{"id":"fcf3ddee-c77a-49bc-8ed3-5b4af8b10f77","arxiv_id":"2411.18924","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Example selection in few-shot prompts changes GPT essay-scoring agreement with human raters, with stronger majority and recency biases in GPT-3.5 than GPT-4.","lead":"This paper tests how the choice and order of example essays in few-shot prompts affect GPT-3.5 and GPT-4 scores on automated essay scoring. It finds that example selection matters most for GPT-3.5, and that one older GPT-4 version outperforms newer versions on this task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Regression models in Eqs. (1)-(2) are rank-deficient because example counts sum to 3 and last-label indicators sum to 1, so the reported t-tests for majority and recency bias are not identifiable.","rationale":"The central claim has two components: (1) example selection affects QWK differently across models, with GPT-3.5 more sensitive; (2) regression analysis reveals majority label and recency bias. Component (1) is grounded in descriptive statistics (mean and SD of QWK across prompt categories) and is not threatened by the collinearity issue. Component (2) rests entirely on the coefficients and t-tests from Eqs. (1) and (2), which are non-identifiable due to perfect multicollinearity. This is not a matter of disagreement with current consensus or a weak effect; it is an internal statistical flaw that invalidates the quantitative evidence for bias. The reader identified the same weakest assumption. The appropriate verdict remains CONDITIONAL because the descriptive contribution and practical recommendations still have value, but the bias regression must be redone with a reduced-rank parameterization or a clearly stated contrast coding before the bias claims can be accepted. No misconduct is suggested; the flaw is methodological and correctable.","tokens_in":10081,"tokens_out":6855,"duration_ms":58928,"concrete_test":"Re-run the regression analysis of §3.2 with a full-rank parameterization: replace the six predictors by two count contrasts (e.g., H_n - L_n and M_n - L_n) and two last-label contrasts (e.g., H_l - L_l and M_l - L_l), or equivalently drop L_n and L_l from the model. Compare the resulting coefficient estimates and p-values with Tables 1 and 2. Additionally, compute the condition number or rank of the correlation matrix of the six standardized predictors to confirm perfect multicollinearity. If the significance patterns for 'majority label bias' and 'recency bias' change materially, the original bias findings should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §3.2, the regression models in Eqs. (1) and (2) include the count predictors H_n, M_n, L_n and the last-label indicators H_l, M_l, L_l. Restricting to 3-shot prompts, every prompt satisfies H_n + M_n + L_n = 3 and exactly one of H_l, M_l, L_l equals 1, so H_l + M_l + L_l = 1. These linear dependencies make the design matrix rank-deficient for every essay-level and QWK-level regression. Consequently, the OLS coefficient vector is not uniquely identified; any vector differing by a nullspace element of X'X gives the same fitted values. The specific coefficients reported in Tables 1 and 2 therefore depend on an arbitrary identifiability constraint (e.g., which variable is dropped or which generalized inverse is used). The one-sample t-tests on these coefficients are not valid hypothesis tests because the coefficients themselves are not well-defined. This directly undermines the quantitative claims of majority label bias and recency bias in §4.2 and §5. The descriptive claim that example selection affects QWK and that GPT-3.5 is more sensitive is supported by raw QWK variation (Fig. 2) and does not rely on the regression.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies how the choice and order of few-shot examples affect automated essay scoring (AES) with several versions of GPT-3.5 and GPT-4. The authors construct 118 prompts (117 few-shot with one to three examples, plus one zero-shot prompt), score 72 TOEFL11 essays with six GPT models, and quantify agreement with human ratings using quadratic weighted kappa (QWK). They then use linear regression on 3-shot prompts to detect 'majority label bias' and 'recency bias' in both raw GPT scores and QWK. The reported findings are that example selection affects QWK, that the effect is stronger for GPT-3.5 than for GPT-4, that majority and recency biases are present and more pronounced in GPT-3.5, and that GPT-4 (Jun23) is the most stable and highest-performing model in their setup. The descriptive, raw-variation part of the paper is clear and useful; however, the regression-based evidence for the bias claims is compromised by a rank-deficient design, as detailed below.","tokens_in":10374,"tokens_out":3930,"duration_ms":40300,"significance":"The descriptive contribution is practically relevant: documenting that example selection has a much larger effect on GPT-3.5 than on GPT-4, and that a minor-version checkpoint (GPT-4 Jun23) outperforms newer checkpoints, is useful for practitioners and for reporting standards in LLM research. The paper also has positive features: it names exact API model