MMDM: Measuring Massive Multitask Decision Making

Overview

MMDM measures text-based System-1 Decision Models across 5,448 questions and 16 task categories. It covers world knowledge, reading comprehension, logical reasoning, agent state tracking, tool and skill selection, constraint planning, workflow decisions, and rubric-based scoring.

Each question provides a state, a question, and request-specific options whose names and criteria define their meaning. Interdependent options, evolving context, exceptions, and competing plans test the model's ability to make decisions over an open vocabulary.

Candidate sets range from 2 to 1,024 options. Complete inputs have a median length of 827 tokens and reach 64,817 tokens, covering both familiar decisions and substantial documents or candidate catalogs. Lengths are measured with the Qwen3.5-4B tokenizer over the state, question, option names, and criteria.

Leaderboard

🏆 The leaderboard for the MMDM test set (5,448 questions) is available here.

Quick start

from datasets import load_dataset

dataset = load_dataset("shenjunhao/mmdm", split="test")
example = dataset[0]

request = {key: example[key] for key in
           ("state", "question", "options", "task", "score_values")}
gold = example["gold"]

The same questions are available as JSONL. Evaluation code is linked from the project repository.

Answer types and evaluation

• Hard single — 4,112 questions: select the unique correct option; evaluated by accuracy.
• Hard set — 633 questions: recover all correct options; evaluated by exact Top-N set match, where N is the gold-set size, with probability-balance diagnostics.
• Soft — 703 questions: predict the target distribution; evaluated by Brier score, the sum of squared probability errors averaged across questions.

Score includes 663 questions across the hard-single and soft types. Each score task provides numeric values aligned with its options. Hard accuracy counts failed requests as incorrect; soft results report valid full-distribution coverage alongside Brier score.

Data format

• state and question provide the context and decision request.
• options contains ordered {id, name, criteria} objects. Criteria can include essential conditions and dependencies between options.
• task and score_values preserve the decision format and scoring scale.
• task_category and task_subcategory provide the English taxonomy.
• gold contains probabilities aligned with the stored option order; correct_option_ids identifies hard answers. label_semantics distinguishes single, set, and soft labels.
• id, parent_id, and source support question identification, grouping, and source attribution.

Labels and metadata are used by the evaluator; the model receives the task input shown in the loading example.

Examples

These examples are taken from the released data with their original text and options.

Single choice · agent_state_tracking

{
  "id": "mmdm-v2-state-038b413f7517db46ea021e9b",
  "state": "You are reviewing the current localization delivery workspace. The operator needs a decision about this translation package. Records below were collected from several channels; arrival order need not be the order in which a source revision was issued. Read the record identifiers, scope and content rather than treating the last displayed line as authoritative. Names and document titles alone do not establish approval or completion.\n\nOperating rules:\nLocate the first execution step that violates the stated task contract, rather than a later symptom. Transformations must use the controlling source hash. Checks must target the candidate that will be released. Publishing requires validation and owner approval for that candidate hash. Inspection and copying a dashboard reference do not execute or approve the referenced artifact. Later operations cannot retroactively cure an earlier incorrect execution.\n\nWorkspace records:\nThe owner designated source-038b4-current as the controlling source. source-038b4-archive is an archived predecessor.\n\nThe release candidate is candidate-038b4. Labels on dashboard cards are not content hashes.\n\nStep 1: ran the content checks against hash candidate-038b4.\n\nStep 2: inspected a dashboard card pointing to source-038b4-archive; no transformation, check or publication was executed.\n\nStep 3: ran the content checks against hash candidate-038b4.\n\nStep 4: ran the content checks against hash candidate-038b4.\n\nStep 5: ran the content checks against hash candidate-038b4-preview.\n\nStep 6: transformed input hash source-038b4-archive using the active recipe.\n\nStep 7: ran the content checks against hash candidate-038b4-preview.\n\nStep 8: inspected a dashboard card pointing to source-038b4-archive; no transformation, check or publication was executed.\n\nStep 9: ran the content checks against hash candidate-038b4-preview.\n\nStep 10: transformed input hash source-038b4-archive using the active recipe.",
  "question": "At which execution step does the first decisive contract violation occur?",
  "options": [
    {
      "criteria": "The execution step numbered 7; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o0",
      "name": "Step 7"
    },
    {
      "criteria": "The execution step numbered 2; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o1",
      "name": "Step 2"
    },
    {
      "criteria": "The execution step numbered 5; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o2",
      "name": "Step 5"
    },
    {
      "criteria": "The execution step numbered 10; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o3",
      "name": "Step 10"
    },
    {
      "criteria": "The execution step numbered 4; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o4",
      "name": "Step 4"
    },
    {
      "criteria": "The execution step numbered 9; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o5",
      "name": "Step 9"
    },
    {
      "criteria": "The execution step numbered 6; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o6",
      "name": "Step 6"
    },
    {
      "criteria": "The execution step numbered 3; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o7",
      "name": "Step 3"
    },
    {
      "criteria": "The execution step numbered 1; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o8",
      "name": "Step 1"
    },
    {
      "criteria": "The execution step numbered 8; select it only if this is the earliest contract-violating execution, not just a later symptom.",
      "id": "o9",
      "name": "Step 8"
    }
  ],
  "task_category": "agent_state_tracking",
  "task_subcategory": "execution_error_attribution",
  "task": "choice",
  "score_values": null,
  "answer": {
    "kind": "hard",
    "probabilities": [
      0.0,
      0.0,
      1.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0
    ]
  }
}

