Reproduce “Context-Grounding Gains Are Mediated by Pre-existing Machinery: Auditing GRPO, SFT, and DPO” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Prakhar Gupta, Vaibhav Gupta
The authors test whether post-training improves context-following by adding new machinery or strengthening mechanisms in the instruction-tuned starting model. They estimate a context-versus-memory direction before training and intervene on it. DPO raises Qwen2.5-3B’s ConFiQA-QA update rate from .362 to .961, while tested GRPO variants yield small grounding gains; this supports a scoped account, not a universal mechanism claim.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.