identifiers, fixes temperature to 0, reports the number of prompts and essays, and makes the raw QWK variation visible in figures. If the bias results were statistically valid, they would extend known findings from sentiment analysis (Zhao et al.) to AES. Unfortunately, the regression equations used to support the majority-bias and recency-bias claims are not identifiable as specified, so the quantitative bias conclusions in the abstract, Section 4.2, and Section 5 are not supported by the reported analysis.","major_comments":[{"comment":"The regression models are not identifiable because of perfect multicollinearity. For every 3-shot prompt, H_n + M_n + L_n = 3 by construction, and exactly one of H_l, M_l, L_l equals 1, so H_l + M_l + L_l = 1. After the stated standardization, these exact linear dependencies remain (the centered versions sum to zero). Thus the six predictors in Eqs. (1) and (2) are linearly dependent, the design matrix is rank-deficient, and the OLS coefficient vector is non-unique. Any coefficient vector differing by a vector in the null space of X'X produces identical fitted values. The specific coefficients and p-values reported in Tables 1 and 2 therefore depend on arbitrary identifiability conventions rather than on the data. The one-sample t-tests on these coefficients are not valid tests of majority label bias or recency bias. The bias claims in Section 4.2 and Section 5 are load-bearing for the paper's central message and need to be re-derived. A concrete fix is to drop one count predictor and one last-label indicator (e.g., use L_n and L_l as baseline categories) or to use sum-to-zero/orthogonal contrast coding, and then to report which parameterization was used.","section":"3.2, Eqs. (1)-(2)"},{"comment":"The QWK regression in Eq. (2) does not specify the observational unit or sample size. If QWK is computed per prompt across the 72 essays, the regression has only 27 distinct 3-shot prompt variants (or 81 if the three random example sets are treated as separate observations), and the t-tests treat these as independent even though the same 72 essays are used in every QWK value. The error terms are therefore correlated across prompts, and the reported p-values in Table 2 are not trustworthy even after resolving the collinearity issue. The paper should state the number of observations used in each regression, and should either use cluster-robust standard errors by essay set or a mixed-effects model with prompt and essay-set as random effects.","section":"3.2, Table 2"},{"comment":"The scores substituted for high/medium/low examples were tuned to [4,3,2] using GPT-3.5 (Jun23) on 200 prior essays, and this fixed mapping was then used for all six models. Since the optimal numerical substitution may differ across model versions, this choice can confound cross-model comparisons of both QWK levels and the size of the bias coefficients. The paper should either re-tune the substitution per model (on a separate validation set) or show that QWK and the inferred bias coefficients are robust to the substitution values. At minimum, the sensitivity of the main results to the [4,3,2] choice should be reported.","section":"3.1"}],"minor_comments":[{"comment":"The abstract states '119 prompts' while Section 3.1 says 117 few-shot prompts plus one zero-shot prompt, for a total of 118. Please correct the inconsistency.","section":"Abstract, 3.1"},{"comment":"The sentence 'Two responses within the category LM by GPT-3.5 (Nov23) did not yield ratings' is unclear: were these API calls that returned empty output, or outputs that could not be parsed? This matters for whether the exclusion is a missing-data mechanism or a model failure.","section":"3.1, GPT Ratings"},{"comment":"The mapping of GPT scores to three levels ('above 3 as high, 3 as medium, below 3 as low') is asymmetric around the example substitution [4,3,2]; please justify this choice and state how ties or non-integer outputs were handled.","section":"3.1, Agreement"},{"comment":"The sentence explaining why a meta-analysis was not possible ('essays where all coefficients were zero') is confusing; coefficients being zero for all predictors would be a special case, not a general obstacle to meta-analysis. Please clarify what was meant.","section":"4.2"},{"comment":"The captions should specify what the error bars represent (standard deviation across the three example sets? across prompts?) and the sample size underlying each average QWK.","section":"Figures 2 and 3"},{"comment":"The paper does not state whether the 72 evaluation essays and the essays used as examples were drawn from the same TOEFL11 essay prompts. If example essays come from the same prompt topics as evaluated essays, prompt-content leakage could inflate QWK for few-shot conditions; please clarify the sampling procedure.","section":"3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a preprint of an already published version (the text says the Version of Record is in Communications in Computer and Information Science, vol. 2150). If this submission is to a journal or another venue that does not allow prior publication, that is a separate editorial issue. My technical assessment is