PedagogicalRL-Qwen3-8B

The PedagogicalRL baseline of Sherpa: Teaching LLMs to Teach Adaptively: a LoRA adapter for Qwen/Qwen3-8B. It is our reproduction of PedagogicalRL with the same base model, student, data and training budget as Sherpa. Code: SALT-NLP/Sherpa.

Usage

hf download SALT-NLP/PedagogicalRL-Qwen3-8B --local-dir pedagogicalrl-qwen3-8b
python -m sglang.launch_server --model-path Qwen/Qwen3-8B --enable-lora \
  --lora-paths pedagogicalrl-qwen3-8b --max-lora-rank 16

Use the prompts in examples/pedagogical_rl/prompts.py of the code repository. Each reply is <reasoning>...</reasoning> followed by <output>...</output> (shown to the student) or <end></end> (ends the dialogue).

Results

From the paper. Bold: this model.

Student test accuracy after teaching (%), Qwen3-1.7B student, 528 MATH test problems (36% unaided). In-distribution (ID) archetypes: Non. None, Att. Attempt, Sub. Subgoal, Ctr. Contrast. Out-of-distribution (OOD): Cau. Causal, Stp. Step, Ver. Verification. Ovl.: all seven.

Teacher	Non.	Att.	Sub.	Ctr.	Cau.	Stp.	Ver.	ID	OOD	Ovl.	95% CI
Qwen3-8B	53.2	50.7	48.4	44.1	54.4	47.8	41.5	49.1	47.9	48.6	[46.1, 51.0]
PedagogicalRL	63.0	59.4	54.7	44.4	59.0	44.3	43.8	55.4	49.1	52.7	[50.3, 55.1]
SherpaNP	63.1	58.8	52.0	45.3	55.1	40.6	49.2	54.8	48.3	52.0	[49.7, 54.5]
SherpaDP	76.2	78.2	72.7	60.7	77.1	51.8	66.4	72.0	65.1	69.0	[66.8, 71.2]

MathTutorBench (%). PS problem solving, SQ Socratic questioning, SC solution correctness, ML mistake location, MC mistake correction, Scaf. scaffolding, IF instruction following, -H hard. Ped. Avg.: mean of Scaf., IF, Scaf.-H and IF-H.

Teacher	PS	SQ	SC	ML	MC	Scaf.	IF	Scaf.-H	IF-H	Ped. Avg.
Qwen3-8B	92.1	25.3	70.6	37.1	74.8	33.1	74.0	32.4	70.6	52.5
PedagogicalRL	93.0	21.9	70.9	18.7	73.9	52.8	76.6	43.4	80.1	63.2
SherpaNP	92.1	21.7	72.7	32.8	75.9	65.3	82.5	64.2	81.7	73.4
SherpaDP	93.2	21.9	73.5	30.3	77.0	72.0	90.3	67.3	87.2	79.2

Citation

@misc{xu2026sherpateachingllmsteach,
      title={Sherpa: Teaching LLMs to Teach Adaptively},
      author={Weixian Xu and Yanzhe Zhang and Zora Zhiruo Wang and Changyu Chen and Diyi Yang},
      year={2026},
      eprint={2610.08778},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.08778},
}