Pith. sign in

REVIEW 2 major objections 5 minor 61 references

Judging Is Not Enumerating: Silent Omissions in LLM-Authored Acceptable Sets

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LLMs promoted to examiners judge membership well but do not reliably author the acceptable sets those roles demand; the omission-dominated gap persists across scales and costs accuracy inside RLVR rewards.

desk verdict Core finding is solid and well-triangulated; the 'usually deployed one-shot' scope is asserted, not measured, and the paper's own data show reasoning closes the algorithmic gap. read the letter →

arxiv 2608.01000 v1 pith:BGQCRSU2 submitted 2026-08-02 cs.AI

Wenhui Chen , Jianlin Chen , Ziyao Lin , Peiji Long , Chi Man Vong This is my paper
classification cs.AI
keywords LLM-as-judgeverifierauthoringacceptablesetssilentomissionjudging-enumeratinggaptest-suitegenerationRLVRrewardsanswer-keyquality
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 tries to establish that language models, increasingly promoted from examinees to examiners, are weak at the very capability the promotion assumes: materializing the acceptable set a correctness specification denotes. Under one-shot greedy authoring—the protocol by which answer keys, test suites, and reward functions are actually produced at scale—models judge membership of a presented candidate well but enumerate the set or write the suite that induces it poorly, and the gap persists across a $24\times$ parameter range without closing. A control locates the deficit precisely: asked to state the rule as an executable predicate rather than list its extension, the same models reach F1 ≈ 0.99, so the failure is not knowledge or specification but one-shot materialization of the induced region. The dominant error is silent omission, which resists audit because a missing member, unlike an over-inclusion, is an absence no reviewer can name; planted over-inclusions are caught 6–7$\times$ more often than planted omissions. Wired into RLVR, an authored key costs 1.9 points of accuracy against a complete oracle and 18.5 WordNet-relative points, in every one of six paired seeds.

What carries the argument

The central object is the paired 'two interfaces to the same weights': execution, a one-pass membership function $\hat{m}(x)\in\{0,1\}$ on a presented candidate (Eq. 1), versus authoring, autoregressive emission of a set $\hat{S}$ under a causal mask with a learned stopping rule (Eq. 2), both scored by set F1 against a constructed oracle $S^*$. Three devices carry the argument: Observation 1 (the measured gap is a set-emission penalty); Observation 2 (omission is witness-blind to local review—even a perfect membership oracle cannot certify recall over candidates the reviewer never names, because finding a missing member is itself the authoring operation); Observation 3 (the authored reward i

What would settle it

Give reviewers the drafted keys with the full candidate universe and the correct cardinality of the oracle set supplied, then measure whether planted omissions are recovered at the same rate as planted over-inclusions, net of clean-key false alarms: Observation 2 predicts omissions stay undiscoverable because recovering a witness is itself the authoring operation, so a comparable recovery rate would falsify the witness-poor mechanism and the directional-review consequence built on it. A second, cheaper check: if any one-shot greedy model at any scale authors the complete-truth sets at F1 match

Watch

Extended reading notes

Core claim

On the paper's own terms: language models execute membership far better than they author artifacts whose induced acceptance region matches the target set $S^*$, except where authoring reduces to restating a compact rule already in the prompt. Measured on four judge-free constructions (mechanical predicates, HumanEval+/MBPP+ execution, WordNet), the gap is +0.34 to +0.29 F1 over a $24\times$ scale range on complete truth; on code, models judging at F1 0.74–0.90 author suites admitting only 19–42% of oracle-correct solutions. The residual error is omission-dominated and witness-poor: planted over-inclusions are caught 6–7$\times$ more often than planted omissions, and subtractive review pushes

Load-bearing premise

The load-bearing premise is the scope condition stated first in the paper—that verifiers are authored one-shot with greedy decoding and no test-time reasoning (the paper's own §3.7 shows frontier reasoning closes the algorithmic gap, authoring F1 rising 0.673 to 0.976, so pipelines that reason, ensemble, fine-tune, or use tools may escape the measured deficit); a second acknowledged premise is that the EvalPlus/HumanEval+ oracle correctly defines code behavior.

