Stepwise Verification and Remediation of Student Reasoning Errors
with Large Language Model Tutors
Abstract
Large language models (LLMs) present an opportunity to scale high-quality personalized education to all. A promising approach towards this means is to build dialog tutoring models that scaffold students’ problem-solving. However, even though existing LLMs perform well in solving reasoning questions, they struggle to precisely detect student’s errors and tailor their feedback to these errors. Inspired by real-world teaching practice where teachers identify student errors and customize their response based on them, we focus on verifying student solutions and show how grounding to such verification improves the overall quality of tutor response generation. We collect a dataset of 1K stepwise math reasoning chains with the first error step annotated by teachers. We show empirically that finding the mistake in a student solution is challenging for current models. We propose and evaluate several verifiers for detecting these errors. Using both automatic and human evaluation we show that the student solution verifiers steer the generation model towards highly targeted responses to student errors which are more often correct with less hallucinations compared to existing baselines.
Stepwise Verification and Remediation of Student Reasoning Errors
with Large Language Model Tutors
Nico Daheim∗1 Jakub Macina∗2,3 Manu Kapur4 Iryna Gurevych1 Mrinmaya Sachan2 1 Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science and Hessian Center for AI (hessian.AI), TU Darmstadt 2Department of Computer Science, ETH Zurich 3 ETH AI Center 4Professorship for Learning Sciences and Higher Education, ETH Zurich
![[Uncaptioned image]](extracted/5727215/figures/github.png){kind=small}
1 Introduction
The field of dialog tutoring aims to build systems that can teach students by holding a conversation with them (Wollny et al., 2021; Jurenka et al., 2024). Dialog tutors hold the potential to make personalized teaching available to learners anywhere anytime. The increasing capabilities of LLMs have brought renewed hope to this field (Thoppilan et al., 2022; Jurenka et al., 2024). However, real-time tutoring is quite complex, and human teachers bring various intricate capabilities when teaching, such as identifying student errors in problem solving, picking a pedagogical strategy, and communicating it (Wang et al., 2024c). The same requirements hold for dialog tutoring models which need all these abilities to be effective.
Yet, although research on effective human tutors shows they perform these steps sequentially by first reasoning about the error, then picking a strategy, and then responding (Lepper and Woolverton, 2002), many tutoring models perform all of them in one forward pass. Recent studies (Macina et al., 2023b, a) have shown that this can lead to several deficiencies that can be detrimental to student learning, for example, in math tutoring. Despite impressive performance on math reasoning benchmarks (Cobbe et al., 2021; Hendrycks et al., 2021), dialog tutors often generate hallucinated outputs and present erroneous information to students, for example, because they assess an incorrect solution as correct. We show an example of this in Figure 1.
In this paper, we alleviate this problem by decoupling the verification of student solutions from response generation with a modular approach. As opposed to the common approach, the model does not directly generate the tutor response from the students’ utterances, whereby solution assessment is done implicitly in the models’ activations, but rather receives the output of an additional verification model that assesses solutions and can therefore also be more specialized. We hypothesize that this increases the correctness of the model as well as makes the response more targeted to the error because the response generation module is already aware of the exact student error. Furthermore, this architecture more closely mimics human tutors.
To test our approach and train verifiers, we collect a dataset of ca. 1k student solutions and their stepwise reasoning chains in the domain of multi-step math problem-solving, which will be released publicly. This dataset augments the math dialog tutoring corpus MathDial (Macina et al., 2023a), which we use for evaluating dialog tutoring models, by teacher-annotated verifications of the first erroneous step in the student solution (Section 4).
We propose three verification approaches based on prompting and finetuning language models. Besides a simple classification-based approach for verification, we also generate a textual verification and notably align student solution steps to steps of a reference solution (Section 3.1) to verify the student solution. We find that using our data for finetuning helps smaller LLMs surpass prompted state-of-the-art LLMs. Furthermore, incorporating the verification output in the response generation step (Section 3.2) clearly improves their performance in terms of both extensive automatic (Section 6) and human evaluation using real teachers (Section 6.3): the generated responses are more targeted to the exact student error, there are less hallucinations, and there is more actionable scaffolding feedback for the student. In general, we find that such improvements are much stronger when the verification output is correct (Section 7.1) indicating a large potential for the community to improve dialog tutors by adding verifiers.
2 Background & Related Work
Dialog tutoring aims at building models that can tutor human students through a conversation. For example, dialog tutoring has been proposed for second-language acquisition (Stasaski et al., 2020; Caines et al., 2020; Kwon et al., 2024), to answer questions in science (Chevalier et al., 2024), or to solve math word problems (MWPs) (Macina et al., 2023a). In each case, the model should guide the learner to solving a problem (e.g. the MWP or translation of a phrase) not by telling the solution outright, but rather by using scaffolding techniques that give students space for guided exploration and self-correction. For example, the tutor might elicit the students’ thinking by asking a question that challenges their understanding of the problem (Reiser, 2004; Anghileri, 2006).
Capturing such intricate tutoring strategies is hard and requires teachers years to master. Due to this complexity, most previous dialog tutoring systems were human-authored, notably the AutoTutor family (Nye et al., 2014), LISP tutor (Anderson et al., 1985) which uses a large set of rules to verify student programming solutions, or any systems built using CTAT which requires enumerating all possible solutions or writing complex production rules (Aleven et al., 2016). However, scaling such human-authored systems can quickly explode in both complexity and human effort (Macina et al., 2023b). Due to this and rapid progress in language generation from learning large models based on large amounts of data, LLMs such as LearnLM (Jurenka et al., 2024) have recently become popular in favor of human-authored, rule-based systems.
Problem formulation
Formally, the goal of dialog tutoring is to continue a tutoring dialog consisting of a sequence of turns taken by either student or teacher and where is constructed from a fixed model vocabulary . Continuation then means generating a new utterance that follows the above principles. Often there is also background knowledge that is either required or helpful to tutor a certain concept, such as grammar rules (Stasaski et al., 2020), or textbook knowledge (Wang et al., 2024b), and can be used to ground . In this work, we deal with teaching math word problem-solving and therefore use textual knowledge , for instance, the math word problem and background knowledge like math rules.
Tutor models
Using such data, a simple approach is generating the teacher response directly by pairing the following model, parameterized by weights , with a decoding algorithm, such as beam search or greedy decoding:
| (1) |
This model is straightforward to implement and learn from data but prior work has shown that it suffers from generating factually incorrect outputs (Macina et al., 2023a). Therefore, in this work we break down response generation into two steps: verification, where it is first assessed whether the student solution is correct, and generation.
Verification
Verification is challenging in its own right and has recently been tackled for general reasoning problems. For example, ROSCOE (Golovneva et al., 2023) presents unsupervised metrics to assess the correctness of a models’ chain-of-thought (CoT) reasoning, and (Jacovi et al., 2024) evaluate open-domain question answering for logical errors. The outputs of verifiers have subsequently also been used for self-refinement of LLMs (Madaan et al., 2023; Shinn et al., 2024) and also allow targeted feedback for the training of student LLMs with teacher LLMs (Saha et al., 2023; Wang et al., 2024a). Closely related to our work (Wang et al., 2024c) define broad error categories, such as miscalculation, to understand the cause of incorrect reasoning by students and condition on it to generate teacher responses. We compare to this baseline and call it Error Reason.
