holds, it gives amplified oversight a concrete mechanism: use AI confidence to find the human-advantage slice, then assist with raw material, not verdicts.

Core claim

The paper's central claim is that on a realistic fact-verification set, combining an AI rater's labels with human labels based on the AI's self-reported confidence produces higher accuracy than either source alone, and that a carefully designed AI assistant can push that hybrid higher still. An AI fact-verification rater with a search tool achieved 87.7% accuracy on the 1,918-sentence evaluation set, while unassisted human raters averaged 75.1% (80.6% by majority vote). Routing every example with AI confidence at or below 0.62 to human raters—280 examples where AI accuracy fell to 60.5% and human majority-vote accuracy was 71.3%—raised whole-set accuracy to 89.3%. Giving those human raters o

What carries the argument

The central mechanism is confidence-based hybridization: use the AI rating when the AI is confident, and route lower-confidence examples to human raters. Confidence is measured as the proportion of 50 sampled model outputs that agree with the majority-vote overall rating. The paper combines this routing rule with a fact-verification assistant that decomposes a sentence into claims, searches the web, and returns verbatim evidence; the effective variant shows humans the search results and selected evidence but withholds the assistant's predicted label, reasoning, and confidence. That evidence-only presentation is what avoids over-reliance and improves human accuracy on the routed slice.

Load-bearing premise

The whole gain rests on the AI rater's confidence scores remaining calibrated on new data, yet the routing threshold T = 0.62 was selected after looking at the evaluation set—the paper itself notes (Section 3.2) that no separate set was used to calibrate it.

What would settle it

Fix T = 0.62 in advance and apply the same hybrid to a fresh fact-verification set with golden labels; if hybrid accuracy does not exceed the AI-alone accuracy on that set, the claimed complementarity was threshold overfitting rather than a stable property.

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

If this is right

  • Confidence-based hybridization can beat AI-alone accuracy even where the AI is the better rater overall, provided confidence is calibrated and human/AI errors are at least partly independent.
  • Evidence-only assistance—raw search results and verbatim quotes—is the only tested form that significantly improves accuracy on the low-confidence slice without causing over-reliance.
  • Displaying the assistant's predicted verdict together with reasoning or confidence produces over-reliance: assisted accuracy is worse than baseline when the AI's overall judgment is wrong.
  • The headline hybrid result (91.3% vs. 89.3%) uses majority-vote aggregation of assisted human ratings; aggregation is what converts individual-rater gains into a larger hybrid gain.
  • Assistance effectiveness depends on rater skill: in later experiments with more experienced raters, evidence assistance no longer improved accuracy above baseline.

Where Pith is reading between the lines

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

  • If confidence calibration generalizes, the same routing recipe could transfer to other oversight tasks—code review, safety critiques, long-form reasoning checks—where LLM self-reported confidence is meaningful.
  • A natural next experiment is calibrating the threshold on a separate held-out set before routing; the paper itself flags this as unvalidated, and its survival would turn the method into a deployable protocol.
  • The evidence-only result suggests a broader design principle: for AI-assisted human evaluation, showing raw material rather than conclusions may preserve the human's complementary judgment while reducing blind deference.
  • Because assistance helps most on the slice where the AI is weak, evaluators should measure assistance methods on routed slices rather than on whole datasets.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper investigates how to combine human and AI judgments for fact-verification of AI-generated sentences, with the goal of improving oversight quality. The authors build an AI fact-verification rater that uses web search and produces confidence scores. They propose 'Confidence-based Hybridization': accept the AI rating when its confidence exceeds a threshold T, otherwise route the example to human raters. In a first study (RQ1), using T=0.62, the hybrid achieves 89.3% accuracy on the evaluation set, higher than the AI rater alone (87.7%) or human majority vote alone (80.6%). In a second set of studies (RQ2), they test different forms of AI assistance for human raters on the low-confidence slice. They find that showing only search results and selected evidence improves human accuracy relative to an unassisted baseline (73.3% vs. 67.3%), while presentations that include judgments and confidence cause over-reliance. A final experiment (§2.4.4) reports that hybridizing with evidence-assisted majority-vote human ratings achieves 91.3% accuracy, significantly higher than the 89.3% obtained with unassisted human ratings. The paper frames these results as evidence for amplified oversight and discusses implications and limitations, including the lack of a separate threshold-calibration set.

