Search2Adapt: Multi-Agent Collaborative Tree Search for Source-Free Unsupervised Domain Adaptation

Abstract

Source-Free Unsupervised Domain Adaptation (SFUDA) aims to adapt a pre-trained source model to unlabeled target domains without accessing source data. Traditional SFUDA methods rely on static pseudo-label generation, which often leads to suboptimal adaptation due to error accumulation and lack of feedback mechanisms. We present Search2Adapt, a novel multi-agent collaborative framework that reformulates SFUDA as a dynamic strategy search problem. Our approach introduces Collaborative Tree Search (CoTS), a plan-level search algorithm that enables efficient exploration of pseudo-label refinement strategies through multi-agent collaboration. The framework employs a closed-loop decision process where a Perceptor analyzes domain characteristics, a Strategist proposes adaptation plans, and an Evaluator provides multi-dimensional feedback without ground-truth labels. A Dynamic Prior Memory accumulates successful strategies across samples for accelerated adaptation. Extensive experiments on medical image segmentation benchmarks demonstrate that Search2Adapt achieves state-of-the-art performance, surpassing existing methods by significant margins while maintaining computational efficiency.

Overview of Search2Adapt. The framework employs a closed-loop multi-agent collaboration where Perceptor analyzes domain characteristics, Strategist proposes adaptation plans, and Evaluator provides feedback. CoTS explores the strategy space through tree search, guided by an anatomy-aware reward mechanism.

Method

Closed-Loop Multi-Agent Framework

Transforms traditional open-loop SFUDA pipelines into feedback-driven decision processes through collaboration between Perceptor, Strategist, and Evaluator.

Plan-Level Tree Search (CoTS)

Operates on complete tool-chain plans rather than individual actions, significantly reducing computational overhead while exploring diverse adaptation strategies.

Anatomy-Aware Reward

Multi-dimensional scoring mechanism that evaluates pseudo-label quality without ground-truth labels, considering anatomical plausibility, morphological consistency, and boundary precision.

Collaborative Tree Search (CoTS). The two-phase search process: Phase 1 uses LLM-based plan evaluation without execution, and Phase 2 verifies top-K plans through actual execution. The diversity-augmented CAVF formula balances exploitation, exploration, and strategy diversity.

Key Features

Diversity-Augmented CAVF

Our Collaborative Action Value Function (CAVF) extends UCT with a diversity bonus to prevent premature convergence:

$$\text{CAVF}(s_i) = V(s_i) + w\sqrt{\frac{\ln N_{\text{parent}}}{N_i}} + \frac{\lambda}{\sqrt{\text{count}(\text{tool}_{\text{sig}_i})}}$$

Exploitation Exploration Diversity Bonus

Dynamic Prior Memory

Five-layer memory structure enables cross-sample knowledge transfer:

Domain Profile — Perceptor's analysis of target domain characteristics
Adaptation Plans — Strategist's candidate plans with success history
Trajectories — Successful tool chains for similar domains
Plan Dialogues — CoTS collaboration records for learning
Evaluation Records — Historical scoring for strategy refinement

Results

82.7%

Mean Dice on AMOS abdominal targets (MR→CT), outperforming all baselines by 60%+.

94.2%

Dice on cardiac MYO (MR→US), matching supervised performance in cross-modality setting.

92.9%

Dice on fundus OD (RIGA→BASE3), achieving competitive performance across all target domains.

13 Domains

Validated across 13 adaptation directions and 24 anatomical targets with consistent improvements.

Abdominal Segmentation (AMOS)

Dice scores (%) on 15 abdominal organs. Best results in bold.

