Benjamin Riviere

We propose DISC, a dynamic decomposition method that adaptively adjusts step sizes during LLM inference to allocate compute more efficiently, significantly improving performance and sample efficiency across reasoning and code generation benchmarks.