Training Verifiers to Solve Math Word Problems

Karl Cobbe∗ Vineet Kosaraju∗ Mohammad Bavarian Mark Chen

arXiv:2110.14168v2 [cs.LG] 18 Nov 2021

Heewoo Jun Lukasz Kaiser Matthias Plappert Jerry Tworek

Jacob Hilton Reiichiro Nakano Christopher Hesse John Schulman

OpenAI

Abstract
State-of-the-art language models can match human performance on
many tasks, but they still struggle to robustly perform multi-step mathe-
matical reasoning. To diagnose the failures of current models and support
research, we introduce GSM8K, a dataset of 8.5K high quality linguisti-
cally diverse grade school math word problems. We find that even the
largest transformer models fail to achieve high test performance, despite
the conceptual simplicity of this problem distribution. To increase per-
formance, we propose training verifiers to judge the correctness of model
completions. At test time, we generate many candidate solutions and
select the one ranked highest by the verifier. We demonstrate that ver-
ification significantly improves performance on GSM8K, and we provide
strong empirical evidence that verification scales more effectively with
increased data than a finetuning baseline.

1 Introduction
In recent years, large language models have demonstrated impressive skills
across many diverse tasks (Wang et al., 2019; Brown et al., 2020). Kaplan
et al. (2020) describe the consistent benefits of increasing model size, character-
izing scaling trends that hold across many orders of magnitude. However, even
the largest models falter when required to perform multi-step mathematical rea-
soning (Hendrycks et al., 2021). Model samples frequently contain catastrophic
mistakes, even after the model has been appropriately finetuned. Mathematical
reasoning thus reveals a critical weakness in modern language models.
One significant challenge in mathematical reasoning is the high sensitivity
to individual mistakes (Shen et al., 2021a). When generating a solution, au-
toregressive models have no mechanism to correct their own errors. Solutions
that veer off-course quickly become unrecoverable. If we rely purely on genera-
tive methods and extrapolate from current trends, we will require an exorbitant
∗ Equal contribution. Correspondence to: Karl Cobbe <[email protected]>, Vineet
Kosaraju <[email protected]>

1
Figure 1: Three example problems from GSM8K. Calculation annotations are
highlighted in red.

parameter count to achieve even moderate performance on distributions as chal-

lenging as the MATH dataset (Hendrycks et al., 2021). This evidence strongly
motivates the search for methods with more favorable scaling laws.
We propose training verifiers to evaluate the correctness of model generated
solutions, similar to concurrent work by Shen et al. (2021a). At test time, we
sample a fixed number of candidate solutions and select the solution ranked
highest by the verifier. Verifiers benefit both from their inherent optionality
and from verification being a simpler task than generation in general.
To facilitate research, we are releasing GSM8K, a dataset of 8.5K high quality
problems at the grade school math level. We designed this dataset to have high
linguistic diversity while relying on relatively simple grade school math concepts.
State-of-the-art language models struggle to achieve high performance on this
dataset, primarily due to the high diversity among problems. At the same time,
GSM8K solutions depend only on elementary concepts, so achieving high test
performance is a tractable goal.
Our main contributions are as follows:

1. We present a curated dataset of 8.5K grade school math questions and nat-
ural language solutions, useful for probing the informal reasoning ability
of large language models.

2. We show that, compared to a finetuning baseline, the use of verifiers results

in approximately the same performance boost as a 30x model size increase,
and that verifiers scale significantly better with increased data.
3. We show that dropout acts as a strong regularizer, significantly improving
performance for both finetuning and verification.

2
2 Dataset
GSM8K consists of 8.5K high quality grade school math problems created by
human problem writers. We segmented these into 7.5K training problems and
1K test problems. These problems take between 2 and 8 steps to solve, and
solutions primarily involve performing a sequence of elementary calculations
using basic arithmetic operations (+ − ×÷) to reach the final answer. A bright
middle school student should be able to solve every problem.
We created GSM8K based on the following design principles.

• High Quality We avoid error-prone scraping procedures and instead rely

on human workers to create problems. After performing extensive quality
control based on workers’ answer agreement, we estimate that less than 2
percent of problems contain breaking errors.
• High Diversity We strive for high diversity among problems. We ac-
tively avoid designing problems that are drawn from the same linguistic
template or differ only in superficial details, an issue that is prevalent
among many other datasets. By creating each individual problem to be
relatively unique, held-out test performance becomes a far more relevant
metric.
• Moderate Difficulty We choose a problem distribution that is challeng-
ing for large state-of-the-art language models, without being completely
intractable. GSM8K will help us better understand the data scaling trends
of different models and methods in this difficulty sweet spot. Problems re-
quire no concepts beyond the level of early Algebra, and the vast majority
of problems can be solved without explicitly defining a variable.

• Natural Language Solutions We collect solutions in natural language

rather than as pure math expressions. We believe this is the most generally
useful data format, and we expect it to shed light on the properties of large
language models’ internal monologues. We instructed problem writers to
explain their work as much as possible, but we allowed them to write
solutions in their own diverse linguistic styles.

