TLDR: This research paper challenges the common practice of parallel reasoning in large language models (LLMs) at test time. It demonstrates that sequential reasoning, where LLMs iteratively refine their answers, consistently outperforms parallel self-consistency methods across various models and benchmarks, even when given the same computational resources. The paper also introduces a new “inverse-entropy weighted voting” method that further boosts accuracy by prioritizing answers the model is more confident about, establishing sequential refinement as a superior strategy for LLM inference-time optimization.
In the rapidly evolving world of artificial intelligence, particularly with Large Language Models (LLMs), optimizing how these models reason and solve problems at inference time is a critical challenge. For years, the dominant approach has been ‘parallel self-consistency,’ where an LLM generates multiple independent answers and then picks the most frequent one. However, a groundbreaking new study challenges this orthodoxy, suggesting that a ‘sequential’ approach, where the model iteratively refines its thinking, is significantly more effective.
The research paper, titled The Sequential Edge: Inverse-Entropy Voting Beats Parallel Self-Consistency at Matched Compute, by Aman Sharma and Paras Chopra, delves into the fundamental question: when given the same computational resources, is it better for an LLM to run many independent thought processes in parallel, or to run fewer, but more refined, sequential steps? Their comprehensive evaluation across five state-of-the-art open-source models and three challenging reasoning benchmarks reveals a clear winner: sequential scaling.
A Shift in Reasoning Paradigms
The study found that sequential reasoning consistently outperformed parallel approaches in an astonishing 95.6% of configurations, with accuracy gains reaching up to 46.7%. This isn’t just a marginal improvement; it’s a fundamental challenge to the way LLM reasoning has been approached since the introduction of self-consistency decoding in 2022. The core idea behind sequential reasoning is that each step explicitly builds upon previous attempts, allowing the model to correct errors, accumulate context, and progressively refine its answer.
Introducing Inverse-Entropy Weighted Voting
Beyond establishing the superiority of sequential reasoning, the researchers also introduced a novel, training-free method called ‘inverse-entropy weighted voting.’ This clever technique further boosts accuracy by weighing answers based on the model’s confidence. Essentially, if the model is very certain about its next token predictions (low entropy), that part of the reasoning chain gets a higher weight in the final decision. This method proved to be optimal in 96.7% of sequential configurations and even improved parallel configurations, making it a universally effective aggregation strategy.
How Sequential Reasoning Works
Imagine solving a complex puzzle. The parallel approach is like having multiple people solve the puzzle independently and then comparing their final solutions. The sequential approach, on the other hand, is like one person working on the puzzle, then reviewing their progress, identifying mistakes, and building upon their current state to get closer to the solution. This iterative refinement allows for mechanisms unique to sequential methods, such as:
- Iterative Error Correction: Models can identify and fix computational mistakes in subsequent steps.
- Progressive Context Accumulation: Each step builds upon insights gathered from previous attempts.
- Answer Verification: Models can validate and refine their initial responses through multiple passes.
The study systematically evaluated seven different sequential voting methods, with inverse-entropy weighted voting emerging as the most effective. Methods that favored later reasoning steps consistently outperformed those that emphasized earlier attempts, highlighting the value of refinement.
Rigorous Evaluation and Key Findings
The experiments were conducted using diverse LLMs like GPT-OSS-20B, GPT-OSS-120B, Qwen3-30B, Qwen3-235B, and Kimi-K2, across benchmarks including the American Invitational Mathematics Examination (AIME-2024/2025) and GPQA-Diamond (graduate-level science questions). Crucially, all comparisons were made under ‘matched compute’ conditions, ensuring that both sequential and parallel approaches used the exact same total token budget, making the comparison fair and robust.
The research identified that a 6-chain configuration (meaning 6 iterative steps in sequential reasoning) offered the optimal balance between computational cost and performance gains. While parallel approaches showed diminishing returns beyond 6 chains, sequential reasoning maintained its efficiency.
Beyond Traditional Reasoning: Creative Tasks and Scaling
The paper also included ablation studies to explore the boundaries of sequential reasoning’s advantages. In creative tasks like joke generation, sequential reasoning led to superior lexical diversity (richer vocabulary), while parallel approaches excelled in semantic diversity (broader conceptual exploration). This suggests a fascinating trade-off, mirroring human creative processes where brainstorming (parallel) is followed by iterative editing (sequential).
Furthermore, when analyzing how these approaches scale with increasing computational budgets, sequential self-refinement consistently achieved higher absolute performance and superior efficiency across all token budgets, from 2K to 16K tokens per question.
Also Read:
- Optimizing LLM Collaboration: A Graph-Based Approach to Test-Time Scaling
- Predicting the Value of Thought: How Re-FORC Optimizes LLM Reasoning
A New Default for LLM Reasoning
The findings advocate for a paradigm shift, positioning sequential refinement as the robust default for modern LLM reasoning. This approach offers enhanced performance without the need for additional training costs, paving the way for more efficient and reliable AI systems. The proposed design principles for future sequential reasoning systems include iterative refinement over independent sampling, entropy-guided confidence weighting, and model-adaptive configuration of chain lengths.