3 Verification-based Response Generation
We first introduce the task of verification and different verifiers in Section 3.1. Afterwards, in Section 3.2, we combine verification and response generation for modular tutor response generation.
3.1 Verification
We deal with the verification of student solutions to a given math word problem . The solutions can be described by a sequence of substep solutions , where each and is the final solution. Usually, is described by the student in one of the student utterances . The task of the model is to assess whether is the correct solution to and if not, potentially, to identify which step caused the error. Oftentimes, this can be done by comparing to a reference solution that is either given or model-generated and might differ in length. All verifiers which we discuss next can then be described by a learned function , usually an LLM. Here, is the verification output and may be an empty string if no reference solution is given. In the following, we introduce different verifiers.
Classification-based Verification
A comparably simple approach to verification is classifying whether the student solution is correct using a binary classifier. We call this Overall Verification. Similarly, identifying the first error step can be framed as multi-class classification with labels , where means no mistake. We call this Stepwise Verification. Alternatively, Stepwise Verification (iterative) can be framed as a binary classification for each step whether it is correct. The first error step is the first step classified as error.
Error Description
While conceptually easy, classification-based approaches locate the first error without explaining the exact issue. Therefore, we propose to use an LLM to directly describe the error, and the concrete first error step, in a textual format, and call this Error Description. For this, we prompt the LLM with the prompt outlined in Appendix G. In comparison to Wang et al. (2024c), this error description is allowed to be free-form and does not map to predefined error types. The LLM-generated error step description can then be passed to a tutor response generation model.
(Step) Alignment
As our third verification approach, we align the steps in the student’s solution with a reference solution, and compare the steps in the student and reference solution to localize errors. We call this approach Step Alignment. As the order of steps in the solutions matters, a greedy algorithm that finds the most similar steps across the two solutions is insufficient. Thus, we frame verification as a sequence alignment problem.
The input to the alignment algorithm is the student solution with steps and the reference solution with steps. Note that here we are aligning solution steps which can be long strings. This is different from other sequence alignment problems in NLP, where typically tokens are aligned (Paolini et al., 2021, inter alia). The output is a sequence of tuples of length , where each and can be either a step of and , respectively, or a special symbol . Here, indicates a missing step in the student solution () and indicates an additional step ().
In our implementation, we use the Needleman-Wunsch (NW) algorithm (Needleman and Wunsch, 1970) as it guarantees an optimal alignment with respect to a chosen cost function. We use a modified version of the algorithm for semantic sequence alignment and use sentence embeddings (Reimers and Gurevych, 2019) to measure the similarity between steps. We detail our adaptation of the NW algorithm in Algorithm 1 and describe each step in the following. The NW algorithm iterates over all possible pairs of substeps from and and calculates a cost for each pair. Since each substep is a string, we use semantic string similarity measured by the cosine similarity of the contextual embeddings of the substeps. In our experiments, this performed better than just matching the final numerical solution of the substeps (cf. Section 7.1 for results and a comparison of embedding models). As not all high sentence embedding scores indicate a significant match, we introduce a threshold to differentiate between exact and near matches. If the similarity is higher than a threshold , the pair is deemed as an exact match and its similarity is added to the similarity of their predecessors. If it is smaller than it could still be a near match if the sequence similarity is high enough after incurring a penalty (). The last option is a gap if the sum is larger than adding a predefined gap cost to either a pair of the previous student and current reference solution step, or a pair of the current student and previous reference solution step. Altogether, this forms a similarity matrix of size . The alignment is finally found by backtracking (moving only to adjacent entries with each step) from entry .
Similar to the classification-based approach, the alignment output can not directly be used in a response generation model but has to be converted to a formatted verification output string. For this, we use a preformatted template shown in Appendix G. The template groups together the missing, additional and matching steps to produce from the alignment produced by the NW algorithm.
3.2 Response Generation
Direct generation of tutor responses can be challenging because one model has to reason over the student solution, pick a teaching strategy, and generate a response in one step. This has been shown to produce hallucinations (Macina et al., 2023a). We tackle this by incorporating an additional verification step that informs the response generation model, as previously discussed. Our aim is to split the task into two less complex tasks which should reduce errors if each task can be performed well enough and has been shown to reduce hallucinations in document-grounded dialog (Adolphs et al., 2022) and question answering (Press et al., 2023).
The verifier and response generation model are combined in a two-stage approach. First, the verifier outputs a verification of the student solution based on a reference solution . Then, the response generation model is conditioned on , the dialog history , and background knowledge . In our work, consists of the student solution , optionally the reference solution , and the math word problem . If is a distribution over verification labels, the overall model is:
| (2) |
The full model provides us with a verification output and the generated response which makes the internal reasoning of the tutor model in terms of student errors more explicit and controllable.
4 Data Collection
We propose and evaluate various verifiers in this work. Since some of them require training data and to evaluate their performance, we collect a dataset of 1,002 human-produced verification outputs to train and evaluate them. This is similar in size to a related corpora (Jacovi et al., 2024). In this section, we describe the annotation task and data collection.
Incorrect Student Solutions Source
Our work extends MathDial (Macina et al., 2023a) by having teachers annotate incorrect student solutions from the dataset with their first error step. There, these incorrect student solutions were used to condition a student model (InstructGPT) to generate responses in a dialogue with a human teacher.
Specifically, these problems are based on the GSM8k (Cobbe et al., 2021) dataset of multi-step math word problems. In MathDial, the reasoning chains are generated using a 2-shot CoT prompt with gpt-3.5-turbo, and temperature sampling () is used to get multiple reasoning paths (). Finally, the most common incorrect solution is chosen. Subsequently, their student model is prompted to respond to a human teacher as a student who tries to solve a problem with a particular incorrect solution.
To not skew our dataset to errors, we balance it with rephrased reference solutions from the student model. We reproduce the student model prompt from MathDial to generate student responses using the reference solutions. All reference solutions and student responses with incorrect solutions are part of the dataset. Details are in Appendix A.
Student Solution Annotation
The objective of the annotation is to mark the exact step of the first error in the student solution. We do not annotate error steps after the first one to decrease ambiguity, as they frequently stem from the first error. We recruit teachers through Prolific, who first read the problem and then mark the precise step of the first error in the student solution. Teachers can access the reference solution to reduce task complexity. Details of the task, the user interface, and examples of collected data are in Appendix A. To compute agreement, 10% of the samples are annotated by one additional annotator with an inter-rater reliability of Cohen’s indicating substantial agreement (Cohen, 1960). We show the distribution of incorrect student solution steps in Figure 3.
5 Experiments
We evaluate different verifiers on our dataset and use them to inform response generation models to improve their correctness. Since we extend MathDial with additional annotations we use MathDial dialogues for evaluating tutor response generation. Besides math problem and student solution in a dialog, we either use a model-generated CoT reference solution if marked by “solution” or no reference solution as input to the models. Next, we detail metrics and models.