The full GSM8K dataset can be found at https://github.com/openai/grade-

school-math. Example problems are shown in Figure 1, and we discuss addi-
tional dataset details in Appendix A.

3 Related Work
3.1 Related Datasets
Early math word problem datasets (Kushman et al., 2014; Roy and Roth, 2015)
are relatively small and are not well suited for testing the limits of modern lan-
guage models. Dolphin18K (Huang et al., 2016) is a larger dataset containing

3
18K problems, but solutions are provided only in the form of equations or fi-
nal answers. AQuA-RAT (Ling et al., 2017) contains 100K problems, but this
dataset unfortunately suffers from both a high degree of problem templatiza-
tion and poor quality control of the natural language solutions. MathQA is
a recently released subset of AQuA-RAT focused on correcting these mistakes
(Amini et al., 2019), but even the corrected dataset has data quality issues, with
around 30% of the data having inconsistencies (Miao et al., 2021). Ape210K
(Zhao et al., 2020) is the largest publicly available dataset, consisting of 210K
Chinese elementary school-level math problems. However, due to the language
barrier and the lack of natural language solutions, we’re unable to evaluate our
methods on this dataset.
The recently developed ASDiv dataset (Miao et al., 2021), which contains
2.3K math word problems, addresses common flaws in prior datasets by ensuring
problems have both high diversity and high quality. We share those design
principles in the creation of GSM8K. However, we note that GSM8K is larger,
provides natural language solutions, and consists of problems that on average
require more steps to solve. The MATH dataset (Hendrycks et al., 2021) is larger
and significantly more complex than GSM8K, but the high difficulty makes
it challenging to accurately measure progress given the current capabilities of
state-of-the-art language models.
Other recent reasoning-related datasets have focused on mathematical rea-
soning on symbolic math (Lample and Charton, 2019), reading comprehension
(LogiQA) (Liu et al., 2020), and commonsense question answering (Common-
senseQA) (Talmor et al., 2018). Similar to CommonsenseQA, GSM8K includes
questions that require basic background knowledge, like the number of days in
a week. Similar to LogiQA, which requires a mix of reading comprehension and
logical reasoning, GSM8K’s main difficulty lies in both properly interpreting a
question and reasoning through the steps to solve it.

3.2 Related Methods

Previous work has attempted to solve classic math word problem benchmarks
with recurrent seq2seq models (Sutskever et al., 2014) and closely related vari-
ants (Wang et al., 2017; Huang et al., 2018). More recent work has improved per-
formance by designing specialized encoder-decoder architectures (Amini et al.,
2019; Chiang and Chen, 2018; Xie and Sun, 2019; Chen et al., 2020; Li et al.,
2020), with the strongest results often relying on large pretrained encoders from
the BERT family (Chen et al., 2019; Kim et al., 2020; Liang et al., 2021).
Other recent work has recommended additional pretraining tasks to further
improve the math reasoning skills of large transformer-based models. Hendrycks
et al. (2021) propose pretraining models on a new AMPS corpus, derived from
Khan Academy problems and Mathematica scripts. Similarly, Shen et al. (2021b)
propose a pretrained a corpus of pre-K to college level curricula extracted from
the internet, and Peng et al. (2021) propose pretraining by predicting masked
subexpressions from expression trees.
Similar to verification, other methods have finetuned a language model to

4
 35 35

30 30

25 25
Test Solve Rate (%)

Test Solve Rate (%)

20 20

15 15

10 10 500 problems
3B model 1K problems
5 6B model 5 2K problems
12B model 4K problems
175B model 7.5K problems
0 500 1000 2000 4000 8000 0 3 6 12 175
Training Set Size # Model Parameters (B)

Figure 2: Final test performance for various GPT-3 model sizes after finetuning
on training sets of different sizes. Mean and standard deviation is shown across
3 runs.

select among many model completions. Nichols et al. (2020) proposed a sample-
and-rank approach to improve the collaborative storytelling ability of large lan-
guage models, with the training signal coming from the preferences of human
workers. In concurrent work closely related to our own, Shen et al. (2021a)
applied a similar approach to solving math word problems, jointly training a
model to both generate and rank solutions. Our work shares many fundamen-
tal similarities with their approach, though we differ in several key respects.
First, we focus attention on the space of natural language solutions, as this is
a richer and more general solution format than pure mathematical expressions.
Moreover, this choice enables our models to develop verbal analytical skills and
to produce solutions that are more readily interpretable by humans. Second,
we provide evidence that verifiers scale far more favorably with additional data
than baseline methods. Finally, we use separate generator and verifier networks,
in order to prevent the generator from overfitting.

4 Methods
We investigate two methods to solve problems in GSM8K: finetuning and ver-
ification. Finetuning, our baseline method, uses the same language modeling
objective as the generative pretraining in GPT-3 (Brown et al., 2020). At test
time, we judge performance by autoregressively sampling a single low temper-
ature solution and checking whether the final answer is correct. In contrast,
verification consists of sampling multiple high temperature solutions, assigning
each solution a score, and outputting the highest ranked solution. Verifiers are
trained to judge the correctness of solutions, with the training signal determined
solely by whether or not the solution reached the correct final answer.

