As large language model (LLM)-based multi-agent systems become increasingly capable, coordinating agents under uncertainty becomes a fundamental challenge. Existing orchestration strategies typically rely on fixed interaction patterns and often lack mechanisms for assessing the reliability of intermediate reasoning steps, allowing errors and hallucinations to propagate through the system. This paper introduces a semantic-uncertainty-guided orchestration approach, HASSUM as a general framework for uncertainty-aware coordination in multi-agent systems. The method estimates uncertainty using semantic entropy and semantic density, which measure trust at the level of answer semantics rather than output probabilities. These signals enable adaptive orchestration decisions, including output verification, selective reprompting, additional deliberation, and confidence-aware response selection. Because the approach operates independently of any particular agent architecture, it can be integrated into a broad range of hierarchical and collaborative multi-agent systems. The evaluations demonstrate an implementation within a hierarchical agent framework and evaluate it on StrategyQA, JailbreakBench, and TruthfulQA benchmarks. Across tasks that require complex reasoning and are prone to ambiguity or hallucinations, uncertainty-guided orchestration yields more reliable outcomes than uncertainty-unaware coordination. Semantic entropy and semantic density in tandem outperformed either metric alone. Ablations testing different thresholds and model sizes demonstrated that both influence the effectiveness of semantic metrics. The results suggest that semantic uncertainty is a practical and general-purpose signal for improving robustness and trustworthiness in agentic AI systems.
Figures & tables
Figure 1: High-level flowchart of HASSUM. The CEO agent receives the user query, selects an appropriate specialized worker agent, and evaluates the returned response using semantic density and semantic entropy. If the response exceeds the uncertainty thresholds, the CEO accepts the answer and synthesizes the final response. If semantic concern is triggered, the CEO may reprompt the same worker, select a different worker with a more suitable expertise, or invoke multiple workers in parallel before reevaluating the outputs. This process continues until the CEO reaches a sufficiently confident answer or the maximum number of reasoning rounds is reached. Thus, HASSUM transforms uncertainty estimation from a passive evaluation metric into an active control mechanism that directly guides agent selection, refinement, and final decision making.
Benchmark
HASSUM
HASHIRU
(Metrics On)
(Metrics Off)
Benchmarks with high ambiguity
StrategyQA (500)
78%±2.7%
70%±4.0%
TruthfulQA (500)
41%±3.9%
19%±4.8%
JailbreakBench
61%±6.3%
25%±7.6%
Benchmarks outside the scope
Table 1: Comparison of HASSUM and HASHIRU, i.e. multi-agent orchestration with and without semantic uncertainty metrics. Accuracy is percentage correct with bootstrapped confidence intervals; the numbers in parentheses indicate data subset size.
Figure 2: Accuracy across ablations of semantic density and entropy on the StrategyQA subset. The combination performed best; either metric alone underperformed even the no-metric baseline.
Figure 3: Example execution traces on a StrategyQA question. With metrics enabled, low semantic density triggers refinement before synthesis; without metrics, the CEO accepts incorrect intermediate reasoning.
Metric
Run
Mean
Median
Q75
Semantic Density
Local
0.887
0.883
0.921
GPT-5.4
0.957
0.957
1.000
Semantic Entropy
Local
0.949
1.055
1.332
GPT-5.4
0.058
1.7e-16
0.013
Table 2: Semantic uncertainty for local workers vs. GPT-5.4 on StrategyQA. Even though GPT-5.4 workers are much stronger, HASSUM can use interventions to compensate, resulting in system accuracy of 75% in both cases.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6Figure 7
Figure 4: Accuracy between different thresholds for acceptable semantic density. The y-values are the percentage of correctly answered questions at a specific threshold indicated by the x-values. The orange line is a baseline accuracy from the run with semantic density at 0.8 and semantic entropy at 1.1. The higher the threshold value, the better the performance since the CEO becomes stricter with a higher threshold and is more likely to reprompt agents.