Significance. If the central claims hold, the paper makes a useful empirical contribution to human-AI complementarity and amplified oversight. The RQ1 result—that a confidence-based routing protocol can beat both AI-alone and human-alone ratings on a realistic fact-verification task—is practically relevant and aligns with prior work on selective prediction and deferral. The paper's strengths include a sizable realistic evaluation set (1918 examples), a trained AI rater with an interpretable evidence trace, multiple assistance-design conditions, mixed-effects logistic regressions, bootstrap confidence intervals, and several control experiments (bonus incentives, time-per-task, rater-pool variation). The authors are also transparent about limitations, explicitly flagging the absence of a separate calibration set for the threshold and the potential role of rater practice effects. However, the headline assistance benefit is less robust: the key comparison in §2.4.4 is confounded by rater experience, and the appendix's own replication (A.3.4.1) shows no significant assistance effect. The paper's contribution therefore hinges primarily on RQ1, and the post-hoc threshold selection must be addressed bef

major comments (3)
  1. [§2.3.2 and §3.2] The threshold T=0.62 is selected on the evaluation set itself, and no separate training/validation set is used for calibration. The paper explicitly admits this in §3.2 ('A key limitation ... we didn't use a separate "train" set to calibrate the threshold'). Consequently, the reported 89.3% vs. 87.7% accuracy gain and the accompanying mixed-effects test are optimistically biased: the threshold and the significance test are evaluated on the same data that determined the threshold. A proper evaluation would either prespecify T, use cross-validation, or report the distribution of hybridized accuracy over thresholds while correcting for multiple comparisons. The claim that 'there are a range of thresholds below T=0.74 for which Hybridized accuracy is higher than AI alone' mitigates but does not eliminate the concern, because the range is still read off the same evaluation set. This issue is
  2. [§2.4.4 with §3.2 and Appendix A.3.4.1] The claim that assistance 'further improves' hybridization (91.3% vs. 89.3%) is not supported by the evidence as presented. The 91.3% result comes from an experiment opened to the entire pool of prior participants (§2.4.4), while the 89.3% comparator uses baseline_t1 data collected earlier from a less experienced, separate sample. The paper itself acknowledges these practice effects (§3.2). Appendix A.3.4.1 reruns the comparison with a concurrent unassisted baseline (baseline_t3) and more skilled raters; there, Search+/Evidence+ assistance does not significantly improve accuracy on the post-hybridized human set (p=.647 at the optimal threshold), and hybridized accuracy with assistance (90.5%) is essentially the same as with unassisted baseline majority vote (90.3%). Thus the only assistance effect that survives a controlled comparison is the within-T2 result for individual raters on the
  3. [§2.4.4 / Appendix A.3.4.1] The comparison underlying the headline 'further improves' claim is not apples-to-apples in another respect: the unassisted hybrid uses majority-vote labels from baseline_t1, while the assisted hybrid uses majority-vote labels from a later, higher-replication experiment with a different rater pool. The appendix itself notes that majority-vote accuracies between T1 and T3 are not comparable because of different ratings-per-example counts. This makes the 2.0 percentage-point gap (91.3% vs. 89.3%) uninterpretable as a pure effect of assistance. A proper test requires the same rater pool and same replication level across the two arms, as attempted in A.3.4.1; that test showed no assistance benefit. This comment is closely tied to the previous one but is worth stating separately because it affects the very framing of the paper's 'combined hybridization + assistance' contribution.
minor comments (5)
  1. [§2.1] Typo: 'across 10 experiments we we selectively filter' should be 'we selectively filter'.
  2. [§2.3.2 vs. Appendix A.3.4.1] Inconsistent reported accuracies for baseline_t1 majority-vote human ratings: §2.3.2 reports 80.6% on the entire Evaluation Set, while Appendix A.3.4.1 reports baseline_t1 majority accuracy as 84.7%. This discrepancy should be reconciled or explained (e.g., different definition of majority vote or different filtering).
  3. [§2.4.2 / Figure 2] Figure 2 caption states 'Baseline unassisted accuracy (69.3%)' but the text reports baseline_t2 accuracy as 67.3% (e.g., 'search, 68.7% vs. baseline, 67.3%'). One of these is a typo.
  4. [§2.4.4] 'showing participants the AI's search results and selected evidencedid increase' should read 'evidence did increase'.
  5. [A.3.4.2] The phrase '𝑏𝑒𝑡𝑎′𝑠' mixes Greek and Latin characters; use the same notation as the main text (e.g., 𝛽). Also 'humanratings' is missing a space.