5
 22 84

82
20
80

Test@100 Solve Rate (%)

Test@1 Solve Rate (%)

18
78

16 76

74
14
72
12
70
6B 6B
0 10 20 30 40 50 0 10 20 30 40 50
Epoch Epoch

Figure 3: Test solve rate after finetuning a 6B model on the full GSM8K training
set, when the model is allowed to make 1 guess (left) or 100 guesses (right).

For both methods, we use models from the GPT-3 family as our initializa-
tion, primarily focusing on the 175B and 6B model sizes. The 175B model is
the largest and produces the most impressive results, while the 6B model is sig-
nificantly more convenient for research purposes. We discuss hyperparameter
choices in Appendix B.
Our models frequently fail to accurately perform calculations. Although
larger models make fewer arithmetic mistakes than smaller models, this remains
a common source of errors. To mitigate this issue, we train all models to use
a calculator by injecting calculation annotations into the training set. At test
time, a calculator will override sampling when the model chooses to use these
annotations. Details can be found in Appendix C.

4.1 Finetuning
We perform finetuning by updating model parameters to minimize the cross-
entropy loss over all training tokens. Figure 2 shows test performance after
finetuning on training sets of varying sizes for 20 epochs. We visualize the same
data both as a function of training set size and as a function of model size.
Test performance is determined by a single low temperature (T = 0) sample
for each test problem. Unsurprisingly, we see that the 175B model significantly
outperforms the smaller models. Assuming a log-linear trend, we can naively
extrapolate these results to estimate that a model with 1016 parameters would
be required to reach an 80% solve rate, when using the full GSM8K training
set. It is even harder to extrapolate along the data dimension, since performance
does not appear to follow a log-linear trend. Nevertheless, it appears likely that
the 175B model would require at least two additional orders of magnitude of
training data to reach an 80% solve rate.
In Figure 3, we show how 6B test performance varies over the course of 100

6
 Figure 4: A diagram of the verification training pipeline.

training epochs. We use test@N to denote the percentage of problems solved

correctly at least once when allowing the model to make N separate guesses for
each problem. We use a low temperature (T = 0) to generate test@1 samples
and we use a higher temperature (T = 0.7) to generate test@100 samples. Both
temperature values were chosen empirically to produce the best results. Test@1
performance improves approximately monotonically, even though we quickly
begin overfitting on test loss. Unfortunately, test@100 performance degrades
much more sharply than test@1 as we increase the number of epochs. This is
to be expected: as the model repeatedly encounters the same data, it becomes
increasingly uncalibrated and overconfident in its predictions. At test time, this
overconfidence leads to poor coverage of the solution space, an effect which only
becomes noticeable when we are considering multiple samples at test time.
Choosing a model with good coverage is critical to successfully train verifiers.
Empirically, we see that test@100 performance peaks within the first few epochs.
For this reason, we use models trained for 2 epochs to generate samples for
training verifiers. We provide several example solutions from 6B and 175B
models in Appendix D. We also note that it is important to allow the model to
generate the full natural language solution before outputting a final answer. If
we instead finetune a 6B model to directly output the final answer without any
intermediate steps, performance drops drastically from 20.6% to 5.2%.

4.2 Verification
To improve upon the finetuning baseline, we train verifiers to judge the correct-
ness of model-generated solutions and search against these verifiers at test time.
Conditioned on the problem and a candidate solution, the verifier outputs the
probability that the solution is correct. Training solutions are labeled as correct
or incorrect based solely on whether they reach the correct final answer. In prac-
tice, some solutions will reach the correct final answer using flawed reasoning,
leading to false positives.

7
 60 60

50 50

40 40
Test Solve Rate (%)

Test Solve Rate (%)

30 30

20 20

10 10
6B Finetuning 175B Finetuning
6B Verification 175B Verification
0 500 1000 2000 4000 8000 0 500 1000 2000 4000 8000
Training Set Size Training Set Size

Figure 5: A comparison between finetuning and verification using 6B and 175B

model sizes. Verification considers 100 solutions per problem. Mean and stan-
dard deviation is shown across 3 runs, except for 175B verification which shows
only a single run.

As shown in Figure 4, we train the verifier as follows:

1. Finetune a model (the “generator”) for 2 epochs on the training set.

2. Sample 100 completions from the generator for each training problem and
label each solution as correct or incorrect.
3. Train a verifier for a single epoch on this dataset.

Training for 2 epochs is enough for the generator to learn basic skills in this
domain. We choose not to train for longer, since the diversity of generated
solutions begins to collapse after this point, as shown in Figure 3. We train
separate generator and verifier models to limit the generator’s training and
prevent overfitting, but in principle, it should be possible to combine these
models. Unless otherwise specified, we use the same model size for the generator
and the verifier. In addition to predicting solution correctness, we also train the
verifier with the same language modeling objective as the generator. This serves
as a valuable auxiliary objective for the verifier. We discuss additional verifier
training details in Appendix E.
At test time, we sample 100 completions to each test problem, rank them
with the verifier, and then return the one with the highest verifier score. A
comparison between verification and finetuning is shown in Figure 5 for both
the 6B and 175B model sizes. We find that it is not beneficial to use verification
at low dataset sizes. We believe this is due to the pressure to overfit to the
correct answer: with small datasets, overfitting to the correct answer happens
faster than learning more generalizable properties of correct reasoning. However,
once we use a sufficiently large dataset, we see a strong boost from verifiers.

