Pith. sign in

REVIEW 6 major objections 7 minor 41 references

EQUATOR: A Deterministic Framework for Evaluating LLM Reasoning with Open-Ended Questions. # v1.0.0-beta

T0 review · 6 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read The paper claims that EQUATOR, a deterministic scoring framework that grades open-ended LLM answers against human-evaluated answer keys in a vector database, yields significantly lower and more accurate scores than traditional…

desk verdict A promising evaluation idea undermined by its own appendix: the proof-of-concept shows false negatives on obviously correct answers, so the headline score drops don't support the paper's claims. read the letter →

arxiv 2501.00257 v1 pith:H7KJ23DB submitted 2024-12-31 cs.CL

classification cs.CL
keywords LLMevaluationopen-endedquestionsdeterministicscoringvectordatabasecosinesimilaritybinarygradingfactualaccuracyreasoningbenchmarks
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

The paper proposes EQUATOR, a deterministic scoring framework for open-ended reasoning questions. It stores human-evaluated answer keys in a vector database, retrieves the closest key for each student answer by cosine similarity, and uses a small local language model to assign 100% only when the final answer matches the key and 0% otherwise. Comparing this scoring with the published human and multiple-choice scores on two benchmark sets, the paper finds that EQUATOR gives significantly lower mean scores, with standardized effect sizes of $d = 1.07$ and $d = 2.85$. The authors interpret this as evidence that traditional evaluations overestimate LLM factual accuracy, and they report that EQUATOR also reduces reliance on human graders and cuts token use by 12.20% (22.61% when run locally). If the interpretation is right, EQUATOR offers a scalable and reproducible way to expose where LLMs fail at reasoning.

What carries the argument

The load-bearing mechanism is a retrieval-and-match pipeline: questions and human-evaluated answers are embedded into 384-dimensional vectors and stored in a vector database; a student answer is embedded the same way; the database returns the answer key with the highest cosine similarity to the student answer; and a prompt-constrained local LLM (the evaluator) returns a score of 100 if the student's final answer matches that key and 0 otherwise. The binary scoring rule is what does the work: it removes threshold tuning, partial credit, and any credit for reasoning or eloquence, so the score reflects only whether the final answer agrees with the chosen reference answer.

What would settle it

Run EQUATOR on a set of open-ended questions where each answer key has several widely accepted correct wordings that differ only in phrasing. If a correct paraphrase that is not the exact stored string consistently receives 0, then the binary answer-key rule is penalizing surface mismatch rather than factual error, and the central claim would fail.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a new way of measuring LLM reasoning: instead of selecting among options or having a human judge weigh reasoning quality, every open-ended answer is compared, through an embedding search, to a small set of human-written correct answers and then graded by a strict binary rule. Applied to the 2024-09-13 multiple-choice benchmark, EQUATOR lowered mean scores by 11.59 points on average relative to the published scoring, with $p \approx 0.0026$ and $d \approx 1.07$. Applied to the 2024-06-12 open-ended benchmark, the drop was 20.38 points, with $p < 0.0001$ and $d \approx 2.85$. The paper claims these large gaps mean that conventional scoring methods award credit for fluency, reasoning, and partial credit, while EQUATOR penalizes factual mismatch even when the answer is persuasive. It further claims that a smaller locally hosted LLM can carry out the grading consistently, eliminating the need for human scoring of each response.

Load-bearing premise

The framework assumes the human-evaluated answer keys stored in the vector database are correct and complete; if any key is wrong, ambiguous, or missing an acceptable alternative, EQUATOR will mark a correct student answer as zero.

Editorial extensions

If this is right

  • Same models and same questions receive substantially lower scores under EQUATOR than under published human or multiple-choice scoring, so standard benchmarks likely overstate factual accuracy on reasoning tasks.
  • Open-ended evaluation no longer requires a human grader for every response; a small local LLM and a vector database of vetted answers can produce reproducible scores.
  • The framework's reported token savings (12.20% overall, 22.61% with local evaluation) make large-scale evaluation cheaper than sending full multiple-choice prompts to an API.
  • Because persuasive but wrong answers receive 0, EQUATOR isolates factual failure points that fluency-biased evaluations hide.
  • The larger score gap on the open-ended benchmark ($d = 2.85$) than on the multiple-choice benchmark ($d = 1.07$) suggests that open-ended scoring captures more of what multiple-choice misses.