Circularity Check

0 steps flagged

No significant circularity: the empirical claims are computed from independent human/AI ratings against external golden labels, not from definitional identities or load-bearing self-citations.

full rationale

The paper reports empirical experiments rather than a derivational framework, and its central claims are not constructed from their inputs. RQ1's confidence-based hybridization is a routing rule: AI ratings are used above threshold T=.62 and human majority-vote ratings below it, with both signal sources measured independently against external golden labels. The threshold is chosen post hoc on the Evaluation Set, a limitation the paper itself acknowledges in Section 3.2 ('A key limitation of our confidence-based hybridization approach was that we didn't use a separate "train" set to calibrate the threshold'), but this is an overfitting/generalization concern, not a definitional reduction: the reported 89.3% accuracy is not equal to the threshold by construction, and Appendix A.3.4.1 re-tests thresholds on a fresh T3 baseline. RQ2's assistance comparison is confounded by rater experience and time, which the paper also flags in Section 3.2 ('These practice effects and human raters improving over time could have been another reason that our human rater accuracy in Section 2.4.4 was so high'), but a confound is a validity threat, not circular reasoning. The only self-citation found is Krakovna et al. (2020), which includes author Vladimir Mikulik; it is cited only for the general idea of rewarding undesirable-looking behavior and is not load-bearing for the hybridization or assistance results. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. The results are therefore self-contained in the sense relevant to a circularity analysis.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

No new physical or theoretical entities are postulated. The central claims rest on a fitted routing threshold, the quality of golden labels, and the calibration of AI confidence. The paper is transparent about the threshold issue but does not provide held-out validation.

free parameters (1)
  • Hybridization confidence threshold T = 0.62
    Selected post hoc as the threshold among a small set where hybridized accuracy was highest on the evaluation set (Section 2.3.2); no held-out set used for calibration (Section 3.2).
axioms (4)
  • domain assumption Golden labels (majority of three expert ratings) correctly represent factual truth of each sentence.
    Section A.2.1: golden labels come from a high-quality trained rating pool; some manually reviewed. All accuracy metrics depend on this ground truth; if it is biased, the comparisons are affected.
  • domain assumption AI confidence scores (proportion of 50 sampled ratings matching the majority) are well-calibrated with accuracy, so routing by confidence identifies slices where humans outperform AI.
    Appendix Fig 10 shows calibration on the evaluation set, but it is measured on the same data used to evaluate the hybrid and is not independently validated.
  • standard math Mixed-effects logistic regression models with random intercepts correctly capture dependencies in the rating data.
    Used for significance tests (Sections 2.3.2, 2.4.2); standard methodology but relies on model assumptions being met.
  • domain assumption Prolific participants with general competency (not fact-verification expertise) are representative of the human raters who will provide oversight.
    Section A.2.4. Later T3 results indicate results depend on rater skill, so this premise is load-bearing.

pith-pipeline@v1.3.0-alltime-deepseek · 26468 in / 13688 out tokens · 123621 ms · 2026-08-04T07:08:23.208706+00:00 · methodology

0 comments
read the original abstract

Human feedback is critical for aligning AI systems to human values. As AI capabilities improve and AI is used to tackle more challenging tasks, verifying quality and safety becomes increasingly challenging. This paper explores how we can leverage AI to improve the quality of human oversight. We focus on an important safety problem that is already challenging for humans: fact-verification of AI outputs. We find that combining AI ratings and human ratings based on AI rater confidence is better than relying on either alone. Giving humans an AI fact-verification assistant further improves their accuracy, but the type of assistance matters. Displaying AI explanation, confidence, and labels leads to over-reliance, but just showing search results and evidence fosters more appropriate trust. These results have implications for Amplified Oversight -- the challenge of combining humans and AI to supervise AI systems even as they surpass human expert performance.

discussion (0)

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

Reference graph

Works this paper leans on

44 extracted references · 17 linked inside Pith

  1. [1]

    Amodei, C

    D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Man \'e . Concrete problems in ai safety. arXiv preprint arXiv:1606.06565, 2016

    Pith/arXiv arXiv 2016

  2. [2]

    Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnon, C. Chen, C. Olsson, C. Olah, D. Hernandez, D. Drain, D. Ganguli, D. Li, E. Tran-Johnson, E. Perez, J. Kerr, J. Mueller, J. Ladish, J. Landau, K. Ndousse, K. Lukosuite, L. Lovitt, M. Sellitto, N. Elhage, N. Schiefer, N. Mercado, N. DasSarma, R. L...

