LLM agreement with human examiners is limited in fast mode and improves in reasoning or thinking modes¶
Claim · All claims
Evidence · 1 study · designr2·q3
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 1 of 1 entry) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q3 i? Fast-mode LLM grading showed low weighted kappas with humans (0.070–0.287); thinking mode raised agreement for ChatGPT (to 0.718), Kimi (0.571) and LeChat (0.186) but not for Gemini, which fell to 0.237. → Brevik A 2026
Evidence¶
Brevik A 2026¶
Brevik A, Jerpseth H and Lafontan SR (2026). Large language models as grading assistants in public health education: a method-comparison study of essay-style exam assessment. Front. Educ. 11:1904452. https://doi.org/10.3389/feduc.2026.1904452
q3 · i? · design · r2
Single-file comparison of four LLMs against human grades on 32 submissions. Table 1 shows fast-mode kappas of 0.269 (ChatGPT), 0.287 (Gemini), 0.070 (LeChat), 0.227 (Kimi); thinking mode raised them for ChatGPT (0.718), Kimi (0.571) and LeChat (0.186) but lowered Gemini's (0.237); Gemini Pro, tested only in thinking mode, reached 0.380. In thinking mode "the LLMs utilized more of the full grading scale".
"The weighted Cohen's kappa measuring inter-rater reliability for grading method comparisons indicated that the agreement between human graders and fast LLM grading was not high."
Discussion¶
Related Claims¶
- ChatGPT in thinking mode shows the strongest agreement with human examiners on essay-style exam grading (weighted kappa 0.718) — a narrower finding that bears on this claim
- Kimi shows reasonable but weaker agreement with human examiners than ChatGPT (weighted kappa 0.571) — a narrower finding that bears on this claim
- LLMs use fewer extreme grades than human examiners, with grade-dependent opposing biases compressing grades toward the middle — related
- Hint-based prompts raise ChatGPT scoring agreement with human experts to 98–100%, while no-hint prompts are unstable (Q2 precision 0.42) — related
- Prior non-LLM automated coding approaches yielded variable and often lower agreement with human coders on teacher knowledge measures, especially PCK — related
- GPT-4 classifies self-affirmation essays reasonably in zero-shot mode but below fine-tuned BERT, and few-shot prompting does not appreciably improve it — related
- Human inter-rater agreement for FRISCO essay scoring was reliable, with mean kappa 0.78–0.84 and percentage of agreement above 75% in all six dimensions — related
- LLM graders show model-specific precision/recall trade-offs, so a single metric is insufficient to judge grading quality — related
- Claude3 and PaLM2 show substantial agreement but fall short on identifying failing grades — related
- Newer LLM generations do not consistently agree better with human essay graders — related
- Gemini and Gemini Pro show the highest internal grading consistency across file upload volumes, with no clear monotonic lost-in-the-middle effect — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
llm-fast-mode-limited-thinking-mode-better-agreement |
title |
LLM agreement with human examiners is limited in fast mode and improves in reasoning or thinking modes |
description |
LLM agreement with human examiners is limited in fast mode and improves in reasoning or thinking modes |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
source_checked |
2026-10-20 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — brevik-a-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.