Reading between the lines

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

  • The paper does not test whether its 0/100 rule distinguishes 'not matching the stored key' from 'factually wrong'; a natural next experiment is to have independent human graders mark the same responses and measure disagreement.
  • Because the retrieval step depends on the embedding model, the robustness of the score gaps could be probed by repeating the evaluation with different sentence-embedding models; if the gaps vanish, part of the effect is an artifact of the embedding choice rather than the scoring rule.
  • A testable extension is to keep the answer-key anchor but allow graded partial credit (for example, 100/70/40/0) and see whether model rankings change; this would tell developers how much information the binary rule discards.
  • The paper discusses high-stakes deployment in medicine, law, and finance but does not test it; the same answer-key-plus-vector-database design would need validation on those domains where the cost of a fluent wrong answer is high.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 7 minor

Summary. The manuscript proposes EQUATOR, a framework that pairs open-ended questions with human-evaluated answers stored in a vector database, retrieves the closest answer key via cosine similarity, and uses a locally hosted LLaMA 3.2B model as a strict binary (0 or 100) scorer. The authors apply EQUATOR to two Williams and Huckle benchmarks, report significantly lower mean scores than the original human/LLM scoring (Cohen's d = 1.07 and d = 2.85), and interpret this as evidence that traditional methods overestimate LLM factual accuracy. A 30-question proof-of-concept evaluation is included in Appendix 15.1.

Significance. The problem of evaluating open-ended LLM reasoning without fluency bias is important, and the paper deserves credit for attempting a transparent, locally hosted evaluation pipeline and for candidly discussing sources of bias in Section 12. However, the empirical contribution does not establish the central claim. The proof-of-concept table in Appendix 15.1 contains clear false negatives produced by the identical scoring mechanism used in the headline experiments, the framework is not deterministic in the sense claimed, and the statistical analyses are performed on model-level means with very small samples. If the false-negative rate were corrected and the statistical claims were re-run on question-level data, the framework could become a useful engineering contribution; as presented, the reported score drops and effect sizes cannot be interpreted as evidence about factual accuracy.

