The Evolution of LLMs: From Imitation to Reasoning
summary
Walmart fine-tunes open models and moves from SFT/RLHF alignment to reasoning methods like DPO and scaled RL (notably GRPO), then applies test-time compute techniques such as tree of thought and model cascades for query routing. Engineering builds include distilling reasoning traces into small language models, quantization experiments on A10/A100-class hardware, and an attribute-extraction pipeline that processes roughly one billion attribute cells with an auto-eval stage that measurably cuts hallucinations.
AI-written summary · about
chapters (7)
- 0:00Introduction to talk
- 1:25Evolution ladder of LLMs
- 7:52Alignment techniques explained
- 11:31Reasoning models breakthrough
- 19:26Test time compute methods
- 25:46Inference optimization strategies
- 28:37Future of LLM engineering
related talks
29mLLMs that think: Demystifying Reasoning Models
28mSymbolic Reasoning in telecom oriented LLMs
27mExploring the Mechanisms of LLM Reasoning Capabilities
31mBeyond LLMs and Maturing RAG use cases
16mBeyond LLMs - The Next Frontier of Machine Reasoning
29mHow to roll your own LLM
42mRDF and the future of LLMs
19mLLMs reasoning and agentic capabilities over time