Training data
gsm-infinite-train.zip from kilian-group/phantom-reasoning:
GSM-Infinite, zero-context, realistic, medium difficulty (2–20 operations). The first 10,000 of the 10,096 problems are used.
Reproduce training
Trained on 4× H100 80GB (about 22 hours):
git clone https://github.com/kilian-group/phantom-reasoning.gitcd phantom-reasoning# Install the environment as described in the repo README, then:./scripts/download_data.sh data accelerate launch --num_processes 4 --config_file recipes/accelerate_configs/zero1.yaml \ src/phantom_reasoner/grpo.py \ --config recipes/Qwen/Qwen3-0.6B/grpo/config_gsminfinite_4gpu.yaml \ --seed 1
This checkpoint was trained at commit 2a83e69
with TRL 0.21.0 and Transformers 4.55.0. The recipe on main has been updated for newer TRL
(dtype replaces torch_dtype, and vllm_max_model_length: 6144 replaces max_prompt_length: 2048), with otherwise identical settings.
Evaluate
From the repo root, the evaluation scripts accept this Hub ID directly:
# GSM-InfiniteMODEL_NAMES="kilian-group/PhantomReasoning-Qwen3-0.6B-gsminf" bash scripts/eval/gsminf_eval_grpo.sh out__eval=gsminf # Real-world multi-hop QA (swap hp500 for 2wiki500, msq500, cofca500, synthrm500)MODEL_NAMES="kilian-group/PhantomReasoning-Qwen3-0.6B-gsminf" bash scripts/eval/wiki_eval_grpo.sh out__eval=wiki hp500 minidev
Citation
@inproceedings{kabra2026learning, title={{Learning from Synthetic Data Improves Multi-hop Reasoning}}, author={Kabra, Anmol and Gong, Albert and Yin, Yilun and Stankevi{\v{c}}i{\=u}t{\.e}, Kamil{\.e} and Go, Dongyoung and Luo, Katie Z and Lee, Johann and Gomes, Carla P and Weinberger, Kilian Q}, booktitle={The Fourteenth International Conference on Learning Representations}, year={2026}, url={https://arxiv.org/abs/2603.02091}}