Parallel reasoning improves the accuracy and robustness of large reasoning models by exploring multiple solution paths, but its computational cost grows with reasoning depth and branch count. Existing methods for managing these parallel paths typically rely on final-answer consensus, local token confidence, or isolated intermediate probes. However, these signals are often delayed, weakly tied to actual reasoning progress, or too noisy for dynamic, branch-level control. To address these limitations, we introduce ParaTempo, a training-free asynchronous parallel reasoning framework. ParaTempo is driven by temporal confidence, a branch-local measure of answer-space convergence. Each branch is periodically probed for a tentative answer probability distribution, and temporal confidence quantifies how sharply the recent intermediate probes concentrate on a dominant answer. Once sufficient evidence has accumulated, ParaTempo drives its entire control process from this single signal: low-confidence branches are pruned, branches that persistently commit to their dominant answer are retired early, freed computation is reallocated by forking new branches, and generation stops globally once the confidence-weighted vote concentrates. Without requiring synchronization among reasoning trajectories, ParaTempo adaptively allocates computation based on branch-level convergence. Experiments on challenging mathematical and scientific reasoning benchmarks show that ParaTempo reduces average latency by 21.8-32.2% and total token usage by 18.1-30.3% while maintaining competitive accuracy. Moreover, temporal confidence exhibits stronger temporal stability and predictive power for future branch convergence than token-level and instantaneous signals.
Figures & tables
Figure 1: Motivating branch-local temporal control. (a) Synchronized control advances all branches through global barriers. (b) ParaTempo asynchronously updates branch-local evidence and triggers local actions: stabilized branches are retired, diffuse branches are pruned, and reclaimed budget is reallocated via forking.
Signal
Volatility ↓
Spearman ∣ρ∣↑
AUC ↑
Mean token entropy
0.54
0.13
0.58
Token perplexity
0.56
0.12
0.57
Inst. answer confidence
0.26
0.41
0.71
Table 1: Evaluation of existing signals by temporal volatility and future answer stability prediction ( h=5 ). Spearman correlation and AUC measure signal stability and predictive performance, respectively.
Figure 2: Overall framework of ParaTempo. The framework periodically probes reasoning branches, estimates temporal confidence, and asynchronously allocates computation through branch control and confidence-weighted voting.
Method
AIME26
HMMT25
HMMT26
GPQA
Acc. ↑
Lat. ↓
Tok. ↓
Seq. ↓
Acc. ↑
Lat. ↓
Tok. ↓
Seq. ↓
Acc. ↑
Lat. ↓
Tok. ↓
Seq. ↓
Acc. ↑
Lat. ↓
Tok. ↓
Seq. ↓
Base model: Qwen3.5-35B-A3B
Zero-shot
72.3
92.2
12.6k
12.6k
64.2
90.6
12.4k
12.4k
42.4
94.7
13.0k
13.0k
82.2
69.0
9.4k
9.4k
SC@16
87.5
250.6
229.7k
15.6k
69.2
257.8
236.8k
16.0k
45.5
254.8
237.4k
15.8k
86.4
225.0
196.7k
14.7k
ESC@16
83.3
279.7
105.4k
27.2k
70.0
333.1
125.9k
32.3k
42.4
369.8
140.0k
35.9k
87.0
218.6
79.6k
21.6k
SAC@16
73.3
216.9
144.8k
12.8k
68.3
217.2
141.9k
13.4k
34.8
216.9
141.6k
13.4k
81.8
144.8
102.8k
9.1k
Table 2: Main results across four benchmarks. Acc., Lat., Tok., and Seq. denote accuracy (%), wall-clock latency (s), total generated tokens, and sequential generated tokens, respectively.
Figure 3: Latency–accuracy scaling curves of ParaTempo and Parallel-Probe with Qwen3.5-35B-A3B. The star marks self-consistency (SC@16 in Table 2 ). ParaTempo achieves a superior Pareto frontier across different operating points.
Figure 4: Analysis of temporal confidence under the preliminary study protocol. (a) Temporal confidence exhibits lower volatility than token-level signals and instantaneous answer confidence. (b) The future stability rate ( h=5 ) increases monotonically with signal quantiles of temporal confidence and top-1 probability.