    Pith/arXiv arXiv 2022

  3. [3]

    Baird and L

    A. Baird and L. M. Maruping. The next generation of research on is use: A theoretical framework of delegation to and from agentic is artifacts. MIS quarterly, 45 0 (1), 2021

    2021

  4. [4]

    Bansal, T

    G. Bansal, T. Wu, J. Zhou, R. Fok, B. Nushi, E. Kamar, M. T. Ribeiro, and D. Weld. Does the whole exceed its parts? the effect of ai explanations on complementary team performance. In Proceedings of the 2021 CHI conference on human factors in computing systems, pages 1--16, 2021

    2021

  5. [5]

    O'Reilly Media, Inc

    S. Bird, E. Klein, and E. Loper. Natural language processing with Python: analyzing text with the natural language toolkit. " O'Reilly Media, Inc.", 2009

    2009

  6. [6]

    Bu c inca, M

    Z. Bu c inca, M. B. Malaya, and K. Z. Gajos. To trust or to think: Cognitive forcing functions can reduce overreliance on ai in ai-assisted decision-making. Proc. ACM Hum.-Comput. Interact., 5 0 (CSCW1), Apr. 2021. doi:10.1145/3449287. URL https://doi.org/10.1145/3449287

    doi:10.1145/3449287 2021

  7. [7]

    Bu c inca, P

    Z. Bu c inca, P. Lin, K. Z. Gajos, and E. L. Glassman. Proxy tasks and subjective measures can be misleading in evaluating explainable ai systems. In Proceedings of the 25th international conference on intelligent user interfaces, pages 454--464, 2020

    2020

  8. [8]

    Christiano, J

    P. Christiano, J. Leike, T. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement learning from human preferences. 06 2017

    2017

  9. [9]

    Christiano, B

    P. Christiano, B. Shlegeris, and D. Amodei. Supervising strong learners by amplifying weak experts. arXiv preprint arXiv:1810.08575, 2018

    Pith/arXiv arXiv 2018

  10. [10]

    K. D. Dvijotham, J. Winkens, M. Barsbey, S. Ghaisas, R. Stanforth, N. Pawlowski, P. Strachan, Z. Ahmed, S. Azizi, Y. Bachrach, L. Culp, M. Daswani, J. Freyberg, C. Kelly, A. Kiraly, T. Kohlberger, S. McKinney, B. Mustafa, V. Natarajan, K. Geras, J. Witowski, Z. Z. Qin, J. Creswell, S. Shetty, M. Sieniek, T. Spitz, G. Corrado, P. Kohli, T. Cemgil, and A. K...

    2023

  11. [11]

    F \"u gener, J

    A. F \"u gener, J. Grahl, A. Gupta, and W. Ketter. Cognitive challenges in human--artificial intelligence collaboration: Investigating the path toward productive delegation. Information Systems Research, 33 0 (2): 0 678--696, 2022

    2022

  12. [12]

    Hemmer, M

    P. Hemmer, M. Westphal, M. Schemmer, S. Vetter, M. V\" o ssing, and G. Satzger. Human-ai collaboration: The effect of ai delegation on human task performance and task satisfaction. In Proceedings of the 28th International Conference on Intelligent User Interfaces, IUI '23, page 453–463, New York, NY, USA, 2023. Association for Computing Machinery. ISBN 97...

    arXiv 2023

  13. [13]

    Hubinger

    E. Hubinger. Ai safety via market making, Jun 2020. URL https://www.alignmentforum.org/posts/YWwzccGbcHMJMpT45/ai-safety-via-market-making

    2020

  14. [14]

    Irving, P

    G. Irving, P. Christiano, and D. Amodei. Ai safety via debate. arXiv preprint arXiv:1805.00899, 2018

    Pith/arXiv arXiv 2018

  15. [15]

