Mr. Priyanshu Sharma

Gleaning: Letting LLMs Mark Their Own Homework

Abstract:

Large Language Models (LLMs) have revolutionized natural language generation, yet their outputs often require manual review and revision—a costly and time-consuming process that doesn't scale. This talk introduces gleaning, a practical inference-time pattern that enables LLMs to iteratively self-improve through structured draft-critique-revise loops.

Gleaning mimics human editing workflows: the model generates an initial draft, critiques its own output using structured feedback, and revises accordingly. This loop continues until a quality gate is passed or a maximum iteration limit is reached. Unlike traditional fine-tuning approaches, gleaning requires no additional model training and incurs only modest token overhead (typically 1.3× the original cost).

We present a 40-line reference implementation that demonstrates how this pattern can be applied to any LLM API. Our evaluation on diverse tasks shows consistent quality improvements: readability scores improve by 15-20%, while human review time decreases by 40-50%. 

The talk provides attendees with a battle-tested pattern they can implement immediately, along with practical guardrails to prevent common failure modes like endless loops or feedback hallucination. By the end, participants will understand how to deploy gleaning in their own workflows to achieve higher-quality LLM outputs with minimal additional cost.

Profile:

Priyanshu Sharma is the co-founder & CEO at ByteBrain LLC (Wyoming, USA) and an AI Researcher specializing in multi-agent systems and sustainable AI. A reviewer for reputed conferences and journals, and author of multiple titles for major publishers, Priyanshu holds two patent-pending AI systems, including MEDRAIL for clinical trial optimization. His research explores ethical AI applications that unite business impact with environmental stewardship.