DIMSUM: Discourse in Mathematical Reasoning as a Supervision Module
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866913723089682432 |
|---|---|
| author | Sharma, Krish Barman, Niyar R Chaturvedi, Akshay Asher, Nicholas |
| author_facet | Sharma, Krish Barman, Niyar R Chaturvedi, Akshay Asher, Nicholas |
| contents | We look at reasoning on GSM8k, a dataset of short texts presenting primary school, math problems. We find, with Mirzadeh et al. (2024), that current LLM progress on the data set may not be explained by better reasoning but by exposure to a broader pretraining data distribution. We then introduce a novel information source for helping models with less data or inferior training reason better: discourse structure. We show that discourse structure improves performance for models like Llama2 13b by up to 160%. Even for models that have most likely memorized the data set, adding discourse structural information to the model still improves predictions and dramatically improves large model performance on out of distribution examples. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2503_04685 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | DIMSUM: Discourse in Mathematical Reasoning as a Supervision Module Sharma, Krish Barman, Niyar R Chaturvedi, Akshay Asher, Nicholas Computation and Language We look at reasoning on GSM8k, a dataset of short texts presenting primary school, math problems. We find, with Mirzadeh et al. (2024), that current LLM progress on the data set may not be explained by better reasoning but by exposure to a broader pretraining data distribution. We then introduce a novel information source for helping models with less data or inferior training reason better: discourse structure. We show that discourse structure improves performance for models like Llama2 13b by up to 160%. Even for models that have most likely memorized the data set, adding discourse structural information to the model still improves predictions and dramatically improves large model performance on out of distribution examples. |
| title | DIMSUM: Discourse in Mathematical Reasoning as a Supervision Module |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2503.04685 |