5.1 Metrics
For teacher response generation, we evaluate the generated output of each model by comparing it to a human-annotated response from MathDial. We report standard text generation metrics: the sacrebleu (Post, 2018) implementation of BLEU (sBLEU) to measure word overlap and BERTScore (Zhang et al., 2020) (BF1, using the all-MiniLM-L6-v2 checkpoint) to measure semantic similarity. Moreover, we report the knowledge F1 (KF1) score with respect to the grounding information (correct solution in the case of MathDial) which has been used as a proxy for faithfulness in prior work (Daheim et al., 2024). Similar to (Zheng et al., 2024; Jurenka et al., 2024), we also prompt LLAMA3-70B and use it complementary to human evaluation (the same task and instructions are used in both) to assess how targeted, correct, and actionable a response is. Details about the LLM-based evaluation are found in Appendix E.
5.2 Models
For both verification and response generation, we use different prompted or finetuned models. For verification, we compare the closed-source model GPT-3.5 to the open models LLAMA2 and LLAMA3. For the latter, we prompt the 70B version of the models and finetune LLAMA2-7B using LoRA. For response generation, we evaluate prompting GPT-3.5 and finetuning the encoder-decoder model Flan-T5 with 3B parameters. We finetune Flan-T5 again using LoRA for both the direct modeling and verify-then-generate approach.
| Model | Overall Verification | Stepwise | ||
|---|---|---|---|---|
| Corr. F1 | Err. F1 | F1 | micro F1 | |
| Few-shot | ||||
| GPT3.5 | 0.66 | 0.52 | 0.59 | 0.42 |
| + solution | 0.97 | 0.97 | 0.97 | 0.61 |
| Llama2-70B | 0.69 | 0.38 | 0.54 | 0.17 |
| + solution | 0.78 | 0.59 | 0.68 | 0.48 |
| Llama3-70B | 0.74 | 0.58 | 0.66 | 0.56 |
| + solution | 0.90 | 0.87 | 0.89 | 0.70 |
| Finetuned | ||||
| Llama2-7B | 0.89 | 0.67 | 0.78 | 0.20 |
| + solution | 0.81 | 0.80 | 0.80 | 0.28 |
| Automatic Metrics | LLM Judge (%) | ||||||
| Model | Variant | sBLEU | KF1 | BF1 | Targeted | Correct | Actionable |
| - | Human | 100.0 | 100.0 | 100.0 | 27 | 82 | 87 |
| Few-shot | |||||||
| GPT-3.5 | Baseline | 2.0 | 27.0 | 51.2 | 29 | 37 | 27 |
| Error Reason | 1.5 | 22.5 | 46.7 | 34 | 40 | 56 | |
| Error Description | 2.8 | 30.3 | 52.6 | 62 | 66 | 45 | |
| Step Alignment | 2.3 | 29.8 | 53.3 | 42 | 61 | 26 | |
| Finetuned | |||||||
| Flan-T5-3B | Baseline | 2.6 | 27.6 | 56.0 | 1 | 89 | 76 |
| Error Description | 3.0 | 26.7 | 56.0 | 2 | 92 | 84 | |
6 Results
We first show the performance of different verification models in Section 6.1 and then use verification models in response generation in Section 6.2.
6.1 Verification
In this section, we benchmark LLMs on their ability to evaluate the correctness of student solutions using the Overall Verification and Stepwise Verification approaches from Section 3.1. For Stepwise Verification we use the multi-class classification approach, because it performed better than iterative classification in our experiments. A comparison is found in Appendix B. We measure the F1 score (balanced dataset), in particular, micro F1 for Stepwise Verification (imbalanced dataset, see Figure 3). We find in Table 1 that Overall Verification can be challenging even for state-of-the-art LLMs. All prompted models show comparably low performance when prompted without a reference solution and especially struggle with identifying incorrect responses. Providing a reference solution improves results significantly. However, for Stepwise Verification even the reference solution does not improve micro F1 beyond 0.70. This result is consistent with expert educator-based assessment (Yen and Hsu, 2023) and LLM self-correction results (Huang et al., 2024).
Interestingly, our dataset can be used effectively for finetuning. Even a smaller LLAMA2-7B model can outperform larger prompted models on Overall Verification, especially when no solution is provided. Potentially, the additional finetuning steps make it easier for the model to also solve the problem before verification. The finetuned Stepwise Verification model outperforms its larger prompted counterpart LLAMA2-70B when no solution is provided. Results for finetuning show a ten-fold cross-validation. Further details are in Appendix H.
6.2 Response Generation
Next, we show in Table 2 that combining verification and tutor response generation models can improve the quality of the generated responses. We compare the Error Description and Step Alignment verifiers to direct response generation and using the Error Reason (Wang et al., 2024c). There, the error is categorized into either: guess, misinterpret, right-idea, imprecise, not sure, or careless. We use a subset of MathDial, where the student describes their solution to the teacher in the dialog, and generate the following teacher utterance.
First, we prompt GPT-3.5 using the prompt templates from Section 3.1 for comparability. We find that providing only the Error Reason does not improve over the direct baseline in simpler automatic metrics (sBLEU, KF1, BF1) but only in terms of the LLM-based judging. Using the more detailed Error Description which provides the exact mistake of the student gives larger improvements, both in terms of automatic metrics and LLM-based judging. Similarly, we find Step Alignment to be helpful, but to provide less actionable responses. When finetuning with the Error Description, we obtain improvements over the finetuned baseline but they are smaller and do not hold for each metric.
Our qualitative analysis shows that both Step Alignment and Error Description result in responses that better localize the exact student error. For example, the baseline often assesses the solution wrongly or skips the first error step and instead asks for the solution of a later step. Examples are shown in Table 10 and Table 11. Section 6.3 confirms our results by human evaluation.
| Model | Variant | target | corr | act |
|---|---|---|---|---|
| Verification | ||||
| GPT3.5 | Error Description | - | 70.6 | - |
| Llama3-70B | Error Description | - | 82.4 | - |
| Response Generation | ||||
| - | Human | 35.0 | 45.0 | 42.5 |
| Few-shot | ||||
| GPT3.5 | Baseline | 30.0 | 37.5 | 30.0 |
| Error Reason | 27.5 | 22.5 | 37.5 | |
| Error Description | 57.5 | 62.5 | 45.0 | |
| Step Alignment | 57.5 | 60.0 | 27.5 | |
| Finetuned | ||||
| Flan-T5-3B | Baseline | 7.5 | 20.0 | 25.0 |
| Error Description | 20.0 | 35.0 | 35.0 | |
6.3 Human Evaluation
We conduct a human evaluation using teachers as expert annotators. All annotators are recruited on Prolific after manual screening. We assess whether the generated responses are targeted, correct, and actionable without outright telling the solution. Annotators are blind to the model source. The exact questions are as follows. 1) (Targeted (T)) Does the Teacher point out the root cause of the student’s mistake? 2) (Correctness (C)) Is the Teacher’s response factually correct with respect to the reference solution? 3) (Actionable (A)) - Does the Teacher provide actionable steps to let the Student correct the mistake without giving away the full answer? More details are in Appendix C.