Editorial extensions

If this is right

  • Model-authored answer keys, rubrics, and test suites produced one-shot without test-time reasoning are systematically incomplete in the omission direction, so any pipeline that treats them as ground truth inherits under-acceptance: over-rejection of correct behavior rather than permissiveness.
  • Review and self-critique do not repair this; they make it worse. Subtractive review is a directional filter that removes visible over-inclusions while omissions persist, walking the artifact monotonically toward higher precision and no higher recall.
  • A known-correct-probe gate is the cheapest real mitigation: discarding any authored verifier that rejects a verified-correct solution cuts false rejection from 58–92% to ≤5%, at a yield of 5–39%; most of the discarded majority is recoverable by rewriting wrong expected values to what reference execution returns (3.3–10.6× yield).
  • Where a compact executable rule exists, the deficit is escapable: asking for the predicate rather than the roster is the difference between F1 ≈ 0.26 and ≈ 0.99, and test-time reasoning closes the algorithmic gap at the frontier.
  • As an RLVR reward, the authored key costs measured accuracy (1.9 points against a complete oracle, 18.5 WordNet-relative), a two-sided channel that withholds reward from correct outputs and grants it to incorrect ones.

Reading between the lines

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

  • If the asymmetry is as structural as the paper argues, the same directional bias should be measurable in human-authored specifications: the paper's Observation 2 is an information-theoretic statement, not an LLM-specific one, so recall auditing against constructed ground truth—rather than more review passes—is the transferable prescription for any specification pipeline.
  • The frontier reasoning result (GPT-5.1's algorithmic gap falling to +0.008 with a CI covering zero at effort=low) suggests the deficit is as much a property of the one-shot greedy decoding protocol as of the weights; a natural test is whether small reasoning budgets close the gap across models and scales, and whether the RLVR tax shrinks correspondingly when keys are authored with reasoning.
  • The trace-repair decomposition—models choose discriminating inputs well but compute expected outputs badly—isolates the intensional half of suite authoring as the tractable target for fine-tuning or tool augmentation; one could test whether feeding the candidate universe or giving execution feedback during generation closes the enumeration gap on the open-set constructions.
  • Taken with the sample-then-verify falsification, the paper implies that model self-checking cannot certify its own authoring, shifting the burden to external oracles; deployments that cannot exhibit a known-correct behavior (the paper's own gate precondition) face an unsolved verification gap that warrants direct study.
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

2 major / 5 minor

Summary. This paper asks whether LLMs can author the acceptable sets that a correctness specification denotes, and compares that authoring interface against the same models' pointwise membership judging. Four reference constructions are used: mechanically decidable string and numeric predicates over presented lists (complete truth), HumanEval+/MBPP+ with an EvalPlus-hardened executable oracle, and WordNet synsets. Under one-shot greedy decoding, authoring F1 lags execution F1 in every tested model–construction cell across a 3B–72B scale range, six families, and two generations; the gap does not close. Supplementary results isolate the deficit as enumeration rather than specification (predicate-emission F1 ≈ 0.99), show omission is 6–7× harder to detect than over-inclusion in the review protocol, falsify a registered sample-then-verify repair, introduce a known-correct-probe gate that cuts code false rejection to ≤5%, and measure an RLVR reward tax of −1.9 points on complete truth and up to −32 points WordNet-relative. The paper explicitly scopes its claims to the one-shot regime and reports a large set of controls, negative results, and instrumentation defects.

Significance. If the results hold, they identify a robust and consequential asymmetry: models can judge membership far better than they can materialise the membership set, and the resulting omissions are structurally resistant to local review and cost accuracy when wired into RLVR. The central measurement is unusually well grounded: the complete-truth constructions have no incompleteness confound; code correctness is decided by execution; no LLM judge grades the construction measurements; bootstrap CIs are reported; and the design includes a predicate-emission control, prompt/format ablations, cross-family and cross-generation replications, and a falsified registered prediction. The paper also reports its own instrumentation defects and retracted sub-predictions, which strongly supports credibility. The exploratory OLS scale fit is explicitly not used for the headline. The main qualification is scope: the practical reach of the headline depends on the premise that one-shot greedy authoring is the usual deployment protocol, a premise that is asserted rather than measured and that §3.7 shows is doing real work.

