LCMP & TAME

This repo provides official implementations for LCMP & TAME:

TAMEing Long Contexts in Personalization:
Towards Training-Free and State-Aware MLLM Personalized Assistant
(KDD 2026 Research Track)

Overview

🔥 LCMP: A long-context MLLM personalization benchmark.

Construction pipeline of LCMP benchmark.

🔥 TAME: A training-free, state-aware personalized MLLM assistant.

Architecture of TAME framework.

🛠️ Environment Setup

The environment is managed with uv (Python 3.13). A prebuilt flash-attn wheel for torch 2.8 / CUDA 12.8 / cp313 / linux-x86_64 is pinned in pyproject.toml, so no compilation is needed.

# install uv (skip if already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# create the venv and install all dependencies (incl. flash-attn)
uv sync

🏗️ LCMP

LCMP is the first benchmark for evaluating Long-Context MLLM Personalization, which is more practical and conceptually different from existing static, single-turn visual-identification-oriented evaluation.

LCMP Content

We will release the complete LCMP benchmark in data/concept, which includes personalized pets, people, and objects, along with their associated dialogues, questions, and personalized images.

Construct LCMP

We also provide code to reproduce LCMP by adding more personalized concepts.

Generate New Personalized Concept for LCMP (Pet Example)

1. Sample an image from the COCO dataset to dataset_maker/concept_pet/$concept_id/img/base.png (.jpg/.jpeg also accepted).

2. Run the following code:

# generate concept and history
uv run python dataset_maker/generate_profile_history_question.py $concept_id
# generate personalized image
uv run python dataset_maker/generate_prompt_images.py --concept_id $concept_id

🚀 TAME

TAME is a training-free and state-aware personalized MLLM assistant powered by double memories.

Run TAME

1. Move the constructed concepts to data/concept

2. Run the following code:

# build memory on all concepts using InternVL3-8B
uv run python method/main.py build --model internvl
# run TAME on all concepts using InternVL3-8B
uv run python method/main.py qa --model internvl

Or with Qwen2.5-VL-7B-Instruct (the default --model):

# build memory on all concepts using Qwen2.5-VL-7B-Instruct
uv run python method/main.py build --model qwenvl
# run TAME on all concepts using Qwen2.5-VL-7B-Instruct
uv run python method/main.py qa --model qwenvl

✅ Evaluation

# run evaluation on generated answers
uv run python evaluator/evaluator.py $input_file $output_file

📚 Project Structure

• dataset_maker/: tools to construct/extend LCMP concepts
• method/: TAME pipeline (memory building + QA)
• evaluator/: evaluation scripts
• data/: benchmark concepts

📖 Citation

@inproceedings{hong2026tameing,
	author = {Hong, Rongpei and Lang, Jian and  Zhong, Ting and Wang, Yong and Zhou, Fan},
	booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
	year = {2026},
	title = {TAMEing Long Contexts in Personalization: Towards Training-Free and State-Aware MLLM Personalized Assistant},
}