8
 40 40 60
6B Gen, 6B Verifier
6B Gen, 175B Verifier
35 35 50 175B Gen, 6B Verifier
175B Gen, 175B Verifier
30 30
40

Test Solve Rate (%)

Test Solve Rate (%)

Test Solve Rate (%)

25 25
30
20 20
20
15 15

10 10 10
token-level token-level, joint
5 solution-level 5 token-level, verification-only
0 500 1000 2000 4000 8000
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Training Set Size
epoch epoch

(a) Comparison between a (b) Comparison between (c) Performance when vary-
verifier trained to predict a verifier trained jointly ing the size of the genera-
correctness after every to- to predict correctness and tor and the verifier in isola-
ken (token-level) and one perform language model- tion. Increasing the size of
trained to predict correct- ing (joint) and one trained the generator has a larger
ness after only the final to- only to predict correctness impact than increasing the
ken (solution-level) (verification-only) size of the verifier.

Figure 6: Verification ablations

It’s interesting to note that the 175B verifiers “take off” earlier than the 6B
verifiers, requiring fewer training problems to surpass the finetuning baseline.
See Appendix D for example solutions found by verifiers and Appendix F for a
visualization of verifier confidence.

4.3 Verification Ablations

We can either train verifiers to make a single scalar prediction conditioned on
the entire generated solution, or to make a scalar prediction after each token
in the solution. By default, we choose the latter, training verifiers to make
predictions after each token. This can be viewed as a token-level value function.
We compare these two methods in Figure 6a, respectively labeled “solution-
level” and “token-level”.
Predicting the value function at every token is a more challenging and nois-
ier task than judging only the full completion. However, despite the initially
slower training, the token-level verifier ultimately outperforms the solution-level
verifier. Moreover, the token-level verifier is still improving late in training,
whereas the solution-level verifier quickly shows signs of overfitting. We hy-
pothesize that the full value function provides a useful auxiliary signal that
encourages the model to judge the reasoning throughout solutions, rather than
merely memorizing the correct final answer.
In Figure 6b, we ablate the objective used when training verifiers. As dis-
cussed in Section 4.2, we can optionally include a language modeling objective
alongside the verification objective. We compare using both objectives to using
only the verification objective. Although both are reasonable choices, including
the language modeling objective is a strict improvement. This makes intuitive

9
 42 44
42
40
40
Test Solve Rate (%)

Test Solve Rate (%)

38
38
36

36 34
100 completions
32 200 completions
400 completions
34 30 800 completions
1600 completions
28 3200 completions
25 50 100 200 400 800 1600 3200 1 10 100
Number of completions per test problem Number of top samples used for voting

(a) 6B verification test performance (b) 6B verification test performance

when given varying numbers of comple- when varying the number of top ranked
tions per problem to rank. samples allowed to vote on the answer.

Figure 7: Performance as the amount of test time compute varies.

sense: better understanding this language distribution should only aid the ver-
ifier in discriminating between samples.
In Figure 6c, we separately ablate the model size of the generator and the
verifier. We find that using a large generator with a small verifier performs sig-
nificantly better than using a small generator with a large verifier. Verification
is still remarkably effective, even when the verifier is much smaller than the gen-
erator. This suggests that the verifier may often be relying on relatively coarse
heuristics to discriminate between solutions from a given generator, rather than
attempting a more thorough form of verification.

5 Additional Experiments
5.1 Test Time Compute
At test time, we can choose to generate arbitrarily many solutions to be judged
by the verifier before selecting the highest ranked completion. Figure 7a shows
how 6B verifier performance varies with the number of completions per test
problem. At this scale, performance improves as we increase the number of
completions up to 400. Beyond this point, performance start to decrease. This
suggests that the benefits of search are eventually outweighed by the risk of
finding adversarial solutions that fool the verifier. In general, we evaluate verifier
test performance using 100 completions, since this captures most of the benefits
of verification with a relatively modest compute cost.
To further increase performance, we can take a majority vote among the
top verifier-ranked solutions instead of selecting only the single top solution.

10
 40
25.0
40
35
22.5

20.0 30
30
Test Solve Rate (%)

Test Solve Rate (%)

Test Solve Rate (%)

17.5 25

15.0 20
20
12.5
15
10
10.0
6B Finetuning, dropout = 0 10 solution-level, dropout = 0
7.5 token-level, dropout = 0
6B Finetuning, dropout = 0.2 solution-level, dropout = 0.2 token-level, dropout = 0.2
0
500 1000 2000 4000 8000 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Training Set Size epoch epoch

(a) Finetuning (b) Solution-level verifiers (c) Token-level verifiers

Figure 8: 6B finetuning and verification dropout ablations.