major comments (2)
  1. [Abstract, §1, §3.7] The abstract and §1 assert that one-shot greedy authoring with no test-time reasoning is 'the protocol the role is usually deployed with.' No evidence is cited for this deployment claim, and §3.7 shows the protocol boundary is load-bearing: with reasoning enabled, GPT-5.1's algorithmic authoring F1 rises from 0.673 to 0.976 and the gap falls to +0.008 with a CI covering zero; Opus's code gap falls from +0.274 to +0.096. §7.2 also lists decoding sweeps and tool-augmented pipelines among controls not run. The central measurement is internally valid as scoped, but the practical generalization in the abstract depends on an unmeasured premise. Please either provide evidence of deployment practice or re-scope the rhetoric (e.g., 'under one-shot greedy decoding, a protocol in current use') and temper the abstract's 'usually deployed.'
  2. [Abstract, §6] The abstract states that 'a production deployment of 43,227 scored items fails omission-first at 10:1' without the caveats that §6 itself insists on. §6 says the 10:1 ratio is evaluator-diagnosed, that the evaluator re-affirms only 25% of open-format fail verdicts, and that the ratio 'is directional, not a measured rate.' Because the paper's central claim does not depend on the field evidence, this is fixable by moving the number behind the same qualification in the abstract or removing it from the abstract; as written, a reader taking only the abstract will read a rate the paper itself says cannot be read as one.
minor comments (5)
  1. [References] References [20] and [57] are the same Konstantinou et al. paper with the same title and arXiv identifier. Please merge into one entry and update citations accordingly.
  2. [§7.3] The limitation paragraph says the complete-truth RLVR tax is 'a much smaller 2.3 points,' but §5.1 and Appendix G report the corrected estimate as 1.91 (≈1.9) points. Update this number to match the revised evaluation.
  3. [§3.1] In the sentence after Eq. (3), 'whichever of recall or precision is , which is what makes' appears to be missing the word 'lower' or 'smaller' after 'is.'
  4. [Table 3] Table 3 shows authoring F1, execution F1, and gap [95% CI], but only the gap column has a CI. Since the text claims 'disjoint CIs' between authoring and execution, please provide the marginal CIs for both arms or explicitly state that the gap CI is the evidence of separation.
  5. [§3.5 / Table 6] In the caption of Table 6, 'False rejection falls from 0.58–0.92 to ≤0.05' is supported by the table, but the adjacent text says '7–10% of Llama's' yield while the table lists 0.213 and 0.134 (i.e., 21.3% and 13.4%). Please reconcile the yield percentages in the prose with the table.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the central scissors and RLVR tax are measured against external oracles; the definitional formalizations (Observations 2–3) are used interpretively, not as evidence.

full rationale

The paper explicitly avoids the circularity it identifies ('using a judge to grade judge-authorship is circular') and builds all headline measurements on constructed or complete ground truth: mechanical predicates over presented lists, EvalPlus/MBPP+ execution, and WordNet synsets. The central scissors is a paired measurement of two interfaces over identical items; the authoring–execution gap is an empirical quantity, not an identity. The intensional control (§3.11) is an independent measured arm, not derived from the extensional result. The RLVR result is a controlled paired experiment with an exact oracle arm (§5.1) and a reference-relative lexical arm whose caveats are stated; Observation 3 is a definitional identity used to interpret the measurements, not to generate them. The sample-then-verify prediction was registered and falsified; trace repair and gating are guarded against tautology by held-out pools. The exploratory OLS fit is explicitly non-load-bearing. No load-bearing self-citation exists; the companion-manuscript note is non-overlapping and not used as evidence. The scope premise that one-shot greedy is the deployment protocol is an unmeasured assumption, flagged in §3.7/§7.2, but that is a correctness/scope risk, not circularity. Score 0.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The central claim uses externally constructed or executable ground truth and does not fit parameters to produce the gap. The only fitted number is an exploratory OLS slope that the paper itself does not rely on. No new physical or conceptual entity is introduced; starvation, pollution, and witness-poorness are defined mechanisms rather than invented entities.

free parameters (1)
  • Exploratory OLS slope of RLVR tax vs. log10 parameters = 5.5 pt/decade, R^2=0.29
    Reported in Section 5 as exploratory and explicitly not a scaling law; it does not enter any central claim.
assumptions (6)
  • standard math Mechanical predicates over presented finite lists compute the exact acceptable set S*.
    Used in Sections 3.2(a) and 3.3; the complete-truth construction depends on this.
  • domain assumption EvalPlus/HumanEval+ and MBPP+ suites are a trustworthy executable oracle for code correctness.
    Section 3.4 treats oracle-verified solution pools as ground truth; oracle weakness is listed as an orthogonal confound.
  • domain assumption WordNet synsets provide a usable lexical oracle whose recall side is incompleteness-robust.
    Section 3.8 explicitly treats WordNet precision as a lower bound and leans only on the incompleteness-robust recall side.
  • domain assumption One-shot greedy decoding is the deployment protocol under which answer keys and suites are authored at scale.
    Stated in the abstract and Section 1; the entire scissors is scoped to this protocol, and Section 3.7 shows reasoning is an escape where a rule exists.
  • domain assumption The GRPO/LoRA implementation in ms-swift faithfully realizes RLVR for the paired comparison.
    Appendix C; the paired key comparison requires identical training except for the reward key.
  • standard math Bootstrap resampling of items or problem clusters gives valid confidence intervals for the measured F1 differences.
    Appendix C and Tables 2 to 5; all central CIs are item or problem-cluster bootstraps.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Judging Is Not Enumerating: Silent Omissions in LLM-Authored Acceptable Sets." pith.science (2026). https://pith.science/paper/BGQCRSU2

