arxiv: 2512.18880 · v2 · submitted 2025-12-21 · 💻 cs.CL · cs.AI· cs.CY

Can LLMs Estimate Student Struggles? Human-AI Difficulty Alignment with Proficiency Simulation for Item Difficulty Prediction

Ming Li , Han Chen , Yunze Xiao , Jian Chen , Hong Jiao , Tianyi Zhou This is my paper

Pith reviewed 2026-05-16 20:13 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.CY

keywords large language modelsitem difficulty predictioneducational assessmenthuman-AI alignmentcognitive simulationproficiency promptingdifficulty estimation

0 comments p. Extension

The pith

Large language models fail to estimate how difficult questions are for human students, even when prompted to simulate lower proficiency.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether LLMs can predict item difficulty by aligning with human judgments of cognitive struggle. It runs a large-scale comparison across more than twenty models on medical and mathematical tasks. Results show consistent misalignment: models do not track human difficulty ratings and instead converge toward a shared internal consensus. Larger scale and explicit role-play prompts for different student levels do not close the gap, and models cannot accurately forecast their own errors. This indicates that strong problem-solving performance does not produce insight into human limitations.

Core claim

The central claim is that LLMs exhibit systematic human-AI difficulty misalignment. Scaling model size does not improve alignment with human ratings; models converge to a machine consensus. High task performance impedes difficulty estimation because models cannot simulate the capability limits of students even under explicit proficiency prompts, and they show no reliable introspection about their own limitations.

What carries the argument

Human-AI Difficulty Alignment measured through proficiency simulation prompts that instruct models to adopt specific student capability levels and then rate item difficulty.

If this is right

Current LLMs cannot be used directly for automated difficulty estimation in educational tests without additional human calibration.
Performance improvements from scaling do not transfer to tasks that require modeling human cognitive limits.
Prompt-based simulation of student perspectives is insufficient for alignment on difficulty judgments.
Models lack the ability to self-assess their own difficulty predictions, limiting their use in adaptive learning systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The misalignment may stem from training data that over-represents expert solutions rather than learner errors, suggesting future datasets could include struggle annotations.
Similar gaps could appear in other domains where models must predict human behavior under capability constraints, such as accessibility or tutoring.
Testing whether retrieval of real student response data during inference improves alignment would be a direct next experiment.

Load-bearing premise

Human-provided difficulty ratings serve as reliable ground truth and that explicit prompting to adopt proficiency levels is a sufficient test of whether an LLM can simulate student struggles.

What would settle it

Demonstrating high correlation between LLM difficulty predictions and actual student error rates or performance data on the same items after improved prompting or fine-tuning would falsify the misalignment result.

Figures

Figures reproduced from arXiv: 2512.18880 by Han Chen, Hong Jiao, Jian Chen, Ming Li, Tianyi Zhou, Yunze Xiao.

**Figure 2.** Figure 2: The Spearman correlation trends when greed [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Heatmap showing the correlation change when applying specific personas compared to the baseline. The impact of individual personas is highly inconsistent and noisy. Despite their profound problem-solving capabilities, LLMs struggle with cold-started IDP. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: The consensus heatmap of the spearman correlation between the models on the USMLE dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

**Figure 5.** Figure 5: The consensus heatmap of the spearman correlation between the models on the CMCQRD dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

**Figure 6.** Figure 6: The consensus heatmap of the spearman correlation between the models on the SAT Math dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

**Figure 7.** Figure 7: The consensus heatmap of the spearman correlation between the models on the SAT Reading dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗

**Figure 8.** Figure 8: The violin plot of the difficulty prediction results of GPT-5 with different personas. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗

**Figure 9.** Figure 9: Prompt templates for difficulty prediction (Low-Proficiency Student) across four tasks. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗

**Figure 10.** Figure 10: Prompt templates for difficulty prediction (Medium-Proficiency Student) across four tasks. [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗

**Figure 11.** Figure 11: Prompt templates for difficulty prediction (High-Proficiency Student) across four tasks. [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗

**Figure 12.** Figure 12: Prompt templates for question answering (Low-Proficiency Student) across four tasks. [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗

**Figure 13.** Figure 13: Prompt templates for question answering (Medium-Proficiency Student) across four tasks. [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗

