SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2110.08207 · ICLR · 2022

Multitask Prompted Training Enables Zero-Shot Task Generalization

Colin Raffel, Stephen Bach, Victor Sanh, Albert Webson, Sutawika Lintang, Zaid Bigscience, Alyafeai, Antoine Chaffin, Irisa Imatag, Arnaud Stiegler

arXiv · PDF · Open in the Atlas

Code that ran

13 of 15 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 1 violated it and 5 have no contract verdict; 7 failures of our instrument; 2 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2110.08207.json · badge: /api/ran/2110.08207.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
bigscience-workshop/promptsource canonical (classifier label)
classifier basis: classifier_2025
3 ran · 0 no recorded run
bigscience-workshop/t-zero canonical (classifier label)
classifier basis not recorded on the link
3 ran · 0 no recorded run
ukplab/arxiv2022-context-injection-stance community (classifier label)
classifier basis not recorded on the link
5 ran · 1 no recorded run
hojjat-mokhtarabadi/promptsource community (classifier label)
classifier basis not recorded on the link
2 ran · 1 no recorded run
FunctionStatusWhere it lives
choice Ran
ran, violated the contract we drafted
bigscience-workshop/promptsource/promptsource/templates.py
code served (permissive licence) · get_code("bb35fcd18dd842fb")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import random

choices = [1, 2, 3, 4, 5]

ARGS = {
    "choices": choices
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
fake_query Ran
ran on a synthesized input
ukplab/arxiv2022-context-injection-stance/src/retrieve/retrieve_t0pp_context.py
code served (permissive licence) · get_code("9b27d71357f65d09")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Analyze the function:
# - prompts: can be str or list of str (function checks type and processes accordingly)
# - model: not used in the function body
# - tokenizer: not used in the function body

# Create stand-in objects for model and tokenizer since they're not used
class StandInModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = torch.nn.Linear(4, 4)
    
    def forward(self, x):
        return self.linear(x)

class StandInTokenizer:
    def __init__(self):
        self.vocab_size = 1000

model = StandInModel()
tokenizer = StandInTokenizer()

# Test with a list of prompts (the else branch)
prompts = ["hello world", "test prompt", "another example"]

ARGS = {
    "prompts": prompts,
    "model": model,
    "tokenizer": tokenizer
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
feature_to_spec Ran
ran; our contract draft was wrong, not the code
bigscience-workshop/t-zero/t0/seqio_tasks/utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from for-ai/parameter-efficient-moe, is pointer only (licence: NONE) · get_code("b7af70f17bfb5e07")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from for-ai/parameter-efficient-moe, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_attribution_stats Ran
ran on a synthesized input
ukplab/arxiv2022-context-injection-stance/src/analysis/analyse_attributions.py
code served (permissive licence) · get_code("aa86817b5ef44488")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import pandas as pd
import numpy as np

torch_seed = 0

# Create a sample DataFrame with 'attribution' and 'fold' columns
# The function expects:
# - frame['attribution']: numeric values to compute mean/std
# - frame['fold']: categorical values (0, 1, 2, 3) for grouping

np.random.seed(0)
data = {
    'attribution': np.random.randn(16),  # 16 samples with random values
    'fold': np.tile([0, 1, 2, 3], 4)  # Repeat fold indices 0-3 four times
}
frame = pd.DataFrame(data)

ARGS = {
    'frame': frame
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_divergence Ran
ran on a synthesized input
ukplab/arxiv2022-context-injection-stance/src/analysis/analyse_topics.py
code served (permissive licence) · get_code("35a244de1ccdbabd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float64, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
from scipy.spatial import distance
from sklearn.feature_extraction.text import CountVectorizer

# Create a simple vectorizer (CountVectorizer has a transform method that returns a sparse matrix)
vectorizer = CountVectorizer(vocabulary={'hello': 0, 'world': 1, 'foo': 2, 'bar': 3})

# Create sample documents
D_p = "hello world hello"
D_q = "foo bar world"

ARGS = {
    "D_p": D_p,
    "D_q": D_q,
    "vectorizer": vectorizer
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy, sklearn — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_indomain_splits Ran
ran on a synthesized input
ukplab/arxiv2022-context-injection-stance/src/analysis/count_dataset_splits.py
code served (permissive licence) · get_code("dd07b1fdb8bc71a0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import pandas as pd
import numpy as np

# Create a small sample DataFrame with the required structure
np.random.seed(0)
n_samples = 20
samples = pd.DataFrame({
    'label': np.random.randint(0, 2, n_samples),
    'feature1': np.random.randn(n_samples),
    'feature2': np.random.randn(n_samples),
})
samples.index = range(n_samples)

ARGS = {
    'samples': samples,
    'seed': 0,
    'stratify': True,
    'dev_test_size': 0.15,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas, sklearn — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
hf_dataset_to_tf_dataset Ran
ran; our contract draft was wrong, not the code
bigscience-workshop/t-zero/t0/seqio_tasks/utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from for-ai/parameter-efficient-moe, is pointer only (licence: NONE) · get_code("de62b3eab371bdf6")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from for-ai/parameter-efficient-moe, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

highlight Ran
ran; our contract draft was wrong, not the code
bigscience-workshop/promptsource/promptsource/templates.py
code served (permissive licence) · get_code("406e3e3a9886da66")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

ARGS = {
    "input": "Hello World"
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
jensen_shannon_divergence Ran
ran on a synthesized input
ukplab/arxiv2022-context-injection-stance/src/analysis/analyse_topics.py
code served (permissive licence) · get_code("f296f02f0a3d87ac")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float64, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
from scipy.spatial import distance