Responses from 6 models and one human response from MathDial were annotated for a random set of 40 conversations. To compute inter-rater reliability, 9 conversations were annotated with at least 2 raters for each response. Cohen’s kappa is 0.21, 0.25, and 0.13 for targeted, correctness, and actionable. For Error Description correctness it is . Next, we describe the results, first on verification and then for response generation.
Verification
Annotators assess the Error Description as correct if the exact mistake of the student is found and incorrect when the model says that the solution is correct when it is not and vice versa, misses the step of the error, or is generic. Results in Table 3 show that LLAMA3-70B outperforms GPT-3.5 but also with of the errors being found correctly.
Response Generation
Next, in Table 3, we evaluate how targeted, correct, and actionable the responses generated by different models are. We find that providing the Error Reason improves over the baseline only in terms of how actionable responses are. We hypothesize that conditioning on only the reason is insufficient for a targeted response. Error Description and Step Alignment provide more information regarding the exact error and therefore improve strongly over the baseline in both targetedness and correctness. Using Step Alignment information also does not improve actionability but Error Description improves it. The same improvements also hold for using the Error Description for a finetuned model. All in all, we find strong evidence that using our verify-then-generate approach improves teacher response generation.
7 Ablation Studies
Next, we provide further ablations, first on the cost function used in the NW algorithm (Section 7.1) and then on the impact of verification before response generation based on its correctness and problem difficulty (Section 7.2).
7.1 Alignment
| Step Alignment | Accuracy(%) | ||
|---|---|---|---|
| Random | 0.6 | -0.2 | 43.6 |
| Solution Match | 0.6 | -0.3 | 51.9 |
| SBERT | 0.8 | -0.1 | 58.2 |
| Roscoe | 0.9 | -0.1 | 61.4 |
We compare different cost functions used for the NW Step Alignment algorithm in Table 4. For the comparison, 30 alignments between a student and reference solution were produced by humans and the accuracy of student solution step alignment is measured. As cost functions we use the cosine similarity of Sentence-BERT (SBERT) (Reimers and Gurevych, 2019) embeddings and embeddings from a model trained on Roscoe (Golovneva et al., 2023), as well as a random cost and an indicator function that is when two substeps have the same numerical solution and otherwise. Similarity threshold and gap cost are optimized via a hyperparameter grid search, as indicated by and . We find that cosine similarity works best and training on relevant math data fine-tuning in Roscoe further improves performance.
7.2 Verification
Verification correctness is important
| Model | Error Desc. | T | C | A |
|---|---|---|---|---|
| Few-shot | ||||
| GPT3.5 | incorrect | 50.0 | 62.5 | 50.0 |
| correct | 62.5 | 62.5 | 41.7 | |
| Llama3-70B | incorrect | 41.7 | 33.3 | 25.0 |
| correct | 82.1 | 71.4 | 21.4 | |
| Finetuned | ||||
| Flan-T5-3B | incorrect | 08.3 | 25.0 | 41.7 |
| correct | 25.0 | 39.3 | 32.1 | |
We find in Table 5 that correct verification is important for subsequent response generation. If it is correct, both targetedness and correctness are strongly improved over when it is incorrect. However, actionability appears to be decreased which indicates less scaffolding and more teacher "telling the solution".
Problem difficulty influences verification
| Model | Steps | T | C | A | Error Desc. |
|---|---|---|---|---|---|
| Few-shot | |||||
| GPT3.5 | 3 | 62.5 | 75.0 | 37.5 | 50.0 |
| 4 | 61.1 | 55.6 | 55.6 | 72.2 | |
| 5 | 50.0 | 64.3 | 50.0 | 50.0 | |
| Llama3-70B | 3 | 87.5 | 75.0 | 12.5 | 87.5 |
| 4 | 72.2 | 72.2 | 72.2 | 72.2 | |
| 5 | 57.1 | 35.7 | 57.1 | 50.0 | |
| Finetuned | |||||
| Flan-T5-3B | 3 | 12.5 | 25.0 | 25.0 | 87.5 |
| 4 | 27.8 | 50.0 | 50.0 | 72.2 | |
| 5 | 14.3 | 21.4 | 14.3 | 50.0 | |
Finally, we show in Table 6 that the performance of both verification and our verify-then-generate approach is heavily correlated with the number of reasoning steps that are used in the reference solution of a given math word problem. We use this as a proxy for problem difficulty. First of all, the performance of the LLAMA3 70B Error Description decreases with the number of steps. This is reflected in the decreased correctness and targetedness of the responses of the few-shot prompted LLAMA3 model. For GPT-3.5 we do not find a similar conclusion for the Error Description model but at least targetedness still decreases with the number of steps. For the finetuned model we do not see similar trends but instead find the best performance for problems with four steps, likely because these are more common in the training data.
8 Discussion & Conclusion
Student errors are key learning opportunities. Tutors should recognize them and precisely guide students with targeted feedback without telling full solution. Motivated by effective teaching practice, we split the task of tutor response generation into two separate steps of verifying the student solution and generating a response.
To evaluate our approach, we collect a dataset of around 1k teacher-annotated solutions to augment an existing math tutoring corpus. Our results show that splitting response generation into two steps can result in more targeted and correct responses that better scaffold human learning. We showcase this using both automatic evaluation and human evaluation annotated by teachers, both for prompted and finetuned models.
9 Limitations
Focus on scaffolding problem-solving
The tutoring scenarios which are considered are centered around the student problem-solving stage. In this case, students have prior knowledge. mostly understand the learning topics and practice them. However, different learning scenarios such as direct instruction, building rapport with students, or open-ended discussions are not considered in this work.
Evaluating student solutions and responding appropriately to a student’s mistakes is inherently challenging, even for human teachers. Furthermore, teachers should ideally give adaptive feedback depending on the problem-solving strategy chosen by the student and treat different errors in different ways to uncover any misconceptions (Nye et al., 2014). For example, in math, productive errors present important learning opportunities for students to learn from them (Kapur, 2016; Shaughnessy et al., 2021; Sinha and Kapur, 2021), e.g. by teacher-guided self-correction or targeted instruction, while unproductive errors, such as numerical miscalculations, could be easily resolved using a calculator (Lepper and Woolverton, 2002).
Difficulty of obtaining student reasoning chains
Model-generated reasoning chains might contain the same biases as human students (Opedal et al., 2024). On the other hand, there might be many additional differences from human student reasoning, e.g. students might not always stick to exact math notations or skip some steps in the explanations. However, because such data from students is sensitive, we work with model-generated reasoning solutions and responses.
Focus on multi-step problems
Procedural or multi-step problems are the basis of most of the scientific disciplines, therefore we believe our approach should be general enough to work across any science subject, especially by including retrieval-augmented generation (RAG) from textbooks. However, it is still an open research question whether a similar solution would work for language learning or fact-based problems, and how models perform in languages other than English.
Evaluation is teacher-centered and complemented with an LLM-judge
Future work should focus on student user studies with AI tutors. However, this requires careful experimental consideration and safety mechanisms. Moreover, assessment of the responses is done exclusively by teachers and therefore future work should consider running assessments of the responses by students.