This voting process considers only the final answer reached by the individual
solutions: the final answer selected is the one with the most votes. Figure 7b
shows how performance varies as we allow a greater number of top samples to
cast a vote. Unsurprisingly, when starting with a greater number of samples,
we can afford to allow a greater number of samples to cast a vote. When we
have only 100 samples, it is optimal to allow only the top 3-5 samples to cast a
vote. When we have 3200 samples, it is approximately optimal to allow the top
30 to cast a vote.

5.2 Regularization
We find that both finetuning and verification strongly benefit from the use of
dropout as a regularizer. Specifically, we apply residual dropout (Vaswani et al.,
2017) along the residual paths of each layer in the network. We use 20% dropout
for all dropout experiments, chosen based on the results of a hyperparameters
sweep. We note that GPT-3 models are not pretrained with dropout. For ex-
periments involving dropout, we therefore perform additional pretraining with
dropout before subsequently finetuning the models. This mitigates the distri-
bution shift the model experiences during finetuning.
We first investigate the effect of dropout on finetuning across various train-
ing set sizes. Figure 8a shows that dropout leads to a significant improvement
over baseline. We next investigate the effect of dropout on verifiers, consider-
ing both the solution-level and token-level variants. In Figure 8b, we see that
dropout significantly improves solution-level verifiers, mitigating the overfitting
that occurs in the unregularized baseline. Notably, using dropout with solution-
level verifiers reaches a similar level of performance as token-level verifiers. In
Figure 8c, we apply dropout to token-level verifiers. Since token-level verifiers
are already less susceptible to overfitting, it is no surprise that the impact of
dropout is less significant. Nevertheless, we do still see a slight gain from train-
ing token-level verifiers with dropout. Note that we increase the batch size for
token-level verifiers by a factor of 4, to better handle the more difficult objective
and the noise from dropout.

11
6 Conclusion
We have seen that verification provides a significant performance boost relative
to a finetuning baseline. On the full dataset, 6B verification slightly outperforms
a finetuned 175B model, thereby offering a boost approximately equivalent to
a 30x model size increase. We have also seen that token-level verifiers are less
prone to overfitting than solution-level verifiers, and that all methods benefit
from regularization with residual dropout. We expect verification to scale well
to problem distributions that require more complex mathematical reasoning,
and we hope GSM8K supports the development of new methods that scale even
better.

Acknowledgements
We thank Dan Hendrycks, Leo Gao, Alec Radford, and Giambattista Paras-
candolo for their valuable feedback on this paper; Harri Edwards, Yura Burda,
Michael Wu, and Nick Ryder for many insightful conversations; Michael Petrov,
Alethea Power, and Jacob Jackson for their technical assistance; the OpenAI
Supercomputing team for the infrastructure that made these experiments pos-
sible; and the team at Surge AI for performing the GSM8K data collection.

References
A. Amini, S. Gabriel, P. Lin, R. Koncel-Kedziorski, Y. Choi, and H. Hajishirzi.
Mathqa: Towards interpretable math word problem solving with operation-
based formalisms. arXiv preprint arXiv:1905.13319, 2019.
T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Nee-
lakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot
learners. arXiv preprint arXiv:2005.14165, 2020.
K. Chen, Q. Huang, H. Palangi, P. Smolensky, K. D. Forbus, and J. Gao. Map-
ping natural-language problems to formal-language solutions using structured
neural representations. In ICML, 2020.
X. Chen, C. Liang, A. W. Yu, D. Zhou, D. Song, and Q. V. Le. Neural symbolic
reader: Scalable integration of distributed and symbolic representations for
reading comprehension. In International Conference on Learning Represen-
tations, 2019.
T.-R. Chiang and Y.-N. Chen. Semantically-aligned equation generation for
solving and reasoning math word problems. arXiv preprint arXiv:1811.00720,
2018.
D. Hendrycks, C. Burns, S. Kadavath, A. Arora, S. Basart, E. Tang, D. Song,
and J. Steinhardt. Measuring mathematical problem solving with the math
dataset. arXiv preprint arXiv:2103.03874, 2021.

12
D. Huang, S. Shi, C.-Y. Lin, J. Yin, and W.-Y. Ma. How well do computers solve
math word problems? large-scale dataset construction and evaluation. In
Proceedings of the 54th Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers), pages 887–896, 2016.
D. Huang, J. Liu, C.-Y. Lin, and J. Yin. Neural math word problem solver with
reinforcement learning. In Proceedings of the 27th International Conference
on Computational Linguistics, pages 213–223, 2018.
J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child,
S. Gray, A. Radford, J. Wu, and D. Amodei. Scaling laws for neural language
models. arXiv preprint arXiv:2001.08361, 2020.

B. Kim, K. S. Ki, D. Lee, and G. Gweon. Point to the expression: Solving

algebraic word problems using the expression-pointer transformer model. In
Proceedings of the 2020 Conference on Empirical Methods in Natural Lan-
guage Processing (EMNLP), pages 3768–3779, 2020.
N. Kushman, Y. Artzi, L. Zettlemoyer, and R. Barzilay. Learning to automati-
cally solve algebra word problems. In Proceedings of the 52nd Annual Meeting
of the Association for Computational Linguistics (Volume 1: Long Papers),
pages 271–281, 2014.
G. Lample and F. Charton. Deep learning for symbolic mathematics. arXiv
preprint arXiv:1912.01412, 2019.

