LLM Degenerate Loop Guardrails

中文说明

A zero-runtime-dependency toolkit for detecting degenerate loops in LLM output and tool calls, producing conservative engineering decisions, and preserving reproducible evidence.

30-second start

Run the included synthetic fixture:

python3 scripts/detect_loop.py \
  --json \
  --timeout 60 \
  --log fixtures/loop_detected.log

The command exits with status 1 when a loop signal is detected. A typical summary looks like this:

{
  "loop_detected": true,
  "details": {
    "type": "consecutive_identical_output"
  }
}

What it detects

The detector is designed for observable signals such as:

• consecutive identical or near-identical output blocks;
• repeated tool calls, including normalized JSON key order;
• suspicious non-consecutive tool-call patterns;
• language drift and related output anomalies when evidence supports them.

Detection results are signals, not model-level explanations. MiMo is the initial case study; observations from MiMo, GLM, and other models must be recorded independently and must not be treated as universal conclusions.

Recovery decisions

Pipe a detector summary into the conservative policy layer:

python3 scripts/detect_loop.py \
  --json \
  --timeout 60 \
  --log fixtures/loop_detected.log \
  > /tmp/loop-summary.json

python3 scripts/recovery_policy.py \
  --summary /tmp/loop-summary.json \
  --retryable \
  --retry-count 0

The policy emits a decision only. The upper-layer runner decides whether to pause, retry, switch models, or request human review under its own safety and idempotency rules.

Installation

The project has no runtime dependencies and currently supports source execution and local CLI installation.

Run from source

git clone https://github.com/xli498/mimo-stable.git
cd mimo-stable
python3 scripts/detect_loop.py --log fixtures/loop_detected.log

Install the CLI from PyPI

python3 -m pip install mimo-stable
mimo-loop-detect --json --timeout 60 --log fixtures/loop_detected.log

Examples

• Basic text detection
• Tool-call detection
• Recovery policy integration

Each example shows inputs, signals, or decisions only. Actual stopping, retrying, model switching, tool execution, and human review remain the responsibility of the upper-layer runner.

Repository layout

scripts/detect_loop.py       # Signal detection
scripts/recovery_policy.py   # Conservative decision output
fixtures/                    # Synthetic regression fixtures
examples/                    # Minimal integration examples
references/                  # Evidence and parameter notes
tests/                       # Behavioral tests

Testing

python3 scripts/check_version.py
python3 tests/test_detector.py
python3 scripts/benchmark_fixtures.py
bash scripts/test_short.sh
bash scripts/test_long.sh

CI tests Python 3.10, 3.11, and 3.12.

Evidence and boundaries

• Fixtures are synthetic or fully redacted.
• Do not commit API keys, tokens, private prompts, user data, private endpoints, or raw production payloads.
• Historical model parameters are case-study evidence, not cross-model defaults.
• The detector and policy layer have no implicit side effects.

License

MIT. See LICENSE.