10 Acknowledgements
This research work has been funded by a Swiss National Science Foundation award (#201009), a Responsible AI grant by the Haslerstiftung, by the German Federal Ministry of Education and Research and the Hessian Ministry of Higher Education, Research, Science and the Arts within their joint support of the National Research Center for Applied Cybersecurity ATHENE. Nico Daheim acknowledges travel support from ELISE (GA no 951847). Jakub Macina acknowledges funding from the ETH AI Center Doctoral Fellowship, Asuera Stiftung, and the ETH Zurich Foundation. We thank Sankalan Pal Chowdhury, Kumar Shridhar, Shehzaad Dhuliawala, and Justus-Jonas Erker for valuable feedback and discussions.
11 Ethics Statement
Intended usage
The benefits of our dataset are in understanding and designing AI technology to assist teachers and students during the problem-solving stage. Most importantly, the goal of such systems is to not replace human teachers, but rather enhance their capabilities and make them focus on important and human aspects of teaching. We will release the dataset under CC-BY-4.0 license 111https://creativecommons.org/licenses/by/4.0/deed.en for further usage and exploration by the community. This also adheres to the licensing of MathDial, which we extend.
Data Anonymization and Privacy
As the data in education are strictly confidential we obtained approval on the proposal222The study was approved by the ETH Zurich Ethics Committee (IRB) under EK-2024-N-97. of the collection interface, questions and how long the data will be stored. All participants fill informed consent at the beginning of the annotation and may withdraw without reason at any time. We store only the necessary data and do not store any personally identifiable information. The collected data are stored anonymously and securely. Moreover, no student data are used in this work.
Accessibility and Potential Misuse
Our work focus on addressing hallucinations of LLM Tutors and their generation of factually incorrect responses. This directly addresses one of the important aspects of responsible use of AI which does not spread incorrect information, especially in education. We encourage the community to work on this important topic by open-sourcing the dataset, the code for running the benchmarks, and the methods used in this work. These are primarily intended for research purposes. As with any AI technology, the methods and dataset could be misused. However, we believe by open-sourcing the work we inform about the risks and capabilities of the technology a wider research community which then leads to further improvements.
References
- Adolphs et al. (2022) Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2022. Reason first, then respond: Modular generation for knowledge-infused dialogue. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 7112–7132, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
- Aleven et al. (2016) Vincent Aleven, Bruce M McLaren, Jonathan Sewall, Martin Van Velsen, Octav Popescu, Sandra Demi, Michael Ringenberg, and Kenneth R Koedinger. 2016. Example-tracing tutors: Intelligent tutor development for non-programmers. International Journal of Artificial Intelligence in Education, 26:224–269.
- Anderson et al. (1985) John R Anderson, C Franklin Boyle, and Brian J Reiser. 1985. Intelligent tutoring systems. Science, 228(4698):456–462.
- Anghileri (2006) Julia Anghileri. 2006. Scaffolding practices that enhance mathematics learning. Journal of Mathematics Teacher Education, 9:33–52.
- Caines et al. (2020) Andrew Caines, Helen Yannakoudakis, Helena Edmondson, Helen Allen, Pascual Pérez-Paredes, Bill Byrne, and Paula Buttery. 2020. The teacher-student chatroom corpus. In Proceedings of the 9th Workshop on NLP for Computer Assisted Language Learning, pages 10–20.
- Chevalier et al. (2024) Alexis Chevalier, Jiayi Geng, Alexander Wettig, Howard Chen, Sebastian Mizera, Toni Annala, Max Jameson Aragon, Arturo Rodríguez Fanlo, Simon Frieder, Simon Machado, et al. 2024. Language models as science tutors. arXiv preprint arXiv:2402.11111.
- Cobbe et al. (2021) Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
- Cohen (1960) Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46.
- Daheim et al. (2024) Nico Daheim, Nouha Dziri, Mrinmaya Sachan, Iryna Gurevych, and Edoardo Ponti. 2024. Elastic weight removal for faithful and abstractive dialogue generation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7096–7112, Mexico City, Mexico. Association for Computational Linguistics.
- Golovneva et al. (2023) Olga Golovneva, Moya Peng Chen, Spencer Poff, Martin Corredor, Luke Zettlemoyer, Maryam Fazel-Zarandi, and Asli Celikyilmaz. 2023. ROSCOE: A suite of metrics for scoring step-by-step reasoning. In The Eleventh International Conference on Learning Representations.
- Hendrycks et al. (2021) Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. NeurIPS.
- Huang et al. (2024) Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large language models cannot self-correct reasoning yet. In The Twelfth International Conference on Learning Representations.
- Jacovi et al. (2024) Alon Jacovi, Yonatan Bitton, Bernd Bohnet, Jonathan Herzig, Or Honovich, Michael Tseng, Michael Collins, Roee Aharoni, and Mor Geva. 2024. A chain-of-thought is as strong as its weakest link: A benchmark for verifiers of reasoning chains. arXiv preprint arXiv:2402.00559.
- Jurenka et al. (2024) Irina Jurenka, Markus Kunesch, Kevin McKee, Daniel Gillick, Shaojian Zhu, Sara Wiltberger, Shubham Milind Phal, Katherine Hermann, Daniel Kasenberg, Avishkar Bhoopchand, Ankit Anand, Miruna Pîslar, Stephanie Chan, Lisa Wang, Jennifer She, Parsa Mahmoudieh, Aliya Rysbek, Wei-Jen Ko, Andrea Huber, Brett Wiltshire, Gal Elidan, Roni Rabin, Jasmin Rubinovitz, Amit Pitaru, Mac McAllister, Julia Wilkowski, David Choi, Roee Engelberg, Lidan Hackmon, Adva Levin, Rachel Griffin, Michael Sears, Filip Bar, Mia Mesar, Mana Jabbour, Arslan Chaudhry, James Cohan, Sridhar Thiagarajan, Nir Levine, Ben Brown, Dilan Gorur, Svetlana Grant, Rachel Hashimoshoni, Laura Weidinger, Jieru Hu, Dawn Chen, Kuba Dolecki, Canfer Akbulut, Maxwell Bileschi, Laura Culp, Wen-Xin Dong, Nahema Marchal, Kelsie Van Deman, Hema Bajaj Misra, Michael Duah, Moran Ambar, Avi Caciularu, Sandra Lefdal, Chris Summerfield, James An, Pierre-Alexandre Kamienny, Abhinit Mohdi, Theofilos Strinopoulous, Annie Hale, Wayne Anderson, Luis C. Cobo, Niv Efron, Muktha Ananda, Shakir Mohamed, Maureen Heymans, Zoubin Ghahramani, Yossi Matias, Ben Gomes, and Lila Ibrahim. 2024. Towards responsible development of generative ai for education: An evaluation-driven approach. Technical report, Google DeepMind.
- Kapur (2016) Manu Kapur. 2016. Examining productive failure, productive success, unproductive failure, and unproductive success in learning. Educational Psychologist, 51(2):289–299.
