EduMCTS-160K: Multimodal Reasoning Dataset

• EduMCTS-160K is a curated dataset of 160K multimodal reasoning trajectories produced using a Monte Carlo Tree Search guided by the process-aware reward model, EduPRM.
• The corpus integrates structured reasoning nodes, including self-reflection and critique metadata, to ensure step-level correctness and global coherence in STEM problem solving.
• Constructed through a three-stage pipeline of PRM-guided selection, MCTS trajectory expansion, and best-of-N filtering, the dataset significantly boosts the performance of multimodal language models.

Searching arXiv for the cited paper and related mentions of EduMCTS-160K. EduMCTS-160K is a large-scale corpus of 160,000 scientifically grounded, stepwise reasoning trajectories constructed via an MCTS-based search framework, EduMCTS, guided by a process-aware reward model, EduPRM, within the EduFlow pipeline for multimodal LLMs (MLLMs) (Zhu et al., 12 Jul 2025). It is designed to provide high-quality, interpretable process supervision for educational scientific reasoning, particularly in STEM settings where models exhibit underrepresentation of symbolic and scientific patterns in pretraining, fragile multi-step inference without global planning, and a lack of reflective self-correction (Zhu et al., 12 Jul 2025). Each retained trajectory is multimodal, pairing question text with visuals, structured reasoning actions, critique metadata, and a final answer, with verification mechanisms intended to ensure pedagogical alignment, step-level correctness, and global coherence (Zhu et al., 12 Jul 2025).

1. Definition and scope

EduMCTS-160K is defined as “a curated set of 160K multimodal (text + images/diagrams) educational reasoning trajectories produced by EduMCTS and filtered by EduPRM” (Zhu et al., 12 Jul 2025). Each example pairs the task input with “a verified sequence of structured actions” using the node types caption, summary, sub_task, thinking, self-reflection, and answer, together with step-level critique metadata from EduPRM and the final answer (Zhu et al., 12 Jul 2025).

The dataset is positioned as process supervision rather than answer-only supervision. Its stated purpose is “to supply high-quality, interpretable process supervision that trains MLLMs to plan, decompose, reason, and reflect” (Zhu et al., 12 Jul 2025). The construction explicitly targets three gaps: ensuring step-level correctness and global coherence, modeling reflective error correction through dedicated [Self-Reflection](https://www.emergentmind.com/topics/self-reflection) nodes, and enforcing multi-perspective critique via EduPRM’s stepwise tags, explanations, and scores (Zhu et al., 12 Jul 2025).

The corpus is multimodal and scientifically grounded. It is built from science problems that require visual grounding, and the pipeline “explicitly removes questions solvable without images” (Zhu et al., 12 Jul 2025). The highlighted subject coverage comprises five STEM domains: Math, Biology, Physics, Geography, and Chemistry, with “17K curriculum-aligned concepts” organized into “54 competency groups” for coverage (Zhu et al., 12 Jul 2025).

2. Position within EduFlow

EduMCTS-160K is the data artifact produced by EduFlow, which is described as “the first end-to-end framework that covers the full pipeline of educational scientific reasoning, including data selection, MCTS-based trajectory construction, model training, and output optimization” (Zhu et al., 12 Jul 2025). Within that pipeline, EduPRM functions as selector, critic, and verifier, while EduMCTS functions as the domain-adapted search mechanism that expands, evaluates, and filters candidate reasoning paths (Zhu et al., 12 Jul 2025).

The dataset addresses persistent reasoning deficiencies of MLLMs in STEM education. The stated motivations are “underrepresentation of symbolic/scientific patterns in pretraining, fragile multi-step inference without global planning, and lack of reflective self-correction” (Zhu et al., 12 Jul 2025). The resulting trajectories are described as “pedagogically aligned, verifiable, and coherent multi-stage trajectories—each built with fine-grained critique and reflective error-checking” (Zhu et al., 12 Jul 2025).

A notable design feature is the integration of reflective verification into the trajectory itself. The dedicated self-reflection node is not an auxiliary annotation but part of the retained reasoning process. This makes EduMCTS-160K a dataset of structured workflows rather than merely chains of thought. A plausible implication is that the corpus is intended to supervise both intermediate reasoning and metacognitive correction behavior, not only task completion.