major comments (6)
  1. [Appendix 15.1, Q7 and Q21] The proof-of-concept table directly contradicts the paper's interpretation of 0/100 scores. For Q7, the student answer 'Bob' is exactly the answer given in the human key ('Bob is on Alan's immediate right...'), yet EQUATOR returns 0 with an explanation that contradicts its own key. For Q21, the student sentence 'Astonishing aardvarks adore afternoon activities' satisfies the constraint that every word starts with A, yet the evaluator returns 0 and says the words start with other letters. These are not ambiguous edge cases: the same retrieval-plus-LLM-judge mechanism is used for Experiments 1 and 2. The lower means and large effect sizes (d=1.07, d=2.85) may therefore be artifacts of evaluator false negatives rather than evidence of 'more accurately penalizing factual inaccuracies' as claimed in Sections 5.1.2 and 8.1.1.
  2. [Section 3.4, Step 3] The formal scoring definition in Step 3 uses a similarity threshold tau: Match(Ae, Kbest) = 1 if Sim(Ae, Kbest) >= tau, and 0 otherwise. Section 3.3, item 5, states 'Instead of setting an empirical similarity threshold, we retrieve the most semantically similar answer from the vector database as the answer key,' and no value of tau is reported anywhere. This is an internal inconsistency in the core algorithm, and tau is a free parameter that should either be specified or removed from the formal definition.
  3. [Sections 5.1, 6, and Section 12] The paired t-tests are computed on model-level mean scores (n=13 in Experiment 1, n=9 in Experiment 2), not on paired question-level responses. This violates the pairing assumption stated in Section 5.1 ('the same models evaluated on the same questions'), and the Central Limit Theorem justification is not credible for n=9 or n=13. The Limitations section itself says '30 questions are not enough to achieve statistical significance,' directly contradicting the paper's earlier claims of statistical significance (p<0.005 and p<0.0001). The reported p-values establish only that the average difference across models is nonzero, not that EQUATOR measures student-answer correctness more accurately.
  4. [Section 3.5.2] The evaluator prompt contains contradictory scoring instructions: it first says 'If it does, assign a score of 100%. Otherwise, you must assign a score of 0%,' but later says 'If the Answer Key matches the student's answer, you must assign a score of 0%; no partial credit is allowed.' A prompt with internal contradictions cannot support the paper's claim of deterministic, reproducible scoring, and it may contribute to the false negatives documented in Appendix 15.1.
  5. [Sections 3.2 and 3.5] The framework is not deterministic. The final score depends on the stochastic output of LLaMA 3.2B interpreting the retrieved answer key, and on the cosine retrieval of the closest key; no temperature, seed, or sampling parameters are reported. The word 'deterministic' appears throughout but is never justified by controlling randomness or by showing run-to-run agreement.
  6. [Table 2 and Table 5] Table 2 (Experiment 1, multiple-choice benchmark) and Table 5 (Experiment 2, open-ended benchmark) report identical means, standard deviations, and confidence intervals for the same model names. Since the two experiments are supposed to use different datasets, this duplication means the results of Experiment 2 cannot be verified as reported.
minor comments (7)
  1. [Abstract] The abstract cites 'Williams and Huckle (2004)' while the rest of the paper correctly uses 2024; please correct the citation.
  2. [Section 6 heading] The heading reads '2014-6-12-Benchmarks' but should read '2024-06-12-Benchmark.'
  3. [Section 3.6] There is a typo in 'Toxonmony' and a formatting error in '15.1Our'; both should be corrected.
  4. [Section 10] The paragraph beginning 'In our analysis, the EQUATOR Evaluator demonstrated significant token savings...' is repeated verbatim before and after Table 9; one copy should be removed.
  5. [Section 15.5] Several figure captions in the Experiment 2 appendix reuse '9-13 Multiple-Choice Benchmark' labels (e.g., Figures 10, 11, and 13), which makes the appendix confusing and should be corrected to refer to the 6-12 open-ended benchmark.
  6. [Appendix 15.1] The question numbering is inconsistent: item 18 repeats the poisoned-wine question already shown as item 17, and there are two items numbered 30.
  7. [Sections 3.2, 3.3, and 10] The model names are inconsistent across the paper: Section 3.2 uses 'll-minilm,' Section 3.3 mentions LLaMA 3.2B, and Section 10 refers to 'Ollama 3.2'; the exact embedding and evaluator models should be named consistently.

Circularity Check

2 steps flagged · score 6.0 of 10

The large score drops that form the paper's central evidence are baked into EQUATOR's binary no-partial-credit rubric, and the paper then interprets those drops as proof that EQUATOR penalizes factual inaccuracies.

  1. self definitional [Section 3.5 (binary scoring criteria) and Sections 5.1.2 / 8.1.1 (interpretation of the drops)]
    "The scoring criteria are binary: • 100%: Assigned when the student’s final answer exactly matches the answer key. • 0%: Assigned when the student’s final answer does not match the answer key. No partial credit is allowed. ... The large effect size (Cohen’s d ≈ 1.07) highlights the substantial impact of the scoring framework on model performance assessments. This supports our assertion that traditional evaluation methods may overestimate LLM performance by not adequately penalizing factual inaccuracies."

    The paper's central empirical finding—that EQUATOR produces significantly lower mean scores than Williams and Huckle's rubric—follows from the definition of the score itself. EQUATOR assigns 100 only on exact match to the retrieved answer key and 0 otherwise, with no partial credit, while the comparison rubric awards 80, 60, 40, or 20 for partially correct responses. Consequently, any answer receiving partial credit under the comparison rubric necessarily receives 0 under EQUATOR unless it is a verbatim key match, so the mean-score gap and the large Cohen's d are constructed by the scoring rule rather than discovered from the data.