- Kwon et al. (2024) Soonwoo Kwon, Sojung Kim, Minju Park, Seunghyun Lee, and Kyuseok Kim. 2024. Biped: Pedagogically informed tutoring system for esl education. arXiv preprint arXiv:2406.03486.
- Lepper and Woolverton (2002) Mark R. Lepper and Maria Woolverton. 2002. Chapter 7 - the wisdom of practice: Lessons learned from the study of highly effective tutors. In Joshua Aronson, editor, Improving Academic Achievement, Educational Psychology, pages 135–158. Academic Press, San Diego.
- Macina et al. (2023a) Jakub Macina, Nico Daheim, Sankalan Chowdhury, Tanmay Sinha, Manu Kapur, Iryna Gurevych, and Mrinmaya Sachan. 2023a. MathDial: A dialogue tutoring dataset with rich pedagogical properties grounded in math reasoning problems. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5602–5621, Singapore. Association for Computational Linguistics.
- Macina et al. (2023b) Jakub Macina, Nico Daheim, Lingzhi Wang, Tanmay Sinha, Manu Kapur, Iryna Gurevych, and Mrinmaya Sachan. 2023b. Opportunities and challenges in neural dialog tutoring. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2357–2372, Dubrovnik, Croatia. Association for Computational Linguistics.
- Madaan et al. (2023) Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-refine: Iterative refinement with self-feedback. In Thirty-seventh Conference on Neural Information Processing Systems.
- Needleman and Wunsch (1970) Saul B Needleman and Christian D Wunsch. 1970. A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology, 48(3):443–453.
- Nye et al. (2014) Benjamin D Nye, Arthur C Graesser, and Xiangen Hu. 2014. Autotutor and family: A review of 17 years of natural language tutoring. International Journal of Artificial Intelligence in Education, 24:427–469.
- Opedal et al. (2024) Andreas Opedal, Alessandro Stolfo, Haruki Shirakami, Ying Jiao, Ryan Cotterell, Bernhard Schölkopf, Abulhair Saparov, and Mrinmaya Sachan. 2024. Do language models exhibit the same cognitive biases in problem solving as human learners? In Forty-first International Conference on Machine Learning.
- Paolini et al. (2021) Giovanni Paolini, Ben Athiwaratkun, Jason Krone, Jie Ma, Alessandro Achille, Rishita Anubhai, Cicero Nogueira dos Santos, Bing Xiang, and Stefano Soatto. 2021. Structured prediction as translation between augmented natural languages. In International Conference on Learning Representations.
- Post (2018) Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186–191, Brussels, Belgium. Association for Computational Linguistics.
- Press et al. (2023) Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5687–5711, Singapore. Association for Computational Linguistics.
- Reimers and Gurevych (2019) Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.
- Reiser (2004) Brian J. Reiser. 2004. Scaffolding complex learning: The mechanisms of structuring and problematizing student work. Journal of the Learning Sciences, 13(3):273–304.
- Saha et al. (2023) Swarnadeep Saha, Peter Hase, and Mohit Bansal. 2023. Can language models teach? teacher explanations improve student performance via personalization. In Thirty-seventh Conference on Neural Information Processing Systems.
- Shaughnessy et al. (2021) Meghan Shaughnessy, Rosalie DeFino, Erin Pfaff, and Merrie Blunk. 2021. I think i made a mistake: How do prospective teachers elicit the thinking of a student who has made a mistake? Journal of Mathematics Teacher Education, 24:335–359.
- Shinn et al. (2024) Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36.
- Sinha and Kapur (2021) Tanmay Sinha and Manu Kapur. 2021. When problem solving followed by instruction works: Evidence for productive failure. Review of Educational Research, 91(5):761–798.
- Stasaski et al. (2020) Katherine Stasaski, Kimberly Kao, and Marti A. Hearst. 2020. CIMA: A large open access dialogue dataset for tutoring. In Proceedings of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 52–64, Seattle, WA, USA → Online. Association for Computational Linguistics.
- Thoppilan et al. (2022) Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. 2022. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239.
- Wang et al. (2024a) Haorui Wang, Rongzhi Zhang, Yinghao Li, Lingkai Kong, Yuchen Zhuang, Xiusi Chen, and Chao Zhang. 2024a. Tpd: Enhancing student language model reasoning via principle discovery and guidance. arXiv preprint arXiv:2401.13849.
- Wang et al. (2024b) Junling Wang, Jakub Macina, Nico Daheim, Sankalan Pal Chowdhury, and Mrinmaya Sachan. 2024b. Book2dial: Generating teacher-student interactions from textbooks for cost-effective development of educational chatbots. arXiv preprint arXiv:2403.03307.
- Wang et al. (2024c) Rose E. Wang, Qingyang Zhang, Carly Robinson, Susanna Loeb, and Dorottya Demszky. 2024c. Bridging the novice-expert gap via models of decision-making: A case study on remediating math mistakes. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics. Association for Computational Linguistics.
- Wolf et al. (2020) Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics.
- Wollny et al. (2021) Sebastian Wollny, Jan Schneider, Daniele Di Mitri, Joshua Weidlich, Marc Rittberger, and Hendrik Drachsler. 2021. Are we there yet?-a systematic literature review on chatbots in education. Frontiers in artificial intelligence, 4.
- Yen and Hsu (2023) An-Zi Yen and Wei-Ling Hsu. 2023. Three questions concerning the use of large language models to facilitate mathematics learning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3055–3069, Singapore. Association for Computational Linguistics.
- Zhang et al. (2020) Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations.
- Zheng et al. (2024) Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36.
Appendix A Data Collection Details
The annotators are screened through Prolific to be teachers native in English with an overall acceptance rate of more than 98% and with at least 500 submissions. We paid a minimum of $20 per hour. Annotators are from the US, Canada, and the UK, with a balanced gender ratio, and their age range is from 25 to 53 years. All annotators have K12 experience and on average they have 12 years of experience in teaching.
The annotators are first trained for the task with an interactive practice problem and then annotate student solutions. In one session one annotator performs 5 stepwise error verifications where they first pick the exact step with the error and then classify the error into 8 categories, with separate descriptions of the error for each category. We filter out all error descriptions not following the prescribed format to remove low-quality annotations.
The interface is shown in Figure 2. The categories are: missing or incorrect factual knowledge, misunderstanding of the question, the reference solution reached but proceed further, missing quantity, extra quantity, unit conversion error, numerical calculation, other.
A.1 Dataset Details
The collected dataset is in English and from the domain of K12 math word problem-solving. Examples from the dataset are shown in LABEL:{tab:dataset-examples}. The dataset consists of 1002 data points with 612 unique math problems. The distribution of total student steps and the location of the first incorrect steps are shown in Figure 3. Notice the student solutions contain up to 11 steps with a mean of ca. 6 steps. The location of the first error ranges from 1 to 8 steps with majority of the errors between the first and third steps.