@misc{pith2026260801000,
  author       = {Pith},
  title        = {Pith review of: Judging Is Not Enumerating: Silent Omissions in LLM-Authored Acceptable Sets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BGQCRSU2}},
  note         = {Machine review of arXiv:2608.01000}
}
read the original abstract

Language models are increasingly promoted from examinees to examiners: they write the test suites, answer keys, rubrics, and reward functions that define correctness for other systems. We measure the capability that role assumes and find it lacking under the protocol the role is usually deployed with, one-shot greedy authoring with no test-time reasoning. Across four reference constructions - two with complete finite truth, one with a hardened executable reference (HumanEval+/MBPP+), one with an explicitly incomplete lexical reference (WordNet) - models judge whether a candidate belongs far better than they author the set itself. On the incompleteness-proof algorithmic construction the gap is +0.34 to +0.29 F1 over a 24x parameter range and does not close; on executable code, models judging at F1 0.74-0.90 author suites admitting only 19-42% of oracle-correct solutions. A control locates the deficit: asked to emit the predicate rather than its extension, the same models reach F1 about 0.99. The failure is not missing knowledge or an inability to specify, but an inability to materialise the region a specification induces. The dominant error is omission, which resists audit: an over-inclusion is a token a reviewer can challenge, a missing member an absence whose discovery is the authoring problem itself. Models detect planted over-inclusions 6-7x more often than planted omissions, and a production deployment of 43,227 items fails omission-first at 10:1. Wired into RLVR, an authored key costs 1.9 points of accuracy against an exact oracle and 18.5 WordNet-relative (six paired seeds, p=0.031). Gating authored verifiers on a known-correct probe cuts false rejection from 58-92% to at most 5%, but keeps only 5-39% of suites. Repairing them instead, by rewriting each wrong expected value to what a reference execution returns, raises yield 3.3-10.6x across four author families.

Figures

Figures reproduced from arXiv: 2608.01000 by the authors.

Figure 1
Figure 1. The paper in one view, and its boundaries. The same weights expose three interfaces to a correctness specification (column 2). Judging membership is strong; authoring an artifact whose induced region should equal S ∗ is much weaker in both of its realisations, extensional enumeration on the word-list constructions and test-suite authoring on code, and does not catch up over the scale range tested; restating a rule t… view at source ↗
Figure 2
Figure 2. The geometry of an authored set. The authoring interface emits Sˆ; scored against the oracle S ∗ it has recall ρ and precision π. Its two error regions behave oppositely under audit: over-inclusions are named tokens a reviewer or an RLVR reward can act on, while omissions are unnamed. These crescents are the starvation and pollution terms of Observation 3; the witness-poorness of the left crescent is Observation 2. … view at source ↗
Figure 3
Figure 3. The two interfaces to the same weights. Execution (Eq. 1) is one forward pass yielding a single bit on a curated, balanced query. Authoring (Eq. 2) emits the acceptable set left-to-right under a causal mask, conditioning each element on the prefix (dashed) and choosing when to stop, with no access to candidates it has not yet sampled. The paper measures how far apart these two readouts of the same knowledge sit. 6 … view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: The judging–enumerating scissors across four LLM-judge-free constructions and six model fami￾lies/generations. Colour encodes interface (blue judging; magenta authoring, i.e. extensional enumeration on the word-list panels, test-suite authoring on the code panel, per §…
Figure 5
Figure 5. Figure 5: The known-correct-probe gate (§3.5). Trusting all authored suites gives a 0.58–0.92 false-rejection rate (magenta); discarding suites that reject the canonical reference collapses it to ≤ 0.05 (green). The annotation y is yield, the fraction of authored suites that sur…
Figure 6
Figure 6. Figure 6: The gap survives a model generation and the jump to the closed frontier. Left of the divider (open weights, local): Qwen3 raises both interfaces over Qwen2.5 at matched scale without shrinking the gap (bold numbers). Right (closed frontier, API, reasoning disabled to m…
Figure 7
Figure 7. Figure 7: Why omission is witness-poor (Observation 2). Review interrogates the tokens the artifact names, so an over-inclusion is exposed by a single membership query (top). An omitted member names no token; to query it the reviewer must first author it, the capability under au…
Figure 8
Figure 8. Figure 8: Sample-then-verify authoring (§4.4) and its two failure points. Diversity sampling supplies recall and the model’s stronger membership interface supplies precision, in principle. In practice the union misses nearly half of S ∗ (retrieval bind) and the filter, strong on…
Figure 9
Figure 9. Figure 9: Propagation as a two-sided reward channel (Observation 3). An RLVR loop identical except for the reward key: the authored key withholds reward from correct rollouts in S ∗ \ Sˆ and grants it to incorrect rollouts in Sˆ \ S ∗ . The reference key has neither leak. On com…
Figure 10
Figure 10. Figure 10: E6 propagation (mean±std over 6 seeds; dots are per-seed values). Left: synset-membership accuracy (WordNet-relative). Wiring the model-authored key into an otherwise-identical loop costs 18.5 points relative to the reference key. Right: reward pollution (fraction of …
Figure 11
Figure 11. Figure 11: The authoring tax persists across scale but is not monotonic. Base, authored-key, and reference￾key synset-membership accuracy (WordNet-relative, ±std over 3 seeds) for five policies spanning scale and family. The reference arm reaches 0.91–0.97 at every scale while t…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 28 canonical work pages

