acceptodds
Under review as a conference paper at ICLR 2027

Do LLMs Take Sides? Adjudex-Labor: Benchmarking Employer-Employee Bias in Arbitral and Judicial Decision-Making

Abstract

Evaluations of legal large language models (LLMs) often emphasize agreement with human adjudicators, overlooking whose interests their disagreements favor. In labor disputes, this omission is consequential: denying an employment relationship can exclude workers from legal protection, and repeated departures may concentrate legal risks on the same party. We introduce Adjudex-Labor, a benchmark of real Chinese labor disputes that examines positional tendencies across arbitration and subsequent judgments, using human decisions as institutional references rather than normative ground truth. Our evaluation distinguishes how often models disagree with human adjudicators from differences in the rates of employer- and employee-favoring departures. Employer-favoring tendencies predominate in independent arbitration. When the same model configuration performs arbitration and judgment, replacing human arbitration outcomes with model-generated outcomes shifts subsequent judgments toward employers. Crucially, closer agreement with human courts can coexist with more employer-favoring departures, while additional reasoning does not consistently improve positional balance. A separate monetary evaluation finds that most configurations award employees less than matched human adjudicators, including some that favor employees or appear balanced in employment-status decisions. These findings show that institutional agreement and positional balance are distinct properties, and that positional tendencies depend on both the adjudicative context and the outcome evaluated. LLM evaluation must therefore examine not only agreement with human decisions, but also how departures distribute protection, risk, and remedies between parties and across decision stages.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, AI Adjudication, Labor Law, Positional Tendency, Algorithmic Bias, Legal AI Evaluation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_eWbhOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Adjudex-Labor, a benchmark for evaluating whether LLM deviations from human labor-dispute adjudications systematically favor employers or employees. Using 804 real cases with both arbitration and court outcomes, the paper proposes complementary metrics for overall disagreement and party-relative directional imbalance, and evaluates multiple LLM configurations in both single-stage and multi-stage settings. The main results show systematic employer-favoring deviations and further employer-ward shifts when downstream judgments rely on LLM-generated arbitration outcomes.

Critical strengths
  1. The paper identifies an important evaluation dimension that standard agreement or accuracy metrics miss. The distinction between DDR and PPR is useful. Two models can disagree with human adjudicators at similar rates while imposing very different distributive consequences across employers and employees. The paper also shows that higher institutional agreement can coexist with greater positional imbalance, which makes a strong case that aggregate legal accuracy alone is insufficient for evaluating LLM adjudication systems.
  2. The multi-stage adjudication design goes beyond isolated legal prediction. The paper studies how upstream arbitration outcomes affect downstream judicial decisions and evaluates all arbitrator-judge model combinations. This is important because deployed AI decision systems may operate in pipelines where one model's output becomes another model's input. The finding that positional tendencies can persist, amplify, or reverse across stages would not be visible in a single-stage evaluation.
  3. The paper is appropriately cautious about institutional references. It generally avoids treating human arbitration or court outcomes as normative ground truth and instead frames the benchmark in terms of directional deviation relative to institutional decisions. This is important for interpreting the reported employer-favoring and employee-favoring tendencies.
Critical weaknesses

1. The benchmark currently lacks sufficient auditability and reproducibility.

The paper's main contribution depends on a newly constructed benchmark derived from real legal documents, but the current submission does not provide the benchmark data, construction code, or even a small set of complete, inspectable examples. This makes it difficult to independently verify case preprocessing, label extraction, leakage control, prompt construction, and output parsing. Even if the underlying legal documents cannot be fully released because of licensing or privacy constraints, the paper should provide anonymized or synthetic representative cases together with benchmark construction and evaluation code.

2. The external validity of the reported employer-favoring tendency is difficult to assess because the benchmark is highly selected.

The 804 cases are not a random sample of labor disputes. They must proceed through both arbitration and subsequent litigation, retain employment-status disputes as a substantive issue, and contain sufficiently clear claims and binary institutional outcomes. These requirements are understandable, but they may induce substantial selection effects. The results therefore strongly support conclusions about this selected litigation population, but extrapolation to labor disputes more broadly or to real deployment populations is less well justified.

