acceptodds
Under review as a conference paper at ICLR 2027

EviDial: When Correct Depression Predictions Lack Evidence

Abstract

Depression dialogue benchmarks typically evaluate models by comparing predicted symptom labels against clinical scale scores. Yet a correct label does not imply that the dialogue contains enough evidence to justify that prediction. We introduce EviDial, a benchmark for auditing this gap between label correctness and evidence-grounded assessment. EviDial spans 382 depression dialogues from three sources and 2,744 participant–symptom cells, characterized under an evidence taxonomy of absent, partial, conflicting, and sufficient evidence. To construct references without assuming a single gold standard, we combine generalizability theory, many-facet Rasch measurement, and random-effects latent class analysis. Our audit reveals a substantial mismatch between benchmark labels and textual evidence: 82.8% of cells lack sufficient evidence for the corresponding symptom assessment. Moreover, among label-correct predictions, 69.2–96.9% are unsupported by sufficient textual evidence across the evaluated model families and prompting strategies. We also find that high evaluator agreement can mask low measurement independence: five same-family LLM judges correspond to only 1.4 effective independent votes despite Fleiss’ κ = .785, and perturbation-based sensitivity checks fail across the tested judges. These findings show that label agreement alone can substantially overstate evidence-grounded performance in depression dialogue evaluation. EviDial provides a measurement layer for distinguishing predictions that match clinical labels from those that are justified by the evidence available in the dialogue.

Primary area:
datasets and benchmarks
Keywords:
LLM, LLM Agents, AI Security

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jUEXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces EviDial, a benchmark that asks whether a depression prediction is supported by the dialogue, rather than simply whether it matches a recorded clinical label. For example, a model might correctly predict “no sleep disturbance” even though sleep was never discussed. The authors call this textually unsupported correctness. They analyze 382 dialogues from three datasets, using 2,744 participant–symptom pairs for the main reference. LLM judges classify evidence as absent, partial, conflicting, or sufficient, with explicit rules for whose symptoms are described, the relevant time window, negation, and retracted statements. A latent-class model combines the judgments. The central contribution is an evaluation framework that separates getting the label right from having enough evidence to justify it. Main findings:

  • 82.8% of symptom assessments lack sufficient evidence under the automated reference, with substantial differences between datasets.
  • Across three models and four prompting strategies evaluated on 87 development cells, 69.2–96.9% of label-correct predictions lack supporting evidence.
  • Evidence-focused prompting does not consistently solve the problem; abstention reduces unsupported predictions at the cost of answering fewer cases.
  • A 120-cell human validation reportedly preserves the main finding, although automated judges perform poorly on controlled evidence perturbations.
Critical strengths

The paper identifies an important limitation of depression-dialogue evaluation: a prediction can match a clinical label without sufficient support in the model-visible transcript. Separating these properties helps prevent benchmark accuracy from being over interpreted as evidence-grounded assessment capability.

Critical weaknesses
  1. The evidence reference is not sufficiently validated to support the headline estimates. No automated judge condition passes the perturbation-validation gate, and latent-class fusion performs worse than plurality on the available anchor comparisons while producing no sufficient-negative or conflicting assignments. The reported human validation is encouraging, but lacks essential details about sampling, pre-adjudication agreement, and overlap with the model audit. Because the central results depend on this reference, these uncertainties materially weaken confidence in the benchmark’s validity.
  2. The empirical evidence supports narrower conclusions than the paper’s framing suggests. The model audit comprises only 87 development cells, with single runs and no completed held-out evaluation. Meanwhile, E-DAIC dominates the reference pool and underwent LLM-based reformatting without independent verification of evidence preservation. Consequently, the submission does not adequately establish how broadly the reported gap generalizes or h
Questions for authors
  1. Clarification: human validation: How were the 120 cells sampled, what were the annotators’ qualifications and pre-adjudication agreement, and how many cells overlapped the model audit? Please provide exact denominators and uncertainty estimates for the human-rescored results. These details are essential to assessing whether human validation supports the headline conclusions.
  2. Clarification: reference robustness: Given fusion’s worse anchor performance than plurality and its absence of sufficient-negative or conflicting assignments, what justifies selecting it as the primary reference? Do the evidence-gap and unsupported-correctness estimates remain similar under plurality and human labels when evaluated on matched cells?
Reviewer_jKqvOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Addressing the evaluation gap regarding "textually unsupported correctness" in mental health-related dialogues, this paper introduces EviDial, a benchmark designed to verify whether depression-symptom predictions are textually well-supported. Its core is decoupling label accuracy from evidential sufficiency, via a four-class evidence taxonomy and four-dimensional binding rules covering subject, time, scope and retraction. They adopt the latent-class RE-LCA alongside psychometric methods (G-theory, MFRM, MTMM) to construct and validate evidence reference labels for LLMs. They evaluate GPT, Gemini, Qwen and diverse prompting strategies on bilingual clinical datasets PDCH and E-DAIC. With metrics including UCR and Hit Caliber, they report that 69–97% of correct predictions lack robust textual support. Additionally, they assess the robustness and validity of the evaluator itself through controlled semantic perturbation stress tests and double-blind human annotation cross-checks.

