AdaptMI : Adaptive Skill-based In-context Math Instructions for Small Language Models

This is the homepage of the paper AdaptMI : Adaptive Skill-based In-context Math Instructions for Small Language Models.

Our work compares the in-context learning of SLM (Small Language Models) with human learning from teachers in a classroom. Instead of feeding in a fixed set of in-context examples, we propose an AdaptMI, an Adaptive approach to selecting skill-based in-context Math Instructions for SLMs.

Inspired by cognitive load theory from human pedagogy, our method only introduces skill-based examples when the model performs poorly. Our method effectively boost the math reasoning accuracy of SLM by up to 6% over naive skill-based prompting strategies.

Stage1: Detection of easy and difficult questions

In this stage, we will label a question as easy or difficult for a Small Language Model.

math-rm/rm_classify.py employs a process reward model to assign scores for each step in the SLM response. We then use thresholds τ1, τ2 (pred_thres1 and pred_thres2 in the code) to classify whether a question q is easy or difficult.

Stage 2: Skill-based selection of in-context examples

• AdaptMI uses skill-based k-shot examples for difficult questions and fixed k-shot examples for easy questions.
• AdaptMI+ focuses only on the skills that the model’s initial response lacks.

Environmental Setup

conda create -n matheval python=3.10
conda activate matheval

cd evaluation/latex2sympy
pip install -e .
cd ..
pip install -r requirements.txt
pip install vllm==0.5.1 --no-build-isolation
pip install transformers==4.42.3

conda create -n stage1 python=3.10.9
conda activate stage1

git clone https://github.com/OpenAccess-AI-Collective/axolotl
cd axolotl
git checkout 55cc214c767741e83ee7b346e5e13e6c03b7b9fa
pip install -e .

pip3 install torch==2.1.2 torchvision torchaudio
pip install flash-attn

git clone https://github.com/lm-sys/FastChat.git
cd FastChat
pip install -e .

git clone https://github.com/WeiXiongUST/RLHF-Reward-Modeling.git
pip install deepspeed

pip install -r math-rm/requirements.txt

conda create -n stage2+ python=3.10
conda activate stage2+
pip install skill_identifier/requirements.txt

Evaluation

AdaptMI

sh scripts/adaptmi.sh

AdaptMI+

sh scripts/adaptmi+.sh

Baseline

[We’re working on this section… Stay tuned!]

Citation

@article{he2025adaptmi,
  title={AdaptMI: Adaptive Skill-based In-context Math Instruction for Small Language Models},
  author={He, Yinghui and Panigrahi, Abhishek and Lin, Yong and Arora, Sanjeev},
  journal={arXiv preprint arXiv:2505.00147},
  year={2025}
}