llm LLM Agents for Scientific Discovery A benchmark for evaluating how well LLM agents generate novel research ideas, formulate rigorous experiment plans, and implement them in code. English to French Machine Translation Compared neural machine translation architectures and adapted Llama 2 7B for English-to-French translation using quantization and QLoRA. Generalizable Agents with Llama 2 Adapted Llama 2 7B into a general-purpose agent using interaction-trajectory instruction tuning and Tree-of-Thought experiments across AgentBench environments. Mini Llama 2 Implementation Implemented the core architecture, generation loop, optimizer, and classification pipeline for a compact 42M-parameter Llama 2 model. Improving Multi-modal Interactions in Memes Studied how vision-language models understand memes, using retrieval-augmented generation and abductive-reasoning fine-tuning to improve multimodal reasoning. RAG Question Answering System on Pittsburgh Built an end-to-end RAG system for factual questions about Pittsburgh and CMU, comparing baseline retrieval, reranking, and multi-query retrieval. ml Real-world Movie Recommendation System A production-style movie recommendation service for 1M+ simulated users, with online evaluation, A/B testing, monitoring, and scalable deployment infrastructure. rl cs
