VecCISC: Improving Confidence-Informed Self-Consistency with Reasoning Trace Clustering and Candidate Answer Selection

Researchers propose VecCISC, an optimization framework for weighted majority voting in large language models that reduces computational costs by 47% while maintaining accuracy. The method filters redundant or hallucinated reasoning traces using semantic similarity before evaluation, addressing the expensive overhead of confidence-scoring multiple candidate answers.

VecCISC addresses a practical bottleneck in scaling inference-time reasoning for language models. Self-Consistency and its weighted variant, Confidence-Informed Self-Consistency (CISC), have emerged as effective techniques for improving LLM reasoning accuracy across domains. However, CISC requires calling a critic LLM on every candidate answer to assign confidence scores, creating substantial computational overhead that limits real-world deployment despite superior performance gains.

The core innovation lies in semantic filtering. Rather than evaluating all candidate reasoning traces, VecCISC uses vector-based similarity measures to identify and remove semantically redundant, degenerate, or hallucinated traces before they reach the critic LLM. This preprocessing step dramatically reduces the number of evaluations needed while preserving the diversity of candidate answers that drives CISC's effectiveness.

The 47% reduction in token usage represents significant cost savings for organizations deploying reasoning-heavy LLM applications, particularly in scientific domains like mathematics, chemistry, and biology where accuracy is critical. The evaluation across five diverse benchmarks—spanning STEM fields to commonsense reasoning—demonstrates the method's robustness and generalizability rather than domain-specific optimization.

For AI infrastructure providers and enterprises implementing reasoning systems, VecCISC makes weighted voting approaches economically viable at scale. The framework maintains accuracy parity with full CISC evaluation, eliminating the traditional performance-cost tradeoff. This efficiency gain becomes increasingly important as organizations scale reasoning inference, where token costs compound rapidly with increased usage volume.

• →VecCISC reduces token usage by 47% while maintaining CISC accuracy through semantic filtering of reasoning traces
• →The method identifies and removes redundant, degenerate, or hallucinated candidate answers before critic evaluation
• →Evaluation across mathematics, chemistry, biology, commonsense reasoning, and humanities demonstrates broad applicability
• →Semantic similarity-based preprocessing makes weighted majority voting economically practical for large-scale inference
• →Framework maintains performance gains without the computational overhead that previously limited CISC adoption