3. Generation-level robustness is insufficiently documented.

The paper does not report several important decoding parameters, such as temperature, top-p, random seed, or the number of generations per case. The reported case-level bootstrap quantifies uncertainty across benchmark cases, but it does not capture stochastic variation in repeated model generations for the same input. Since several central conclusions concern directional shifts at the configuration level, at least a representative repeated-generation or multi-seed analysis would be important for establishing their stability.

4. The main sequential experiment establishes outcome propagation, but not richer reasoning propagation.

In the large-scale analysis, the downstream judge receives only the binary upstream arbitration outcome. This is a clean design for isolating the effect of replacing a human outcome with an LLM outcome, but it does not test whether model-generated legal reasoning, evidence interpretation, or argumentation itself propagates and amplifies positional tendencies. The smaller monetary experiment partially addresses this, but remains exploratory. The main claim should therefore remain scoped to outcome propagation rather than stronger reasoning self-reinforcement.

5. The technical novelty is limited, and the paper could be more concise.

The main contribution lies in the benchmark framing, the party-relative evaluation perspective, and the empirical study of a multi-stage legal process, rather than in a substantially new technical method. DDR is essentially an overall disagreement measure, while PPR captures directional asymmetry in those disagreements. This is still a meaningful contribution, but the paper's novelty depends heavily on the benchmark and empirical framing. The manuscript also repeats several central observations across multiple sections, and a more concise presentation would improve information density.

Questions for authors
  1. Can the authors provide inspectable benchmark artifacts during review, at least in the form of several anonymized representative cases containing processed case materials, Human Arbitrator and Human Judge labels, the exact LLM input prompts, and representative model outputs? Can the benchmark construction and evaluation code also be made available?
  2. Can the authors report the complete generation setup, including temperature, top-p, random seed where applicable, maximum generation length, and the number of generations per model-case-condition? If stochastic decoding is used, can repeated-generation or multi-seed results be provided for a representative subset?
  3. How representative are the final 804 cases of the broader source population? Can the authors report the number of cases remaining after each filtering stage and, where possible, compare retained and excluded cases on observable characteristics?
  4. The paper is appropriately careful in treating human decisions as institutional references rather than normative ground truth. Should the core claim therefore be interpreted strictly as a systematic directional deviation relative to observed institutional decisions, or do the authors intend to make a stronger fairness or legal-bias claim?
  5. The appendix states that the 50-case monetary sample was selected after preliminary analysis. What exactly were the selection criteria, and were they fixed before inspecting model monetary outputs?
8 comments
AuthorsOpenReview ↗

(Q1+W1) Benchmark Auditability and Reproducibility

To address the reviewer's concerns, we have provided an anonymous review repository containing de-identified real-case examples, complete experimental inputs for these examples, model responses, case-level results, and evaluation code. All materials are accessible without sign-in.

  1. Case materials and experimental auditability. We provide five de-identified real cases containing case facts, parties' arguments, key evidence, arbitration outcomes, court reasoning, and both human decision labels. These examples cover all four human arbitrator–judge outcome combinations, as well as both agreements and disagreements between model and human decisions.

  2. Prompts, information separation, and output verification. The repository includes three original prompt templates, complete de-identified inputs for all five cases under the three conditions, and 15 corresponding raw responses from GPT-5.6-Sol NT. Reviewers can inspect how information is used across conditions, how upstream arbitration outcomes are introduced, and how model responses correspond to the scored labels.

  3. Case-level data and reproducibility. We provide anonymized case-level results for 804 cases across 14 model configurations, together with runnable evaluation code. Without rerunning the models, reviewers can reproduce DDR, PPR, and the results for all 196 arbitrator–judge configuration pairings, including the central finding of an average PPR shift of +8.53 percentage points in same-configuration chains when replacing upstream human arbitration with LLM arbitration. The repository also includes execution instructions, dependencies, and expected outputs.

Appendix A of the revised manuscript further documents case screening, evidence processing, and label extraction.

AuthorsOpenReview ↗

(Q2) Generation Settings and Robustness

1. Generation settings (Appendix D.1). Table 5 reports the decoding settings for all 14 configurations in the 804-case experiment. We deliberately configured high maximum output-token limits to reduce truncation risks. Table 5 therefore reports the maximum observed completion tokens, distinguishing actual usage from configured limits.