# Create small probability distributions for testing jensen_shannon_divergence
# The function expects p and q to be 1D arrays that sum to 1 (probability distributions)

np.random.seed(0)

# Generate two random probability distributions
p = np.array([0.2, 0.3, 0.3, 0.2])
q = np.array([0.1, 0.4, 0.3, 0.2])

# Normalize to ensure they sum to 1
p = p / np.sum(p)
q = q / np.sum(q)

ARGS = {
    "p": p,
    "q": q
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
most_frequent Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
bigscience-workshop/promptsource/promptsource/templates.py
code served (permissive licence) · get_code("10c4b8cfda850b4c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned list, ndim 1, dtype family int; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from collections import Counter, defaultdict

# Test case 1: List with multiple items, some appearing more frequently
items_test = [1, 2, 3, 2, 4, 2, 5, 3, 3]

# Test case 2: List with tie for most frequent
items_tie = ['a', 'b', 'a', 'b', 'c']

# Test case 3: Empty list
items_empty = []

# Test case 4: All items appear once
items_all_once = [10, 20, 30, 40]

# Test case 5: Single item repeated
items_single = ['x', 'x', 'x']

# Using the first test case for the main ARGS
ARGS = {
    'items': items_test
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family int — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
removeHyphen Ran
ran; our contract draft was wrong, not the code
hojjat-mokhtarabadi/promptsource/promptsource/utils.py
code served (permissive licence) · get_code("0b1cb631e6e315b0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    "example": {
        "key-with-hyphen": "value1",
        "normal_key": "value2",
        "another-hyphenated-key": 42,
        "no_hyphen": [1, 2, 3]
    }
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
renameDatasetColumn Ran
ran; our contract draft was wrong, not the code
hojjat-mokhtarabadi/promptsource/promptsource/utils.py
code served (permissive licence) · get_code("a77010010d56e618")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Dataset. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from datasets import Dataset

# Create a mock dataset with column names containing hyphens
data = {
    'col-one': [1, 2],
    'col-two': [3, 4],
    'normal_col': [5, 6]
}

dataset = Dataset.from_dict(data)

ARGS = {
    'dataset': dataset
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Dataset — a Dataset output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Dataset output the recorded ndim 1 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
strip_whitespace Ran
ran; our contract draft was wrong, not the code
bigscience-workshop/t-zero/t0/seqio_tasks/tasks.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from for-ai/parameter-efficient-moe, is pointer only (licence: NONE) · get_code("11b64b0a85dfb964")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from for-ai/parameter-efficient-moe, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_dataset_builder Not yet run
unverified (harvested, no recorded run)
hojjat-mokhtarabadi/promptsource/promptsource/utils.py
code served (permissive licence) · get_code("e14d51a44fec9477")
How it ran

no invocation recorded.

query Not yet run
unverified (harvested, no recorded run)
ukplab/arxiv2022-context-injection-stance/src/retrieve/retrieve_t0pp_context.py
code served (permissive licence) · get_code("279ad86d58c62ffb")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 distinct functions with an identity, 9 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 9 later) and 6 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
removeHyphen 3 other papers
3 later
2211.01786 (2022-11)
Crosslingual Generalization through Multitask Finetuning
2310.20707 (2023-10)
What's In My Big Data?
2408.03617 (2024-08)
Is Child-Directed Speech Effective Training Data for Language Models?
choice 2 other papers
2 later
2310.20707 (2023-10)
What's In My Big Data?
2408.03617 (2024-08)
Is Child-Directed Speech Effective Training Data for Language Models?
get_dataset_builder 2 other papers
2 later
2310.20707 (2023-10)
What's In My Big Data?
2408.03617 (2024-08)
Is Child-Directed Speech Effective Training Data for Language Models?
highlight 2 other papers
2 later
2310.20707 (2023-10)
What's In My Big Data?
2408.03617 (2024-08)
Is Child-Directed Speech Effective Training Data for Language Models?
most_frequent 2 other papers
2 later
2310.20707 (2023-10)
What's In My Big Data?
2408.03617 (2024-08)
Is Child-Directed Speech Effective Training Data for Language Models?
renameDatasetColumn 2 other papers
2 later
2310.20707 (2023-10)
What's In My Big Data?
2408.03617 (2024-08)
Is Child-Directed Speech Effective Training Data for Language Models?
feature_to_spec 1 other paper
1 later
2309.05444 (2023-09)
Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning
hf_dataset_to_tf_dataset 1 other paper
1 later
2309.05444 (2023-09)
Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning
strip_whitespace 1 other paper
1 later
2309.05444 (2023-09)
Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Large language models have recently been shown to attain reasonable zero-shot generalization on a diverse set of tasks (Brown et al., 2020). It has been hypothesized that this is a consequence of implicit multitask learning in language models' pretraining (Radford et al., 2019). Can zero-shot generalization instead be directly induced by explicit multitask learning? To test this question at scale, we develop a system for easily mapping any natural language tasks into a human-readable prompted form. We convert a large set of supervised datasets, each with multiple prompts with diverse wording. These prompted datasets allow for benchmarking the ability of a model to perform completely held-out tasks. We fine-tune a pretrained encoder-decoder model (Raffel et al., 2020;Lester et al., 2021) on this multitask mixture covering a wide variety of tasks. The model attains strong zero-shot performance on several standard datasets, often outperforming models up to 16× its size. Further, our approach attains strong performance on a subset of tasks from the BIG-bench benchmark, outperforming models up to 6× its size.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2110.08207")
get_code_for_paper("2110.08207")
have("2110.08207")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2110.08207.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2110.08207.svg)](https://syntology.ai/paper/2110.08207)

Connect an agent — have() is free.