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 | |