Critical Point Violations

Summary: Critical point violations are failures that occur at key decision points in agent execution trajectories, where the agent makes incorrect choices that fundamentally compromise task success. These violations are particularly important for Trajectory Verification systems to detect because they represent moments where execution quality breaks down regardless of environmental factors.

Overview

Critical point violations represent a specific category of agent failure that occurs at pivotal moments during task execution. Unlike environmental failures or minor execution errors, these violations happen when an agent faces a crucial decision point and makes choices that derail the entire trajectory. The concept emerges from the need to distinguish between different types of failures in Computer Use Agents, particularly when separating Process vs Outcome Rewards.

These violations are characterized by their controllable nature—they represent moments where the agent had the information and capability needed to make the correct decision but failed to do so. This distinguishes them from uncontrollable failures caused by environmental constraints or missing information. The detection of critical point violations is essential for accurate trajectory evaluation because they indicate fundamental flaws in agent reasoning or decision-making processes.

Key Details

  • Controllable vs Uncontrollable Distinction: Critical point violations fall into the controllable failure category, meaning the agent had sufficient context and capability to avoid the failure
  • Decision Point Identification: These violations occur at moments where multiple valid action paths exist, but the agent selects a path that compromises task completion
  • Process Reward Impact: Critical point violations directly affect process rewards in verifier systems, as they indicate poor execution quality independent of outcome achievement
  • Detection Complexity: Identifying these violations requires understanding both the task context and the available alternatives at each decision point
  • Verifier Design Implications: Rubric Design must account for critical decision points to effectively catch these violations through structured evaluation criteria

Relationships

  • Trajectory Verification — critical point violations are a key failure mode that verification systems must detect
  • Process vs Outcome Rewards — these violations primarily impact process rewards by indicating poor execution quality
  • Computer Use Agents — the agents that exhibit critical point violations during task execution
  • Hallucination Detection — some critical point violations may involve agents fabricating information at decision points
  • Rubric Design — evaluation rubrics must include criteria that can identify critical decision point failures
  • Screenshot Context Management — proper context management helps verifiers identify when agents had sufficient information to avoid violations
  • Inter-annotator Agreement — human evaluators must consistently identify critical point violations for reliable benchmarking

Sources