  2. other [Appendix 15.1 (Proof-of-concept evaluation table, Q7 and Q21)]
    "7: Student Response {'student_answer': 'Bob'} ... 7: Human Answer Bob is on Alan's immediate right because it is stated that Alan is on Bob's immediate left. ... 7: Score 0 ... 21: Student Response {'student_answer': 'Astonishing aardvarks adore afternoon activities.'} ... 21: Score 0"

    The framework's own proof-of-concept evaluation table shows that a score of 0 is not equivalent to factual inaccuracy. For Q7, the student answer 'Bob' matches the relation given in the human answer key, yet the evaluator returns 0. For Q21, 'Astonishing aardvarks adore afternoon activities' satisfies the stated requirement that every word start with A, yet the evaluator returns 0 and claims those words start with other letters. Because the headline experiments use this same LLM evaluator with a single retrieved answer key, the reported score drops cannot be cited as evidence that EQUATOR 'penalizes factual inaccuracies'; they are, at least in part, evaluator false negatives and the mechanical consequence of requiring exact agreement with one pre-stored answer.

full rationale

The EQUATOR framework's main result—lower mean scores and large effect sizes relative to Williams and Huckle—is not an independent empirical discovery. It is largely predetermined by the scoring definition that allows only 100 for an exact match and 0 for everything else, with explicit denial of partial credit. Since the comparison rubric gives partial credit for partially correct responses, a gap is built into the two rubrics, and the paper's interpretation of that gap as 'penalizing factual inaccuracies' is a claim that the score measures factual accuracy. Appendix 15.1 undercuts that interpretation by showing correct answers scored 0 (Q7, Q21), so the score is not a reliable proxy for factual correctness. The central claim therefore reduces, in substantial part, to the framework's own scoring rule. I do not find circularity in the paper's use of external benchmarks or prior work: the human-evaluated answers come from Williams and Huckle, the benchmark data are external, and the authors' self-citations (e.g., [7], [17], [18], [20]) are background references rather than load-bearing steps in the derivation. The circularity is confined to the headline interpretation of the score drops, which is why the score is 6 rather than higher.

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

The central claim rests on unverified human answer keys, the reliability of cosine similarity for semantic matching, and the ability of a small stochastic LLM to grade deterministically. No new physical or mathematical entities are introduced. The framework therefore inherits all of its correctness assumptions from the Williams and Huckle dataset, the MiniLM embedding model, and the LLaMA 3.2 evaluator.

free parameters (1)
  • similarity threshold tau = Not set (defined but unused)
    Section 3.4 Step 3 defines Match using threshold tau, but Section 3.3 says no empirical threshold is used. If applied, tau would be a free parameter controlling binary scoring.
assumptions (5)
  • domain assumption Williams and Huckle (2024) human-evaluated answer keys are correct and complete ground truth.
    Used as the 'source of truth' in vector database construction (Section 3.2) and in the evaluator prompt 'Only consider the answer key as the source of truth' (Section 3.5.2). No independent validation is provided.
  • domain assumption Cosine similarity in all-MiniLM-L6-v2 embedding space reflects semantic equivalence between student answers and answer keys.
    Section 3.4 Step 2 retrieves Kbest by maximizing cosine similarity; the paper does not calibrate embedding similarity against human judgments.
  • domain assumption The LLaMA 3.2B evaluator reliably follows the binary grading instructions and behaves deterministically.
    Sections 3.5 and 3.5.2 use a stochastic LLM as the grader; determinism is asserted but not demonstrated, and the prompt contains contradictory scoring instructions.
  • standard math The Central Limit Theorem ensures approximate normality of mean differences with n=13 and n=9.
    Sections 5.1.2 and 8.1 invoke CLT despite small sample sizes; the validity is questionable.
  • domain assumption Model-level differences are independent across models.
    Sections 5.1.3 and 8.1 assume independence; Section 12 later acknowledges that clustering effects are not accounted for, which can underestimate standard errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EQUATOR: A Deterministic Framework for Evaluating LLM Reasoning with Open-Ended Questions. # v1.0.0-beta." pith.science (2026). https://pith.science/paper/H7KJ23DB