that the descriptive QWK findings are plausible and useful, but the regression-based bias findings require a full reanalysis because of the rank deficiency in Eqs. (1)-(2). The presence of a published version may also limit the scope of revision, so the editor may wish to consider whether a major-revision route is feasible for this submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"One thing to know upfront: the paper's quantitative evidence for majority-label and recency bias in AES does not survive scrutiny. The stress-test is right. In §3.2, for 3-shot prompts the predictors satisfy H_n+M_n+L_n=3 and H_l+M_l+L_l=1, so the design matrix is rank-deficient. The reported coefficients and t-tests in Tables 1 and 2 are therefore not identifiable without an arbitrary constraint. That means the specific claims in §4.2 and §5 about which biases are significant are not supported by the analysis as presented.\n\nThat said, the descriptive core of the paper is real and useful. The raw QWK variation across prompt categories (Fig. 2) is enough to show that example selection matters and that GPT-3.5 models are more sensitive than GPT-4 models, with GPT-4 (Jun23) being the most stable. The comparison across six model versions is a genuinely practical addition to the AES literature, and the observation that the latest model is not the best is worth keeping in mind. The paper also cleanly connects to Zhao et al.'s known few-shot biases and asks whether they transfer to essay scoring, which is a reasonable and novel question for this task.\n\nThe soft spots beyond the regression flaw: the evaluation set is 72 essays, which limits the precision of the QWK estimates even if the ranking of models is plausible. The example score substitution [4,3,2] was tuned on a validation set with the same GPT-3.5 (Jun23) model used in the main experiments; that is a minor hyperparameter choice, not a circularity, but it should be disclosed more clearly. The exact prompts and example essays are not released, which makes replication harder.\n\nThe regression problem is fixable: use effect coding or drop one category and interpret coefficients relative to a reference. As it stands, the bias analysis needs to be redone before the paper can be fully trusted. But the descriptive findings are worth preserving, and the paper engages honestly with its limitations.\n\nIf an editor asked me whether to send this to review: yes, it deserves referee time, but only with the expectation that the regression analysis will have to be substantially revised. The practical question about GPT-3.5 vs GPT-4 in AES is important enough that a corrected version would be a real contribution.","headline":"The descriptive QWK results are worth a look, but the regression-based bias claims collapse on perfect multicollinearity.","tokens_in":10873,"tokens_out":2319,"would_cite":false,"duration_ms":21008,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Example selection in few-shot prompts shifts GPT essay scores, with GPT-3.5 most affected.","keywords":["automated essay scoring","large language models","few-shot prompting","majority label bias","recency bias","quadratic weighted kappa","GPT-3.5","GPT-4"],"falsifier":"Because every 3-shot prompt satisfies $H_n+M_n+L_n=3$ and $H_l+M_l+L_l=1$, the reported coefficients are not uniquely identified; a re-estimation with one count and one indicator omitted would settle whether the bias effects are real.","tokens_in":9874,"feed_emoji":"📝","tokens_out":7756,"duration_ms":68558,"temperature":0.7,"pith_summary":"This study asks whether the choice and order of scored examples placed in a prompt change how GPT models grade essays. Using more than a hundred few-shot prompt variants applied to TOEFL11 essays, it finds that example selection moves quadratic weighted kappa, the agreement between GPT and human raters, and that the movement is larger for GPT-3.5 than for GPT-4. The paper also reports majority label bias and recency bias in generated scores and in QWK, with GPT-3.5 showing both more strongly. Its practical conclusion is that deliberate example selection can let cheaper GPT-3.5 models beat some GPT-4 models, and that model evaluations need to report minor version details.","feed_headline":"Example choice biases GPT essay scores, GPT-3.5 most","feed_subtitle":"Few-shot examples shift automated essay scores and human agreement; June 2023 GPT-4 stays steadiest.","key_machinery":"The central machinery is a controlled family of prompts combined with regression. Each few-shot prompt is named by its sequence of example scores such as H, ML, or LMH, with all one-, two-, and three-shot combinations represented, and QWK is computed after mapping ratings to high, medium, and low. To isolate bias, the paper regresses GPT scores and QWK on the counts of high, medium, and low examples ($H_n, M_n, L_n$) and on indicators for the last example's label ($H_l, M_l, L_l$). The signs and significance of these coefficients are what the majority-label and recency-bias claims rest on.","core_discovery":"The central discovery is that the examples in a few-shot prompt are not a neutral scaffold for AES: the score labels included and their order