UCLA's PTTS Gives AI Parallel Reasoning Branches Separate Plans, Gaining 13.4 Points
A planner-executor framework that coordinates parallel reasoning branches instead of independent sampling, lifting pass@64 on math benchmarks by up to 13.4 points.
- PTTS replaces independent repeated sampling with a planner that jointly generates k diverse solution outlines for an executor.
- Improves pass@64 by up to 6.7 points zero-shot and 13.4 with an RL-trained planner on five math benchmarks.
- PTTS-RL matches repeated sampling's pass@64 with only half the sampling budget across Qwen3-1.7B and 4B.
- Truncating executor rollouts to 4k tokens during training sharpens the reward signal and beats a 10k budget by 7.8 points.
- Small 1.7B planners transfer to larger 4B executors, enabling cheap planner training with scaled inference.
- Code available at github.com/shirley-wu/planned-test-time-scaling; complements diversity-aware RL executors like e3.
PTTS gives parallel reasoning samples separate plans
Researchers at UCLA and Amazon propose Planned Test-Time Scaling, or PTTS, which assigns different solution outlines to parallel inference branches. Across five math benchmarks, the authors report gains of up to 6.7 pass@64 points without planner training and 13.4 points with reinforcement learning.
Standard test-time scaling draws many responses from the same model distribution, so independent samples often follow similar high-probability strategies and repeat similar mistakes. PTTS adds a short planning stage before execution while keeping the reasoning model frozen, making it an orchestration change rather than an executor-training requirement.
Give every branch a route
Given a budget of completed solutions, PTTS generates several high-level outlines and conditions executor rollouts on those plans. The inference pipeline has four steps:
- Send the problem and requested branch count to a planner.
- Generate a set of distinct solution outlines.
- Parse the outlines and pass each one to the frozen executor.
- Run the executor branches, then verify or rank their answers.
A pass@k score measures the probability that at least one of k completed solutions is correct. Because the planner can emit empty outlines, the PTTS policy class includes ordinary repeated sampling. Its best achievable pass@k therefore cannot be lower than the baseline’s, although a particular trained planner can still underperform because of optimization or parsing failures.
Why extra samples echo each other
The paper’s toy model assigns every problem a hidden category and gives each reasoning strategy a category-specific success rate. Optimizing a single response for pass@1 favors the dominant strategy, leaving repeated samples vulnerable on categories that require a less common approach.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