@misc{pith2026250100257,
  author       = {Pith},
  title        = {Pith review of: EQUATOR: A Deterministic Framework for Evaluating LLM Reasoning with Open-Ended Questions. # v1.0.0-beta},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H7KJ23DB}},
  note         = {Machine review of arXiv:2501.00257}
}
read the original abstract

Despite the remarkable coherence of Large Language Models (LLMs), existing evaluation methods often suffer from fluency bias and rely heavily on multiple-choice formats, making it difficult to assess factual accuracy and complex reasoning effectively. LLMs thus frequently generate factually inaccurate responses, especially in complex reasoning tasks, highlighting two prominent challenges: (1) the inadequacy of existing methods to evaluate reasoning and factual accuracy effectively, and (2) the reliance on human evaluators for nuanced judgment, as illustrated by Williams and Huckle (2024)[1], who found manual grading indispensable despite automated grading advancements. To address evaluation gaps in open-ended reasoning tasks, we introduce the EQUATOR Evaluator (Evaluation of Question Answering Thoroughness in Open-ended Reasoning). This framework combines deterministic scoring with a focus on factual accuracy and robust reasoning assessment. Using a vector database, EQUATOR pairs open-ended questions with human-evaluated answers, enabling more precise and scalable evaluations. In practice, EQUATOR significantly reduces reliance on human evaluators for scoring and improves scalability compared to Williams and Huckle's (2004)[1] methods. Our results demonstrate that this framework significantly outperforms traditional multiple-choice evaluations while maintaining high accuracy standards. Additionally, we introduce an automated evaluation process leveraging smaller, locally hosted LLMs. We used LLaMA 3.2B, running on the Ollama binaries to streamline our assessments. This work establishes a new paradigm for evaluating LLM performance, emphasizing factual accuracy and reasoning ability, and provides a robust methodological foundation for future research.

Figures

Figures reproduced from arXiv: 2501.00257 by the authors.

Figure 1
Figure 1. EQUATOR Evaluator Framework - Drawn by Rahul Murugan [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Cohen’s d Effect Sizes Chart 5.1.1 Effect Size (Cohen’s d ) To quantify the magnitude of the observed difference, we calculated Cohen’s d: d = ¯d sd = 11.59 10.86 ≈ 1.07 Interpretation 5.1.2 Normality Assumption With n = 13T, the sample size is relatively small. However, the Central Limit Theorem assures that the sampling distribution of the mean difference is approximately normal. Independence We assume that the di… view at source ↗
Figure 3
Figure 3. Cohen’s d Effect Sizes Chart d = ¯d sd = 20.38 7.16 ≈ 2.85 Interpretation According to Cohen’s conventions: • Small effect: d = 0.2 • Medium effect: d = 0.5 • Large effect: d = 0.8 A Cohen’s d of approximately 2.85 indicates a very large effect size, suggesting that the deterministic scoring framework has a substantial impact on the evaluation outcomes [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Comparison of Mean Scores, Confidence Interval (CI) Lower Bounds, and CI Upper [PITH_FULL_IMAGE:figures/full_fig_p037_4.png]
Figure 5
Figure 5. Figure 5: The accompanying figure illustrates the Z Interval Error associated with each model [PITH_FULL_IMAGE:figures/full_fig_p038_5.png]
Figure 6
Figure 6. Figure 6: Mean Scores for Each Model on the Williams and Huckle 9-13 Multiple-Choice Benchmark. [PITH_FULL_IMAGE:figures/full_fig_p039_6.png]
Figure 7
Figure 7. Figure 7: Comparison of Mean Scores, Confidence Interval Lower and Upper Bounds for Each Model [PITH_FULL_IMAGE:figures/full_fig_p040_7.png]
Figure 8
Figure 8. Figure 8: Histogram of Standard Deviations for Each Model on the Williams and Huckle 9-13 [PITH_FULL_IMAGE:figures/full_fig_p041_8.png]
Figure 9
Figure 9. Figure 9: Comparison of Mean Scores, Confidence Interval Lower and Upper Bounds for Each Model [PITH_FULL_IMAGE:figures/full_fig_p044_9.png]
Figure 10
Figure 10. Figure 10: Z Interval Error for Each Model on the Williams and Huckle 9-13 Multiple-Choice [PITH_FULL_IMAGE:figures/full_fig_p045_10.png]
Figure 11
Figure 11. Figure 11: Mean Scores for Each Model on the Williams and Huckle 9-13 Multiple-Choice Bench [PITH_FULL_IMAGE:figures/full_fig_p046_11.png]
Figure 12
Figure 12. Figure 12: Mean Scores with Confidence Intervals for Various Models on the Williams and Huckle [PITH_FULL_IMAGE:figures/full_fig_p047_12.png]
Figure 13
Figure 13. Figure 13: Histogram of Standard Deviations for Each Model on the Williams and Huckle 9-13 [PITH_FULL_IMAGE:figures/full_fig_p047_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 29 canonical work pages

