# Qwen2.5-3B: SFT on atomic skills (SFT_Mem+Ctx) This model is released with the paper [**From Atomic to Composite: Reinforcement Learning Enables Generalization in Complementary Reasoning**](https://arxiv.org/abs/2512.01970). - Project page: https://sitaocheng.github.io/atomic/ - Code: https://github.com/sitaocheng/from_atomic_to_composite - Data: [sitao/From_atomic_to_composite](https://huggingface.co/datasets/sitao/From_atomic_to_composite) ## What this model is Full-parameter SFT on the Parametric (MEM) and Contextual (CTX) training data only. The model sees no Complementary (COMP) example. It is the starting point of the atomic-first recipe. Base model: [Qwen/Qwen2.5-3B](https://huggingface.co/Qwen/Qwen2.5-3B). SFT used LLaMA-Factory. RL used GRPO in verl. The paper reports this model in Figure 8 (model scaling). ## Results on the Complementary (COMP) test set Accuracy (%) on the 4,468-item COMP test set, by generalization level. The numbers come from the validation log of the training run, as reported in the paper. | IID | Composition | Zero-shot | |---:|---:|---:| | 21.17 | 27.42 | 22.11 | ## The four models at this size | Model | Repository | |---|---| | SFT on atomic skills (SFT_Mem+Ctx) | [sitao/atomic-to-composite-qwen2.5-3b-sft-atomic](https://huggingface.co/sitao/atomic-to-composite-qwen2.5-3b-sft-atomic) | | SFT on the composite task (SFT_Comp) | [sitao/atomic-to-composite-qwen2.5-3b-sft-composite](https://huggingface.co/sitao/atomic-to-composite-qwen2.5-3b-sft-composite) | | Atomic-first recipe (SFT_Mem+Ctx → RL_Comp) | [sitao/atomic-to-composite-qwen2.5-3b-atomic-first-rl](https://huggingface.co/sitao/atomic-to-composite-qwen2.5-3b-atomic-first-rl) | | Composite-first recipe (SFT_Comp → RL_Comp) | [sitao/atomic-to-composite-qwen2.5-3b-composite-first-rl](https://huggingface.co/sitao/atomic-to-composite-qwen2.5-3b-composite-first-rl) | ## Usage The models answer with a reasoning chain and end with `So, the answer is: `. Use the same prompt format as in training: the question, then the context paragraphs after `PARAGRAPH:`. ```python from transformers import AutoModelForCausalLM, AutoTokenizer repo = "sitao/atomic-to-composite-qwen2.5-3b-sft-atomic" tok = AutoTokenizer.from_pretrained(repo) model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto") question = "A classmate of the parent of Tonya Serrano likes to eat what?" context = "..." # the biography paragraphs of the test item msg = "Answer the following question.\n" + question + "\nPARAGRAPH: " + context prompt = tok.apply_chat_template([{"role": "user", "content": msg}], add_generation_prompt=True, tokenize=False) out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=256, do_sample=False) print(tok.decode(out[0], skip_special_tokens=True)) ``` The models are trained on synthetic biographies. They are research artifacts for studying skill composition, not general-purpose assistants. ## Citation ```bibtex @article{cheng2025atomic, title = {From Atomic to Composite: Reinforcement Learning Enables Generalization in Complementary Reasoning}, author = {Cheng, Sitao and Yin, Xunjian and Zhou, Ruiwen and Li, Yuxuan and Wang, Xinyi and Pan, Liangming and Wang, William Yang and Zhong, Victor}, journal = {arXiv preprint arXiv:2512.01970}, year = {2025} } ```

Full Model Card (README)