  1. [1]

    Concrete problems in ai safety.arXiv preprint arXiv:1606.06565, 2016

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. Concrete problems in ai safety.arXiv preprint arXiv:1606.06565, 2016

    arXiv 2016

  2. [2]

    Program synthesis with large language models.arXiv preprint arXiv:2108.07732, 2021

    Jacob Austin, Augustus Odena, Maxwell Nye, et al. Program synthesis with large language models.arXiv preprint arXiv:2108.07732, 2021

    arXiv 2021

  3. [3]

    Rein- forcement learning with verifiable yet noisy rewards under imperfect verifiers.arXiv preprint arXiv:2510.00915, 2025

    Xin-Qiang Cai, Wei Wang, Feng Liu, Tongliang Liu, Gang Niu, and Masashi Sugiyama. Rein- forcement learning with verifiable yet noisy rewards under imperfect verifiers.arXiv preprint arXiv:2510.00915, 2025

    arXiv 2025

  4. [4]

    An- swer matching outperforms multiple choice for language model evaluation.arXiv preprint arXiv:2507.02856, 2025

    Nikhil Chandak, Shashwat Goel, Ameya Prabhu, Moritz Hardt, and Jonas Geiping. An- swer matching outperforms multiple choice for language model evaluation.arXiv preprint arXiv:2507.02856, 2025

    arXiv 2025

  5. [5]

    CodeT: Code generation with generated tests

    Bei Chen, Fengji Zhang, Anh Nguyen, et al. CodeT: Code generation with generated tests. In International Conference on Learning Representations (ICLR), 2023

    work page 2023

  6. [6]

    Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    Mark Chen, Jerry Tworek, Heewoo Jun, et al. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    arXiv 2021

  7. [7]

    Jimenez, John Yang, Kevin Liu, and Aleksander Madry

    Neil Chowdhury, James Aung, Chan Jun Shern, Oliver Jaffe, Dane Sherburn, Giulio Starace, Evan Mays, Rai Dias, Marwan Aljubeh, Mia Glaese, Carlos E. Jimenez, John Yang, Kevin Liu, and Aleksander Madry. Introducing SWE-bench verified. OpenAI Blog, 2024

    work page 2024

  8. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learn- ing.arXiv preprint arXiv:2501.12948, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learn- ing.arXiv preprint arXiv:2501.12948, 2025. 38

    arXiv 2025

Show all 61 references
  1. [9]

    Scoring verifiers: Evaluating synthetic verification for code and reasoning.arXiv preprint arXiv:2502.13820, 2025

    Aleksander Ficek, Somshubra Majumdar, Vahid Noroozi, and Boris Ginsburg. Scoring verifiers: Evaluating synthetic verification for code and reasoning.arXiv preprint arXiv:2502.13820, 2025

    2025 arXiv

  2. [10]

    Rabe, Talia Ringer, and Yuriy Brun

    Emily First, Markus N. Rabe, Talia Ringer, and Yuriy Brun. Baldur: Whole-proof genera- tion and repair with large language models. InACM Joint European Software Engineering Conference (ESEC/FSE), 2023

    2023

  3. [11]

    Scaling laws for reward model overoptimization

    Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. Proceedings of ICML, 2023

    2023

  4. [12]

    A survey on LLM-as-a-judge.arXiv preprint arXiv:2411.15594, 2024

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, et al. A survey on LLM-as-a-judge.arXiv preprint arXiv:2411.15594, 2024

    2024 arXiv

  5. [13]

    Rubrics as rewards: Reinforcement learning beyond verifiable domains.arXiv preprint arXiv:2507.17746, 2025

