Research map / Large language models / AI and society

Socialized Coevolution: Advancing a Better World through Cross-Task Collaboration

ICML 2025 · Large language models › AI and society · approach: Social-science view of AI evaluation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Socialized Coevolution (SC) and DISC, a cross-task collaboration framework where classification and detection models interact via dynamic hierarchical (feature) and dynamic selective (logit) collaboration modules, evaluated on CIFAR100 and VOC07+12.

Problem

Existing knowledge distillation and multi-task learning transfer knowledge between models in a way that either blindly imitates teachers or sacrifices task-specific expertise, so models cannot both learn new tasks and improve existing ones through interaction.

Approach

Models specialized in different downstream tasks (classification and detection) form a society and dynamically exchange hierarchical features and sample-weighted logits, with the main model using auxiliary features under a progressive schedule and weights adapting to each model's per-sample loss.

Key results

• DISC reaches 72.28% CLS and 84.92% DET, averaging 78.60%, versus 67.55-68.10% average for SwAV, DeepClusterV2, and MoCo v2 in the Table 2 comparison.
• Best prior KD/MTL combination (CrossKD(DET)+MTL) averages 68.49%, so DISC gains about 10.1 points on the average metric.
• Ablation: DHC alone raises average from 77.01% to 77.76%, DSC alone to 77.60%, and the combination to 78.60%.

Measured effects reported in the paper

Surprising findings

• Fine-tuning and standard KD improve new-task ability only by sacrificing performance on the original task (e.g., Expert(CLS)+FT drops detection from 0.00 to 0.00 while Expert(DET)+FT drops classification to 0.00 after evolution), suggesting blind imitation erodes existing expertise.
• Feature-level and logit-level interactions at extreme weights both hurt performance, indicating an intermediate weighting is needed.

Limitations

• Evaluation limited to two downstream tasks (classification and detection) on two datasets under a data-efficient 10% training setting.
• Impact statement notes model heterogeneity in open environments may complicate real-world deployment.
• Theoretical results (generalization bound, information-theoretic bounds) are argued with loose assumptions and not empirically validated in depth.

Experimental setup

Datasets: CIFAR100 VOC07+12

Benchmarks: CIFAR100 VOC07+12

Metrics: top-1 accuracy (CLS %) detection mAP (DET %) average (AVG %)

Baselines: SwAV DeepClusterV2 MoCo v2 CLIP LSKD+MTL CrossKD+MTL PPAL+MTL Fine-Tune KD MTL

Builds on: Knowledge Distillation Multi-Task Learning LSKD CrossKD PPAL

Closest papers by meaning

• Socialized Learning: Making Each Other Better Through Multi-Agent Collaboration (ICML 2024 · similarity 0.90)
• Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · similarity 0.85)
• Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.85)
• Tapes Together Strong: The Co-evolution of Computation and Cooperation (NeurIPS 2026 · similarity 0.85)
• Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.85)
• Beyond Single Stationary Policies: Meta-Task Players as Naturally Superior Collaborators (NeurIPS 2024 · similarity 0.85)