Method	Spleen	R.Kidney	L.Kidney	Gallbl.	Esoph.	Liver	Stom.	Aortic	V.Cava	Pancr.	R.Adre.	L.Adre.	Duod.	Bladder	Pros.	Mean
UPL	23.9	7.7	30.2	0.0	12.5	23.6	10.7	9.3	0.7	8.6	0.0	0.0	2.0	0.0	0.0	8.6
ProtoContra	5.6	11.9	7.3	4.3	37.5	40.1	25.1	38.6	29.2	39.1	0.0	0.0	17.6	0.0	0.0	17.1
IAPC	8.7	4.5	1.9	1.3	4.5	15.2	14.3	9.7	4.6	2.0	0.0	0.0	2.2	19.8	7.2	6.4
◇ DFG	0.7	28.7	4.1	0.0	35.1	35.0	16.6	13.0	22.0	29.7	0.0	0.0	21.2	0.0	0.0	13.7
◇ IPLC	0.0	0.0	0.0	0.0	0.0	65.4	10.5	6.6	4.1	3.8	0.0	0.0	0.0	0.0	0.0	6.0
◇ SRPL	0.0	0.0	0.0	0.0	41.0	47.6	42.2	66.0	26.2	46.6	0.0	0.0	29.4	0.0	0.0	19.9
◇ Search2Adapt	96.5	95.3	95.0	82.4	79.1	97.0	86.9	94.6	45.3	81.2	74.2	76.1	78.0	78.8	80.5	82.7

Qualitative Results

Qualitative results across diverse domain adaptation directions. Rows 1-2: abdominal targets; Row 3: cardiac, polyp, and brain targets.

Cardiac & Polyp Segmentation

Dice (%) and ASD (mm) on cross-modality cardiac and cross-dataset polyp targets.

Method	Cardiac MR→US				Cardiac US→MR				Polyp Kvasir→CVCDB		Polyp CVCDB→Kvasir
Method	LV	MYO	LV_ASD	MYO_ASD	LV	MYO	LV_ASD	MYO_ASD	Dice	ASD	Dice	ASD
UPL	2.8	0.0	90.5	N/A	0.5	2.9	88.9	85.6	67.1	2.2	60.2	11.4
ProtoContra	1.2	17.0	54.9	56.4	5.2	4.9	72.9	74.8	78.0	0.9	68.9	9.5
IAPC	18.4	3.1	51.3	97.9	25.8	5.6	34.1	42.6	73.3	4.4	61.2	13.7
◇ DFG	4.6	4.5	51.6	63.1	2.3	6.9	69.3	66.5	19.1	97.5	26.8	87.5
◇ IPLC	1.7	5.8	50.5	48.3	2.9	11.5	48.4	26.1	34.8	34.7	45.3	19.2
◇ SRPL	0.7	0.9	83.1	79.8	8.5	9.1	30.1	45.9	33.0	31.4	44.9	18.9
◇ Search2Adapt	87.9	94.2	3.0	2.2	89.7	95.2	0.3	0.2	86.0	7.1	87.4	10.6

Fundus Segmentation (RIGA → BASE)

Method	Dice (%)						ASD (mm)
Method	BASE1 OD	BASE1 OC	BASE2 OD	BASE2 OC	BASE3 OD	BASE3 OC	BASE1 OD	BASE1 OC	BASE2 OD	BASE2 OC	BASE3 OD	BASE3 OC
◇ SRPL	90.5	86.3	88.7	87.7	89.6	89.9	5.9	6.7	6.0	7.0	5.9	5.7
◇ Search2Adapt	92.3	84.9	89.3	85.4	92.9	88.4	6.5	7.8	6.7	8.0	3.9	6.4

Fundus qualitative results from RIGA to BASE1, BASE2, and BASE3.

Ablation Studies

Component Analysis (MR→CT)

Variant	Spleen	Liver	Pancr.	V.Cava	Mean Dice
w/o Perceptor	96.1	97.2	81.0	5.2	80.2
w/o CoTS	95.7	96.6	79.4	14.8	79.9
w/o CoT Diversity	96.5	97.2	81.5	3.8	80.5
w/o CoT Reflection	95.6	95.7	74.8	15.3	76.2
w/o VFM Postprocess	96.0	96.1	78.2	10.1	78.9
Search2Adapt (Full)	96.5	97.0	81.2	45.3	82.7

BibTeX

@inproceedings{shi2026search2adapt,
  title={Search2Adapt: Multi-Agent Collaborative Tree Search for Source-Free Unsupervised Domain Adaptation},
  author={Yulong Shi and Wenwen Zhang and Ziyi Li and Lin Qi},
  booktitle={Proceedings of the 34th ACM International Conference on Multimedia},
  year={2026}
}