Method
Ditto-8B is trained with DITTO, a reinforcement learning method that uses verbal
feedback as the learning signal. After each output, the model receives descriptive feedback
and produces an improved version; both are jointly optimized with GRPO. This distills the
verbal guidance into the policy, so no feedback is needed at inference time.
Results
Primary metric for each benchmark (higher is better).
Table with columns: Dim, Benchmark, GPT 5.5, Gemini 3.1 Pro, Claude Opus 4.7, Qwen 3.6 Plus, Others*, Qwen3 8B Inst, Ditto-8B| Dim | Benchmark | GPT 5.5 | Gemini 3.1 Pro | Claude Opus 4.7 | Qwen 3.6 Plus | Others* | Qwen3 8B Inst | Ditto-8B |
|---|
| CONV | UserLLM | 65.3 | 67.7 | 57.6 | 72.1 | 44.6 | 46.0 | 91.5 |
| CONV | MirrorBench | 56.7 | 48.3 | 63.7 | 48.0 | 45.4 | 54.0 | 73.4 |
| CONV | Humanual-Chat | 28.2 | 21.0 | 22.6 | 22.2 | 25.8 | 24.7 | 21.0 |
| CONV | SimArena-Doc | 83.4 | 83.0 | 83.5 | 82.4 | 83.5 | 83.6 | 84.4 |
| SS | Sotopia-Hard | 31.9 | 27.8 | 32.4 | 28.3 | 31.7 | 27.7 | 45.8 |
| COG | Fantom | 93.0 | 93.0 | 80.0 | 89.0 | 70.0 | 23.0 | 92.0 |
| COG | Hitom | 82.0 | 86.0 | 93.0 | 73.0 | 56.0 | 62.0 | 79.0 |
| COG | Paratomi | 99.0 | 97.0 | 90.0 | 94.0 | 75.0 | 67.0 | 95.0 |
| COG | Social-R1 | 69.0 | 79.0 | 67.0 | 67.0 | 47.0 | 54.0 | 50.0 |
| ROLE | Coser | 66.2 | 62.1 | 66.5 | 55.9 | 30.3 | 43.5 | 64.4 |
| ROLE | Lifechoices | 91.0 | 84.0 | 92.0 | 79.0 | 67.0 | 70.0 | 70.0 |
| ROLE | Twinvoice | 74.0 | 86.0 | 83.0 | 71.0 | 40.0 | 42.0 | 71.0 |
| ROLE | BehaviorChain | 95.0 | 92.0 | 96.0 | 85.0 | 36.0 | 41.0 | 44.0 |
| ROLE | SimArena-Math | 68.5 | 71.5 | 68.7 | 70.9 | 70.5 | 68.9 | 69.6 |
| ROLE | Mistakes | 72.0 | 73.0 | 74.0 | 67.0 | 56.0 | 27.0 | 36.0 |
| ROLE | Humanual-Email | 50.1 | 46.9 | 50.4 | 47.9 | 42.8 | 43.7 | 40.8 |
| ROLE | Humanual-News | 40.2 | 42.3 | 41.3 | 41.8 | 33.1 | 32.5 | 27.5 |
| ROLE | Humanual-Politics | 42.0 | 32.5 | 43.5 | 31.6 | 34.2 | 33.2 | 29.7 |
| EVAL | AlignX | 71.2 | 73.4 | 71.6 | 69.8 | 66.8 | 68.6 | 67.4 |
| EVAL | Humanllm | 45.7 | 46.9 | 44.2 | 42.7 | 35.2 | 34.1 | 33.1 |
| EVAL | Socsci210 | 77.2 | 78.0 | 77.2 | 74.5 | 75.2 | 73.6 | 72.5 |
| EVAL | Humanual-Book | 57.6 | 62.4 | 61.4 | 58.4 | 50.2 | 53.6 | 53.4 |
| EVAL | Humanual-Opinion | 39.8 | 36.0 | 46.2 | 34.2 | 37.4 | 37.2 | 30.3 |
* Others: best result among other specialized human-simulation models (HumanLM-8B, Sotopia-RL-7B, UserLM-8B, Coser-8B).
Note. The released Ditto-8B is a single generalist distilled from a set of task-specific DITTO experts via rejection sampling on the training set.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "sunweiwei/Ditto-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Citation
@article{sun2026ditto,
title = {Reinforcing Human Behavior Simulation via Verbal Feedback},
author = {Sun, Weiwei and Zhou, Xuhui and Liu, Jiarui and Du, Weihua and Sun, Haojia and Xie, Yiqing and Ma, Qianou and Chen, Sihao and Wan, Mengting and Yang, Longqi and Zhou, Pei and Wu, Sherry and Welleck, Sean and Neubig, Graham and Yang, Yiming and Sap, Maarten},
year = {2026},
eprint = {2605.20506},
archivePrefix = {arXiv},
url = {http://arxiv.org/abs/2605.20506}
}