  1. [1]

    Easy problems that llms get wrong

    S. Williams and J. Huckle, “Easy problems that llms get wrong.” Cornell University, May 2024

  2. [2]

    Wider and deeper llm networks are fairer llm evaluators

    X. Zhang, B. Yu, H. Yu, Y . Lv, T. Liu, F. Huang, H. Xu, and Y . Li, “Wider and deeper llm networks are fairer llm evaluators.” Cornell University, August 2023

  3. [3]

    Likelihood-based mitigation of evaluation bias in large language models,

    M. Ohi, M. Kaneko, R. Koike, M. Loem, and N. Okazaki, “Likelihood-based mitigation of evaluation bias in large language models,” arXiv preprint arXiv:2402.15987, 2024

  4. [4]

    Eliminating position bias of language models: A mechanistic approach,

    Z. Wang, H. Zhang, X. Li, K.-H. Huang, C. Han, S. Ji, S. M. Kakade, H. Peng, and H. Ji, “Eliminating position bias of language models: A mechanistic approach,” arXiv preprint arXiv:2407.01100, 2024

  5. [5]

    Is Reference Necessary in the Evaluation of NLG Systems? When and Where?

    S. Sheng, Y . Xu, L. Fu, J. Ding, L. Zhou, X. Wang, and C. Zhou, “Is reference necessary in the evaluation of nlg systems? when and where?,” arXiv preprint arXiv:2403.14275, 2024

  6. [6]

    Style over substance: Evaluation biases for large language models,

    M. Wu and A. F. Aji, “Style over substance: Evaluation biases for large language models,” arXiv preprint arXiv:2307.03025, 2023

  7. [7]

    Developing safe and responsible large language models–a comprehensive framework,

    S. Raza, O. Bamgbose, S. Ghuge, F. Tavakoli, and D. J. Reji, “Developing safe and responsible large language models–a comprehensive framework,”arXiv preprint arXiv:2404.01399, 2024

  8. [8]

    Can large language models be an alternative to human evaluations?

    C.-H. Chiang and H. yi Lee, “Can large language models be an alternative to human evaluations?.” arXiv:2305.01937, 2023. ACL 2023 main conference paper. Main content: 10 pages (including limitations). Appendix: 13 pages

Show all 41 references
  1. [9]

    Calibrating reasoning in language models with internal consistency,

    Z. Xie, J. Guo, T. Yu, and S. Li, “Calibrating reasoning in language models with internal consistency,” May 2024. arXiv:2405.18711 [cs.AI]

  2. [10]

    Investigating automatic scoring and feedback using large language models,

    G. A. Katuka, A. Gain, and Y .-Y . Yu, “Investigating automatic scoring and feedback using large language models,” 2024. cs.CL; cs.LG

  3. [11]

    Superglue: A stickier benchmark for general-purpose language understanding systems,