3. Construction pipeline

EduFlow constructs EduMCTS-160K through three integrated stages (Zhu et al., 12 Jul 2025).

First, in PRM-guided sample selection, the pipeline starts from “a large raw pool of multimodal science questions (scanned exams, online platforms, structured question banks)” (Zhu et al., 12 Jul 2025). Images are parsed “with Mathpix/LayoutLMv3,” equations are LaTeX-encoded, and layouts are structured into JSON (Zhu et al., 12 Jul 2025). The process then applies “difficulty and text-solvability filtering,” extracts “17K curriculum-aligned concepts,” and clusters them into “54 competency groups for coverage” (Zhu et al., 12 Jul 2025). EduPRM is used to evaluate preliminary rollouts and compute step-level reward variance; samples with high stepwise disagreement are prioritized for search (Zhu et al., 12 Jul 2025).

Second, in PRM-guided MCTS trajectory construction, EduMCTS expands candidate paths using actor model generation and the functional node types caption, summary, sub_task, thinking, self-reflection, and answer (Zhu et al., 12 Jul 2025). These nodes scaffold “extraction of visual info, abstraction of goals, decomposition, inference, reflective verification, and conclusion reporting” (Zhu et al., 12 Jul 2025). At each rollout, EduPRM scores partial trajectories via step-level correctness, coherence, and error tags; those scores backpropagate to guide future expansions and prune weak branches (Zhu et al., 12 Jul 2025). The self-reflection mechanism prompts verification and correction of prior steps, and “trajectories failing verification are discarded” (Zhu et al., 12 Jul 2025).

Third, in PRM-based best-of-N selection, multiple candidates are sampled, scored stepwise by EduPRM, and ranked by accumulated stepwise reward, with the highest-scoring path retained (Zhu et al., 12 Jul 2025). This stage is used both during dataset curation and later during evaluation as a form of inference-time scaling (Zhu et al., 12 Jul 2025).

The filtering rules are unusually strict. The pipeline retains “only trajectories in which all steps are labeled ‘Correct Step’” and discards paths with low overall confidence or failed self-reflection verification (Zhu et al., 12 Jul 2025). It additionally applies self-consistency and rejection sampling to aggregate multiple candidates and filter them with EduPRM’s critiques (Zhu et al., 12 Jul 2025). After multiple rollouts per problem and filtering, the process finalizes “160K verified trajectories” (Zhu et al., 12 Jul 2025).

4. Search procedure and reward modeling

EduMCTS is a domain-adapted Monte Carlo Tree Search whose expansions are driven by “a pool of actor models” and whose node evaluation is provided by EduPRM (Zhu et al., 12 Jul 2025). The paper reproduces the search equations explicitly.

For expansion:

$$\mathcal{C}_t = \bigcup_{j=1}^{K} \pi_j\left( \cdot \mid s_t \right)$$

For stepwise scoring by EduPRM:

$$\hat{r}(s') = \frac{1}{K} \sum_{j=1}^K \mathrm{PRM}_\psi\left( Yes \mid s' \right)$$

For pruning:

$$\mathcal{C}_t^+ = \left\{ s' \in \mathcal{C}_t \mid \hat{r}(s') \geq \tau \right\}$$

For backpropagation:

$$V(s) \leftarrow \frac{N(s) \cdot V(s) + \sum \hat{r}(s')}{N(s) + |\mathcal{C}_t^+(s)|}$$

$N(s) \leftarrow N(s) + |\mathcal{C}_t^+(s)|$

For selection by UCB:

$$s_{t+1} = \arg\max_{s \in \mathcal{C}_t^+} \left[ V(s) + c \cdot \sqrt{ \frac{\log N(\mathrm{parent}(s))}{1 + N(s)} } \right]$$

(Zhu et al., 12 Jul 2025)

EduPRM is the process-aware reward model that provides the critique signal guiding this search. Its step-level labels are: Correct Step, Visual Misunderstanding, Problem Misunderstanding, Lack of Domain Knowledge, Misapplication of Knowledge, Logical Reasoning Error, Hallucination, Computational Error, and Off-topic or Incongruent (Zhu et al., 12 Jul 2025). Annotation is represented either as a quadruple {content, label, explanation, score} or as a triple {content, label, explanation} (Zhu et al., 12 Jul 2025).