Variant
Acc ↑
Lat ↓
Tok ↓
Seq ↓
ParaTempo
73.3
205.7
166.6k
10.6k
w/o Prune
71.7
233.0
179.1k
11.4k
w/o Retire
66.7
224.6
175.4k
10.4k
w/o Fork
70.0
195.7
145.5k
10.4k
Table 3: Ablation results of ParaTempo on HMMT25 with Qwen3.5-35B-A3B. Acc., Lat., Tok., and Seq. denote accuracy (%), latency (s), total tokens, and sequential tokens (K).
Scaling test-time reasoning has substantially improved the problem-solving ability of large language models (LLMs), but standard autoregressive decoding still executes long reasoning traces sequentially, creating severe latency for difficult tasks (up to days and weeks). Parallel reasoning offers a natural remedy. However, prior systems primarily focus on Subtask Parallelism, where the model learns to decompose a high-level task into smaller chunks that can be solved independently. This approach overlooks another pervasive form of parallelism: Trial Parallelism, where multiple speculative attempts explore, verify, and aggregate competing hypotheses in parallel. In this paper, we introduce Parason, which reveals and learns both forms of parallelism in LLM reasoning. Our analysis identifies Trial Parallelism as the majority of parallelizable reasoning computation (65.5% in DeepSeek-V4's reasoning steps in HLE), and it becomes increasingly dominant on hard problems. Guided by this taxonomy, Parason converts sequential reasoning traces into structured parallel trajectories with a context-free grammar, then trains models with Parallelism-Aware Group Relative Policy Optimization (PA-GRPO), whose reward jointly balances accuracy, latency, and the two parallelism ratios. At inference time, Parason executes the learned parallel structure through tool calls, translating theoretical savings to real-world wall-clock acceleration. Experiments on mathematical reasoning benchmarks including AIME24 and AIME25 show that Parason achieves an average acceleration about 1.7× while maintaining competitive accuracy.
Zhengyang Zhang, Zijian Zhang, Jiaxuan Gao +4
Tsinghua University · NVIDIA · Independent Researcher +1
Scaling inference-time computation has enabled Large Language Models (LLMs) to achieve strong reasoning performance, but their inherently sequential decoding incurs substantial latency, motivating parallelization of the generation process. However, existing parallel reasoning approaches suffer from performance degradation compared to their sequential counterparts, and often rely on specialized inference engines. We introduce ThreadWeaver, a framework for adaptive parallel reasoning that matches the accuracy of comparably sized sequential reasoning models while significantly reducing inference latency via three key innovations: 1) a two-stage parallel trajectory generator that produces high-quality parallel chain-of-thought data for supervised fine-tuning; 2) a trie-based rollout design that enables parallel reasoning on any off-the-shelf autoregressive inference engine; and 3) a parallelization-aware reinforcement learning framework that trains the model to balance reasoning accuracy with effective parallelization. Across six challenging math reasoning benchmarks, ThreadWeaver trained on top of Qwen3-8B achieves performance on par with cutting-edge sequential reasoning models (79.9% on AIME24 and 71.9% on average) while delivering up to 1.53x speedup in token latency, establishing a new Pareto frontier between accuracy and efficiency.
Long Lian, Sida Wang, Felix Juefei-Xu +7
Meta Superintelligence Labs (MSL), Menlo Park, CA, United States · UC Berkeley, Berkeley, CA, United States · UCSF, San Francisco, CA, United States
Test-time scaling improves the reasoning performance of large language models but often results in token-inefficient overthinking, where models continue reasoning beyond what is necessary for a correct answer. Existing dynamic early-exit methods typically rely on single-step confidence signals, which are often unreliable for detecting reasoning convergence in multi-step settings. To mitigate this limitation, we propose TRACE, a training-free framework for efficient test-time scaling that determines when to terminate reasoning based on temporal aggregation of multi-step evidence rather than instantaneous signals. TRACE detects reasoning convergence over time by aggregating two complementary signals across recent reasoning steps: answer consistency, capturing the persistence of predicted answers, and confidence trajectory, modeling the temporal evolution of model confidence. Benefiting from these two factors, TRACE can accurately determine whether the reasoning process has converged, thereby promptly halting inference and effectively avoiding redundant reasoning steps. Extensive experiments on multiple challenging benchmarks show that TRACE reduces reasoning token usage by 25-30% on average while maintaining accuracy within 1-2% of full-length reasoning, consistently outperforming existing dynamic reasoning methods.
Jiakun Li, Xingwei He, Kefan Li +3
School of Computer Science and Engineering, Beihang University · 3Hangzhou Innovation Institute, Beihang University · 2Qingdao Research Institute, Beihang University