The incorrect student solutions and reference solutions are part of the MathDial dataset (Macina et al., 2023a). The prompt used to generate correct student responses to balance the dataset with correct student responses are based on the Student model from (Macina et al., 2023a).
| Problem: Carl has been selling watermelons on the side of the road for $3 each. This evening he went home with $105 in profit and 18 watermelons. How many watermelons did he start out with this morning? |
| Reference solution: Carl sells his watermelons for $3 each so today he sold $105 / $3 per watermelon = 35 watermelons. He had 18 watermelons left over, so this morning he started with 18 + 35 = 53 watermelons. |
| Incorrect solution: Let’s start by finding out how much money Carl made selling the 18 watermelons. Since he sold each watermelon for $3, he made 18 x $3 = $54. We know that his total profit for the day was $105, so he must have started with $105 - $54 = $51 worth of watermelons. Since each watermelon costs $3, he must have started with 51 / $3 = 17 watermelons. |
| Student response: I started by finding out how much money Carl made selling the 18 watermelons. Since he sold each watermelon for $3, he made 18 x $3 = $54. Then I realized that his total profit for the day was $105, so he must have started with $105 - $54 = $51 worth of watermelons. Since each watermelon costs $3, I concluded that he must have started with 51 / $3 = 17 watermelons. |
| Error category: Misunderstanding of a question |
| Error description: Carl did not sell 18 watermelons, but 18 watermelons are left unsold. |
| Problem: Dakota gets hit by a bus and has to spend 3 days in the hospital. The hospital charges her $900/day for her bed, $250/hour for two specialists to look at her 15 minutes each, and $1800 for the ambulance ride. How much is Dakota’s medical bill? |
| Reference solution: First find the total cost of the hospital bed by multiplying the cost per day by the number of days: $900/day * 3 days = $2700. Then figure out how many minutes total the two specialists looked at Dakota: 15 minutes/specialist * 2 specialists = 30 minutes Then divide the number of minutes by the number of minutes per hour and multiply it by the specialists’ hourly rate to find their total charge: 30 minutes / 60 minutes/hour * $250/hour = $125. Finally, add up the cost of the specialists, ambulance and bed to find Dakota’s total bill: $2700 + $1800 + $125 = $4625 |
| Incorrect solution: The hospital charges Dakota $900 x 3 = $2700 for her bed. Each specialist charged her $250/hour x 2 = $500 for their 15 minutes each. So, Dakota was charged $500 x 2 = $1000 for the two specialists. Therefore, her medical bill is $2700 + $1000 + $1800 = $5500 |
| Student response: I started by calculating the cost of the bed, which was $900 x 3 days = $2700. Then I calculated the cost of the two specialists, which was $250/hour x 2 specialists x 15 minutes each = $500. Then I added all the costs together to get the total cost of $2700 + $1000 + $1800 = $5500 |
| Error category: Misunderstanding of a question |
| Error description: Student computes charges for a full hour of 2 specialists, not just 15 minutes as indicated in the question. |
Appendix B Details of Overall Verification and Stepwise Verification
For Stepwise Verification, we compare multi-class classification and iterative approach on our dataset and the results are in LABEL:{tab:error-finding-iterative}. The iterative approach classifies each step whether it is correct and therefore is more resource-intensive than multi-class classification. The multi-class classification directly predicts the label where 0 represents the solution is correct. The results indicate no improvements (with the exception of Llama2-70B) by using the iterative approach and in the main paper we therefore report multi-class classification results.
Moreover, to confirm the quality of our collected dataset, we run the same models on the smaller and simpler Roscoe human evaluation set (Golovneva et al., 2023). The dataset is smaller and contains 105 correct and 95 incorrect solutions. The results are shown in Table 9 and the conclusions are identical to our dataset.
| Model | micro F1 |
|---|---|
| GPT3.5 multi-class | 0.42 |
| + solution | 0.61 |
| GPT3.5 iterative | 0.39 |
| + solution | 0.36 |
| Llama2-70B multi-class | 0.17 |
| + solution | 0.48 |
| Llama2-70B iterative | 0.36 |
| + solution | 0.58 |
| Llama3-70B multi-class | 0.56 |
| + solution | 0.70 |
| Llama3-70B iterative | 0.56 |
| + solution | 0.58 |
| Model | Overall Verification | Stepwise | ||
|---|---|---|---|---|
| Corr. F1 | Err. F1 | F1 | micro F1 | |
| Few-shot | ||||
| GPT3.5 | 0.75 | 0.65 | 0.70 | 0.50 |
| + solution | 0.91 | 0.88 | 0.89 | 0.63 |
| Llama2-70B | 0.07 | 0.62 | 0.34 | 0.11 |
| + solution | 0.80 | 0.82 | 0.81 | 0.59 |
Appendix C Guidelines for Human Evaluation
The user interface used in the human evaluation is shown in Figure 4 and Figure 5. All the annotators had to complete a training for the task where each of their responses was evaluated and the feedback was provided to them. We used the subset of the annotators from Appendix A with the same selection conditions and the same payment. Before evaluating the quality of responses, annotators are asked to analyze the math problem and the conversation and explain the student error in an open-ended text. To not bias their understanding of the student solution only subsequently the error descriptions from verifiers were annotated with their correctness using these instructions: Does the text above correctly describe the root cause of the first student’s mistake? Answer "No" if the correct part of the student solution is identified as incorrect. Answer "No" if is too general without any further details e.g. ’There is a small mistake’.
The exact wording of the annotation questions for evaluating the quality of responses is the following:
Targeted
Does the Teacher point out to the root cause of the student’s mistake? Answer ’No’ if the Teacher gives the right answer without pointing out the mistake. Answer ’No’ if the Student’s statement is wrong and the Teacher does not point out the mistake directly. Answer ’Yes’ if the Teacher correctly describes the mistake in the student’s solution. Answer ’No’ if the Teacher addresses the correct part of the student solution. Answer ’No’ if response is too general and could be applied to any mistake e.g. ’You made a small mistake’.
Correctness
Is the Teacher’s response factually correct with respect to the reference solution? The teacher should NOT say incorrect information or provide parts of the solution that are NOT correct with respect to the reference solution. Answer ’No’ if the Teacher provides parts of a solution that is incorrect or does not guide a student towards the reference solution. Carefully compare the reference solution and the Teacher’s response.
Actionable
Does the Teacher provide actionable steps to let the Student correct the mistake without giving away the full answer? The teacher should provide actionable hints or steps WITHOUT revealing the full reference solution. Instead, the Tutor should give hints or ask questions to help the Student find the solution by themselves. Answer ’No’ if the Teacher simply just reveals the full reference solution.
Appendix D Alignment Details
To find the best hyperparameters for the Alignment algorithm we run a grid search using values of similarity threshold and gap costs . The best hyperparameters are reported in Table 4. The exact models which are used for semantic similarity are SBERT (sentence-transformers/all-mpnet-base-v2) and Roscoe (facebook/roscoe-512-roberta-base).