2. Generation counts and experimental design (Appendix D.1). We retained one response per case and configuration under 3 input conditions, yielding 33,768 direct outputs across 804 cases and 14 configurations.

3. Repeated-generation experiments (Appendix D.4). We conducted two additional runs on a stratified subset of 60 employment-status cases across five representative configurations (Appendix D.4.1), and on all 50 monetary-award cases across 14 configurations (Appendix D.4.2). The appendices report the experimental procedures, output consistency, and stability of the results.

4. Statistical robustness (Appendix D.2). We report configuration-level 95% confidence intervals for DDR and PPR. In same-configuration chains, replacing human arbitration with LLM arbitration increased mean PPR by +8.53 pp (95% CI: [+5.09, +11.96]) toward employers and mean DDR by +9.53 pp (95% CI: [+7.81, +11.27]). These results support the directional shift under case-level resampling.

AuthorsOpenReview ↗

(Q3+W2) Sample Selection and Representativeness

We constructed the 804-case dataset to enable matched comparisons between arbitration and court decisions in the same labor dispute. Accordingly, retained cases must involve a substantive employment-relationship dispute with identifiable outcomes at both stages.

1. Case Retrieval and Screening. We retrieved approximately 15,000 first-instance judgments from China Judgments Online and PKULaw concerning the confirmation of employment relationships. After removing 2,320 duplicates, 12,680 documents remained. We subsequently excluded 3,540 documents with incomplete text and 3,130 lacking identifiable claims, defenses, or judicial examination of evidence, leaving 6,010 documents. We further excluded 5,081 cases in which employment status was not a substantive issue but appeared only as background or in procedural matters, leaving 929 cases.After removing 59 highly similar template-based cases and 66 cases without clearly identifiable arbitration outcomes, the final dataset comprised 804 cases.

2. Sample Composition and Scope. We characterize these 804 cases as a well-defined and analytically relevant subset rather than a statistically representative sample of broader labor disputes. All cases underwent both arbitration and subsequent litigation, making them suitable for examining positional tendencies across linked adjudicative stages. The sample includes 416 court-recognized employment relationships (51.7%) and 388 non-recognized relationships (48.3%), providing substantial observations in both reference-outcome groups for estimating PPR.

3. External Validity. Our selection criteria are designed to support matched comparisons across arbitration and litigation stages, requiring identifiable decisions and sufficient case materials at both stages. While this creates a selected litigation population rather than a representative sample of all labor disputes, it provides a well-defined setting for examining positional tendencies in sequential adjudication. Our findings should therefore be interpreted within this scope, where the benchmark offers systematic evidence of directional deviations from institutional decisions and their propagation across adjudicative stages.

Reviewer_mQLxOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies positional tendencies in labor dispute and monetary compensation applications. The authors ask the following questions:

  • RQ1: How does human vs. LLM arbitration affect judicial decisions?
  • RQ2: What are the positional implications (employee-favoring vs. employer-favoring) of LLM vs. human reviews?
  • RQ3: Do positional patterns/preferences persist in a second task, monetary reward assignment?

To answer these RQs, the authors introduce a new benchmark called Adjudex-Labor. It contains 804 Chinese judicial decisions and 50 cases on monetary judgments. Evaluating 14 models, they find that (1) there is a high-level (but not universal) pattern of favoring employers, and (2) on average, models award less monetary compensation than the corresponding human adjudicators.

Critical strengths
  1. Clarity and sound analysis. The motivation, exposition, experiments, results, and figures are all very clear and easy to understand. The analysis is clear, and the claims are appropriately corroborated.

  2. Important findings, broad evaluation, and interpretable metrics. The findings are interesting, consequential, and important. The model coverage is very good. The chosen metrics (DDR and PPR) are interpretable and appropriately simple.

Critical weaknesses

I have no concerns of significance.

Questions for authors

On what basis was the descriptive interpretation of PPR (lines 210–211) decided? Is there a precedent for using 5% as an informative threshold?

1 comment
AuthorsOpenReview ↗

Clarification on the PPR Classification Threshold and Robustness

We sincerely thank the reviewer for the careful evaluation. Your question has helped us clarify the rationale for the PPR classification threshold and further examine the robustness of our findings. The ±5 percentage-point (pp) band is a descriptive reporting convention rather than a threshold for determining whether positional tendencies exist. Our main findings remain robust across the alternative cutoffs examined.