shift the distribution of GPT ratings and the agreement with human raters. The paper names two mechanisms: majority label bias, where more examples of a particular score pull ratings toward that score, and recency bias, where the last example's score has extra influence. These appear in GPT-3.5 snapshots; GPT-4 shows weaker effects, and the June 2023 GPT-4 version is both the most accurate and least sensitive to example variation. The paper argues that this makes example selection a prompt-engineering lever, especially for cost-sensitive deployments of lower-tier models.","pith_inferences":["One extension the paper does not run is a within-prompt permutation study: holding the multiset of example labels fixed while permuting their order would isolate recency from majority effects, which the current prompt sets vary together.","The same audit could be applied to other constructed-response scoring tasks and to non-GPT LLM families, since neither the prompt mechanism nor the bias hypotheses are GPT-specific.","In practice, the findings suggest a cheap calibration step: evaluate a few label and order combinations on a small validation set and freeze the best prompt for each model snapshot before large-scale scoring."],"forward_implications":["Prompt design becomes a tunable factor in AES: the set and order of examples can be optimized per model rather than treated as fixed.","Cost savings are available: with careful example selection, GPT-3.5 can match or beat some GPT-4 configurations.","Reports of LLM-based AES should state exact minor versions, since behavior shifts between releases such as 0613, 1106, and 0125.","Examples affect models differently, so conclusions about example selection from one model cannot be assumed to transfer to another.","Because recency effects are comparable to or larger than majority effects, simply reordering examples may be a low-cost intervention."],"supporting_citations":[{"why":"Provides the prompt template and the earlier finding that scoring examples improve GPT AES performance, which this study varies.","marker":"[7]"},{"why":"Establishes few-shot prompting as a general in-context learning method that the experimental design relies on.","marker":"[9]"},{"why":"Source of the majority label bias and recency bias hypotheses that the paper tests in the AES setting.","marker":"[11]"},{"why":"Supplies evidence that in-context majority label bias affects LLMs, making the bias tests relevant.","marker":"[16]"},{"why":"Supplies the TOEFL11 corpus and the expert three-level scores used to compute QWK.","marker":"[30]"},{"why":"Represents the recent-model quality claims that the June 2023 GPT-4 result is contrasted with.","marker":"[33]"}],"fun_headline_variants":["Few-shot examples skew GPT essay scores","Prompt examples bias GPT grading, GPT-3.5 worst","GPT-3.5 more affected by few-shot examples than GPT-4","Example selection shifts GPT essay scores","June 2023 GPT-4 best resists few-shot example bias"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The bias findings assume the regression predictors are linearly independent, so each coefficient can be uniquely estimated; in the 3-shot prompts the count variables sum to 3 and the last-label indicators sum to 1, making the coefficients non-identifiable as specified.","fun_headline_variants_meta":{"raw":{"variants":["Few-shot examples skew GPT essay scores","Prompt examples bias GPT grading, GPT-3.5 worst","GPT-3.5 more affected by few-shot examples than GPT-4","Example selection shifts GPT essay scores","June 2023 GPT-4 best resists few-shot example bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000621,"raw_usage":{"total_tokens":2888,"prompt_tokens":964,"completion_tokens":1924,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1844}},"tokens_in":580,"tokens_out":1924,"duration_ms":12728,"temperature":1.0,"reasoning_tokens":1844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:43:55.076753+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Because every 3-shot prompt satisfies $H_n+M_n+L_n=3$ and $H_l+M_l+L_l=1$, the reported coefficients are not uniquely identified; a re-estimation with one count and one indicator omitted would settle whether the bias effects are real.","supporting_citations":[{"cited_title":"In: Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications, pp","cited_arxiv_id":null,"evidence_quote":"Provides the prompt template and the earlier finding that scoring examples improve GPT AES performance, which this study varies."},{"cited_title":"In: Advances in Neural Infor- mation Processing Systems 33 (NeurIPS 2020), pp","cited_arxiv_id":null,"evidence_quote":"Establishes few-shot prompting as a general in-context learning method that the experimental design relies on."},{"cited_title":"In: Proceedings of the 38th International Con- ference on Machine Learning (ICML 2021), pp","cited_arxiv_id":null,"evidence_quote":"Source of the majority label bias and recency bias hypotheses that the paper tests in the AES setting."},{"cited_title":"ETS Research Report Series","cited_arxiv_id":null,"evidence_quote":"Supplies the TOEFL11 corpus and the expert three-level scores used to compute QWK."}],"review_version":1}