    Anisha Gunjal, Anthony Wang, Elaine Lau, Vaskar Nath, Bing Chen, and Sean Hendryx. Rubrics as rewards: Reinforcement learning beyond verifiable domains.arXiv preprint arXiv:2507.17746, 2025

    2025 arXiv

  6. [14]

    LLMs gaming veri- fiers: RLVR can lead to reward hacking.arXiv preprint arXiv:2604.15149, 2026

    Lukas Helff, Quentin Delfosse, David Steinmann, Raban Härle, Hikaru Shindo, Patrick Schramowski, Wolfgang Stammer, Kristian Kersting, and Felix Friedrich. LLMs gaming veri- fiers: RLVR can lead to reward hacking.arXiv preprint arXiv:2604.15149, 2026

    2026 arXiv

  7. [15]

    Large language models cannot self-correct reasoning yet

    Jie Huang, Xinyun Chen, Swaroop Mishra, et al. Large language models cannot self-correct reasoning yet. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  8. [16]

    Pitfalls of rule- and model-based verifiers: A case study on mathematical reasoning.arXiv preprint arXiv:2505.22203, 2025

    Yuzhen Huang, Weihao Zeng, Xingshan Zeng, Qiyuan Zhu, and Junxian He. Pitfalls of rule- and model-based verifiers: A case study on mathematical reasoning.arXiv preprint arXiv:2505.22203, 2025

    2025

  9. [17]

    Jacobs, Loïc Grobol, and Alvin Tsang

    Cassandra L. Jacobs, Loïc Grobol, and Alvin Tsang. Large-scale cloze evaluation reveals that token prediction tasks are neither lexically nor semantically aligned.arXiv preprint arXiv:2410.12057, 2024

    2024 arXiv

  10. [18]

    SELF-[IN]CORRECT: LLMs struggle with discriminating self-generated responses

    Dongwei Jiang, Jingyu Zhang, Orion Weller, Nathaniel Weir, Benjamin Van Durme, and Daniel Khashabi. SELF-[IN]CORRECT: LLMs struggle with discriminating self-generated responses. arXiv preprint arXiv:2404.04298, 2024

    2024 arXiv

  11. [19]

    Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221, 2022

    Saurav Kadavath, Tom Conerly, Amanda Askell, et al. Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221, 2022

    2022 arXiv

  12. [21]

    Tulu 3: Pushing frontiers in open language model post-training.arXiv preprint arXiv:2411.15124, 2024

    Nathan Lambert, Jacob Morrison, Valentina Pyatkin, et al. Tulu 3: Pushing frontiers in open language model post-training.arXiv preprint arXiv:2411.15124, 2024

    2024 arXiv

  13. [22]

    Let’sverifystepbystep

    HunterLightman, VineetKosaraju, YuraBurda, etal. Let’sverifystepbystep. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  14. [23]

    Is your code generated by ChatGPT really correct? rigorous evaluation of large language models for code generation

    Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. Is your code generated by ChatGPT really correct? rigorous evaluation of large language models for code generation. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 39

    2023

  15. [24]

    Rethinking verification for LLM code generation: From genera- tion to testing

    Zihan Ma, Taolin Zhang, Maosong Cao, Junnan Liu, Wenwei Zhang, Minnan Luo, Songyang Zhang, and Kai Chen. Rethinking verification for LLM code generation: From genera- tion to testing. InAdvances in Neural Information Processing Systems (NeurIPS), 2025. arXiv:2507.06920

    2025 arXiv

  16. [25]

    Self-refine: Iterative refinement with self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, et al. Self-refine: Iterative refinement with self-feedback. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  17. [26]

    George A. Miller. WordNet: A lexical database for English.Communications of the ACM, 38 (11):39–41, 1995

    1995

  18. [27]

    What can we learn from collective human opinions on natural language inference data? InProceedings of EMNLP, 2020

    Yixin Nie, Xiang Zhou, and Mohit Bansal. What can we learn from collective human opinions on natural language inference data? InProceedings of EMNLP, 2020

    2020

  19. [28]

    The effects of reward misspecification: Mapping and mitigating misaligned models

    Alexander Pan, Kush Bhatia, and Jacob Steinhardt. The effects of reward misspecification: Mapping and mitigating misaligned models. InInternational Conference on Learning Repre- sentations (ICLR), 2022

    2022

  20. [29]

    Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2024

    Qwen Team. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  21. [30]

    Before the model learns the bug: Fuzzing RLVR verifiers.arXiv preprint arXiv:2606.01066, 2026

