UCAgents: Unidirectional Convergence for Visual Evidence Anchored Multi-Agent Medical Decision-Making¶

🕒 Published (v1): 2025-12-02 07:20 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

UCAgents is a 3-tier hierarchical multi-agent framework for medical Visual Question Answering (VQA) that replaces open-ended debate with structured, unidirectional convergence to keep reasoning anchored to visual image evidence. It formalizes the tension between visual noise and textual noise as a "dual-noise bottleneck" via information theory, then eliminates it through fixed agent roles and a one-round targeted inquiry protocol. On PathVQA it achieves 71.3% accuracy (+6.0% over MDAgents) while using 87.7% fewer tokens.

Problem¶

Existing multi-agent frameworks for medical VQA (e.g., MDAgents) simulate Multidisciplinary Team (MDT) debates, but open-ended multi-round discussion inflates textual noise $N_t$ and causes "rhetorical drift"—agents change positions through persuasion rather than visual evidence. Single-agent VLMs conversely suffer from visual noise $N_v$ (no cross-validation). The dual-noise bottleneck means suppressing one noise source worsens the other, and neither paradigm maintains tight alignment between textual reasoning and observable image features, undermining clinical trust.

Method¶

UCAgents enforces a strict diversify → verify → converge pipeline across three tiers, with no agent permitted to change its stance:

Tier 1 – Initial Independent Diagnosis: Two identical expert agents ($A_{1-1}$, $A_{1-2}$) independently analyze the image at temperature $\tau=0.7$ with zero inter-agent communication. Divergence is measured as $D = \mathbb{1}[H_{1-1} \neq H_{1-2}]$. Consensus ($D=0$) routes to Tier 2; disagreement ($D=1$) skips to Tier 3.

Tier 2 – Consensus Purification: A single Guidance Expert Agent ($A_2$, $\tau=0.5$) performs bidirectional validation: visual scan → evidence verification → logic verification → independent hypothesis $H_2$. If $H_2 = H_1$ the case terminates; otherwise it escalates to Tier 3.

Tier 3 – Unidirectional Risk Auditing: Two Critical Analyst Agents ($C_1$, $C_2$) are each exclusively assigned to challenge one hypothesis via adversarial "devil's advocate" risk reports $R_i^{risk}$. A Leader Agent $A_L$ then issues a single targeted inquiry $Q_i$ to each analyst (one round only, $\tau=0.1$) to probe specific ambiguities, receives supplemented responses $\hat{R}_i^{risk}$, and performs final arbitration: $$Y^* = A_L(\hat{R}_a^{risk},\, \hat{R}_b^{risk},\, R,\, V,\, T)$$

The framework is grounded in Fano's Inequality bounding diagnostic error $P_e \geq \frac{H(Y) - I(Y;I) - 1}{\log|H|}$, motivating maximization of $I(Y;V)$ and $I(Y;T|V)$ while suppressing entropy inflation from auxiliary agent interactions $M$.

Key Contributions¶

Formalizes the dual-noise bottleneck (visual ambiguity $N_v$ vs. textual drift $N_t$) via information theory, showing how both single-agent and open-debate multi-agent systems fail to jointly minimize them.
Proposes a 3-tier hierarchical architecture with fixed, non-overlapping agent roles and strict unidirectional information flow, preventing position changes and open debate.
Introduces a one-round leader-directed inquiry mechanism that probes ambiguities without triggering entropy inflation, replacing iterative debate with a single question-response exchange per hypothesis.
Achieves superior accuracy on 4 medical VQA benchmarks with significantly fewer API calls and tokens.

Results¶

PathVQA (GPT-4): 71.3% ± 1.3 vs. MDAgents 65.3% ± 3.9 (+6.0%); vs. best single-agent (CoT-SC) 61.2% (+10.1%).
MIMIC-CXR-VQA (GPT-4): 60.3% ± 0.5 vs. MDAgents 55.9% ± 9.1 (+4.4%).
Qwen2.5VL-3B average gain over single-agent baseline: +11.4% across 4 benchmarks.
Qwen2.5VL-72B average gain over MDAgents: +3.5%.
SLAKE-VQA: up to 80.0% with Qwen2.5VL.
VQA-RAD: +8.5% (full-set) and +10.6% (closed-set) gains with Qwen2.5VL.
Token efficiency: 87.7% lower token cost vs. MDAgents; UCAgents uses ~5.38 API calls vs. MDAgents' 10.15 (Qwen2.5VL-7B).
Ablation (LLaVA-7B on VQA-RAD): removing the one-round inquiry step drops Tier-3 accuracy by 27.23% and overall by 15.60%; Tier-2 supervision contributes +3.54%.

Limitations¶

Evaluated only on medical VQA benchmarks; generalization to other domains or open-ended generation tasks is not demonstrated.
Tier-3 introduces 4–5 agent calls with a fixed structure; cases requiring more nuanced multi-round deliberation may be ill-served by the rigid one-round inquiry cap.
Temperature hyperparameters ($\tau \in \{0.1, 0.5, 0.7\}$) are set heuristically with no ablation over temperature schedules.
The route-selection logic (D=0 → Tier 2, D=1 → Tier 3) is binary; no mechanism handles edge cases where both agents agree incorrectly on a clearly ambiguous image.
Open-ended subset evaluation requires GPT-4 to generate answer options, introducing a potential confound in measuring true generalization.

Relevance to Harnesses / Meta-Harnesses¶

UCAgents is a concrete instantiation of a structured multi-agent meta-harness: it does not task any single model with end-to-end reasoning but instead composes a fixed pipeline of specialized agent roles (divergence, verification, adversarial auditing, arbitration) with deterministic routing logic. The information-theoretic framing of when to escalate ($D=0$ vs. $D=1$, $H_2 = H_1$ vs. not) is directly analogous to meta-harness control flow—deciding at runtime which sub-agents to invoke based on intermediate outputs. The 87.7% token reduction demonstrates that harness-level routing (skipping unnecessary tiers) can yield efficiency gains orthogonal to model-level improvements. For harness designers, the one-round inquiry constraint is a notable design principle: bounding inter-agent communication to a single structured exchange per hypothesis controls downstream entropy, a generalizable technique for any multi-agent pipeline where open debate would degrade signal fidelity.