    H. Kaur, H. Nori, S. Jenkins, R. Caruana, H. Wallach, and J. Wortman Vaughan. Interpreting interpretability: understanding data scientists' use of interpretability tools for machine learning. In Proceedings of the 2020 CHI conference on human factors in computing systems, pages 1--14, 2020

    2020

  16. [16]

    Kenton, N

    Z. Kenton, N. Y. Siegel, J. Kramár, J. Brown-Cohen, S. Albanie, J. Bulian, R. Agarwal, D. Lindner, Y. Tang, N. D. Goodman1, and R. Shah. On scalable oversight with weak llms judging strong llms. arXiv preprint arXiv:2407.04622, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    A. Khan, J. Hughes, D. Valentine, L. Ruis, K. Sachan, A. Radhakrishnan, E. Grefenstette, S. R. Bowman, T. Rockt \"a schel, and E. Perez. Debating with more persuasive llms leads to more truthful answers. arXiv preprint arXiv:2402.06782, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Krakovna, J

    V. Krakovna, J. Uesato, M. R. Vladimir Mikulik, T. Everitt, R. Kumar, Z. Kenton, J. Leike, and S. Legg. Specification gaming: the flip side of ai ingenuity, apr 2020. URL https://deepmind.google/discover/blog/specification-gaming-the-flip-side-of-ai-ingenuity/

    2020

  19. [19]

    Lai and C

    V. Lai and C. Tan. On human predictions with explanations and predictions of machine learning models: A case study on deception detection. In Proceedings of the conference on fairness, accountability, and transparency, pages 29--38, 2019

    2019

  20. [20]

    J. D. Lee and N. Moray. Trust, self-confidence, and operators' adaptation to automation. International journal of human-computer studies, 40 0 (1): 0 153--184, 1994

    1994

  21. [21]

    J. D. Lee and K. A. See. Trust in automation: Designing for appropriate reliance. Human factors, 46 0 (1): 0 50--80, 2004

    2004

  22. [22]

    Leike, D

    J. Leike, D. Krueger, T. Everitt, M. Martic, V. Maini, and S. Legg. Scalable agent alignment via reward modeling: a research direction. arXiv preprint arXiv:1811.07871, 2018

    Pith/arXiv arXiv 2018

  23. [23]

    J. Li. A comparative study on annotation quality of crowdsourcing and llm via label aggregation. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6525--6529. IEEE, 2024

    2024

  24. [24]

    S. Ma, Y. Lei, X. Wang, C. Zheng, C. Shi, M. Yin, and X. Ma. Who should i trust: Ai or myself? leveraging human and ai correctness likelihood to promote appropriate trust in ai-assisted decision-making. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1--19, 2023

    2023

  25. [25]

    are you really sure?

    S. Ma, X. Wang, Y. Lei, C. Shi, M. Yin, and X. Ma. “are you really sure?” understanding the effects of human self-confidence calibration in ai-assisted decision making. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI '24, New York, NY, USA, 2024. Association for Computing Machinery. ISBN 9798400703300. doi:10.1145/3613...

    arXiv 2024

  26. [26]

    Manzini, G

    A. Manzini, G. Keeling, N. Marchal, K. R. McKee, V. Rieser, and I. Gabriel. Should users trust advanced ai assistants? justified trust as a function of competence and alignment. In The 2024 ACM Conference on Fairness, Accountability, and Transparency, pages 1174--1186, 2024

    2024

  27. [27]

    McAleese, R

    N. McAleese, R. M. Pokorny, J. F. C. Uribe, E. Nitishinskaya, M. Trebacz, and J. Leike. Llm critics help catch llm bugs, 2024. URL https://arxiv.org/abs/2407.00215

    Pith/arXiv arXiv 2024

  28. [28]

    Meinke, B

    A. Meinke, B. Schoen, J. Scheurer, M. Balesni, R. Shah, and M. Hobbhahn. Frontier models are capable of in-context scheming. arXiv preprint arXiv:2412.04984, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Michael, S

    J. Michael, S. Mahdi, D. Rein, J. Petty, J. Dirani, V. Padmakumar, and S. R. Bowman. Debate helps supervise unreliable experts. arXiv preprint arXiv:2311.08702, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    R. Ngo, L. Chan, and S. Mindermann. The alignment problem from a deep learning perspective. arXiv preprint arXiv:2209.00626, 2022

