Reproduce “From "Aha Moments" to Controllable Thinking: Toward Meta-Cognitive Reasoning in Large Reasoning Models via Decoupled Reasoning and Control” 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
Rui Ha, Rui Pu, Chaozhuo Li, Li Sun, Sen Su
MERA trains reasoning models to alternate problem-solving segments with control segments assessing whether to continue, revise, or stop. It constructs training traces from existing reasoning and optimizes control text separately. On five math benchmarks, a 7B model reports 76.0% accuracy and 4,680 average generated tokens, versus DEER's 69.2% and 5,484; these measure answer accuracy and token counts, not runtime.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.