Temporal Self-Consistency Voting

• Temporal self-consistency voting is a framework that aggregates outputs through iterative refinement, using prior reasoning chains to enforce stability and correct errors.
• It leverages a sequential reasoning process where each iteration conditions on previous outputs, enabling robust context accumulation and dynamic error correction.
• The approach employs inverse-entropy weighted voting and committee overlap optimization to balance accuracy, efficiency, and resistance to disruptive change.

Temporal self-consistency voting encompasses a family of aggregation and decision protocols in which outputs—whether by AI systems across iterative reasoning chains or by committees evolving under dynamic voter preferences—are explicitly shaped by their alignment to preceding sequences or stages. This framework prioritizes not only accuracy or representativeness at each iteration but also the stability, reliability, and overlap of outputs over time, reflecting systemic aversion to disruptive change or the exploitation of temporal error-correction.

1. Formal Definitions and Core Paradigms

Temporal self-consistency voting arises in two principal domains:

• LLM Reasoning (Sequential Self-Consistency, SSC): Given a problem input $x$, an initial solution chain $c^{(1)}\sim p(c|x)$ is generated. For each subsequent step $t=2,...,K$, a new chain $c^{(t)}\sim p(c|x, c^{(1)},...,c^{(t-1)})$ is sampled, conditioned on all prior chains. Final answers $y^{(t)}$ are extracted and aggregated, typically through a voting mechanism, producing a temporally self-consistent output that incorporates iterative refinement and context accumulation (Sharma et al., 4 Nov 2025).
• Committee Voting with Dynamic Preferences: In two-stage elections, an initial committee $S$ is selected under voters' initial preferences, which may then change, yielding a new preference profile and inducing re-election to a second-stage committee $S'$. The overlap $|S\cap S'|$ is maximized, or the distance $d(S,S')=k-|S\cap S'|$ is minimized, directly enforcing aversion to change and temporal stability (Zech et al., 2024).

These mechanisms contrast with purely parallel approaches, wherein independent solutions or committees are produced at each time step without explicit conditioning or overlap optimization between stages.

2. Sequential Self-Consistency and Iterative Refinement in LLMs

Sequential self-consistency (SSC) in LLM reasoning, as formalized in (Sharma et al., 4 Nov 2025), is characterized by stepwise iterative refinement:

• At each refinement step, the model receives all previous reasoning chains as input (the “continuation prompt”) and outputs a new chain and corresponding answer.
• This process defines a Markovian structure over chains, enabling both error correction (by revisiting and amending earlier mistakes) and context accumulation (as each chain potentially integrates information from all predecessor chains).

Pseudocode for K-stage sequential refinement:

$t=2,...,K$1

This mechanism is fundamentally distinct from parallel self-consistency (PSC), where $k$ independent chains are sampled as $c^{(1)}\sim p(c|x)$0 and combined only at the aggregation stage, without inter-chain information flow.

3. Inverse-Entropy Weighted Voting

Inverse-entropy weighted (IEW) voting is a central contribution for temporally self-consistent aggregation in LLMs (Sharma et al., 4 Nov 2025). It adjusts the influence of each reasoning chain during output voting according to the model's uncertainty, as follows:

• For each chain $c^{(1)}\sim p(c|x)$1, compute the average Shannon entropy $c^{(1)}\sim p(c|x)$2 over its sequence of top-k log-probabilities.
• Assign a weight $c^{(1)}\sim p(c|x)$3, with $c^{(1)}\sim p(c|x)$4 for numerical stability.
• Normalize the weights: $c^{(1)}\sim p(c|x)$5.
• Aggregate answers using these weights:

$c^{(1)}\sim p(c|x)$6

The rationale is that lower entropy indicates higher model confidence; thus, IEW voting down-weights noisy (high-entropy) chains, producing an aggregation that empirically correlates with higher answer accuracy compared to unweighted majority voting. Majority voting treats all chains equally, irrespective of their certainty.

4. Temporal Self-Consistency in Multiwinner Elections

In multiwinner voting scenarios with temporal aversion to change (Zech et al., 2024):

• The objective is to select, after dynamic preference shifts, a committee $c^{(1)}\sim p(c|x)$7 at stage two that overlaps maximally with the original $c^{(1)}\sim p(c|x)$8.
• Theoretical results show a full complexity dichotomy for Thiele rules:
  • Approval Voting (AV): Overlap optimization is tractable (polynomial time).
  • All other Thiele rules (including PAV, Chamberlin–Courant): Overlap-optimal re-election is NP-hard even for minimal voter changes, and W[1]-hard with respect to committee size $c^{(1)}\sim p(c|x)$9.
• Greedy variants (e.g., Greedy-PAV) are also intractable for the overlap objective, necessitating heuristics or restricted settings for practical deployment.

