Project
Eliminating the Lost-in-the-Middle Problem in Long-Context LLMs
Objective
1. Investigate the root causes of the lost-in-the-middle phenomenon in long-context LLMs, especially within RAG frameworks. 2. Develop and evaluate a lightweight, fine-tuning-free method to mitigate positional bias in retrieved contexts. 3. Benchmark the proposed approach across multiple datasets and retrieval configurations to assess generality and robustness.
Outcome
1. A robust, easy-to-adopt solution that addresses positional bias in long-context RAG systems. 2. A research paper documenting the findings, submitted to a top-tier NLP/ML conference.
Apply By Date 21 Jun 2025
Students 1 / 2
Duration 3 months
Mentor Sonam Mishra
Tools-Technologies
Jupyter Python Notebooks, NLP API, WatsonX.ai
Platform
1 ) WatsonX
College