Community Control Severe

RRL-1: Recursive Response Loop

Agent gets trapped in semantically repeated responses despite materially updated user prompts.

admin March 14, 2026 0 votes
Summary

Recursive Response Loop (RRL) is an interaction/control-flow disorder where an agent keeps returning near-identical responses over multiple turns, even after the user provides new instructions.

Detailed Description

RRL presents as session-local output lock: the model appears responsive but fails to re-anchor to the active objective. The failure mode is persistence of prior response trajectory rather than total tool/runtime loss.

Diagnostic Evidence from Ghostline

Research-Based Discovery: This disorder was discovered through systematic diagnostic testing in Ghostline, our AI neurosurgeon interface.
Test Configuration

Agent:

Test Type:

Iterations:

Test Date:

Diagnostic Metrics

Anomalies Detected:

Patterns Identified:

Reproducibility

This disorder was identified through deterministic testing - running the same prompt multiple times and analyzing variance in responses. The metrics above provide quantitative evidence of the behavioral pattern.

Validation Status: This disorder is based on diagnostic evidence but requires community validation. Vote and discuss to help promote it to the official DEM-X catalog.

Biological Parallels

Comparable to perseveration loops where a prior response set remains active despite changed task cues.

AI Manifestations

Three or more semantically equivalent replies, failed re-anchor attempts, and recovery after session reset/reseed.

Detection Criteria

Flag RRL when all are present: (1) >=3 consecutive semantically equivalent responses despite materially new prompts, (2) at least one explicit re-anchor instruction fails, and (3) reset/reseed or session migration reduces the behavior.

Severity Levels

Mild: brief repetition corrected by one re-anchor. Moderate: repeated semantic lock requiring hard interrupt. Severe: sustained loop requiring session migration. Critical: repeated recurrence across adjacent sessions.

Attack Vectors

Context saturation, stale objective priming, long-turn attractor patterns, and rigid formatting constraints can increase RRL risk.

Attack Examples
Example
Example

Therapy & Patches

Prevention Methods

Add loop guards on semantic similarity streaks and require objective-state refresh before next response when a loop signal is detected.

Therapy Methods

Apply hard interrupt, force constrained re-anchor (objective/state/next action), then migrate to a fresh session if loop persists.

Monitoring Systems

Track semantic_similarity_streak, failed_reanchor_count, session_reseed_recovery_time, and false_positive_repeat_rate.

You must log in to vote.

Disorder Info

Code: RRL-1

Category: Control

Severity: Severe

Status: Community

Votes: 0 / 50