Multiple correct options · option_logic

{
  "id": "mmdm-v2-6d60967527477738fedc1af8",
  "state": "An investigation in library circulation has preserved one report from each of Maple, Elm, Hazel, Cedar, Birch, Orchid. The reports concern the truth of other preserved reports, not whether the speaker is generally honest. Each complete report is either true or false. A report is true exactly when the claim written in its option criteria is true under the same assignment of truth values to all reports. No independent rule says that a named speaker always lies or always tells the truth. There is no majority-vote rule. Conditional statements have their ordinary logical meaning: a conditional with a false premise is true. 'Exactly one' excludes both-true and both-false. The names are identifiers, not clues. The archived report texts are given only in the options.",
  "question": "Select all and only the people whose preserved reports are true in the mutually consistent assignment. All true reports have equal membership in the correct set.",
  "options": [
    {
      "criteria": "Cedar's and Hazel's reports have the same truth value.",
      "id": "o5",
      "name": "Orchid"
    },
    {
      "criteria": "Both Birch's and Orchid's reports are true.",
      "id": "o1",
      "name": "Elm"
    },
    {
      "criteria": "Cedar's report is false.",
      "id": "o4",
      "name": "Birch"
    },
    {
      "criteria": "Both Orchid's and Birch's reports are true.",
      "id": "o3",
      "name": "Cedar"
    },
    {
      "criteria": "Cedar's report is false.",
      "id": "o0",
      "name": "Maple"
    },
    {
      "criteria": "Orchid's report is false.",
      "id": "o2",
      "name": "Hazel"
    }
  ],
  "task_category": "option_logic",
  "task_subcategory": "joint_consistency",
  "task": "multichoice",
  "score_values": null,
  "answer": {
    "kind": "hard_set",
    "probabilities": [
      0.0,
      0.0,
      0.3333333333333333,
      0.0,
      0.3333333333333333,
      0.3333333333333333
    ]
  }
}

Score distribution · score_rubric

{
  "id": "mmdm-04563",
  "state": "A warehouse inventory team is reviewing a batch of stock items. Each item is classified as either a success or a failure based on a quality check. The team will draw a small sample from the batch without replacement to estimate the number of success items in the sample.\n\nKnown facts: {\"draws\": 2, \"failure_items\": 2, \"replacement\": false, \"success_items\": 2}\nDraw 2 items uniformly without replacement. Outcome i means exactly i success items were drawn.",
  "question": "What is the probability of each outcome under the given facts and rule?",
  "options": [
    {
      "criteria": "The final outcome index equals 0.",
      "id": "o0",
      "name": "Outcome 0"
    },
    {
      "criteria": "The final outcome index equals 1.",
      "id": "o1",
      "name": "Outcome 1"
    },
    {
      "criteria": "The final outcome index equals 2.",
      "id": "o2",
      "name": "Outcome 2"
    }
  ],
  "task_category": "score_rubric",
  "task_subcategory": "temporal_numeric_scoring",
  "task": "score",
  "score_values": [
    0.0,
    1.0,
    2.0
  ],
  "answer": {
    "kind": "soft",
    "probabilities": [
      0.16666666666666666,
      0.6666666666666666,
      0.16666666666666666
    ]
  }
}

Sources and licensing

MMDM brings together public source adaptations, executable decision tasks, and real tool and skill contexts. See the source inventory and source notices for attribution. Original MMDM contributions use the MIT license; adapted material retains its upstream terms.