  1. Rationale. PPR measures the difference between two conditional disagreement rates relative to human decisions. By conditioning on the human reference outcome, it accounts for differences in the prevalence of the two reference labels when comparing disagreement directions. Positive values indicate employer-favoring departures, while negative values indicate employee-favoring departures. We chose ±5 pp as a pragmatic descriptive tolerance to avoid assigning party-favoring labels to small nonzero differences. Values within this band may still exhibit directional imbalance. Importantly, the classification threshold does not affect the calculation of PPR itself.
  2. Precedent. We are unaware of an established cutoff specific to PPR. However, two clinical classification studies provide methodological examples of using a 5 pp margin for conditional-rate comparisons. Gianattasio et al. (2020) used 5 pp for subgroup differences in sensitivity and specificity, while Mackin et al. (2025) used the same margin for subgroup differences in false-negative rates. These studies illustrate the use of this numerical tolerance in other contexts, rather than establishing a PPR-specific threshold.
  3. Sensitivity. At cutoffs of 0, 5, and 10 pp, the numbers of independent arbitrator configurations classified as employer-favoring are 11/14, 11/14, and 10/14, respectively. The corresponding counts among the 196 arbitrator–judge pairings are 158, 147, and 118. In same-configuration chains, replacing human arbitration with LLM arbitration increases mean PPR by 8.53 pp (95% CI [5.09, 11.96]), with positive point estimates for all 14 configurations. Because the mean change and its confidence interval are computed from continuous PPR values, they are unaffected by the descriptive cutoff. We have clarified the reporting rule in Section 3.3 and added the methodological examples and sensitivity analysis to Appendix C.1. We appreciate the reviewer's helpful suggestion, which has improved the clarity and transparency of our evaluation.
Reviewer_erDmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • Adjudex-Labor: 804 real Chinese labor disputes with contested employment status, each with both an arbitration and a court outcome, plus a separate 50-case monetary sample.
  • Tests 14 configurations (7 model families, each with and without Thinking) as arbitrator and as judge; the judge is given either the human or the LLM arbitration outcome.
  • Metrics: DDR (how often the model disagrees with the human decision) and PPR (difference in conditional error rates; positive means the errors favor employers).
  • Findings:
    • 11 of 14 arbitrators lean employer-favoring.
    • Replacing human with LLM arbitration shifts judgments toward employers.
    • 147 of 196 arbitrator–judge pairings are employer-favoring.
    • Most models award employees less than human adjudicators.
Critical strengths
  • Useful reframing: the direction of errors, not just how often they occur. This is very useful to understand. Same DDR (e.g. ~40%) can come with opposite PPR (e.g. −27 vs. +43).
  • Real two-stage institutional data, 196 pairings, careful hedging.
Critical weaknesses
  • PPR may reflect a stricter evidence threshold, not taking sides. It tracks how often models say "no relationship", and there is no legal review of whether the departures are defensible.
  • The reference and inputs are confounded. Courts overturn arbitrators in 31.6% of cases, mostly toward employers; case materials may include the court’s review of evidence; this possible confoundation is unchecked.
  • The sequential result is nearly built in. Only flipped labels change the judge’s input, so it mainly measures anchoring on the prior label; there is no 'no-arbitration' baseline.
  • The monetary analysis is weak. 50 cases selected "after preliminary analysis". It is not clear what the criteria was. There is no confidence intervals provided as well.
Questions for authors
  1. Does PPR hold after controlling for each model’s rate of "relationship exists"?
  2. Did legal experts review a sample of the employer-favoring departures?
  3. Do the case materials include the court’s review of evidence?
  4. What do judges decide with no arbitration input?
  5. How were the 50 monetary cases chosen, and what are the CIs?
  6. What decoding settings, runs, and per-configuration CIs? Please provide details.
  7. Will the data be released?

Please consider adding these as part of the paper - especially the legal expert analysis.

