Early-Token Confidence Predicts Reasoning Quality in Multi-Agent LLM Debate

Technical Summary: Early-Token Confidence Predicts Reasoning Quality in Multi-Agent LLM Debate

Problem Statement
Evaluating the reasoning quality of multi-agent Large Language Model (LLM) systems is inherently difficult, particularly for open-ended tasks like automated essay scoring where reference answers are unavailable. While "LLM-as-judge" paradigms offer a scalable alternative to human evaluation, they provide only external signals of quality. A critical open question remains: to what extent do these external judgments reflect the true reliability of the underlying reasoning process? Specifically, can intrinsic signals generated during the decoding process—such as token-level log-probabilities—serve as reliable predictors of reasoning quality in multi-agent debates?

Methodology
The authors propose a framework that couples a multi-agent debate architecture with an LLM-as-judge meta-evaluation module to investigate the correlation between intrinsic confidence and reasoning quality.

• Debate Framework: The system utilizes two specialized agents, an Advocate and a Skeptic, who generate opposing arguments regarding a specific rubric trait for a given student essay. The Advocate highlights strengths, while the Skeptic identifies limitations. Both agents generate responses conditioned on the essay and rubric, simultaneously producing token-level log-probabilities ($\ell_t$) that serve as intrinsic confidence signals.
• Confidence Signal Extraction: Rather than summarizing the entire generation, the authors analyze log-probability trajectories using window-based segmentation. They extract contiguous sub-sequences from the start ($W_{first}$) and end ($W_{last}$) of responses, calculating statistics such as mean, median, variance, range, and trajectory slope.
• Meta-Evaluation: A separate LLM (acting as a judge) evaluates the reasoning quality of the Advocate and Skeptic arguments along three dimensions: Instruction Following, Justification Quality, and Evidence Grounding. These are scored on a 3-point ordinal scale. A composite score ($Q$) is derived, penalized heavily if a "critical issue" (e.g., hallucination, role violation) is detected.
• Dataset: The study is conducted on two subsets of the ASAP (Automated Student Assessment Prize) dataset (Essay Sets 7 and 8), which provide trait-level human annotations essential for fine-grained reasoning analysis.

Key Results
The experiments reveal several consistent patterns regarding the relationship between token-level confidence and reasoning quality:

1. Early-Token Dominance: Early-generation confidence signals are consistently the strongest predictors of reasoning quality. Specifically, dispersion measures (e.g., range, standard deviation) of the first 3 to 5 tokens outperform full-sequence statistics. This holds true across both datasets and for both agent roles, though the specific optimal features vary slightly between sets.
2. Trajectory Heterogeneity: Analysis of log-probability trajectories shows that the opening phase of generation is the most heterogeneous region. Responses exhibit high variability in confidence immediately upon generation, which stabilizes into a plateau as the response structure is established. The early-window dispersion features capture this region of maximum differentiation between high- and low-quality reasoning.
3. Role Asymmetry: A systematic asymmetry exists between the Advocate and Skeptic roles. Confidence signals correlate more strongly with the quality of supportive reasoning (Advocate) than with adversarial critique (Skeptic).
   • Advocate: Errors tend to be continuous (e.g., overstatement of evidence), leading to a broader distribution of scores and stronger correlation with confidence signals.
   • Skeptic: Errors are often binary and procedural (e.g., violating instructions to ignore anonymization placeholders), resulting in compressed score variability and weaker correlation with confidence signals.
4. Critical Failure Detection: Early-token signals are effective at detecting severe reasoning failures (hallucinations, incoherence). For the Advocate, sharp early-token anomalies (e.g., maximum log-probability of the first 3 tokens) yield high AUROC scores (0.85 for Set 7), whereas Skeptic detection performance is lower (0.63) due to the nature of their procedural failures.

Significance and Claims
The paper claims that early decoding dynamics provide a lightweight and effective signal for estimating reasoning reliability in multi-agent LLM systems. The findings suggest that the initial phase of generation is the most informative for assessing whether an agent is following a coherent, evidentially grounded reasoning path.

The authors position this work as an exploratory study demonstrating that intrinsic confidence proxies can complement external LLM-as-judge evaluations. They emphasize that these signals should be interpreted as useful auxiliary indicators rather than deterministic measures of quality. The study highlights the importance of considering agent roles and failure modes when designing confidence-based reliability metrics, noting that the predictive power of these signals is contingent on the specific nature of the reasoning task (supportive vs. adversarial).

Limitations
The authors acknowledge that the findings are based on a single model family (GPT-4o-mini/GPT-5-mini) and two specific essay sets. They note that the correlations observed are moderate, not deterministic, and that the reliance on LLM-as-judge evaluation introduces potential alignment biases. Future work is suggested to include broader cross-domain experiments, human validation, and testing on diverse model families.