**Figure 14.** Figure 14: Prompt templates for question answering (High-Proficiency Student) across four tasks. [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗

read the original abstract

Accurate estimation of item (question or task) difficulty is critical for educational assessment but suffers from the cold start problem. While Large Language Models demonstrate superhuman problem-solving capabilities, it remains an open question whether they can perceive the cognitive struggles of human learners. In this work, we present a large-scale empirical analysis of Human-AI Difficulty Alignment for over 20 models across diverse domains such as medical knowledge and mathematical reasoning. Our findings reveal a systematic misalignment where scaling up model size is not reliably helpful; instead of aligning with humans, models converge toward a shared machine consensus. We observe that high performance often impedes accurate difficulty estimation, as models struggle to simulate the capability limitations of students even when being explicitly prompted to adopt specific proficiency levels. Furthermore, we identify a critical lack of introspection, as models fail to predict their own limitations. These results suggest that general problem-solving capability does not imply an understanding of human cognitive struggles, highlighting the challenge of using current models for automated difficulty prediction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LLMs fail to align with human difficulty ratings on education items and explicit student-proficiency prompts do not fix it.

read the letter

The core result is that LLMs across more than twenty models show consistent misalignment with human difficulty judgments in medical and math items. Scaling does not improve alignment; instead the models converge on a shared machine view, and prompting them to adopt specific student proficiency levels still leaves them unable to simulate human struggles. They also cannot accurately flag their own limitations. This is a straightforward empirical observation rather than a derived claim, and the scale of the comparison is the main contribution. It documents patterns that smaller prior studies had only hinted at. The direct comparison to human ratings keeps the argument grounded without obvious circularity. The work is useful for anyone building automated testing or adaptive systems who might assume general capability transfers to difficulty estimation. On the downside, the abstract leaves sample sizes, exact prompting templates, and inter-rater reliability for the human judgments unspecified, so the strength of the misalignment numbers depends on how those were handled in the full methods. If those details are thin or the human ratings have high variance, the central claim weakens. The proficiency-simulation test is reasonable but not exhaustive; other elicitation approaches might behave differently. Overall the evidence is empirical and falsifiable, so the paper is worth a referee's time even if revisions are needed on the reporting of human data and statistical tests. I would bring it to a reading group focused on LLM evaluation in applied domains.

Referee Report

2 major / 2 minor

Summary. The paper conducts a large-scale empirical analysis of Human-AI Difficulty Alignment across over 20 LLMs in domains including medical knowledge and mathematical reasoning. It claims systematic misalignment between model-generated difficulty estimates and human ratings, that model scaling does not reliably improve alignment (instead producing convergence to a shared machine consensus), that high problem-solving performance impedes simulation of student struggles even under explicit proficiency-level prompting, and that models exhibit a lack of introspection about their own limitations. The central conclusion is that general problem-solving capability does not imply an understanding of human cognitive struggles, limiting the use of current LLMs for automated item difficulty prediction.

Significance. If the empirical results hold after addressing methodological gaps, the work is significant for AI-assisted educational assessment. It provides concrete evidence that superhuman LLM performance on tasks does not translate to accurate modeling of human learner difficulties, which could steer research away from naive prompting-based simulation toward more targeted alignment methods or hybrid systems. The observation of machine consensus and introspection failures offers falsifiable predictions that could be tested in follow-up studies.

major comments (2)

[Methods] Methods section: The manuscript provides no details on the number of items per domain, the number of human raters, inter-rater reliability statistics (e.g., Krippendorff’s alpha or ICC), or the exact statistical tests and effect sizes used to establish misalignment and convergence to machine consensus. These omissions are load-bearing because the central claims rest entirely on direct empirical comparisons between model outputs and human ratings.
[Results (Proficiency Simulation)] Proficiency simulation experiments: The claim that explicit prompting to adopt specific proficiency levels fails to align models with human difficulty judgments requires quantitative before/after comparisons and controls for prompt sensitivity; without these, it is unclear whether the observed failure stems from inherent model limitations or from the particular prompting protocol employed.

minor comments (2)

[Results] Include a summary table listing all evaluated models, their parameter counts, and baseline accuracies on the target tasks to allow readers to assess the scaling claims directly.
[Discussion] Define 'machine consensus' operationally (e.g., via pairwise agreement or clustering of model outputs) in the text or a dedicated subsection.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful for the referee's insightful comments, which have helped us identify areas for improvement in our manuscript. We address each major comment below and commit to revising the paper accordingly.

read point-by-point responses

Referee: [Methods] Methods section: The manuscript provides no details on the number of items per domain, the number of human raters, inter-rater reliability statistics (e.g., Krippendorff’s alpha or ICC), or the exact statistical tests and effect sizes used to establish misalignment and convergence to machine consensus. These omissions are load-bearing because the central claims rest entirely on direct empirical comparisons between model outputs and human ratings.

