Sequential Scaling Beats Majority Voting at Matched Compute (arxiv.org)

Researchers demonstrate that, given the same token and compute budget, running fewer reasoning chains that iteratively refine answers (sequential scaling) outperforms the dominant parallel self-consistency approach (many independent chains + majority vote). Across five state-of-the-art open-source LLMs and three challenging reasoning benchmarks, sequential refinement beat parallel sampling in 95.6% of tested configurations, with accuracy improvements as large as 46.7%. The comparison was done at matched compute/token budgets to isolate the effect of chaining style rather than extra resources, directly challenging the prevailing inference-time orthodoxy established by self-consistency decoding. They also introduce inverse-entropy weighted voting, a simple training-free aggregation that weights candidate answers by the inverse entropy of their reasoning chains (i.e., lower-entropy chains get more weight). This mechanism further amplifies sequential scaling’s advantage over majority voting, suggesting that chain-level confidence/consensus is a useful proxy for correctness. Implications are practical and immediate: inference strategies should favor sequential refinement and entropy-aware aggregation when optimizing for reasoning accuracy under fixed compute, prompting a rethink of deployment choices (chain length vs. parallel samples), decoding algorithms, and test-time ensembling in LLM-based reasoning systems.