SESSE Decomposes LLM-as-Judge Evaluation for Deeper Insights.

Dae Lee, Mihai Delgeanu, Adel Youssef· August 20, 2026 View original

Key takeaways

  • SESSE provides a training-free method for structured LLM-as-a-Judge evaluation.
  • It decomposes holistic judgments into interpretable sub-questions.
  • The framework helps diagnose label ambiguity and judge failure modes.
  • SESSE achieves competitive performance with specialized fine-tuned models.

Who benefits

AI/ML EngineeringProduct ManagementSoftware DevelopmentQuality AssuranceContent Moderation

Summary

SESSE (Sketch, Expand, Sort, Summarize, Evaluate) is a training-free framework that decomposes holistic LLM-as-a-Judge evaluations into structured sub-questions. This method provides an interpretable audit trail, helping diagnose label ambiguity and judge failure modes, and achieves competitive performance with specialized fine-tuned models.

This paper introduces SESSE, a novel, training-free framework designed to enhance the interpretability and diagnostic capabilities of LLM-as-a-Judge evaluations. Traditional LLM-as-a-Judge methods often provide only a single holistic preference choice, making it difficult to understand the specific quality dimensions driving the judgment or to differentiate between model errors and genuine label ambiguity. SESSE addresses this by decomposing the holistic judgment into structured sub-questions. The framework operates through a sequence of steps: Sketch, Expand, Sort, Summarize, and Evaluate. Crucially, it does not require oracle responses, task-specific rubrics, or fine-tuning, as the sub-questions are mined directly from the judge's own error cases. On the RewardBench dataset, SESSE demonstrated performance comparable to chain-of-thought baselines and competitive with specialized fine-tuned models like RISE-Judge-32B (92.7%). The per-criterion vote evidence generated by SESSE offers a valuable, interpretable audit trail, enabling better diagnosis of evaluation issues that are otherwise obscured by a single output token.

Why it matters

Professionals can use SESSE to gain deeper, more actionable insights from LLM-as-a-Judge evaluations, moving beyond simple A/B preferences to understand *why* a judgment was made and diagnose underlying issues.

How to implement this in your domain

  1. 1Integrate the SESSE framework into existing LLM-as-a-Judge pipelines to obtain more granular and interpretable evaluation results.
  2. 2Utilize the per-criterion vote evidence from SESSE to diagnose specific failure modes of generative AI models or identify areas of label ambiguity.
  3. 3Apply SESSE in A/B testing scenarios to understand the qualitative differences between model outputs, not just overall preferences.
  4. 4Explore using SESSE as a tool for developing more robust and nuanced evaluation rubrics for LLM applications.

Original post by Dae Lee, Mihai Delgeanu, Adel Youssef

"arXiv:2608.18303v1 Announce Type: new Abstract: LLM-as-judge evaluation reduces response quality assessment to a single holistic A/B preference choice, providing no mechanism to isolate which quality dimensions drove the preference or distinguish model errors from genuine label a…"

View on X

Originally posted by Dae Lee, Mihai Delgeanu, Adel Youssef on X · view source

Want to go deeper?

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

Explore courses