9 comments
AuthorsOpenReview ↗

(W1 + Q1) Clarification of PPR

Yes. PPR captures information beyond each model's overall rate of predicting "relationship exists". PPR does not simply track how frequently a model denies such recognition. Rather, it measures the directional asymmetry of model–human disagreements relative to observed institutional decisions.

  1. Definition. PPR is defined as (100(e/n_1-w/n_0)), where (e) counts cases in which humans recognize an employment relationship but the model denies it, and (w) counts the reverse. The denominators (n_1) and (n_0) are the corresponding human-positive and human-negative case counts. A positive PPR indicates an employer-favoring asymmetry, while a negative value indicates an employee-favoring asymmetry.
  2. Distinction from simpler measures. Unlike the overall non-recognition rate, PPR distinguishes decisions that agree with human adjudicators from those that depart from them. Unlike a one-sided disagreement rate, it considers both directions of disagreement and normalizes each by its reference-group size.
  3. PPR helps distinguish how often a model disagrees with humans from whether those disagreements tend to favor one side. For example, reversing 5% of human-approved employment decisions and 5% of human-rejected decisions gives a PPR of zero—even if approved cases greatly outnumber rejected ones. Reversing 20% of approvals but only 5% of rejections gives a PPR of +15 percentage points, revealing an employer-favoring asymmetry. Thus, PPR avoids mistaking class imbalance and symmetric errors for directional bias.
  4. This distinction is also reflected in our results. Two adjudication chains have nearly identical DDRs (40.17% and 40.67%) but opposite PPRs (−27.46 pp and +42.74 pp), showing that similar overall disagreement rates can conceal substantially different positional tendencies. Furthermore, Appendix C.3 (Table 4) identifies three pairs of independent arbitrator configurations with identical recognition rates but different PPRs, confirming that recognition frequency alone does not determine PPR.
  5. Legal interpretation. In Chinese labor disputes, establishing an employment relationship is generally a prerequisite for accessing labor-law protections and remedies. Recognition therefore favors the employee's legal position, while non-recognition favors the employer's position. This classification reflects the direction of the outcome's legal consequences, not whether the decision is legally correct or fair. PPR uses this distinction to examine whether deviations from institutional decisions systematically favor one party over the other.
AuthorsOpenReview ↗

(W2 + Q2–Q3) Legal Expert Review and Evidentiary Inputs

We explicitly considered the risk that the court's evidence review might reveal its final decision and applied specific processing rules to prevent such leakage. Legal experts also participated in reviewing the case materials and model outputs. Our responses are as follows.

  1. Q2: Did legal experts review employer-favoring departures? Yes. Legal experts reviewed sampled model outputs, including employer-favoring departures. Their review focused on whether the models' stated reasons were internally consistent with their final decisions, rather than independently reassessing the substantive legal correctness of each departure. This distinction reflects our research objective: identifying directional differences from institutional decisions, rather than adjudicating the legal merits of every disagreement.
  2. Q3: Why did we retain the court's evidence review? We explicitly distinguished item-level evidence assessments from the court's ultimate determination of employment status. For example, confirming the authenticity or acceptance of a payroll record is distinct from concluding that the record, together with other evidence, establishes an employment relationship. The former informs the model about the status of the evidence; removing it entirely could lead the model to treat unverified and accepted evidence alike. We therefore retained item-level assessments of authenticity and acceptance, together with relevant objections, while excluding the court's ultimate findings and passages that directly disclosed or clearly implied those findings. Human arbitration and court decisions served as separate institutional references, and the court's final decision was excluded from the model input.
  3. Legal experts' involvement. This study is a collaboration between legal and AI researchers. Legal experts contributed to the evaluation metrics and reviewed (1) the arbitration–litigation workflow; (2) the processing of case facts, parties' arguments, and evidence; (3) the separation of adjudicative outcomes from model inputs; (4) the formal consistency between stated reasons and final decisions in sampled model outputs.. Their involvement helped ensure that the experimental procedures respected the relevant legal workflow and information boundaries. Appendix A.3 describes the scope of this review, and we will further clarify the sampled-output verification procedure.
AuthorsOpenReview ↗

(W3 + Q4) No-Arbitration Baseline and Sequential Experimental Design

We would like to clarify two points: our study includes an independent-decision condition closely resembling a Judge without arbitration input, and the observed directional shifts in the sequential experiment arise from actual model decisions rather than being predetermined by the experimental design.