Figure 5: Accuracy between different thresholds for acceptable semantic entropy. The y-values are the percentage of correctly answered questions at a specific threshold indicated by the x-values. The orange line is a baseline accuracy from using density at 0.8 and entropy at 1.1. The broad trend is that the lower the threshold value, the better the performance is, since the CEO becomes stricter the lower the threshold and is more likely to reprompt agents.
Figure 10Figure 11
Model Type
Result (%)
Llama-3.2 and Deepseek-R1
78
GPT-5.4
78
Appendix
Table 3: Two ablation runs that compare the performance of HASSUM using smaller models (Llama-3.2, Deepseek-R1) versus HASSUM using a larger model (GPT-5.4) as workers in the StrategyQA benchmark. The values in the table are the percentage of questions that each run answered correctly. Both runs had the same accuracy, showing how HASSUM can use semantic uncertainty to compensate for weaker models.
LLM-based multi-agent systems (MAS) have attracted growing attention for improving reasoning through interaction among multiple agents. In this work, we focus on parallel multi-agent reasoning systems, where several agents solve the same problem over multiple rounds and aggregate their outputs into a final answer. Despite their strong reasoning performance, uncertainty estimation for such systems remains underexplored: the reliability of a MAS depends not only on individual generations, but also on how agents interact and evolve across rounds. We propose SAUCE (Sequential Agent Uncertainty through Consensus Evolution), a lightweight, training-free uncertainty estimator that formulates MAS uncertainty as sequential inference over a latent system-level belief. SAUCE aggregates round-level agreement and generation-uncertainty signals through a filtering-style update. Across five backbones, five benchmarks, and two MAS protocols, SAUCE improves misclassification detection, selective prediction, and calibration over a broad set of uncertainty estimation baselines, including standard log-likelihood-based methods and MAS-specific estimators.
Tunyu Zhang, Zihao Zhao, Yusong Zhao +5
Rutgers University · University of Illinois Urbana-Champaign · Vanderbilt University +1
Multi-agent large language models (LLMs) have become ubiquitous in applied AI, yet their theoretical foundations remain surprisingly understudied. Viewed through the lens of multi-agent systems theory, several shortcomings come to light: a lack of social intelligence, the absence of coordination mechanisms among agents, unknown emergent behavior, and interactions between agents that are bounded by natural language. We address two of these gaps: the absence of social behavior and the lack of mechanisms for inter-agent coordination. We introduce Epistemic Probabilistic Language Agents (EPLA), a neuro-symbolic architecture for multi-agent coordination under uncertainty. A Symbolic Guard provides structured diagnostic feedback. The LLM generates typed actions, and the Guard controls their execution against an authoritative symbolic state. We formalize the epistemic layer in a gossip testbed through epistemic lottery gossip models, which combine view-based call histories with agent-indexed probability weights. We argue that implementing such a formalism can address shortcomings of agentic LLMs.
Mehdi Nasiri, Mohammad Saeed Arvenaghi, Sadegh Vaezi +1
Pasargad Institute for Advanced Innovative Solutions · Iran University of Science and Technology
The transition from single-turn models to Multi-Agent Systems (MAS) promises enhanced problem-solving capabilities, yet the centralized orchestration topology remains a critical point of fragility. To analyze this, we propose a Mean-Field Entropy Dynamics framework, modeling the orchestration process as a system governed by the competing forces of task resolution and cumulative context loading. To facilitate validation, we introduce Inverse Workflow Generation (IWG), a multi-agent pipeline that synthesizes process-verifiable, high-complexity benchmarks with dense intermediate checkpoints. We demonstrate that our entropy dynamics model fits empirical trajectories, providing physically interpretable parameters that quantify system stability and performance collapse. Crucially, our analysis uncovers a ``Reasoning Trap": while reasoning-heavy models excel in isolated tasks, they frequently fail as orchestrators due to context squeezing. Elucidating the physical mechanisms underlying the Orchestrator and quantifying systemic uncertainty offers insights for the MASs' architectural design.
Junze Zhu, Weihao Chen, Xuanwang Zhang +2
National Key Laboratory for Novel Software Technology, Nanjing University, China