Vittle (L) is the learnable-prior variant of Vittle (NeurIPS 2025), a VLM instruction tuning framework that improves robustness to distribution shifts via variational information bottleneck.

Paper: arXiv:2505.13946
Code: deeplearning-wisc/vittle

Usage

python

import torch
from vittle.model.language_model.vittle_llama import VittleLlamaForCausalLM
from transformers import AutoTokenizer

model = VittleLlamaForCausalLM.from_pretrained(
    "changdae/vittle-7b-L",
    torch_dtype=torch.bfloat16,
    device_map="cuda:0",
)
tokenizer = AutoTokenizer.from_pretrained("changdae/vittle-7b-L", use_fast=False)

Refer to the evaluation guide for full inference instructions.

Model Details

Property	Value
Base model	`lmsys/vicuna-7b-v1.5`
Vision encoder	`openai/clip-vit-large-patch14-336`
Bottleneck layer	24
Interpolation coefficient (alpha)	0.5
KLD strength (beta)	0.1
Learnable prior	Yes
Training dtype	bfloat16

Citation

bibtex

@inproceedings{
oh2025visual,
  title={Visual Instruction Bottleneck Tuning},
  author={Changdae Oh and Jiatong Li and Shawn Im and Sharon Li},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
  url={https://openreview.net/forum?id=yzHiEmLSk8}
}