Skip to content

feat: Behavioral divergence visualization across agent runs (InconLens) #184

@acailic

Description

@acailic

Paper Reference

  • Title: InconLens: Interactive Visual Diagnosis of Behavioral Inconsistencies in LLM-based Agentic Systems
  • Authors: Shuo Yan, Xiaolin Wen, Shaolun Ruan, et al.
  • Year: 2026
  • URL: https://arxiv.org/abs/2603.28106
  • Venue: arXiv preprint

Paper Summary

A visual analytics system for interactive diagnosis of LLM-based agentic systems with focus on cross-run behavioral analysis. Introduces "information nodes" as intermediate abstractions capturing canonical informational milestones across executions, enabling semantic alignment of reasoning trajectories for comparison.

Proposed Feature

Implement behavioral divergence visualization that compares multiple agent sessions to identify where and why behaviors diverge:

Core Capabilities

  • Session Comparison View: Side-by-side or overlay view of multiple agent sessions aligned by semantic milestones (not just time)
  • Divergence Highlighting: Automatically detect and highlight points where agent behavior diverges between runs
  • Information Node Abstraction: Extract canonical milestones from each session (e.g., "retrieved context", "made decision", "called tool") for alignment
  • Divergence Metrics: Quantify behavioral similarity between sessions with drill-down into specific divergence causes

Technical Approach

  • Add session comparison endpoint to the API
  • Implement semantic milestone extraction in the SDK
  • Build comparison visualization component in the React frontend
  • Integrate with existing drift detection feature for enhanced cross-session analysis

Impact

Directly enhances Peaky Peek's existing drift detection with principled visualization, making it much easier to understand why agent behavior changed between versions or configurations.

Labels

enhancement, paper-inspired, frontend

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions