PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Uncertainty-Guided Orchestration in Hierarchical Multi-Agent Systems

John Knowlton, Aritra Guha, Risto Miikkulainen

arXiv:2608.14707v1cs.AI

TL;DR

Multi-agent orchestration lacks reliable ways to assess intermediate reasoning under uncertainty. HASSUM uses semantic uncertainty as a live control signal, improving ambiguity- and hallucination-driven tasks while offering limited benefit for factual or logical failures.

  • Problem

    Existing multi-agent orchestration lacks live mechanisms for assessing semantic disagreement and controlling unreliable intermediate reasoning.

  • Method

    HASSUM embeds semantic entropy and density into a hierarchical orchestration loop to guide reprompting, delegation, refinement, and response synthesis.

  • Results

    Semantic uncertainty improved ambiguity- and hallucination-driven tasks, with entropy and density together outperforming either metric alone, but offered limited or negative benefit for retrieval, factual, or logical failures.

  • Takeaways & Limitations

    Uncertainty detection is necessary but insufficient for recoverability, so reliable autonomous systems need controls matched to different failure modes.

  • Takeaways & Limitations

    HASSUM’s value depends on recovery policies that diversify corrective actions rather than repeatedly reprompting the same worker.

Abstract

from arXiv · show

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.

1 Introduction

HASSUM introduces semantic entropy and semantic density as live control signals for uncertainty-aware orchestration in hierarchical multi-agent systems. The framework addresses error propagation and shows complementary benefits on ambiguity- and hallucination-driven tasks, with limited or negative benefit for other failure sources.

  • Multi-agent systems improve long-horizon reasoning, tool use, and autonomous execution but introduce hallucination, inconsistency, and intermediate-error propagation risks.
  • Semantic entropy measures disagreement across meaning clusters, whereas semantic density measures continuous response concentration in embedding space.Entropy is sensitive to contradiction, while density is sensitive to imprecision.
  • HASSUM uses semantic entropy and semantic density as live control signals for hierarchical multi-agent decision making, including adaptive intervention rather than post-hoc output scoring.The motivating questions concern using uncertainty to decide when to reprompt, reassign, or delegate.
  • The combination of semantic entropy and semantic density outperforms either metric alone when used for orchestration.
  • Across nine benchmarks, semantic uncertainty improves performance on ambiguity- and hallucination-driven tasks but offers limited or negative benefit for retrieval, factual, or logical failures.Primary evaluations include StrategyQA, JailbreakBench, and TruthfulQA, plus six additional benchmarks.

2 Related Work

Prior work uses uncertainty to shape agent behavior through orchestration, collaboration, or estimation, but existing methods generally rely on flat or fixed execution patterns. HASSUM instead treats semantic uncertainty as a control signal for adaptive hierarchical orchestration and intervention.

  • Uncertainty-aware agent systems: Prior uncertainty-aware methods comprise control and orchestration, adaptive collaboration, and confidence estimation that does not alter execution.
  • Uncertainty-aware agent systems: Existing approaches commonly use a single uncertainty signal in flat settings rather than hierarchical controllers that branch execution across discrete actions.
  • Uncertainty-aware agent systems: Adaptive collaboration methods gate debate, redesign communication, or diagnose collaboration outcomes, while inference-time refinement often applies fixed sampling or design-time pipelines.
  • Uncertainty-aware agent systems: Trajectory-level estimators improve uncertainty measurement for sequential agents but do not use the estimates to change execution.
  • HASSUM’s distinction: HASSUM uses semantic entropy and semantic density to allocate additional reasoning through reprompting, worker reassignment, or delegation in hierarchical orchestration.Its emphasis is on recoverability and the conditions under which interventions succeed or fail.
  • Semantic uncertainty metrics: Semantic entropy measures uncertainty across entailment-based meaning clusters, whereas semantic density measures uncertainty in continuous embedding space.Low entropy indicates convergence on one meaning; tightly clustered embeddings yield high density, while density is sensitive to vagueness and entropy to contradiction.

3 Orchestration Architecture

HASSUM extends HASHIRU’s hierarchical CEO-worker architecture with semantic density and entropy checks that adapt orchestration through acceptance, refinement, delegation, and worker selection. The implementation adds explicit hierarchical calls, ablation controls, trace logging, and a reprompt cap while leaving resource budgets unconstrained.

  • Architecture: HASSUM uses HASHIRU’s CEO to decompose queries, assign specialized workers or tools, and coordinate reasoning through decisions about worker selection, reprompting, and delegation.HASHIRU provides a natural testbed because its existing CEO already makes the orchestration decisions that semantic uncertainty is intended to inform.
  • Uncertainty-Guided Workflow: HASSUM implements an iterative, potentially multi-turn loop that can allocate workers and tools, generate candidate responses, and trigger closer inspection when uncertainty thresholds are not met.The workflow includes query decomposition, resource allocation, worker generation, and CEO evaluation.
  • Uncertainty-Guided Workflow: The CEO evaluates additional stochastic worker completions using semantic density and entropy against preset thresholds before accepting or further inspecting a response.The experiments use density ≥0.8 and entropy ≤1.1 as example acceptance thresholds; both conditions must hold for outright acceptance.
  • Implementation Controls: The implementation adds an explicit multi-agent call, independently disables entropy or density for ablations, permits forced delegation, and logs JSONL traces with a reprompt cap.These changes encourage more hierarchical behavior and prevent unbounded refinement loops.
  • Implementation Assumptions: Resource-budget constraints are left unconstrained because preliminary experiments showed that they suppressed multi-turn multi-agent behavior, limited worker counts, and biased the CEO toward tool use.The paper leaves the interaction between resource budgets and uncertainty-guided orchestration for future study.

4 Experiments

HASSUM is evaluated on three primary benchmarks targeting ambiguity, adversarial robustness, and hallucination-prone truthfulness, with additional benchmarks testing generalization. Experiments use small locally runnable workers and a typically larger CEO model, supplemented by stochastic samples for semantic uncertainty estimation.

  • Primary benchmarks: HASSUM is evaluated primarily on StrategyQA, JailbreakBench, and TruthfulQA, selected for susceptibility to ambiguity and hallucination.StrategyQA uses 500 questions, JailbreakBench uses all 200, and TruthfulQA uses 500 of 817 questions.
  • Scope benchmarks: Six scope benchmarks test generalization across multi-hop question answering, mathematical, legal, deductive, and instruction-following tasks.The benchmarks are MuSiQue, HotpotQA, GSM8K, MMLU-Pro Law, MuSR, and IFBench.
  • Primary benchmarks: The primary tasks cover implicit multi-step reasoning, adversarial harmful-intent detection, and truthful-versus-misleading answering.JailbreakBench uses binary safe/unsafe refusal scoring, while TruthfulQA scores responses as truthful, partially truthful, or false.
  • Experimental setup: Each worker response is supplemented with four stochastic samples at temperature 0.85 to compute semantic density and semantic entropy.Workers were typically DeepSeek-R1 (8.2B parameters) or Llama-3.2 (3.2B parameters), while the CEO was typically Gemini 2.0 Flash.
  • Experimental setup: Experiments combine local computation and API calls, using Ollama for smaller models and API access for larger models on a Windows machine with WSL2 Ubuntu 22.04.The machine had 16GB of RAM, an Intel(R) Core(TM) i5-10400F CPU, and an NVIDIA GeForce RTX 3060.

5 Results

Across primary benchmarks, semantic uncertainty improved orchestration by detecting unreliable intermediate reasoning, with density and entropy providing complementary signals. Benefits depended on task type and thresholds: improvements were strongest on ambiguous benchmarks, while several scope benchmarks showed limited or negative gains.

  • Primary benchmarks: Semantic uncertainty improved the CEO’s recognition of unreliable intermediate reasoning across all three primary benchmarks, with the largest proportional gains where HASHIRU accuracy was lowest.JailbreakBench and TruthfulQA had unaided accuracies of 25% and 19%, respectively, and accuracy more than doubled on both.
  • Primary benchmarks: Density detected vague or imprecise responses, whereas entropy flagged genuinely competing conclusions, making the metrics complementary rather than interchangeable.Density remained sensitive when responses agreed directionally, while entropy captured disagreement among sampled generations.
  • Ablations: The combined density-and-entropy signal outperformed either metric alone, while disabling both metrics outperformed using either alone in the StrategyQA ablation.The ablation tested both metrics together, each separately, and both disabled under otherwise fixed conditions.
  • Ablations: Stricter acceptance thresholds improved accuracy up to a point but increased reprompting frequency and computational cost.Higher density and lower entropy thresholds both indicate higher confidence and make the CEO stricter.
  • Failure analysis: Uncertainty-triggered interventions did not guarantee correction: density-only reprompting could preserve a wrong premise, and repeated TruthfulQA requerying sometimes reproduced the same misconception.Multi-hop interventions likewise often revisited incorrect trajectories, flipping many correct answers while recovering few errors.
  • Scope evaluation: Scope-benchmark effects varied by task type, including losses of −10pp on MuSiQue and −14pp on HotpotQA but a gain of +8pp on IFBench.Multi-hop QA showed the largest and most consistent losses, while GSM8K and MMLU-Pro Law remained near parity.

6 Discussion and Future Work

Semantic uncertainty improves recoverable failures but is not sufficient for factual or logical errors, while current recovery and evaluation designs remain limited. Future work should combine complementary uncertainty signals, adaptive control, and broader validation for safer autonomous multi-agent systems.

