CEDI Framework Improves MLLM Evaluation with Dynamic Interactions

Yijiang Li, Huiqi Zou, Bingyang Wang, Ziang Xiao· July 17, 2026 View original

Summary

CEDI is a new framework that evaluates Multi-modal Large Language Models (MLLMs) through dynamic, multi-turn interactions, moving beyond static benchmarks. It reveals significantly more visual hallucinations and identifies how they accumulate over long contexts, offering a more realistic assessment of MLLM capabilities.

Despite significant progress on benchmarks, the real-world performance of Multi-modal Large Language Models (MLLMs) remains uncertain due to a mismatch between static evaluations and dynamic, interactive applications. To bridge this gap, a new framework called CEDI (Contextualized Evaluations of MLLMs through Dynamic, multi-round Interactions) has been introduced. CEDI redefines evaluation as a three-party interaction involving the MLLM, an automated examiner, and a grader. The examiner engages in multi-turn, semi-structured conversations guided by a graph-based task representation. This allows CEDI to employ diverse strategies, including clarification requests and adversarial probes, to thoroughly test the MLLM's performance. Applying CEDI to visual hallucinations, experiments showed it uncovered significantly more hallucinations than traditional static methods. These identified hallucinations more closely resembled those encountered in practical use cases, often accumulating over extended dialogue histories. The research highlights MLLMs' particular vulnerability to questions requiring premise rejection or refusal, underscoring CEDI's value in providing realistic and systematic assessments.

Why it matters

For professionals deploying MLLMs, CEDI offers a more robust and realistic evaluation method, helping to identify and mitigate critical issues like visual hallucinations before models reach production, thereby improving reliability and user trust.

How to implement this in your domain

  1. 1Adopt dynamic, multi-turn interaction frameworks like CEDI for evaluating MLLMs instead of relying solely on static benchmarks.
  2. 2Prioritize testing MLLMs for hallucination accumulation over long conversational contexts.
  3. 3Develop specific test cases that require MLLMs to reject premises or refuse inappropriate requests to assess their robustness.
  4. 4Integrate contextualized evaluation into MLLM development pipelines to ensure real-world effectiveness and safety.

Who benefits

AI DevelopmentAutonomous SystemsCustomer ServiceContent ModerationRobotics

Key takeaways

  • Static benchmarks are insufficient for evaluating MLLM real-world effectiveness.
  • CEDI provides a dynamic, multi-turn interactive evaluation framework for MLLMs.
  • Contextualized evaluations reveal significantly more and more realistic hallucinations.
  • MLLMs are vulnerable to hallucination accumulation and premise rejection challenges.

Original post by Yijiang Li, Huiqi Zou, Bingyang Wang, Ziang Xiao

"arXiv:2607.14499v1 Announce Type: new Abstract: Multi-modal Large Language Models (MLLMs) have made substantial advances on benchmarks, yet their real-world effectiveness remains uncertain. This gap stems from the fundamental misalignment between benchmarks in controlled, static…"

View on X

Originally posted by Yijiang Li, Huiqi Zou, Bingyang Wang, Ziang Xiao on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research