Tsinghua's DeepPrune Slashes Parallel AI Reasoning Costs by 88% Without Losing Accuracy
A Tsinghua team's DeepPrune framework cuts parallel reasoning token costs by 65 to 88 percent by killing redundant chain-of-thought traces before they finish.
Takeaways
- Tsinghua's THU-KEG team introduces DeepPrune, a framework cutting parallel reasoning tokens by 65-88 percent.
- Analysis shows over 80 percent of parallel chain-of-thought traces produce identical final answers, wasting compute.
- A fine-tuned Qwen3-4B judge predicts answer equivalence from partial traces at 0.87 AUROC.
- Online greedy clustering prunes redundant traces during generation, then majority-votes on the largest cluster.
- Tested on AIME24, AIME25, GPQA across DeepSeek-8B, Qwen3-32B, GPT-OSS-20B, staying within 3 points of cons@512 accuracy.
- Code, judge model, and dataset are released on GitHub and Hugging Face.
When large language models tackle hard reasoning problems, one of the most reliable tricks is to run them many times in parallel and vote on the answer. It works, but it is expensive. A team from Tsinghua University's THU-KEG lab has now shown that most of that compute is wasted, and they have released a framework called DeepPrune that kills off duplicate reasoning attempts before they finish, cutting token consumption by 65 to 88 percent on major math and science benchmarks without meaningfully hurting accuracy.
The framework, along with a trained judge model and dataset, is available on GitHub, with the judge model on Hugging Face.
The 80 percent problem hiding in parallel scaling
Parallel scaling, also called best-of-N sampling or self-consistency, has become a workhorse for pushing reasoning benchmarks. You generate many chain-of-thought traces for the same question, then either pick the best or take a majority vote. The catch is that total token cost can balloon past 100 million tokens per benchmark run.

The Tsinghua team started by measuring how much of that computation is actually useful. They collected 16 parallel traces per problem across four reasoning models including DeepSeek-R1-Distill-Llama-8B, Qwen3-4B-Thinking-2507, GLM-4.5-Air and QwQ-32B, then compared every pair of traces for the same problem. The result was stark: same-answer pairs dominated across all models, with ratios exceeding 80 percent in most cases, and GLM-4.5-Air producing 94.5 percent same-answer pairs.
When you run 16 parallel reasoning attempts, most of them are grinding through the exact same logic to reach the exact same conclusion. The authors call this inter-trace redundancy