    Pith/arXiv arXiv 2022

  31. [31]

    Parasuraman and V

    R. Parasuraman and V. Riley. Humans and automation: Use, misuse, disuse, abuse. Human factors, 39 0 (2): 0 230--253, 1997

    1997

  32. [32]

    Parrish, H

    A. Parrish, H. Trivedi, N. Nangia, V. Padmakumar, J. Phang, A. S. Saimbhi, and S. R. Bowman. Two-turn debate doesn't help humans answer hard reading comprehension questions. arXiv preprint arXiv:2210.10860, 2022 a

    Pith/arXiv arXiv 2022

  33. [33]

    Parrish, H

    A. Parrish, H. Trivedi, E. Perez, A. Chen, N. Nangia, J. Phang, and S. R. Bowman. Single-turn debate does not help humans answer hard reading-comprehension questions. arXiv preprint arXiv:2204.05212, 2022 b

    Pith/arXiv arXiv 2022

  34. [34]

    Petridis, B

    S. Petridis, B. D. Wedin, J. Wexler, M. Pushkarna, A. Donsbach, N. Goyal, C. J. Cai, and M. Terry. Constitutionmaker: Interactively critiquing large language models by converting feedback into principles. In Proceedings of the 29th International Conference on Intelligent User Interfaces, pages 853--868, 2024

    2024

  35. [35]

    Pinski, M

    M. Pinski, M. Adam, and A. Benlian. Ai knowledge: Improving ai delegation through human enablement. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI '23, New York, NY, USA, 2023. Association for Computing Machinery. ISBN 9781450394215. doi:10.1145/3544548.3580794. URL https://doi.org/10.1145/3544548.3580794

    arXiv 2023

  36. [36]

    Rafailov, A

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024

    2024

  37. [37]

    Saunders, C

    W. Saunders, C. Yeh, J. Wu, S. Bills, L. Ouyang, J. Ward, and J. Leike. Self-critiquing models for assisting human evaluators. 2022. URL https://arxiv.org/abs/2206.05802

    Pith/arXiv arXiv 2022

  38. [38]

    Sharma, M

    M. Sharma, M. Tong, T. Korbak, D. Duvenaud, A. Askell, S. R. Bowman, N. Cheng, E. Durmus, Z. Hatfield-Dodds, S. R. Johnston, et al. Towards understanding sycophancy in language models. arXiv preprint arXiv:2310.13548, 2023

    Pith/arXiv arXiv 2023

  39. [39]

    C. Si, N. Goyal, S. T. Wu, C. Zhao, S. Feng, H. Daum \'e III, and J. Boyd-Graber. Large language models help humans verify truthfulness--except when they are convincingly wrong. arXiv preprint arXiv:2310.12558, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Tharwat and W

    A. Tharwat and W. Schenck. A survey on active learning: State-of-the-art, practical challenges and research directions. Mathematics, 11 0 (4), 2023. ISSN 2227-7390. doi:10.3390/math11040820. URL https://www.mdpi.com/2227-7390/11/4/820

    doi:10.3390/math11040820 2023

  41. [41]

    Vaccaro, A

    M. Vaccaro, A. Almaatouq, and T. Malone. When combinations of humans and ai are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8 0 (12): 0 2293--2303, 2024

    2024

  42. [42]

    Vasconcelos, M

    H. Vasconcelos, M. J \"o rke, M. Grunde-McLaughlin, T. Gerstenberg, M. S. Bernstein, and R. Krishna. Explanations can reduce overreliance on ai systems during decision-making. Proceedings of the ACM on Human-Computer Interaction, 7 0 (CSCW1): 0 1--38, 2023

    2023

  43. [43]

    X. Wang, H. Kim, S. Rahman, K. Mitra, and Z. Miao. Human-llm collaborative annotation through effective verification of llm labels. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1--21, 2024

    2024

  44. [44]

    J. Wei, C. Yang, X. Song, Y. Lu, N. Hu, D. Tran, D. Peng, R. Liu, D. Huang, C. Du, et al. Long-form factuality in large language models. arXiv preprint arXiv:2403.18802, 2024

    Pith/arXiv arXiv 2024