GEMS: Agent-Native Multimodal Generation with Memory and Skills

Dimensions overall score 7.0

• The GEMS framework enables the lightweight 6B model Z-Image-Turbo to surpass the state-of-the-art Nano Banana 2 on the GenEval2 benchmark.result 90%
  Evidencepartial

  Most notably, it enables the lightweight 6B model Z-Image-Turbo to surpass the state-of-the-art Nano Banana 2 on GenEval2

  Implicationpartial

  Directly stated in the abstract and repeated in the analysis with specific model and benchmark names.

  Verificationpartial

  partial

• GEMS achieves significant average performance gains of 14.22 on mainstream benchmarks and 14.03 on downstream tasks when using the Z-Image-Turbo backend.result 90%
  Evidencepartial

  GEMS yielded significant average performance gains of 14.22 on mainstream benchmarks and 14.03 on downstream tasks.

  Implicationpartial

  Specific numeric gains are provided in the analysis section.

  Verificationpartial

  partial

• The GEMS framework is built upon three core components: Agent Loop, Agent Memory, and Agent Skill.method 95%
  Evidencepartial

  GEMS is built upon three core components. Agent Loop introduces a structured multi-agent framework... Agent Memory provides a persistent, trajectory-level memory... Agent Skill offers an extensible collection of domain-specific expertise

  Implicationpartial

  Explicitly and clearly stated in the abstract as the core architecture.

  Verificationpartial

  partial

• Agent Loop introduces a structured multi-agent framework that iteratively improves generation quality through closed-loop optimization.method 90%
  Evidencepartial

  Agent Loop introduces a structured multi-agent framework that iteratively improves generation quality through closed-loop optimization.

  Implicationpartial

  Direct definition of the component provided in the abstract.

  Verificationpartial

  partial

• Agent Memory provides a persistent, trajectory-level memory that hierarchically stores factual states and compressed experiential summaries.method 90%
  Evidencepartial

  Agent Memory provides a persistent, trajectory-level memory that hierarchically stores both factual states and compressed experiential summaries

  Implicationpartial

  Direct definition of the component provided in the abstract.

  Verificationpartial

  partial

• GEMS was evaluated across five mainstream benchmarks (GenEval, GenEval2, DPG-Bench, OneIG, WISE) and four specialized downstream tasks.result 85%
  Evidencepartial

  We evaluate our system across five mainstream benchmarks, including GenEval, GenEval2, DPG-Bench, OneIG, and WISE

  Implicationpartial

  Specific benchmark names are listed in the analysis and result details sections.

  Verificationpartial

  partial

• The framework's generalizability was verified across multiple generative backends, including Z-Image-Turbo and Qwen-Image-2512.result 85%
  Evidencepartial

  Our framework’s generalizability was verified across multiple generative backends. Specifically, leveraging the lightweight, distilled Z-Image-Turbo... We further validated our framework on another mainstream open-source model, Qwen-Image-2512

  Implicationpartial

  Explicitly stated in the analysis with specific model names mentioned.

  Verificationpartial

  partial

• GEMS enables agentic reasoning and domain-specific expertise to push beyond the inherent boundaries of foundational models.technical 80%
  Evidencepartial

  demonstrating that agentic reasoning and domain-specific expertise can effectively push beyond the inherent boundaries of foundational models.

  Implicationpartial

  Claim is a conclusion drawn from the results, strongly implied by the performance gains described.

  Verificationpartial

  partial

Competitive landscape

GEMS is an agent-native multimodal generation framework that enhances foundational models with structured multi-agent loops, hierarchical memory, and domain-specific skills to achieve significant performance gains on complex and specialized tasks.