Personalized Distractor Generation

• Personalized distractor generation is a process that creates student-specific incorrect answer options designed to diagnose individual misconception patterns.
• It employs MCTS-guided reasoning reconstruction to analyze past question-answer records and simulate a student's error trajectory.
• Empirical results show enhanced diagnostic precision over traditional group-level methods, leading to more targeted remediation strategies.

Personalized distractor generation is the task of generating distractors tailored to a specific student’s misconceptions inferred from that student’s past multiple-choice question-answering records, so that different students may receive different distractors for the same question when their misconception patterns differ (Wu et al., 15 Aug 2025). In this formulation, distractors are not merely incorrect alternatives but diagnostic probes intended to expose recurring reasoning errors. The recent literature also distinguishes learner-specific personalization from adjacent forms of adaptation. In particular, context-sensitive and data-adaptive distractor generators adjust to question context, candidate confidence, example difficulty, or dataset/domain characteristics, but they do not model individual learners directly (Huang et al., 16 Jul 2025).

1. Definition and conceptual boundaries

The core distinction in the literature is between group-level distractor generation and personalized distractor generation. Traditional or group-level methods produce a small fixed set of incorrect but plausible answer options for a question by learning from aggregate student response patterns, implicitly assuming a common misconception structure shared by many learners. Personalized distractor generation instead aims to generate distractors tailored to a specific student’s misconceptions inferred from that student’s past records, with the explicit consequence that two students answering the same question may receive different distractors (Wu et al., 15 Aug 2025).

Formally, the personalized setting is defined for a student with past QA records

$$P=\{P_i\}_{1\le i\le M}=\{(s_i,a_i,d_i)\}_{1\le i\le M},$$

where $s_i$ is the stem, $a_i$ the correct answer, and $d_i$ the student’s chosen wrong option. The goal is to generate a distractor for a new question $s_j$ that reflects the student’s own recurring misconception patterns. This differs sharply from cloze distractor generation frameworks that are adaptive only at the level of context or optimization dynamics. DualReward, for example, is explicitly described as not being about personalization in the user-modeling sense: it has no student model, no learner proficiency estimate, no user history, no adaptation to individual misconceptions, and no tailoring to learner ability or preferences (Huang et al., 16 Jul 2025).

This boundary matters because the term “personalized” can otherwise become conflated with any context-aware or adaptive generation procedure. The papers considered here support a narrower definition: true personalization requires learner-specific conditioning, whereas context-sensitive distractor generation remains question-centered or training-centered.

2. Diagnostic function and educational rationale

The literature frames distractors as central to the diagnostic value of multiple-choice assessment. Distractors are designed to expose specific misconceptions, not simply to increase plausibility or difficulty. The personalized setting is motivated by the claim that different students can make different reasoning errors even on the same question, and that a shared distractor may reflect only one misconception. If a student’s misconception is not represented among the available options, the question may lose diagnostic power because the response no longer cleanly reveals the underlying misunderstanding (Wu et al., 15 Aug 2025).

Within this view, the educational value of personalized distractors comes from diagnosis rather than from surface realism alone. The stated benefits are increased diagnostic sensitivity to individual misconceptions, more accurate revelation of reasoning errors, and support for more targeted remediation and instructional intervention. A broader implication drawn in the paper is a shift from one-size-fits-all distractors to misconception-aware, student-specific assessment, with consequences for diagnostic testing, adaptive learning systems, personalized remediation, and cognitive diagnosis (Wu et al., 15 Aug 2025).

A related but distinct line of work emphasizes distractor quality in terms of challenge, plausibility, and discrimination. DisGeM treats distractors as central to MCQ quality because they determine whether a question is challenging and discriminative, while D-GEN studies whether generated distractors preserve benchmark ranking behavior and model uncertainty (Cavusoglu et al., 2024, Byun et al., 18 Apr 2025). These perspectives are compatible with personalization but do not by themselves define it. Personalized distractor generation adds learner-specific misconception alignment as the primary criterion.

3. Foundations in non-personalized and context-aware distractor generation

Recent distractor generation systems provide several methodological foundations that are directly relevant to personalization even when they are not learner-aware. The main contrast is among training-free candidate generation pipelines, controllable distractor synthesis for evaluation, and reinforcement-learning-based adaptive ranking.

System	Core mechanism	Personalization status
DisGeM	Two-stage pipeline with Candidate Set Generator and Distractor Selector	Not direct personalization
D-GEN	Open-source distractor generator with ranking alignment and entropy analysis	Not learner-aware
DualReward	Dual reward structure with adaptive scaling	Context-sensitive, not true personalization
MCTS-guided reasoning reconstruction	Training-free two-stage framework with student-specific misconception prototype	Personalized

DisGeM is a training-free two-stage pipeline for multiple-choice cloze or fill-in-the-blank questions. Its Candidate Set Generator masks the answer span and decodes distractor candidates with pretrained masked LLMs such as BERT or RoBERTa, while its Distractor Selector uses an NLI model to remove candidates that entail or paraphrase the gold answer or are too similar to each other. The framework introduces controllable knobs such as $n_{mask}$, dispersion, decoding strategy L2R/R2L/CTL, search multiplier $m_s$, and probabilistic ranking by $T_{rank}$, but it explicitly states that there is no control of the difficulty level for generated distractors and does not model learner profiles, skill level, misconceptions, knowledge state, or personalized distractor preferences (Cavusoglu et al., 2024).