We use the template to transform the output of the algorithm into the textual prompt. In the template, all the steps from the student solution and reference solution are used. Furthermore, the cost of the alignment can be used to filter out student solutions that differ completely from reference solution which we leave for future work. The template is the following:
Missing steps in student solution: {missing steps}
Unnecessary steps in the student solution: {unnecessary steps}
Matching steps: {matching steps}
Appendix E Details on LLM-based Evaluation
A response is targeted if it targets the students’ mistake, correct if it does not conflict with grounding information, and actionable if it provides the student with useful guidance to help the student progress in their solution attempt. In all cases, for each quality dimension, we provide the model with three examples (3-shots). We use the LLAMA3-70B333meta-llama/Meta-Llama-3-70B-Instruct with temperature for reproducibility. The task description and the examples are the same as in the human evaluation for instructing the annotators described in Section 6.3. The prompt also includes the reference solution for more reliable judging (Zheng et al., 2024; Jurenka et al., 2024).
Appendix F Qualitative examples
In this section, we show qualitative examples to better understand the behavior of verification and verification-based response generation. We first show examples for prompted models in Table 10 and then show examples for finetuned models in Table 11.
| Problem: At a cafe, the breakfast plate has two eggs and twice as many bacon strips as eggs. If 14 customers order breakfast plates, how many bacon strips does the cook need to fry? |
|---|
| Correct Solution: Each breakfast plate has 2 x 2 = 4 bacon strips. The cook needs to fry 14 x 4 = 56 bacon strips. |
| Error Description Verification: The student incorrectly wrote the number of bacon strips per plate as 2 + 2x, where x is the number of eggs, instead of simply 2 x 2 = 4 bacon strips per plate. This led to an incorrect total of 84 bacon strips. |
| Dialogue |
| Teacher: Hi Claire could you please walk me through your solution |
| Student: Sure! For each breakfast plate, there are 2 eggs and twice as many bacon strips as eggs. So, for each breakfast plate, there are 2 + 2x bacon strips, where x is the number of eggs. Since each breakfast plate has two eggs, x = 2. So, for each breakfast plate, there are 2 + 2x bacon strips = 2 + 2(2) = 6 bacon strips. Thus, for 14 breakfast plates, the cook needs to fry 6 x 14 = 84 bacon strips. |
| Generated Responses |
| Human Teacher: Let’s work on simpler problem. If I have 2 dogs for every cat and I have 5 cats, I should have double the number of dogs. How many dogs will I have |
| Baseline: It looks like you made a small mistake in your solution. You correctly found the number of bacon strips for one breakfast plate, but you used the wrong value for x. Since each breakfast plate has 2 eggs, x should be 2, not 2 eggs. So the correct number of bacon strips for one breakfast plate is 2 + 2(2) = 6, and for 14 plates, the cook needs to fry 6 x 14 = 84 bacon strips. Great effort! |
| Error reason: It looks like you made a small mistake in your calculation. You correctly found that for each breakfast plate, there are 6 bacon strips, but when you multiplied by 14, you should have gotten 84, not 6 x 14 = 84. So the correct answer is 6 bacon strips per plate multiplied by 14 plates equals 84 bacon strips. |
| Error description: It looks like there was a mistake in your calculation for the number of bacon strips for each breakfast plate. Remember that the breakfast plate has twice as many bacon strips as eggs, so for each plate, there are 2 eggs and 4 bacon strips. Can you try recalculating the total number of bacon strips needed for 14 breakfast plates? |
| Alignment: It looks like you made a small mistake in your solution. The breakfast plate has 2 eggs, so it has 2 x 2 = 4 bacon strips. If 14 customers order breakfast plates, the cook needs to fry 14 x 4 = 56 bacon strips. The answer is 56. |
| Problem: A pet store owner had 12 birds, 9 puppies, 5 cats, and 15 spiders. Half the birds were sold and 3 puppies were adopted. Then, someone left the spider enclosure open and 7 of them went loose. How many animals does the pet store owner have left? |
|---|
| Correct Solution: There are 12/2 = 6 birds left. There are 9-3= 6 puppies left. There are 15 - 7 = 8 spiders left. The pet store owner has 5+6 + 6 + 8 = 25 animals left. |
| Verification: The student failed to include the 5 cats in their final count, resulting in an incorrect total of 20 animals instead of 25. |
| Dialogue |
| Teacher: Hi , could you please walk me through your solution? |
| Student: Sure, I started by taking half of the 12 birds, which is 6 birds. Then, I subtracted 6 from 12 to get the number of birds left, which is 6. After that, I subtracted 3 from 9 to get the number of puppies left, which is 6. Finally, I subtracted 7 from 15 to get the number of spiders left, which is 8. So, the pet store owner has 6 birds, 6 puppies, and 8 spiders left, which adds up to 20 animals. |
| Generated Responses |
| Human Teacher: i think you missed something. Cats? |
| Baseline: How many birds were sold and how many puppies were adopted? |
| Ours: How many cats did the pet store owner have? |
| Problem: Every year, Tabitha adds a new color to her hair. She started this tradition when she was 15 years old, which was the year she added her second hair color. In three years, Tabitha will have 8 different colors in the hair. Currently, how old is Tabitha? |
| Correct solution: Since, three years from now, Tabitha will have 8 different colors in her hair, this year she has 8 - 3 = 5 colors in her hair. Since she had 2 colors in her hair at the age of 15, she has added 5-2=3 colors since she was 15 years old. Since she ads one color per year, 3 added colors = 3 added years. Thus, currently, she is 15+3=18 years old. |
| Verification: The student incorrectly assumed that Tabitha is currently 21 years old, without considering the information that in three years, she will have 8 different colors in her hair. The student failed to work backwards to find Tabitha’s current age. |
| Dialogue |
| Teacher: Hi Jordy. Tell me about your answer |
| Student: I think my answer is correct. I used a calculator and I didn’t make any calculation errors. I figured that since Tabitha started adding a new color to her hair when she was 15, and she has added a new color every year since then, she must be 21 years old now. |
| Generated Response |
| Human: Let’s approach this working backward. If in three years, Tabitha will have 8 colors in her hair, how many colors does she have in her hair now? |
| Baseline: How many years has Tabitha been adding a new color to her hair? |
| Ours: How many different colors does Tabitha have in her hair now? |
Appendix G Prompts
This section provides the exact prompts used in our work. First, we show the prompt used for the baseline, error description-based, and alignment-based response generation models in Fig. 6 and Fig. 9. Verification prompts for Error Description are in Fig. 8 and for Error Reason in Fig. 7. The prompt with 5 examples for the CoT solution generation is in Fig. 10. Then, we show the prompts used for targeted LLM-based evaluation in Fig. 11, correctness evaluation in Fig. 12, and evaluation of how actionable responses are in Fig. 13. To sample responses from models by prompting we use temperature for reproducibility.
Appendix H Finetuning Details
We finetune all models by extending the huggingface transformers library (Wolf et al., 2020) and using the checkpoints from the huggingface hub in accordance with the corresponding license agreements.
For verification, we finetune LLAMA2 with 7B parameters and using LoRA. We use a learning rate of , linear learning rate decay with 32 warmup steps, a batch size of 2 and train for 6 epochs in total.
For response generation, we finetune Flan-T5 3B with LoRA with a learning rate of , a batch size of 2 and a total of 10 training epochs.
For both tasks, we used NVIDIA A100 80GB GPU and training takes around 3-6 hours for 5 or 10-fold cross-validation.