Experimental analysis demonstrates that even modest changes (1% of approved ballots) typically induce substantial committee turnover (replacing ≈2/10 members), and tie-breaking dramatically affects overlap; optimal-overlap tie-breaking can significantly outperform naïve lexicographic methods.

5. Quantitative Results and Experimental Insights

Extensive evaluation on LLM settings (Sharma et al., 4 Nov 2025) included large models (GPT-OSS series, Qwen3, Kimi-K2) and challenging benchmarks (AIME-2024/2025 for integer math, GPQA-Diamond for graduate-level science MCQ). Key findings:

• Accuracy Gains:
  • Sequential scaling with IEW outperforms parallel self-consistency in 95.6% of tested configurations, with accuracy improvements up to 46.7 percentage points.
  • IEW provides an additional average gain of 1.7 percentage points over sequential majority; its advantage is consistent and never degrades performance compared to simple majority voting within the sequential setting.
• Token Efficiency:
  • Optimal observed efficiency occurs at $t=2,...,K$0 sequential steps: ~13.8 pp gain per 1K tokens (compared to parallel self-consistency’s ~11.7 pp/1K).
  • Sequential scaling retains a robust accuracy margin (+6.6–8.9 pp) over parallel aggregation across a wide range of token budgets.
• Diversity:
  • Sequential methods yield higher lexical diversity (more varied expressions), while parallel methods exhibit greater semantic diversity (different reasoning paths).
• Committee Voting:
  • Even approval-based committees in multiwinner voting are sensitive to small preference perturbations, with on average ≈5/10 replacements for 10% ballot changes in Greedy-CC and Greedy-PAV scenarios (Zech et al., 2024).

The tables below summarize representative results from (Sharma et al., 4 Nov 2025):

Model/Benchmark	Parallel Majority	Sequential IEW	Maximum Gain
Qwen3-235B/AIME-2025	30.0%	76.7%	+46.7 pp
GPT-OSS-120B/GPQA	72.2%	74.2%	+2.0 pp
Kimi-K2/GPQA	73.7%	74.8%	+1.1 pp

6. Design Challenges and Theoretical Landscape

Temporal self-consistency introduces nontrivial complexity-theoretic and design issues:

• LLM Context Accumulation: Sequential reasoning enables history-aware correction but can increase computational requirements per chain. The Markov structure of stepwise conditioning is crucial for leveraging error correction and context integration.
• Voting Rule Complexity: AV’s tractability for overlap-optimal committee selection contrasts with prohibitive complexity for stronger proportional rules (PAV, CCAV) and their greedy counterparts (Zech et al., 2024). Parameterized tractability with respect to limited committee or voter counts is possible, but unrestricted real-world elections require heuristic or domain-specific approaches.
• Tie-Breaking in Committees: Stability is highly sensitive to tie-breaking; naïve lexicographic methods can catalyze unnecessary replacement, while overlap-optimal tie-breaking can substantially mitigate disruption.
• Fallback and Robustness: Systems should revert to majority voting if log-probabilities for entropy computation are unavailable or ill-behaved (Sharma et al., 4 Nov 2025).

7. Open Problems and Prospects

Key open questions and future directions include:

• Axiomatic Characterization: Formalizing stability axioms that capture desirable temporal self-consistency; determining which voting rules satisfy these axioms or inherently support aversion to change (Zech et al., 2024).
• Multi-Stage Extensions: Analyzing and designing temporal self-consistency mechanisms for applications with more than two stages or with partially observed/stochastic preference evolution.
• Algorithmic Improvements: Developing efficient heuristics for proportional multiwinner rules given the intractability of the optimal overlap objective.
• LLM Reasoning Paradigms: Adopting sequential refinement with entropy-guided aggregation as the default paradigm for complex LLM inference; further characterizing tradeoffs between diversity and accuracy, and scaling such methods to broader classes of reasoning and creativity tasks (Sharma et al., 4 Nov 2025).
• Empirical Assessment: Broadening experimental studies to new datasets, LLM architectures, committee election scenarios, and alternative tie-breaking or aggregation schemes.

The field continues to evolve along axes defined by both computational feasibility and the nuanced interplay between stability, accuracy, and adaptability in temporally evolving decision systems.