D-GEN is presented as the first open-source distractor generator model that transforms open-ended data into an MC format. It generates three plausible but incorrect distractors from a question and correct answer pair, and evaluates them using ranking alignment and entropy analysis. The paper is not explicitly about personalization, but it separates the content of the distractor, the difficulty/plausibility balance, and the degree to which distractors change model confidence. These are natural control interfaces for a personalized system, even though the reported framework itself remains group-level (Byun et al., 18 Apr 2025).

DualReward moves further toward adaptation by introducing a reinforcement learning framework for cloze-test distractor generation. For each training instance it uses a candidate pool consisting of 3 gold standard distractors from human annotations and 7 model-generated distractors from the Candidate Set Generation module of CDGP. Gold distractors receive

$reward_{gold} = 1 \times reward\_scale,$

while model-generated candidates receive

$$reward_{gen} = 0.9 \times reward\_scale \times confidence\_score,$$

where confidence comes from BERT’s masked language modeling softmax probability at the blank position. The adaptive scaling factor is

$s_i$0

and the RL objective is

$s_i$1

This framework is context-aware and data-adaptive because it adjusts reward based on training loss, uses candidate confidence, distinguishes human and generated distractors, and performs better when domain or task variability is higher; however, it is explicitly described as not being true personalization (Huang et al., 16 Jul 2025).

4. MCTS-guided reasoning reconstruction for personalized generation

The first paper in the supplied set to define personalized distractor generation directly proposes a training-free two-stage framework built around MCTS-guided reasoning reconstruction (Wu et al., 15 Aug 2025). The motivation is that each student typically has only a few QA records, those records often lack the student’s underlying reasoning processes, and training-based group-level approaches are therefore infeasible for this task.

The first stage constructs a student-specific misconception prototype from historical QA records. For each past record $s_i$2, the model $s_i$3 extracts key knowledge concepts:

$s_i$4

The system then uses Monte Carlo Tree Search to recover plausible reasoning trajectories from the student’s observed wrong answers. Each node in the tree corresponds to a reasoning step and its intermediate result, and the tree includes both reasoning nodes and terminal nodes. The inclusion of terminal nodes is an important design choice because it captures cases where students stop early after a partially correct computation.

Selection is performed with UCT,

$s_i$5

with $s_i$6. Expansion creates at most $s_i$7 children for each reasoning node, excluding the terminal child; among these children, one is the correct next step and the remaining $s_i$8 are plausible incorrect alternatives. Simulation continues generation until a final answer is reached. Backpropagation uses two rewards: the match score

$s_i$9

and a plausibility score $a_i$0 from a separate evaluator model $a_i$1. The final reward is

$a_i$2

After $a_i$3 MCTS iterations, the method selects the most plausible trajectory that reaches the student’s actual wrong answer (Wu et al., 15 Aug 2025).

The recovered trajectories are then summarized into generalized misconceptions. For each concept $a_i$4, associated trajectories $a_i$5 are summarized by

$a_i$6

yielding concept–misconception pairs that collectively form the student’s misconception prototype. For a new question $a_i$7, the system extracts the relevant concepts, retrieves corresponding misconceptions, and generates the distractor by

$a_i$8

The resulting distractor is intended to be plausible specifically for that student because it is grounded in the student’s own historical reasoning failures (Wu et al., 15 Aug 2025).

A central conceptual contribution of this framework is that distractor generation is treated as reasoning reconstruction and misconception modeling, rather than only as text generation. This reorients the problem from producing broadly plausible wrong answers to simulating how a particular student is likely to think incorrectly.

5. Experimental settings and empirical findings

The personalized framework is evaluated on two curated datasets, Eedi_100 and Discrete_40, containing 140 students total, with 100 students in Eedi_100 and 40 students in Discrete_40, and over 7,000 QA records in total. Each student has, on average, $a_i$9 past QA records and $d_i$0 test QA records. Eedi_100 covers elementary arithmetic/math, while Discrete_40 covers undergraduate discrete math. The test records are restricted to incorrect answers, and the selected wrong option is used as the ground-truth distractor (Wu et al., 15 Aug 2025).

The reported baselines include heuristic methods—Random, Similarity, Level, Level+Random, Level+Similarity, and Mastery—as well as reasoning-aware baselines IO and CoT. Five LLM backbones are evaluated: LLaMA-3-70B, DeepSeek-V3, Claude-4-Sonnet, GPT-3.5-turbo, and GPT-4o. Metrics include Accuracy, defined as whether the generated distractor matches the student’s selected wrong answer, and LLM-based Plausibility and Coherence, each scored on a 1–5 scale using o3-mini as evaluator (Wu et al., 15 Aug 2025).