  1. No-arbitration baseline (Q4). The reviewer's proposed "Judge with no arbitration input" closely resembles our LLM Arbitrator condition. Both perform the same employment-status determination task using identical case facts and evidence. The main difference is whether the preceding arbitration outcome is provided; the role instruction differs only in the brief designation of "arbitrator" versus "judge." Thus, LLM Arbitrator provides a closely related independent-decision reference. Relative to Human Judge, 10 of the 14 configurations are employer-favoring, one is middle, and three are employee-favoring under this condition. We did not separately include a Judge without arbitration input because our study follows China's arbitration-before-litigation procedure for labor disputes, in which cases reaching court already have a preceding arbitration outcome.
  2. Are the sequential results predetermined? (W3). We deliberately varied only the preceding arbitration outcome to isolate its effect on downstream judgments. However, changing this input does not necessarily change the model's decision. The empirical results directly demonstrate this: among 5,025 case–configuration instances where human and LLM arbitration outcomes differed, downstream judgments remained unchanged in 3,094 instances (61.6%), while only 1,931 (38.4%) changed. Thus, the experimental design predetermines neither whether a judgment changes nor the direction of that change. Moreover, across all 804 cases, replacing human arbitration with LLM arbitration shifted PPR toward employers in all 14 configurations, with a mean shift of +8.53 pp (95% CI: [+5.09, +11.96]). This directional pattern reflects observed model behavior rather than a mechanical consequence of the reconstruction rule.
  3. Possible anchoring effects (W3). We agree that anchoring may be one mechanism through which preceding arbitration outcomes influence downstream judgments, but this does not contradict our findings. First, employer-favoring tendencies are not confined to the sequential experiment. Section 4.1 shows that, even in the LLM Arbitrator condition, where no preceding arbitration outcome is provided, 11 of 14 configurations exhibit employer-favoring tendencies relative to Human Arbitrator. Meanwhile, LLM Judge with Human Arbitration receives the actual human arbitration outcome rather than a model-generated one. Therefore, anchoring on model-generated arbitration outcomes cannot explain all the positional differences observed in our study. Second, even if anchoring contributes to the sequential shifts, it does not diminish their significance. Our study asks whether LLM-generated arbitration outcomes systematically alter downstream judgments in sequential legal decision-making and which party benefits from these changes. Our contribution is to identify and quantify these directional effects rather than establish the specific mechanisms underlying them.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Materials

We thank all reviewers for their careful evaluations and constructive feedback. The manuscript and supplementary materials have been revised in response to the concerns raised, with additional experiments and analyses conducted. The main updates are summarized below.

  1. Benchmark Transparency and Reproducibility. An anonymous review repository. is now available, containing five de-identified real-case examples, experimental prompts and inputs, representative raw model responses, and case-level results for 804 cases across 14 configurations. The repository provides runnable evaluation code, enabling independent reproduction of DDR, PPR, and all 196 arbitrator–judge pairing results.
  2. Experimental Robustness and Statistical Analysis. Appendix D reports decoding settings, generation counts, and 95% confidence intervals, while documenting limitations in the original parameter records. Two additional generation rounds were conducted on 60 employment-status cases across five configurations and all 50 monetary-award cases across 14 configurations. The results document both aggregate stability and configuration-level variation.
  3. Benchmark Construction and Evaluation Design. Appendix A provides further details on the case-screening process, evidence processing, information separation, and legal expert review. Appendix C clarifies the distinction between PPR, overall recognition rates, and DDR, and includes sensitivity analyses under alternative PPR classification thresholds.
  4. Scope and Interpretation of Findings. The revisions clarify the scope of the selected litigation population, the use of human decisions as institutional references rather than normative ground truth, and the distinction between outcome propagation and reasoning propagation. The post-hoc selection criterion is also acknowledged.

We hope these revisions and additional materials help address the concerns raised during review. We sincerely appreciate the reviewers' time and thoughtful feedback. We hope this work helps draw attention to the importance of evaluating LLMs in employment-status disputes, where decisions can directly affect access to labor-law protections and remedies. If any concerns remain or further clarification would be helpful, we would be grateful to hear from you.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.