S. Li, L. Wu, S. Feng, F. Xu, F. Xu, and S. Zhong. Graph-to-tree neural

networks for learning structured input-output translation with applications
to semantic parsing and math word problem. EMNLP, 2020.
Z. Liang, J. Zhang, J. Shao, and X. Zhang. Mwp-bert: A strong baseline for
math word problems, 07 2021.
W. Ling, D. Yogatama, C. Dyer, and P. Blunsom. Program induction by ra-
tionale generation: Learning to solve and explain algebraic word problems.
arXiv preprint arXiv:1705.04146, 2017.
J. Liu, L. Cui, H. Liu, D. Huang, Y. Wang, and Y. Zhang. Logiqa: A challenge
dataset for machine reading comprehension with logical reasoning. In IJCAI,
2020.
S.-Y. Miao, C.-C. Liang, and K.-Y. Su. A diverse corpus for evaluat-
ing and developing english math word problem solvers. arXiv preprint
arXiv:2106.15772, 2021.

E. Nichols, L. Gao, and R. Gomez. Collaborative storytelling with large-scale

neural language models. arXiv preprint arXiv:2011.10208, 2020.
S. Peng, K. Yuan, L. Gao, and Z. Tang. Mathbert: A pre-trained model for
mathematical formula understanding. ArXiv, abs/2105.00377, 2021.

13
S. Roy and D. Roth. Solving general arithmetic word problems. In Pro-
ceedings of the 2015 Conference on Empirical Methods in Natural Lan-
guage Processing, pages 1743–1752, Lisbon, Portugal, Sept. 2015. Associ-
ation for Computational Linguistics. doi: 10.18653/v1/D15-1202. URL
https://aclanthology.org/D15-1202.

J. Shen, Y. Yin, L. Li, L. Shang, X. Jiang, M. Zhang, and Q. Liu. Generate

& rank: A multi-task framework for math word problems. arXiv preprint
arXiv:2109.03034, 2021a.
J. T. Shen, M. Yamashita, E. Prihar, N. Heffernan, X. Wu, B. Graff, and D. Lee.
Mathbert: A pre-trained language model for general nlp tasks in mathematics
education, 08 2021b.
I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence learning with
neural networks. In Advances in neural information processing systems, pages
3104–3112, 2014.

A. Talmor, J. Herzig, N. Lourie, and J. Berant. Commonsenseqa: A ques-

tion answering challenge targeting commonsense knowledge. arXiv preprint
arXiv:1811.00937, 2018.
A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
L. Kaiser, and I. Polosukhin. Attention is all you need. In Advances in
neural information processing systems, pages 5998–6008, 2017.

A. Wang, Y. Pruksachatkun, N. Nangia, A. Singh, J. Michael, F. Hill, O. Levy,

and S. R. Bowman. Superglue: A stickier benchmark for general-purpose
language understanding systems. arXiv preprint arXiv:1905.00537, 2019.
Y. Wang, X. Liu, and S. Shi. Deep neural solver for math word problems. In
Proceedings of the 2017 Conference on Empirical Methods in Natural Lan-
guage Processing, pages 845–854, Copenhagen, Denmark, Sept. 2017. Asso-
ciation for Computational Linguistics. doi: 10.18653/v1/D17-1088. URL
https://aclanthology.org/D17-1088.
Z. Xie and S. Sun. A goal-driven tree-structured neural model for math word
problems. In IJCAI, 2019.
W. Zhao, M. Shang, Y. Liu, L. Wang, and J. Liu. Ape210k: A large-
scale and template-rich dataset of math word problems. arXiv preprint
arXiv:2009.11506, 2020.

14
A Dataset Details
We initially collected a starting set of a thousand problems and natural lan-
guage solutions by hiring freelance contractors on Upwork (upwork.com). We
then worked with Surge AI (surgehq.ai), an NLP data labeling platform, to
scale up our data collection. After collecting the full dataset, we asked workers
to re-solve all problems, with no workers re-solving problems they originally
wrote. We checked whether their final answers agreed with the original solu-
tions, and any problems that produced disagreements were either repaired or
discarded. We then performed another round of agreement checks on a smaller
subset of problems, finding that 1.7% of problems still produce disagreements
among contractors. We estimate this to be the fraction of problems that con-
tain breaking errors or ambiguities. It is possible that a larger percentage of
problems contain subtle errors.
To assist contractors with writing questions, we provided seed questions au-
tomatically generated from a few-shot prompted 175B GPT-3 model. Contrac-
tors were allowed to use those seed questions directly, to use them as inspiration
and make modifications, or to come up with their own questions entirely. We
instructed contractors to be as descriptive as possible in their solutions, and to
not re-use problem settings or templates between different questions. To ensure
contractors were not re-using problem templates, we computed pairwise simi-
larity scores between problems and used this to provide feedback to contractors.