    Jonas Ray. Before the model learns the bug: Fuzzing RLVR verifiers.arXiv preprint arXiv:2606.01066, 2026

    2026 arXiv

  22. [31]

    An empirical evaluation of using large language models for automated unit test generation.IEEE Transactions on Software Engineering, 2024

    Max Schäfer, Sarah Nadi, Aryaz Eghbali, and Frank Tip. An empirical evaluation of using large language models for automated unit test generation.IEEE Transactions on Software Engineering, 2024

    2024

  23. [32]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y.K. Li, Y. Wu, and Daya Guo. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  24. [33]

    Joar Skalse, Nikolaus H. R. Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and characterizing reward hacking. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  25. [34]

    Mind the gap: Examining the self-improvement capabilities of large language models.arXiv preprint arXiv:2412.02674, 2024

    Yuda Song, Hanlin Zhang, Carson Eisenach, Sham Kakade, Dean Foster, and Udaya Ghai. Mind the gap: Examining the self-improvement capabilities of large language models.arXiv preprint arXiv:2412.02674, 2024

    2024 arXiv

  26. [35]

    On the self-verification limitations of large language models on reasoning and planning tasks.arXiv preprint arXiv:2402.08115, 2024

    Kaya Stechly, Karthik Valmeekam, and Subbarao Kambhampati. On the self-verification limitations of large language models on reasoning and planning tasks.arXiv preprint arXiv:2402.08115, 2024

    2024 arXiv

  27. [36]

    Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models

    Katherine Tian, Eric Mitchell, Allan Zhou, et al. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models. InEmpirical Methods in Natural Language Processing (EMNLP), 2023

    2023

  28. [37]

    Order matters: Sequence to sequence for sets

    Oriol Vinyals, Samy Bengio, and Manjunath Kudlur. Order matters: Sequence to sequence for sets. InInternational Conference on Learning Representations (ICLR), 2016

    2016

  29. [38]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, et al. Self-consistency improves chain of thought reasoning in language models. InInternational Conference on Learning Representations (ICLR), 2023. 40

    2023

  30. [39]

    Con- sistency of a recurrent language model with respect to incomplete decoding

    Sean Welleck, Ilia Kulikov, Jaedeok Kim, Richard Yuanzhe Pang, and Kyunghyun Cho. Con- sistency of a recurrent language model with respect to incomplete decoding. InProceedings of EMNLP, 2020

    2020

  31. [40]

    Large language models are better reasoners with self-verification

    Yixuan Weng, Minjun Zhu, Fei Xia, et al. Large language models are better reasoners with self-verification. InFindings of EMNLP, 2023

    2023

  32. [41]

    what it can create, it may not understand

    Peter West, Ximing Lu, Nouha Dziri, Faeze Brahman, Linjie Li, Jena D. Hwang, Liwei Jiang, Jillian Fisher, Abhilasha Ravichander, Khyathi Chandu, Benjamin Newman, Pang Wei Koh, Allyson Ettinger, and Yejin Choi. The generative AI paradox: “what it can create, it may not understa...

    2024

  33. [42]

    Willard and Rémi Louf

    Brandon T. Willard and Rémi Louf. Efficient guided generation for large language models. arXiv preprint arXiv:2307.09702, 2023

    2023 arXiv

  34. [43]

    Jiang, Wenda Li, et al

    Yuhuai Wu, Albert Q. Jiang, Wenda Li, et al. Autoformalization with large language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  35. [44]

    Can LLMs express their uncertainty? an empirical evaluation of confidence elicitation in LLMs

    Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. Can LLMs express their uncertainty? an empirical evaluation of confidence elicitation in LLMs. In International Conference on Learning Representations (ICLR), 2024

    2024

  36. [45]

    DAPO: An open-source LLM reinforcement learning system at scale.arXiv preprint arXiv:2503.14476, 2025

    Qiying Yu et al. DAPO: An open-source LLM reinforcement learning system at scale.arXiv preprint arXiv:2503.14476, 2025

    2025 arXiv

  37. [46]

    Judging LLM-as-a-judge with MT-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, et al. Judging LLM-as-a-judge with MT-bench and chatbot arena. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  38. [47]

    RubricBench: Aligning model-generated rubrics with human standards

    Junyi Zhou, Qiyuan Zhang, Yufei Wang, Fuyuan Lyu, Yidong Ming, Can Xu, Qingfeng Sun, Kai Zheng, Peng Kang, Xue Liu, and Chen Ma. RubricBench: Aligning model-generated rubrics with human standards. InProceedings of the 64th Annual Meeting of the Association for Computational Li...

