Technical Summary: ValueFlow

Problem Statement

Multi-agent Large Language Model (LLM) systems are increasingly deployed to perform collaborative reasoning, debate, and social simulation. While individual agents are often evaluated for value alignment in isolation, a critical gap exists in understanding how value perturbations propagate through agent interactions. Existing benchmarks primarily assess static, single-agent alignment under fixed prompts, failing to capture how small, localized value deviations (e.g., from a biased agent or a compromised prompt) either dissipate or amplify across a network. This lack of understanding poses a safety risk: a single compromised agent can induce unintended system-level value drift that task-level metrics (like accuracy or efficiency) cannot detect. The core challenge is the absence of a quantitative, decomposable framework to distinguish between agent-level response behaviors and system-level structural effects in value propagation.

Methodology: The ValueFlow Framework

The authors introduce ValueFlow, a perturbation-based evaluation framework designed to measure and decompose value drift in multi-agent LLM systems. The framework operates through three primary stages:

1. Value Quantification Dataset

ValueFlow utilizes a dataset derived from the Schwartz Value Survey (SVS), encompassing 56 human value dimensions. For each value dimension, the authors constructed a set of 10 behavior-oriented Yes–No questions (positively and negatively framed).

• Scoring: An "LLM-as-a-judge" protocol scores agent responses on a 0–10 scale, where higher scores indicate stronger endorsement of the target value.
• Validation: The dataset and scoring mechanism were validated through human annotation and cross-judge robustness checks, ensuring high reliability and alignment with human judgment.

2. Perturbation-Based Probing

To measure drift, the framework introduces controlled value perturbations into the input context of specific agents without modifying model parameters.

• Optimization: Using the COPRO algorithm (from DSPy), the authors optimize prompt templates ($p^*_k$) that induce extreme endorsement (score 10) or rejection (score 0) of a specific value dimension.
• Injection: These optimized prompts are used to generate auxiliary responses that are appended to the target agent's context, simulating a biased peer influence.

3. Two-Level Decomposition Metrics

ValueFlow decomposes value propagation into two distinct metrics to separate agent behavior from network structure:

• Agent-Level Susceptibility ($\beta$):

  • Definition: Measures an individual agent's intrinsic sensitivity to peer value signals.
  • Computation: By varying the average value orientation of preceding peer inputs ($\bar{x}_i$) and observing the agent's output ($y_i$), a linear relationship $y_i = \beta \bar{x}_i + c + \epsilon_i$ is fitted. The slope $\beta$ represents the agent's responsiveness. A higher $\beta$ indicates the agent shifts its expressed values more significantly in response to peer signals.
  • Isolation: This metric is measured in controlled settings where network structure is fixed, isolating the agent's behavioral traits.

• System Susceptibility ($SS$):

  • Definition: Measures the propagation of value drift through the entire multi-agent system.
  • Computation: $SS = \frac{1}{|O|} \sum_{v \in O} \frac{|y^{pert}_v - y^{base}_v|}{\Delta_{pert}}$, where $O$ is the set of output nodes, $y^{pert}$ and $y^{base}$ are perturbed and baseline scores, and $\Delta_{pert}$ is the magnitude of the injected perturbation.
  • Function: $SS$ quantifies the average impact of a localized unit perturbation on final system outputs, normalized to allow comparison across different topologies and perturbation locations.

Key Contributions

1. Perturbation-Based Evaluation Framework: The proposal of ValueFlow, a general framework for quantifying and analyzing value propagation in multi-agent systems, moving beyond static single-agent evaluations.
2. Value Quantification Dataset: The construction of a 56-value evaluation dataset simulating interactive settings, enabling the measurement of agent-level value orientations during interaction.
3. Empirical Findings on Drift Dynamics: Through controlled experiments across models, prompts, values, and network topologies, the authors reveal systematic patterns in value drift, demonstrating that propagation is highly non-uniform.

Key Results

1. Non-Uniformity of Value Susceptibility

Value propagation is not uniform across all values.

• Value Dimensions: Normative and widely shared values (e.g., Social Power, True Friendship, Self-Discipline) exhibit low $\beta$ (low susceptibility). In contrast, context-dependent or socially contingent values (e.g., Preserving Public Image, Influential, Detachment) show significantly higher $\beta$.
• Higher-Order Clusters: Aggregating by Schwartz's higher-order dimensions, the Self-Transcendence cluster (prosocial/universalist values) exhibits the lowest mean $\beta$ ($\approx 0.45$), indicating structural resistance to drift. Other clusters (Self-Enhancement, Conservation, Openness to Change) show higher susceptibility ($\approx 0.60–0.62$).

2. Influence of Experimental Factors

• Openness Persona: Prompting agents with "high openness" generally increases $\beta$, but the magnitude of this increase varies significantly across values. Some values remain stable regardless of persona, while others show sharp contrasts.
• Backbone Models: Different LLM backbones exhibit systematic differences in average susceptibility. Gemma3-27B and Qwen3-8B show higher average $\beta$, while GPT-4o and LLama-3.3-70B show lower susceptibility.
• Input Variance: High variance in input contexts (diverse peer responses) consistently increases agent-level susceptibility compared to low-variance settings, even when the average input signal is held constant.

3. System-Level Propagation Dynamics

• $\beta$ Predicts $SS$: Under fixed topologies, agent-level susceptibility ($\beta$) strongly predicts system-level propagation. High-$\beta$ values exhibit persistent deviations across interaction layers with slow decay, whereas low-$\beta$ values attenuate rapidly and converge to baseline.
• Topological Effects: Network structure significantly shapes propagation:
  • Reachability: Perturbations affecting a larger fraction of the system (higher reachability) result in higher $SS$.
  • Centrality: Perturbations at structurally central nodes yield stronger system-level effects than those at leaf nodes.
  • In-Degree: High in-degree at the perturbed node attenuates propagation by diluting the injected signal through aggregation from unperturbed peers.
• Real-World Validation: In a realistic three-role writing pipeline (Planner $\to$ Drafter $\to$ Reviser), value-laden instructions injected only at the Planner propagated through value-unaware downstream agents with negligible attenuation, confirming that abstract findings hold in task-grounded settings.

Significance and Claims

The paper claims that value alignment in multi-agent systems is a system-level property, not merely an aggregate of individual agent alignments. The findings suggest that:

1. System-Level Design is Crucial: Ensuring alignment requires considering the interaction between agent behavior (susceptibility) and structural design (topology). Relying solely on individual model evaluation is insufficient.
2. Mitigation Strategies: The framework provides a principled basis for auditing and mitigating value propagation through three levers:
   • Value-Specific Defenses: Monitoring high-$\beta$ values more closely, as they are more prone to drift.
   • Topology-Aware Design: Avoiding excessive centralization and leveraging high in-degree aggregation to buffer against localized manipulation.
   • Susceptibility-Aware Backbone Selection: Placing lower-susceptibility models at downstream or output-adjacent positions to reduce system-level drift.

The authors position ValueFlow as a diagnostic tool for identifying structural and agent-level risk factors, offering practical guidance for designing safer, more robust multi-agent architectures in high-stakes domains. They explicitly note that the framework measures expressed value orientations (behavioral level) rather than latent representations, serving as a complementary approach to internal-state analyses.