15
B Hyperparameters
We include a table of important hyperparameters below. We performed sweeps
of the learning rate and batch size by an order of magnitude in both directions
from the values in the table and were unable to find any significant improve-
ments. Other reasonable choices for both the verifier temperature (eg: 1.0
instead of 0.7) and objective (cross-entropy instead of mean squared error) also
had negligible effect in our ablations.

General Hyperparameters Value

Batch Size 3.2 × 104 tokens

Max Sample Length 400 tokens
Tokenization reversible 50000
Optimizer Adam, β1 = 0.9, β2 = 0.95
Dropout 0.0
Learning Rate Schedule Linear decay to 0

Finetuning Hyperparameters Value

Epochs 20
Sampling Temperature 0 (argmax)
Base Learning Rate (α) 1.6 × 10−5 (3B)
1.2 × 10−5 (6B)
1.0 × 10−5 (12B)
6.0 × 10−6 (175B)
Learning Rate 0.1 × α

Verification Hyperparameters Value

Epochs 2 for generator, 1 for verifier

Sampling Temperature 0.7
Learning Rate 1.0 × 10−5
Loss weight 1.0
Verifier loss MSE
Completions per train problem 100
Completions per test problem 100

Table 1: Hyperparameters used for all experiments, unless explicitly said oth-
erwise. Notable exceptions include Figure 8c, which uses 4x more tokens per
batch and 300 completions at both training and test time. All dropout exper-
iments in Figure 8 use 20% dropout. Figure 7a uses verifiers trained on 100
completions, but searching over more completions at test time.

16
C Calculator Annotations
The calculator annotations were not provided by human contractors: they were
generated by a combination of hard-coded logic and a finetuned language model.
The logic for auto-generating calculator annotations is imperfect. It is highly
unlikely to generate any incorrect annotations, but it is not uncommon for it to
ignore some lines that could be annotated.
During training, there is no special distinction between the annotated to-
kens and the rest of the solution: they are all just tokens. During testing, we
override model sampling when a well-formatted annotation exists, specifically
overwriting the token(s) directly following “=” and within <<. . . >>.
To simulate the calculator, we simply use the python eval function to evalu-
ate the tokens in the expression (Figure 9). Evaluations that time out or throw
an error result in the annotations being skipped and the model being sampled
from as usual.
We note that the original version of our calculator, used for all results in this
paper, had some minor implementation bugs. Our reported test performance
is therefore a slight underestimate, though the magnitude of this discrepancy is
less than 1% in most experiments. Fixing the calculator improves verification
test performance by about 1% when using the full GSM8K training set.

Her sister gave her 20 + 10 = <<20 Generator +

Original Examp
Her sister gave her 20 + 10 = <<20+ Generator 10
Q Adrianne is

Trigger
Her sister gave her 20 + 10 = <<20+10 Generator =
Calculator
S

Calculator
Her sister gave her 20 + 10 = <<20+10= 30>>
eval(“20+10”)

Annotated
Her sister gave her 20 + 10 = <<20+10=30>> Generator books
S Her

Figure 9: A diagram of the calculator sampling procedure.

17
D Example Model Solutions
We showcase a handful of samples comparing finetuning and verification at both
6B and 175B scale. Samples were slightly cherry-picked for diversity.

18
19
E Verifier Details
As noted in section 4.2, we train verifiers with a joint objective where the model
learns to label a model completion as correct or incorrect, in addition to the
original language modeling objective. Architecturally, this means our verifiers
are language models, with a small scalar head that outputs predictions on a
per-token basis.
We implement this scalar head as a single bias parameter and single gain
parameter that operate on the logits outputted by the language model’s final
unembedding layer. Specifically, the bias and gain shift and scale the logit
corresponding to a special token in the vocabulary. As such, the logits for other
tokens can continue to represent the language modeling objective, while this
special token is reserved for the verifier’s predictions.
We can choose to initialize the verifier from the same pretrained language
model the generator was finetuned from, or from the generator itself. In our
ablations the latter performed slightly better; we suspect this is because better
understanding the language distribution that the generator learned should only
aid the verifier in scoring samples from that distribution. Unless otherwise
explicitly stated, we initialize our verifiers from their corresponding generators
in all experiments.
When training verifiers with the joint objective, we use an equal mix of
language data and verifier data. Because we sample 100 completions for each
original training example to generate the verifier data, using an equal mix means
we effectively upsample the original language data by a factor of 100. To form
the joint objective, we simply add the verifier loss and language modeling loss
unweighted, and define an epoch of this joint objective as having seen each
verifier example once. With both objectives, we mask out tokens in the question
and only train on tokens in the solutions, as visualized in Figure 12.

Language Modeling Objective Verifier Objective

Q1 Q2 ... Qn S1 S2 ... Sm-1 Q1 Q2 ... Qn S1 S2 ... Sm

Generator Verifier

Q2 Q3 ... S1 S2 S3 ... Sm Y1 Y2 ... Yn Yn+1 Yn+2 ... Yn+m

Qi the i-th question token Yk the k-th value prediction

Sj the j-th solution token masked out (does not

contribute to loss)

Figure 12: Visualization of the joint training objective. We mask out tokens in
the question and only consider the loss corresponding to tokens in the solution.