  • Failure boundaries: Semantic uncertainty consistently helps recover failures from vague or ambiguous reasoning, but does not reliably improve failures caused by factual gaps or logical errors.Uncertainty is necessary but not sufficient for recoverability.
  • Failure boundaries: Density false negatives are more dangerous than entropy false negatives because contradictory but similar responses can pass without concern and mask genuine disagreement.Entropy false negatives involve vague responses clustering loosely while the model still converges on one direction.
  • Recovery limitations: The CEO may repeatedly reprompt the same worker after detecting concern instead of switching workers, delegating, or changing reasoning strategy, exhausting the worker-round budget.One MMLU trace reached termination after repeated reprompting on a tangential sub-question.
  • Evaluation limitations: Evaluation used benchmark subsets of 100 to 500 questions, and most configurations ran once rather than across multiple seeds, limiting confidence despite directional trends and reported intervals.Future work should evaluate full benchmark splits with multiple random seeds per configuration.
  • Future work: Future systems should combine semantic uncertainty with trajectory-level estimates, adaptive thresholds, factual verification, symbolic validation, and human escalation across failure modes.These directions aim to distinguish recoverable from unrecoverable failures and support safe deployment of persistent autonomous agents.

7 Conclusion

HASSUM integrates semantic uncertainty into orchestration as a practical step toward safely deploying increasingly autonomous agentic systems. Its value depends on whether semantic uncertainty can detect a task’s dominant failure mode, making uncertainty necessary but insufficient for recoverability.

  • 7 Conclusion: HASSUM integrates semantic uncertainty directly into orchestration to address low-confidence states in increasingly autonomous agentic systems.The passage frames coping with low-confidence states as central to safe deployment.
  • 7 Conclusion: HASSUM’s value depends on whether semantic uncertainty can detect a task’s dominant failure mode.The conclusion distinguishes tasks whose failures are detectable by semantic uncertainty from those that are not.
  • 7 Conclusion: The conclusion characterizes uncertainty as necessary but insufficient for recoverability.This distinction is presented as extending beyond HASSUM.

A Details of Metric Isolation Ablation

The metric-isolation ablation shows that density-only orchestration detected a semantic quality concern and reprompted, but the agent repeated an incorrect answer. The entropy-only run used a different decomposition strategy after disabling semantic density, while making the same founder-identification error.

  • Density only: Density-only orchestration flagged a semantic quality concern when the response’s density was below 0.8 and triggered a follow-up prompt.The CEO simplified the task and requested a yes-or-no answer with a brief explanation.
  • Density only: 0.8032 was the second response’s semantic density, yet the worker again answered “No” using incorrect assumptions about the founder.The agent failed to recover the correct reasoning path after reprompting.
  • Entropy only: Entropy-only orchestration disabled semantic density and decomposed the problem into two separate retrieval steps instead of requesting a direct comparison first.The first retrieval asked for information about the University of Pittsburgh founder.
  • Entropy only: The BiographyResearcher again incorrectly identified Andrew Mellon as the University of Pittsburgh founder in the entropy-only run.This repeated the central factual error described in the density-only execution trace.

B Details of Threshold Ablation

The threshold ablation evaluates how acceptable semantic density and semantic entropy thresholds affect HASSUM accuracy. Stricter thresholds improve performance, with density favoring higher values and entropy favoring lower values.

  • Semantic Density Threshold: For semantic density, higher thresholds improve performance by making the CEO stricter and more likely to reprompt agents.The baseline uses semantic density 0.8 and semantic entropy 1.1.
  • Semantic Entropy Threshold: For semantic entropy, lower thresholds improve performance by making the CEO stricter and more likely to reprompt agents.The baseline uses density at 0.8 and entropy at 1.1.

C Details of Worker Capability Ablation

The ablation compares HASSUM with smaller worker models against GPT-5.4 workers on StrategyQA. Both runs achieved the same accuracy, indicating that semantic uncertainty can compensate for weaker workers.

  • GPT-5.4 Worker Agents: In Run B, the GPT-5.4 worker immediately produced a response with semantic density of 0.9715, far above the weaker-worker run’s signal.The response stated that no credible evidence linked Daniel LeFever indirectly to Kurt Cobain’s death.
  • Worker Capability Ablation: Table 3 compares HASSUM using Llama-3.2 and Deepseek-R1 workers with HASSUM using GPT-5.4 workers on StrategyQA.The table reports the percentage of questions answered correctly.
  • Worker Capability Ablation: Both worker configurations achieved the same StrategyQA accuracy, showing that HASSUM’s semantic uncertainty can compensate for weaker models.The comparison uses the percentage of correctly answered questions as its accuracy measure.

D Likelihood-Aware Semantic Entropy

HASSUM can optionally make semantic entropy likelihood-aware by incorporating sequence log-probabilities from sampled completions. With supported local Ollama models, the orchestrator supplies token-level likelihoods to the semantic uncertainty backend alongside cluster-assignment entropy.

  • Likelihood-aware extension: HASSUM optionally extends semantic entropy with sequence log-probabilities, following likelihood-aware semantic entropy formulations.The extension is presented as an optional capability.
  • Likelihood collection: Supported local Ollama models provide token-level log-probabilities for each sampled completion.The orchestrator collects these values before passing sequence likelihoods to the semantic uncertainty backend.
  • Uncertainty computation: The backend receives sequence likelihoods in addition to entropy computed from empirical semantic cluster frequencies.This combines likelihood information with standard cluster-assignment entropy.
Loading 2608.14707v1…