Critical strengths

The paper raises a valuable research question: agreement between model outputs and clinical labels does not guarantee sufficient textual evidence. Distinguishing label correctness from evidential support is insightful, particularly for mental health assessment. But in my view, the current experiments and materials fail to validate its core quantitative claims.

Critical weaknesses
  1. The study relies on a reference standard without independent span‑ and rule‑level human gold annotations. It uses an LLM judge and RE‑LCA for labeling, making it hard to separate true evidence insufficiency from extraction, alignment or aggregation errors. Perturbation tests and audits on 120 samples cannot locate stage‑specific errors or confirm pipeline stability.
  2. Critically, RE‑LCA produces no SUFFICIENT‑negative or CONFLICTING labels across 2744 units, automatically marking all valid negative predictions as evidence‑deficient. This structural asymmetry, together with uneven cross‑language and cross‑judge coverage, may bias UCR and Hit Caliber and undermine the reliability of the proposed framework.
Questions for authors
  1. Figures 3 and 4 have labeling errors. Figure 3 contains only panels (a) and (b), yet the caption refers to (a)–(e). In Figure 4, the legends for (b) and (c) are swapped: (b) shows Hit Caliber and Coverage, while (c) plots UCR with its confidence intervals. Please check all figures and captions across the manuscript.
  2. The entry for Wang et al. (2025) does not match the official publication record. It should be: Computers and Education: Artificial Intelligence, Vol. 9, Article 100481; authors: Yuehan Wang, Jinyan Huang, Lun Du, Yuxin Guo, Ying Liu, and Rong Wang (DOI: 10.1016/j.caeai.2025.100481). Because citation errors raise doubts about overall bibliography reliability, please double-check all references.
  3. In Table 2, evaluation samples are unevenly distributed across languages and LLM judges, without explanation of why or how this affects RE-LCA reference label quality. Crucially, across all 2,744 evaluation units, RE-LCA produced zero SUFFICIENT-negative or CONFLICTING MAP labels. This forces all valid negative predictions to be classified as lacking evidence, creating an asymmetry that likely biases UCR and Hit Caliber. Please explain these missing classes and verify whether negative labels stem from genuine lack of evidence or systematic bias in the judge/RE-LCA.
  4. Relying on an LLM judge without snippet-level or binding-rule human gold standards makes it difficult to validate evidence extraction and binding. Errors can compound down the pipeline. Because current perturbation tests and manual spot-checks only evaluate end-to-end outputs, they cannot isolate stage-wise errors or separate intrinsic evidence gaps from pipeline-induced noise. Please add staged ground-truth annotations and controlled experiments to quantify per-module error propagation.
  5. Section 5 (Human Validation) reports RE-LCA agreement (76.5%) and plurality agreement (73.1%) over 120 human validation units. If both rates use the full 120 units as the denominator, these values cannot come from exact integer counts, so please clarify the calculation method.
  6. The anonymous GitHub repository in the reproducibility statement has permission issues: the directory tree is visible, but all files include README cannot be opened without extra access. Please fix permissions so reviewers can view the replication materials.
Reviewer_xWKoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

With the depression dialogue benchmarks, the paper studies whether correct model predictions are supported by evidence in the dialogue. Using assessment cells, the paper introduces an evidence taxonomy and constructs reference labels for detailed analysis. The paper also studies the reliability of these labels through experiments. The experiments show that evidence availability varies across benchmark sources, with substantially more sufficient evidence available in some than in others. However, the paper has presentation issues, and the reliability of its LLM-generated evidence labels is not sufficiently established.

Critical strengths
  1. The paper studies an important question: when a prediction matches the benchmark label, it may not be well supported by the dialogue text. Distinguishing label correctness from evidence support is useful and provides additional insights.
  2. The paper proposes a four-state taxonomy to analyze evidence support, which provides a more detailed breakdown of the evidence in the dialogue. The paper also examines the reliability of LLM evidence judges using controlled perturbations.
Critical weaknesses
  1. The paper uses LLMs to produce evidence labels, but the experimental results in Section 4.3 show that all tested LLM judges perform poorly on the controlled perturbation tests. This raises concerns about the reliability of the labels and the main results. Human validation in Section 5 provides some support, but it includes only 120 assessment cells. Their coverage across symptom dimensions and data sources remains unclear, making it difficult to assess whether this validation establishes reliability across the benchmark. This weakens the paper, as evidence support is its central topic.
  2. The paper is hard to follow in general. Understanding the paper requires careful reading and piecing together information from different sections. The presentation requires significant improvement. For example, the main method presents too many implementation details without first giving a clear overview. The relationship between the current reference construction and the legacy panel analysis is unclear. Some notation is also used without being clearly defined, e.g., C and R_{c}.
Questions for authors

Why do LLMs predict the correct labels when evidence is judged insufficient? Is it possible that the dialogue contains implicit cues or negation context that the evidence judges fail to recognize? These possibilities are not clearly discussed, and addressing them would strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.