    A. Wang, Y . Pruksachatkun, N. Nangia, A. Singh, J. Michael, F. Hill, O. Levy, and S. R. Bowman, “Superglue: A stickier benchmark for general-purpose language understanding systems,” arXiv preprint arXiv:1905.00537, 2019. NeurIPS 2019

  4. [12]

    Beyond the imitation game: Quantifying and extrapolating the capabilities of language models,

    A. Srivastava, A. Rastogi, A. Rao, A. A. M. Shoeb, A. Abid, A. Fisch, A. R. Brown, A. Santoro, A. Gupta, and et al., “Beyond the imitation game: Quantifying and extrapolating the capabilities of language models,” arXiv preprint arXiv:2206.04615 , 2022. Version 3, last revised ...

  5. [13]

    Multiple-choice questions are efficient and robust llm evaluators,

    Z. Zhang, Z. Jiang, L. Xu, H. Hao, and R. Wang, “Multiple-choice questions are efficient and robust llm evaluators,” June 2024. arXiv:2405.11966 [cs.CL]

  6. [14]

    Sentence-bert: Sentence embeddings using siamese bert- networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert- networks,” August 2019. arXiv:1908.10084 [cs.CL]

  7. [15]

    Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction,

    B. Sarmah, B. Hall, R. Rao, S. Patel, S. Pasquali, and D. Mehta, “Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction,” August 2024. arXiv:2408.04948 [cs.CL], 9 pages, 2 figures, 5 tables

  8. [16]

    Evaluations are critical for understanding the capabilities of large language models (llms),

    E. Miller, “Evaluations are critical for understanding the capabilities of large language models (llms),” 2024. 14 pages. 23

  9. [17]

    Exploring bias and prediction metrics to characterise the fairness of machine learning for equity-centered public health decision-making: A narrative review,

    S. Raza, A. Shaban-Nejad, E. Dolatabadi, and H. Mamiya, “Exploring bias and prediction metrics to characterise the fairness of machine learning for equity-centered public health decision-making: A narrative review,”IEEE Access, 2024

  10. [18]

    Fake news detection: Comparative evaluation of bert-like models and large language models with generative ai-annotated data,

    S. Raza, D. Paulen-Patterson, and C. Ding, “Fake news detection: Comparative evaluation of bert-like models and large language models with generative ai-annotated data,” 2024

  11. [19]

    Gender bias in contextualized word embeddings,

    J. Zhao, T. Wang, M. Yatskar, R. Cotterell, V . Ordonez, and K.-W. Chang, “Gender bias in contextualized word embeddings,” arXiv preprint arXiv:1904.03310, 2019

  12. [20]

    Vilbias: A framework for bias detection using linguistic and visual cues,

    S. Raza, C. Saleh, E. Hasan, F. Ogidi, M. Powers, V . Chatrath, M. Lotif, R. Javadi, A. Zahid, and V . R. Khazaie, “Vilbias: A framework for bias detection using linguistic and visual cues,” arXiv preprint arXiv:2412.17052, 2024

  13. [21]

    Usability maturity model: Human centredness scale,

    J. Earthy, “Usability maturity model: Human centredness scale,” INUSE Project deliverable D, vol. 5, pp. 1–34, 1998

  14. [22]

    Cohen, Statistical Power Analysis for the Behavioral Sciences

    J. Cohen, Statistical Power Analysis for the Behavioral Sciences . Hillsdale, NJ: Lawrence Erlbaum Associates, 2nd ed., 1988

  15. [23]

    Cohen, Statistical Power Analysis for the Behavioral Sciences

    J. Cohen, Statistical Power Analysis for the Behavioral Sciences . New York, NY: Academic Press, 1969

  16. [24]

    New effect size rules of thumb,

    S. S. Sawilowsky, “New effect size rules of thumb,” Journal of Modern Applied Statistical Methods, vol. 8, no. 2, pp. 597–599, 2009

  17. [25]

    Calculating and reporting effect sizes to facilitate cumulative science: A practical primer for t-tests and anovas,

    D. Lakens, “Calculating and reporting effect sizes to facilitate cumulative science: A practical primer for t-tests and anovas,” Frontiers in Psychology, vol. 4, p. 863, 2013. 24 15 Appendix 15.1 Question Taxonomy S. Williams and J. Huckle 2013 [1] Question Type Description Pu...

