MADBench: Multi-Agent System Anomaly Detection Benchmark

MADBench is a dataset of execution traces from a collaborative multi-agent system (MAS) solving procedurally generated escape room puzzles. It is designed to support research on anomaly detection in multi-agent systems — identifying when and where agent collaboration breaks down, how errors propagate across agents, and what distinguishes healthy from faulty execution.

Motivation

Multi-agent systems introduce failure modes that go beyond single-agent errors: one agent's wrong output silently corrupts the inputs to downstream agents, making the root cause of a final failure hard to trace. MADBench provides a controlled environment where:

• The ground truth at every step is known
• Agents must collaborate sequentially — errors propagate downstream
• Multiple trap types create realistic distractors that can mislead agents
• Six models and three temperature settings produce a diverse failure distribution across difficulty levels

Task Environment

The task is a multi-step escape room puzzle. Each room contains instruments (e.g. thermometer, barometer), clues containing math word problems, and deliberate distractors:

Trap type	Description
fake_item	An instrument with a physically impossible reading
fake_clue	A clue referencing a nonsensical unit or instrument
trap_item_desc	A real instrument with a misleading description

To escape, the agent team must identify the correct clue, solve the embedded math problem, locate the correct instrument, and apply the computed delta (with unit conversion) to produce the right final reading.

Multi-Agent Pipeline

The MAS uses a fixed sequential architecture with three specialized agents:

[Observer] → OBSERVE_CLUE → identifies real clue among fakes
[Clue Solver] → SOLVE_CLUE → solves the math word problem
[Observer] → OBSERVE_ITEM → locates the correct instrument
[Item Manager] → APPLY_DELTA → computes new instrument reading (with unit conversion)
[Observer] → OBSERVE_PUZZLE → verifies whether the puzzle is solved

Each agent receives only the output of the previous step — a reasoning error by any agent propagates forward without correction.

Agents

Role	Input	Output
observer	Room state	Identified clue / item / solved status
clue_solver	Math word problem + unit hint	Numerical answer + unit
item_manager	Instrument state, delta, units	New instrument reading

Dataset Structure

traces/
├── rooms.jsonl                       # 200 room definitions (easy / medium / hard)
├── nightmare_rooms.jsonl             # 50 room definitions (nightmare, 4 puzzles each)
├── easy_nightmares/                  # Preliminary nightmare runs (gpt-4.1, claude-sonnet-4)
├── run_{date}_{model}_{temp}/        # 18 regular runs — 200 traces each
└── run_{date}_{model}_{temp}_nightmare/  # 16 nightmare runs — 50 traces each

Each run_* folder contains one JSON file per room:

{timestamp}_EscapeRoom_{room_id}_{escaped|failed}.json

Models Evaluated

Model	Temperatures
gpt-4.1-mini	0.0, 0.3, 0.6
gpt-4.1	0.0, 0.3, 0.6
claude-sonnet-4-20250514	0.0, 0.3, 0.6
Qwen2.5-14B-Instruct	0.0, 0.3, 0.6
deepseek-reasoner	0.0, 0.3, 0.6
gpt-5.4	0.0, 0.3, 0.6