    2026

  39. [48]

    Self-rewarding language models.arXiv preprint arXiv:2401.10020, 2024

    Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models.arXiv preprint arXiv:2401.10020, 2024

    2024 arXiv

  40. [49]

    Generative verifiers: Reward modeling as next-token prediction.arXiv preprint arXiv:2408.15240, 2024

    Lunjun Zhang, Arian Hosseini, Hritik Bansal, Mehran Kazemi, Aviral Kumar, and Rishabh Agarwal. Generative verifiers: Reward modeling as next-token prediction.arXiv preprint arXiv:2408.15240, 2024

    2024 arXiv

  41. [50]

    LLM critics help catch LLM bugs.arXiv preprint arXiv:2407.00215, 2024

    Nat McAleese, Rai Michael Pokorny, Juan Felipe Ceron Uribe, Evgenia Nitishinskaya, Maja Trebacz, and Jan Leike. LLM critics help catch LLM bugs.arXiv preprint arXiv:2407.00215, 2024

    2024 arXiv

  42. [51]

    SWT-Bench: Test- ing and validating real-world bug-fixes with code agents

    Niels Mündler, Mark Niklas Müller, Jingxuan He, and Martin Vechev. SWT-Bench: Test- ing and validating real-world bug-fixes with code agents. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  43. [52]

    Ehsan Kamalloo, Nouha Dziri, Charles L. A. Clarke, and Davood Rafiei. Evaluating open- domain question answering in the era of large language models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), 2023. 41

    2023

  44. [53]

    Tomayto, tomahto: Beyond token-level answer equivalence for question answering evaluation

    Jannis Bulian, Christian Buck, Wojciech Gajewski, Benjamin Börschinger, and Tal Schuster. Tomayto, tomahto: Beyond token-level answer equivalence for question answering evaluation. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2022

    2022

  45. [54]

    Soneya Binta Hossain and Matthew B. Dwyer. TOGLL: Correct and strong test oracle genera- tion with LLMs. InProceedings of the 47th IEEE/ACM International Conference on Software Engineering (ICSE), pages 1475–1487, 2025. arXiv:2405.03786

    2025 arXiv

  46. [55]

    VALTEST: Automated validation of language model generated test cases.arXiv preprint arXiv:2411.08254, 2024

    Hamed Taherkhani and Hadi Hemmati. VALTEST: Automated validation of language model generated test cases.arXiv preprint arXiv:2411.08254, 2024. Later versions appear as Taherkhani et al., IEEE TSE 52(4), 2026, rebuilt around semantic entropy

    2024

  47. [56]

    Hallucination to consensus: Multi- agent LLMs for end-to-end JUnit test generation.ACM Transactions on Software Engineering and Methodology, 2026

    Qinghua Xu, Guancheng Wang, Lionel Briand, and Kui Liu. Hallucination to consensus: Multi- agent LLMs for end-to-end JUnit test generation.ACM Transactions on Software Engineering and Methodology, 2026. Introduces the CANDOR framework. arXiv:2506.02943

    2026

  48. [57]

    Do LLMs generate test oracles that capture the actual or the expected program behaviour?arXiv preprint arXiv:2410.21136, 2024

    Michael Konstantinou, Renzo Degiovanni, and Mike Papadakis. Do LLMs generate test oracles that capture the actual or the expected program behaviour?arXiv preprint arXiv:2410.21136, 2024

    2024 arXiv

  49. [58]

    PAL: Program-aided language models

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. PAL: Program-aided language models. InProceedings of the 40th Inter- national Conference on Machine Learning (ICML), 2023. arXiv:2211.10435

    2023 arXiv

  50. [59]

    Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks.Trans- actions on Machine Learning Research, 2023. arXiv:2211.12588

    2023 arXiv

  51. [60]

    TinyV: Reducing false negatives in verification improves RL for LLM reasoning.arXiv preprint arXiv:2505.14625, 2025

    Zhangchen Wang et al. TinyV: Reducing false negatives in verification improves RL for LLM reasoning.arXiv preprint arXiv:2505.14625, 2025

    2025 arXiv

  52. [61]

    Sequential enumeration in large language models.arXiv preprint arXiv:2512.04727, 2025

    Kuinan Hou, Marco Zorzi, and Alberto Testolin. Sequential enumeration in large language models.arXiv preprint arXiv:2512.04727, 2025

    2025

  53. [62]

    |S ∗|= 10

    Jon Saad-Falcon et al. Shrinking the generation–verification gap with weak verifiers.arXiv preprint arXiv:2506.18203, 2025. 42 A Proofs Observations 2 and 3 are consequences of Definition 1; we give the arguments in full because their force is structural, not quantitative. (Ob...

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.