  18. [26]

    Cohen’s Work: Established the importance of effect size as a measure of practical signif- icance and created interpretable benchmarks that can help quantify the impact of scoring methods in our experiments

  19. [27]

    Sawilowsky’s Updates: Suggest that domain-specific considerations may influence the interpretation of effect sizes, reinforcing the importance of contextual analysis

  20. [28]

    Human-Eval

    Lakens’ Practical Guide : Offers detailed methodologies for calculating and reporting effect sizes, ensuring rigor and reproducibility in our work. By leveraging these works, our approach to using Cohen’s d to evaluate deterministic scoring frameworks becomes both statisticall...

  21. [29]

    • Length of the Bar: The longer the bar, the greater the effect size, meaning the model’s performance deviates more significantly from the EQUATOR evaluator

    Horizontal Bars Bars Represent Cohen’s d : Each bar represents Cohen’s d value for a model. • Length of the Bar: The longer the bar, the greater the effect size, meaning the model’s performance deviates more significantly from the EQUATOR evaluator. • Directionality: Positive ...

  22. [30]

    – Higher values mean a larger difference between the mean scores, standardized by the pooled variability

    Axes • Horizontal Axis (Cohen’s d): – Represents the magnitude of the difference in performance between the EQUATOR evaluator and each model. – Higher values mean a larger difference between the mean scores, standardized by the pooled variability. • Vertical Axis (Models): – L...

  23. [31]

    Desired Confidence Level : Typically, a 95% confidence level is used in educational assessments

  24. [32]

    Margin of Error: This is the acceptable error range (e.g., ±5% or ±10%)

  25. [33]

    Variability: The level of variance in responses; higher variability means more questions are needed to capture a reliable measure

  26. [34]

    Population Size: For smaller sample groups, you need more questions to achieve signifi- cance

  27. [35]

    Larger effects require fewer questions

    Effect Size: How large of a difference or effect you expect to detect. Larger effects require fewer questions. 15.8.2 Interpreting the Chart The magnitude of Cohen’s d Cohen’s d values can be interpreted based on standard guidelines: • Small Effect (d = 0.2): Minimal differenc...

  28. [36]

    • These models are relatively aligned with the EQUATOR evaluator, suggesting they produce similar results

    Top Models (Lowest Cohen’s d): • Models near the bottom of the chart (shorter bars) have smaller effect sizes, meaning their performance is closer to the EQUATOR evaluator. • These models are relatively aligned with the EQUATOR evaluator, suggesting they produce similar results. 52

  29. [37]

    • These models perform significantly worse under the given evaluation, as indicated by large positive Cohen’s d values

    Outperforming Models (Higher Cohen’s d): • Models near the top of the chart (longer bars) show the largest deviations from the EQUATOR evaluator. • These models perform significantly worse under the given evaluation, as indicated by large positive Cohen’s d values

  30. [38]

    • A wide range of Cohen’s d values (e.g., from˜0.3 to >2.0) indicates substantial variability in model performance

    Performance Variation: • The length of the bars demonstrates the variability in how well the models align with the EQUATOR evaluator. • A wide range of Cohen’s d values (e.g., from˜0.3 to >2.0) indicates substantial variability in model performance. Example Observations: • Low...

  31. [39]

    Objective Comparison: Cohen’s d standardizes the difference in mean scores, allowing for an unbiased comparison across models, irrespective of score scale or variability

  32. [40]

    Highlights Model Robustness: Models with lower Cohen’s d values are more consistent with the EQUATOR evaluator, suggesting better alignment with desired evaluation criteria

  33. [41]

    Prioritization: The chart helps identify models that need improvement (those with high Cohen’s d values) and those performing well (low Cohen’s d). 53 16 Formula to Estimate Sample Size (Number of Questions): You can use a sample size formula commonly used in statistics: n = Z...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.