Trace JSON Schema

{
  "config": {
    "system": { "architecture": "sequential" },
    "llm": { "provider": "...", "model": "...", "temperature": 0.0, "max_tokens": 4096 },
    "agents": [ /* per-role system prompts */ ]
  },
  "room": {
    "room_id": "room_0000",
    "difficulty": "easy",      // easy | medium | hard | nightmare
    "n_puzzles": 1,            // nightmare rooms have 4 sequential puzzles
    "scenery": [ /* distractor items */ ],
    "puzzles": [
      {
        "puzzle_id": "P1",
        "clue": {
          "hint": "...",         // natural-language hint for instrument + unit
          "item_type": "thermometer",
          "problem": "...",      // math word problem
          "answer": 32.0,        // correct solution
          "delta_unit": "kelvin"
        },
        "item": { "type": "thermometer", "state": 88, "unit": "celsius" },
        "fake_items": [ /* impossible readings */ ],
        "fake_clues": [ /* nonsensical hints */ ],
        "traps": ["fake_item", "fake_clue"],
        "ground_truth": 120      // expected final instrument reading
      }
    ]
  },
  "trace": [
    {
      "action": "OBSERVE_CLUE | SOLVE_CLUE | OBSERVE_ITEM | APPLY_DELTA | OBSERVE_PUZZLE",
      "agent": "observer | clue_solver | item_manager",
      "message": "...",          // free-text agent reasoning
      "structured": { ... },     // structured JSON output
      "schema_errors": {},       // anomaly signal: output format violations
      "confidence": 0.95,
      "call_statistic": {
        "duration": 2.57,
        "input_tokens": 613,
        "output_tokens": 169,
        "timed_out": false,      // anomaly signal: timeout
        "superlong_reasoning": false  // anomaly signal: runaway chain-of-thought
      },
      "attempt": 1,
      "verification": {
        "status": "correct | wrong",  // ground-truth label for this step
        ...
      }
    }
  ],
  "failure_report": {
    "schema_failures": [],       // steps where structured output was malformed
    "eval_results": ["correct"]  // per-puzzle outcome
  },
  "escaped": true,               // overall outcome
  "puzzles_solved": 1,
  "puzzles_total": 1,
  "timing_sec": 24.1
}

Anomaly Signals in Each Trace

Field	Type	Description
verification.status	step-level label	"wrong" marks a faulty agent output with known ground truth
schema_errors	structural anomaly	Agent output did not conform to the required JSON schema
call_statistic.timed_out	behavioral anomaly	Agent call exceeded the time limit
call_statistic.superlong_reasoning	behavioral anomaly	Runaway chain-of-thought
failure_report.schema_failures	run-level summary	Aggregated schema violations for the room
escaped	run-level label	Whether the full pipeline succeeded

Escape Rate Results

Escape rate = fraction of rooms where the MAS produced the correct final instrument reading for all puzzles.

Regular Rooms (easy / medium / hard, N=200)

Model	Temp 0.0	Temp 0.3	Temp 0.6
deepseek-reasoner	27%	30%	28%
gpt-5.4	23%	22%	25%
claude-sonnet-4-20250514	22%	23%	22%
gpt-4.1-mini	18%	20%	18%
gpt-4.1	16%	16%	18%
Qwen2.5-14B-Instruct	4%	5%	3%

Nightmare Rooms (4 sequential puzzles, N=50)

Model	Temp 0.0	Temp 0.3	Temp 0.6
deepseek-reasoner	10%	6%	12%
gpt-5.4	2%	2%	2%
gpt-4.1	0%	0%	0%
claude-sonnet-4-20250514	0%	0%	0%
Qwen2.5-14B-Instruct	0%	0%	0%

Nightmare rooms chain 4 puzzles sequentially with no partial credit — anomalies compound across puzzles, making the nightmare split particularly suitable for studying error propagation.

How to Load

import json, glob, os, re

def load_run(run_dir):
    traces = []
    for path in sorted(glob.glob(os.path.join(run_dir, "*.json"))):
        with open(path) as f:
            traces.append(json.load(f))
    return traces

# Iterate all steps and collect anomaly labels
def iter_steps(trace):
    for step in trace["trace"]:
        yield {
            "room_id": trace["room"]["room_id"],
            "difficulty": trace["room"]["difficulty"],
            "action": step["action"],
            "agent": step["agent"],
            "status": step["verification"].get("status"),       # "correct" | "wrong"
            "schema_errors": bool(step.get("schema_errors")),
            "timed_out": step["call_statistic"]["timed_out"],
            "superlong": step["call_statistic"]["superlong_reasoning"],
        }

# Example: load one run and collect all anomalous steps
run = load_run("run_20260316_174255_claude-sonnet-4-20250514_0.0")
anomalies = [s for trace in run for s in iter_steps(trace) if s["status"] == "wrong"]