Parallel reasoning stops writing every path to the end: PreG scores short prefixes first and saves tokens in 54 of 60 settings
Synopsis
Parallel test-time scaling for large language models normally generates every reasoning path to completion before internal confidence can score it, so compute on low-quality paths is already spent. PreG instead generates short prefixes in parallel, ranks them by internal confidence, and completes only the most promising candidates before aggregating as before. On Qwen3-8B and Llama-Nemotron-4B across math competition sets and GPQA-Diamond, it reaches the accuracy of the original method at a 1000k-token budget with fewer tokens, saving tokens in 54 of 60 configurations and up to 80%.
Interpretation
The way parallel test-time scaling spends compute changes: instead of writing every sampled path to the end before judging it, a small slice of the budget goes to prefixes and most of it to a few favored candidates. Earlier confidence-weighted parallel methods computed confidence only after each path was fully generated, so more paths meant more tokens spent on low-quality paths that were later discarded. Across Qwen3-8B and Llama-Nemotron-4B, four math competition datasets plus GPQA-Diamond, and budgets of 100k, 300k and 1000k tokens, PreG gave the best result among compared methods in 13 of 15 dataset-budget combinations for Qwen and 14 for Llama.
The procedure splits one generation into two stages: write fixed-length prefixes in parallel, rank them by the model's own confidence, then complete only the top fraction. The confidence signal comes from inside the model, so no process reward model or verifier is needed and the same scheme can be applied across models. Prefixes use the first 4096 tokens and completion keeps either Borda voting or the top 10% of paths with weighted voting; on CMIMC at the 1000k budget, SCert-PreG with W-10% on Qwen beats all baselines by 5.4%.
The savings show up in the token account: to reach the accuracy the original method attains at a 1000k-token budget, PreG usually needs fewer tokens. The original method had to spend its whole budget on complete paths to reach the same accuracy. PreG used fewer tokens than the corresponding original method in 54 of 60 configurations, and under half the tokens in 27 of them; savings reach 80% on BRUMO, and Qwen saves over 40% on all but one math configuration.
Perspective
The evaluation covers two reasoning models, Qwen3-8B and Llama-3.1-Nemotron-Nano-4B-v1.1, on the AIME25, BRUMO25, HMMT25-Feb and CMIMC25 math competition datasets plus GPQA-Diamond, at budgets of 100k, 300k and 1000k tokens; the system-level measurement uses vLLM on four NVIDIA H200 GPUs.
Effectiveness is bounded by the quality of the underlying confidence signal, gains are limited when that signal is weak, and AUROC on GPQA-Diamond is lower than on math data, so adding confidence sometimes degrades performance; under very small token budgets PreG may not beat the original method, and W-10% voting roughly matches the baseline at 100k for Qwen while degrading slightly for Llama.