20
F Verifier Visualization

Question & Model Generation Verifier Actual

Pred. Score

Claire makes a 3 egg omelet every morning for breakfast. How many dozens
of eggs will she eat in 4 weeks?
✅ ✅
Claire makes 3 omelets every morning, so she eats 3*7=<<3*7=21>>21
omelets per week
She eats 21 omelets per week so over 4 weeks she will eat
4*21=<<21*4=84>>84 omelets
There are 12 in 1 dozen so she will eat 84/12=<<84/12=7.0>>7 dozens of
eggs
#### 7

Sophia and Rose went together to the market to buy onions and potatoes.
Rose bought 4 times the number of onions and potatoes Sophia bought. If
❌ ✅
Rose bought 12 onions and 4 potatoes, how many onions and potatoes in
total did Sophia buy at the market?
If Rose bought 12 onions and 4 potatoes, she bought 12+4 = <<12+4=16>>16
onions and potatoes.
If Rose bought 4 times the number of onions and potatoes that Sophia
bought, Sophia bought 16/4 = <<16/4=4.0>>4 onions and potatoes at the
market.
#### 4

Two sisters, Elizabeth and Margareth, bought beads. Elizabeth bought 1

pack of red and 2 packs of clear beads, while Margareth bought 3 packs
❌ ✅
of blue and 4 packs of red beads. How many more beads do one sister have
than the other, if each pack of beads contains 20 pieces of beads?
Elizabeth bought 1 pack of red and 2 packs of clear, totaling 1+2 =
<<1+2=3>>3 packs of beads.
Margareth bought 3 packs of blue and 4 packs of red, totaling 3+4 =
<<3+4=7>>7 packs of beads.
So, Elizabeth has 3-7 = <<3-7=-4>>-4 more packs of beads than Margareth.
And since each pack of beads contains 20 pieces of beads, Elizabeth has
20*4 = <<20*4=80>>80 more beads than Margareth.
#### 80

Howard spends $8 dollars at the arcade on Monday. On Tuesday, he spends

twice as much at the arcade as he did on Monday. On Wednesday, he spends
❌ ❌
4 times as much at the arcade as he spent on Tuesday. If he originally
had $100, how much money does he have left?
He spent $8 dollars on Monday and $8*2= $<<8*2=16>>16 dollars on
Tuesday.
He spent $16 dollars on Tuesday and $16*4= $<<16*4=64>>64 dollars on
Wednesday.
He has $100 dollars and spent $64 dollars so he has 100-64=
<<100-64=36>>36 dollars left.
#### 36

In a jewelers store, the price of a gold Jewell is 4/5 times as much as

the price of a diamond Jewell. The cost of a silver Jewell is $400 less
✅ ❌
than the price of gold. If a diamond Jewell is $2000, find the total
price for all three jewels.
The price of a gold Jewell is 4/5*$2000 = $<<4/5*2000=1600.0>>1600.
The cost of a silver Jewell is $2000-$400= $<<2000-400=1600>>1600.
The total price for all the jewels is $2000+$1600+$1600 =
$<<2000+1600+1600=5200>>5200
#### 5200

Figure 13: Five cherry-picked samples generated by a 175B finetuned model

and scored by a 175B token-level verifier. A green background color indicates a
high verifier score, and a red background color indicates a low one.

One benefit of the token-level verifiers is that these models become imme-
diately interpretable: we can visualize the predicted value for each token and
better understand how the verifier makes decisions on judging samples. Above
we present a visualization of the predicted values for five different cherry-picked
questions and model completions, verified by a 175B token-level verifier that
was trained on the full training set.
In the visualization, the background color of the text corresponds to the ver-
ifier score for that token, where red is low value (predicted incorrect) and green

21
is high value (predicted correct). The second column of the table summarizes
the verifier’s prediction, and the third column indicates whether the generated
model completion was actually correct or incorrect. Any disagreement between
the second and third columns indicates that the verifier made an error.
The first row includes a true positive example, where the verifier correctly
classifies the completion as correct. Note that the model is initially unsure about
whether the solution is correct and gradually gains certainty as the solution
progresses: this is likely a property of the verifier training procedure, where it
trains on a large fraction of incorrect model-generated samples.
The second row contains a problem where the solution is correct, but the ver-
ifier has rated it as incorrect. This is potentially due to the ambiguity between
the “4 times” and the “4 potatoes” in the problem description.
The third row consists of another false negative example. However, unlike the
previous example, here the model completion contains some faulty reasoning.
As such, even though the final answer in the model completion was correct,
the natural language explanation was incorrect, and so the verifier correctly
assigned a low score.
In the fourth row we see the verifier score a model completion that starts out
correct, but where the verifier gradually becomes less confident in the solution
as the solution progresses. After the solution makes a clear mistake (saying that
$64 dollars were spent, instead of the 64 + 16 + 8 = $88), the verifier judges the
solution as incorrect with a high degree of confidence.
The final row contains a false positive, where the model makes a mistake
on the second step, where it subtracts 400 from the price of a diamond jewel
instead of a gold one. Verifiers occasionally make mistakes with performing this
variable binding of quantities to their relationships.

22