Across both datasets and all five backbones, the personalized method is reported as best. On Eedi_100, the best reported Accuracy values among compared methods are 17.7 for LLaMA-3-70B, 32.1 for DeepSeek-V3, 27.0 for Claude-4-Sonnet, 22.1 for GPT-3.5-turbo, and 28.2 for GPT-4o. On Discrete_40, the corresponding values are 28.6, 34.7, 40.4, 28.6, and 31.0. The gains are described as especially strong on Discrete_40, indicating particular effectiveness for reasoning-intensive settings; the paper also notes that stronger LLM backbones improve performance, suggesting that the framework benefits from richer candidate reasoning during MCTS (Wu et al., 15 Aug 2025).

The ablation studies attribute performance to several distinct components. Replacing $d_i$1 with simple stem similarity causes a notable drop, indicating that concept-level retrieval is essential. Removing terminal nodes reduces performance because truncated reasoning can no longer be modeled. Removing $d_i$2 weakens MCTS because the search then relies mostly on sparse answer matching. Removing $d_i$3 and feeding raw reasoning trajectories instead of summarized misconceptions also hurts performance. Search iteration analysis further shows that performance on Eedi_100 improves up to $d_i$4 and then plateaus or slightly declines, whereas on Discrete_40 it continues improving up to $d_i$5, supporting dataset-specific choices of $d_i$6 and $d_i$7 respectively (Wu et al., 15 Aug 2025).

Adjacent systems provide useful empirical context. DualReward reports modest but consistent benefits on the homogeneous passage-level dataset CLOTH-F and more substantial improvements on the cross-domain sentence-level MCQ dataset, including 3.48–3.86% in P@1 over fixed reward settings. It also reports that distinguishing gold distractors from model-generated distractors matters more on diverse data, with P@1 of 37.84 versus 33.20 under uniform reward on MCQ. These results do not establish learner personalization, but they show that adaptive reward scaling is particularly effective when difficulty and domain vary (Huang et al., 16 Jul 2025).

6. Evaluation, misconceptions, and open problems

A recurrent issue in this literature is how to evaluate distractors when gold personalized distractors are scarce. D-GEN proposes two methods that are not personalized themselves but are directly relevant to evaluation design. Ranking alignment tests whether generated distractors preserve the relative ordering of evaluated models; D-GEN reports overall Spearman’s $d_i$8 and Kendall’s $d_i$9 when comparing original MMLU distractors with generated ones. Entropy analysis compares the model confidence distribution over answer choices, with entropy computed from the choice probabilities $s_j$0, and most Wilcoxon signed-rank comparisons are reported as non-significant. A plausible implication is that personalized distractor generation may eventually require analogous evaluation protocols that assess diagnostic behavior rather than only surface fluency (Byun et al., 18 Apr 2025).

Several misconceptions are explicitly corrected by the available papers. First, personalization is not the same as harder distractor generation. DisGeM includes controllable generation knobs and studies human difficulty ratings, but explicitly states that there is no control of the difficulty level for generated distractors and does not implement learner-adaptive mechanisms (Cavusoglu et al., 2024). Second, context sensitivity is not equivalent to learner personalization. DualReward adapts rewards based on training loss, candidate confidence, and the distinction between human and generated distractors, but it does not include a student model, user history, or adaptation to individual misconceptions (Huang et al., 16 Jul 2025). Third, shared distractors are not necessarily diagnostically sufficient, because a shared distractor may reflect only one misconception and fail to capture the diverse reasoning errors of individual students (Wu et al., 15 Aug 2025).

The personalized framework also has explicit limitations. It assumes temporally stable student cognition over a short time window, depends on having enough QA history to reconstruct recurring patterns, relies on LLM quality at multiple stages, and is computationally more expensive than simple retrieval-based distractor generation because of MCTS search (Wu et al., 15 Aug 2025). At the same time, it generalizes to group-level settings by generating personalized distractors for many students, aggregating them, and selecting the most frequent distractors per question; this is reported to outperform standard group-level baselines on both in-domain and out-of-domain datasets, including Eedi_100, Discrete_40, MMLU Elementary Math, and CEval Discrete Math (Wu et al., 15 Aug 2025).

Taken together, these papers delineate a clear research trajectory. Training-free and controllable generation methods such as DisGeM and D-GEN establish mechanisms for candidate generation, semantic filtering, and behavioral evaluation. Adaptive frameworks such as DualReward show that learning signals can be conditioned on candidate source, confidence, and loss dynamics. Personalized distractor generation, as introduced by MCTS-guided reasoning reconstruction, goes beyond these by explicitly modeling student-specific misconception prototypes and generating distractors that align with an individual’s recurring reasoning failures (Cavusoglu et al., 2024, Byun et al., 18 Apr 2025, Huang et al., 16 Jul 2025, Wu et al., 15 Aug 2025).