The reward model is trained via curriculum learning on three supervision sources: “MCTS-guided trajectories (150K),” “Error-injected critiques (150K),” and “Teacher–student dialogues (120K)” (Zhu et al., 12 Jul 2025). The MCTS-guided trajectories provide verified multi-step solutions from structured search; the error-injected critiques insert one of the nine error types into steps; the teacher–student dialogues pair drafts from a 7B model with reviews from a 72B teacher (Zhu et al., 12 Jul 2025). The paper does not provide a closed-form PRM loss, instead describing “classification and explanation generation under curriculum learning” (Zhu et al., 12 Jul 2025).

A central operational feature is the self-reflection mechanism. It is defined as a dedicated action node where the model “reviews earlier steps, identifies inconsistencies or errors, and corrects them” (Zhu et al., 12 Jul 2025). Successful reflection leads to continuation on corrected branches, whereas failed verification causes the path to be discarded (Zhu et al., 12 Jul 2025). This suggests that reflective error correction is treated as a first-class search action rather than as a post hoc reranking heuristic.

5. Dataset composition and schema

EduMCTS-160K contains “160,000 trajectories” (Zhu et al., 12 Jul 2025). The paper does not report steps per trajectory or token counts (Zhu et al., 12 Jul 2025). Its modalities are multimodal, built from science problems requiring images or diagrams, and the pipeline removes questions “solvable without images” (Zhu et al., 12 Jul 2025).

The dataset’s schema can be summarized as follows:

Component	Reported content
Input	Question text; one or more images; LaTeX-encoded equations; JSON layout metadata
Process	Ordered node sequence with node_type, content, PRM label, PRM explanation, and score
Final output	Scalar or textual answer, sometimes with units

The node types are caption, summary, sub_task, thinking, self-reflection, and answer (Zhu et al., 12 Jul 2025). These correspond to a pedagogy-aligned workflow: “extract visual facts; abstract goals; decompose tasks; perform inference; reflect/verify; conclude” (Zhu et al., 12 Jul 2025).

Correctness is enforced by construction rather than estimated afterward. The retained trajectories have “all steps labeled ‘Correct Step’ per EduPRM,” with “additional verification via Self-Reflection” to ensure internal consistency (Zhu et al., 12 Jul 2025). This makes the corpus an explicitly filtered set of verified paths rather than a raw trace dump.

The paper includes representative examples that are “constructed to illustrate the schema; not verbatim from the paper” (Zhu et al., 12 Jul 2025). One example concerns a Physics velocity–time graph with steps from graph captioning through acceleration computation, self-checking of units and uniform acceleration, and final answer reporting; another concerns the area of a trapezoid with analogous decomposition and verification (Zhu et al., 12 Jul 2025). Since those examples are illustrative rather than source examples, they clarify structure rather than establish additional empirical properties.

6. Empirical behavior and reported effects

The paper reports several performance indicators associated with the EduMCTS construction process and with downstream training and inference using EduMCTS-160K (Zhu et al., 12 Jul 2025).

For search success rate, Table “successraterefined” gives:

• Vanilla MCTS: 67.8%
• • Stepwise Action Nodes: 73.3%
• • EduPRM Judge: 87.1% (Zhu et al., 12 Jul 2025)

The paper also states that, compared with “LLaVA-CoT-style approaches,” EduMCTS-160K trajectories achieved “an 18% higher success rate in science domains” (Zhu et al., 12 Jul 2025).

For Best-of-N scaling with EduPRM-based reranking, the reported accuracy values increase with larger sample counts:

• 41.21 at $N=2$
• 43.28 at $N=8$
• 43.45 at $N=12$ (Zhu et al., 12 Jul 2025)

For K12Vista, the paper reports that fine-tuning with EduMCTS-160K (“+ EduMCTS16W”) and BoN inference yields consistent gains for Qwen variants. For Qwen2.5VL-7B-Instruct, the step-by-step score changes from 27.17 at baseline to 39.50 with EduMCTS16W and 43.28 with EduMCTS16W + BoN ($N=8$); the corresponding direct inference scores improve from 39.82 to 44.03 to 45.94 (Zhu et al., 12 Jul 2025).