Authors: We agree with the referee that these details should be explicitly reported. We will revise the Methods section to provide the number of items per domain, the number of human raters, inter-rater reliability statistics, and the exact statistical tests and effect sizes used. These details are available from our experimental logs and will be added to ensure the claims are fully supported. revision: yes
Referee: [Results (Proficiency Simulation)] Proficiency simulation experiments: The claim that explicit prompting to adopt specific proficiency levels fails to align models with human difficulty judgments requires quantitative before/after comparisons and controls for prompt sensitivity; without these, it is unclear whether the observed failure stems from inherent model limitations or from the particular prompting protocol employed.

Authors: We appreciate this suggestion for strengthening the proficiency simulation results. In the revised version, we will add quantitative comparisons of alignment metrics (e.g., correlation with human ratings) before and after applying the proficiency-level prompts. Additionally, we will include an analysis of prompt sensitivity by testing multiple prompt variations and reporting the variance in outcomes. Our data shows that the improvement is marginal across prompts, reinforcing that the misalignment is not merely due to prompting protocol but reflects deeper model limitations. We will present this in a new figure and table. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper conducts a large-scale empirical study comparing LLM difficulty estimates against human ratings across 20+ models and multiple domains. All load-bearing claims (misalignment, scaling not helping, failure of proficiency prompting, lack of introspection) are presented as direct observations from these comparisons rather than derived from equations, fitted parameters renamed as predictions, or self-citation chains. No self-definitional steps, ansatzes, or uniqueness theorems appear in the reported methodology or results. The work is self-contained against external human benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The study is purely empirical and relies on standard domain assumptions from educational measurement rather than introducing new free parameters, axioms, or invented entities.

axioms (1)

domain assumption Human difficulty ratings serve as reliable ground truth for student cognitive struggles.
The entire alignment analysis treats aggregated human judgments as the reference standard without independent validation of their accuracy.

pith-pipeline@v0.9.0 · 5486 in / 1153 out tokens · 31621 ms · 2026-05-16T20:13:22.541162+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

BEAGLE: Behavior-Enforced Agent for Grounded Learner Emulation
cs.AI 2026-02 unverdicted novelty 6.0

BEAGLE uses a semi-Markov model, Bayesian knowledge tracing with injected flaws, and decoupled strategy-code actions to make LLM agents produce authentic student learning trajectories that humans cannot distinguish fr...

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages · cited by 1 Pith paper · 2 internal anchors

[1]

Wanyong Feng, Peter Tran, Stephen Sireci, and An- drew S Lan

Missing premise exacerbates overthinking: Are reasoning models losing critical thinking skill? InSecond Conference on Language Modeling. Wanyong Feng, Peter Tran, Stephen Sireci, and An- drew S Lan. 2025. Reasoning and sampling- augmented mcq difficulty prediction via llms. In International Conference on Artificial Intelligence in Education, pages 31–45. ...

work page 2025
[2]

In2021 ieee 33rd international conference on tools with artificial intel- ligence (ictai), pages 1398–1402

Automatically predict question difficulty for reading comprehension exercises. In2021 ieee 33rd international conference on tools with artificial intel- ligence (ictai), pages 1398–1402. IEEE. Fu-Yuan Hsu, Hahn-Ming Lee, Tao-Hsing Chang, and Yao-Ting Sung. 2018. Automated estimation of item difficulty for multiple-choice tests: An application of word embe...

work page 2018
[3]

GPT-4o System Card

Question difficulty prediction for reading prob- lems in standard tests. InProceedings of the AAAI conference on artificial intelligence, volume 31. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card.arXiv preprint arXiv:2410.21276...

work page internal anchor Pith review Pith/arXiv arXiv 2024
[4]

HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal

Leveraging llm respondents for item evalua- tion: A psychometric analysis.British Journal of Educational Technology, 56(3):1028–1052. Frederic M Lord. 2012.Applications of item response theory to practical testing problems. Routledge. Anastassia Loukina, Su-Youn Yoon, Jennifer Sakano, Youhua Wei, and Kathy Sheehan. 2016. Textual complexity as a predictor ...

work page internal anchor Pith review Pith/arXiv arXiv 2012
[5]

practice the art of tutoring,

More diverse dialogue datasets via diversity- informed data collection. InProceedings of the 58th annual meeting of the association for computational linguistics, pages 4958–4968. John Sweller. 1988. Cognitive load during problem solving: Effects on learning.Cognitive science, 12(2):257–285. John Sweller. 2011. Cognitive load theory. InPsychol- ogy of lea...

work page arXiv 1988