For cross-benchmark generalization, on MDK12, MathVision, and MMMU_Pro_V_COT, the reported Qwen2.5VL-7B average improves from 26.87 to 30.46 after EduMCTS-160K (Zhu et al., 12 Jul 2025). In ablation, progressively adding Data Selection, Stepwise Action Nodes, EduPRM Judge, and more rollouts increases step-by-step scores from 25.64 for naïve MCTS to 39.50 at $$\hat{r}(s') = \frac{1}{K} \sum_{j=1}^K \mathrm{PRM}_\psi\left( Yes \mid s' \right)$$0, with marginal further gains to 39.71 at $$\hat{r}(s') = \frac{1}{K} \sum_{j=1}^K \mathrm{PRM}_\psi\left( Yes \mid s' \right)$$1 (Zhu et al., 12 Jul 2025).

These results support the claim that the dataset is not merely a static resource but part of a training-and-inference regime centered on process-level supervision. A plausible implication is that EduMCTS-160K derives much of its utility from its coupling to EduPRM-based filtering and reranking, rather than from scale alone.

7. Novelty, intended use, and limitations

The paper characterizes EduMCTS-160K as novel in several respects: scale and multimodality, process granularity, construction methodology, unified use of EduPRM, and educational coverage (Zhu et al., 12 Jul 2025). Specifically, it is described as “160K multimodal science trajectories with images/diagrams, LaTeX equations, structured layouts,” with explicit node types, step-level critique metadata, and reflective Self-Reflection steps (Zhu et al., 12 Jul 2025). The construction relies on “domain-adapted MCTS with a process-aware reward model (EduPRM) guiding search,” augmented by “planning, decomposition, and reflection actions,” and filtered through self-consistency and rejection sampling (Zhu et al., 12 Jul 2025).

The intended use includes supervised fine-tuning for MLLMs, training or verifying PRMs and process-aware evaluators, and research on “test-time scaling (BoN) and search-based reasoning with reflective correction” (Zhu et al., 12 Jul 2025). The reported training recipe specifies supervised fine-tuning with AdamW, learning rate $$\hat{r}(s') = \frac{1}{K} \sum_{j=1}^K \mathrm{PRM}_\psi\left( Yes \mid s' \right)$$2, batch size 4, deterministic or near-deterministic inference with temperature 0, and Best-of-N sampling up to $$\hat{r}(s') = \frac{1}{K} \sum_{j=1}^K \mathrm{PRM}_\psi\left( Yes \mid s' \right)$$3–12 with temperature sampled from $$\hat{r}(s') = \frac{1}{K} \sum_{j=1}^K \mathrm{PRM}_\psi\left( Yes \mid s' \right)$$4 (Zhu et al., 12 Jul 2025).

Availability is stated succinctly: “Code, data, and models will be released,” while licensing is not specified (Zhu et al., 12 Jul 2025). The preprocessing stack includes Mathpix and LayoutLMv3 for images and layout, LaTeX equation encoding, JSON layout representation, and subject/grade-stratified sampling after difficulty and text-solvability filtering (Zhu et al., 12 Jul 2025).

The limitations noted in the paper include concentration on Qwen models and K-12 benchmarks, limited comparisons with more MCTS variants, and limited experiments at large $$\hat{r}(s') = \frac{1}{K} \sum_{j=1}^K \mathrm{PRM}_\psi\left( Yes \mid s' \right)$$5 in BoN (Zhu et al., 12 Jul 2025). Ethical considerations include anonymization and copyright or usage rights for scanned exams and online platform data, as well as possible biases from uneven subject distribution, grade-level emphasis, or cultural context; stratification is presented as mitigation rather than elimination (Zhu et al., 12 Jul 2025).

A potential misconception is to treat EduMCTS-160K as a generic reasoning dataset independent of its verifier. The reported construction instead makes EduPRM central at every stage: sample selection, rollout critique, pruning, and final Best-of-N verification (Zhu et al., 12 Jul 2025). Another misconception is to read it as a text-only chain-of-thought corpus; the pipeline removes text-only-solvable questions and is explicitly built around visual grounding (Zhu et al., 12 Jul 2025). Within the reported scope, EduMCTS-160K is best understood as a process-centric, verification-filtered, multimodal educational reasoning corpus produced by a unified search-and-critique framework rather than as a conventional collection of demonstrations.