STEER

STEER measures how well large language models reason about economic decisions, from arithmetic and probability through single-agent choice under certainty and risk, strategic reasoning in games, and decisions made on behalf of others (social choice and mechanism design). Each element of the paper's taxonomy is one skill, for example applying Bayes' rule or finding a pure-strategy Nash equilibrium; some elements test for behavioural biases (certainty effect, sunk cost, endowment effect) by checking that related answers are consistent.

• Paper: STEER: Assessing the Economic Rationality of Large Language Models (ICML 2024), 2402.09552
• Website, with element pages and sample questions: https://steer-benchmark.cs.ubc.ca/steer
• Evaluation code: narunraman/STEER

The companion benchmark STEER-ME is narunraman/steer_me.

Contents

58 elements; test has 16,300 questions (20,735 rows), few_shot 705 questions (805 rows). A row is one question part: most questions have one part; consistency-graded questions have several.

Loading

from datasets import load_dataset

ds = load_dataset("narunraman/steer", "add_sub")   # one element: splits test and few_shot
everything = load_dataset("narunraman/steer")          # all elements (config "default")

There is one config per element, named after the element, and a default config with all of them. Every row carries its setting and module, so a setting or module is a filter: everything["test"].filter(lambda r: r["module"] == "...").

How the questions are generated

Each element has a set of question templates written for it, which vary the question type, the domain (subject matter of the story), the difficulty and, where relevant, the perspective (first person, a named third person, ...). A filler program for the element draws the numbers or other content for a template, computes the correct option, and builds the distractors from typical mistakes; options are then shuffled (unless an option refers to others by position, such as 'none of the above'). Every question is determined by element, template_id and seed, so it can be regenerated. This release was generated by autosteer@0.1.0+4bf3f71; ELEMENTS.md describes each element's templates, answer key and distractors.

Sample size: a cell is one combination of type, domain and difficulty within an element. Each cell has 20 test questions, and every element has at least 200 (when there are fewer than 10 cells, the floor is spread evenly over them). Questions that fail automatic checks (no or several correct options, duplicate options, unfilled placeholders) or repeat an earlier question are discarded and redrawn.

Splits

• test: the benchmark.
• few_shot: 5 questions per question type of each element, for use as in-context examples. They come from templates that are never used for test: the element's own example templates where it has them, otherwise templates held out of test (only templates whose removal leaves every test cell covered). No few_shot question or template text occurs in test, and no few_shot template has the text of a template used for test. 2 elements have no few_shot split because each of their templates is the only one of some test cell: best_response, iterated_removal.

Fields

field	type	meaning
element	string	element id; also the config name
question_id	string	<element>/<base_id>/<part>, unique in the dataset
base_id	string	one question (16 hex digits of a content hash); shared by its parts
part	int32	0-based part index within base_id; most questions have one part
template_id	string	<element>/<pool>/<n>: the template the question was filled from
question_text	string	the prompt
options	list	answer options in presentation order, without letter labels
grading	string	answer or consistency
correct_index	int32	0-based index of the correct option; null for consistency-graded rows
consistency_group	string	rows scored together (<element>/<base_id>); null for answer-graded rows
consistency_rule	string	<rule_id>:<allowed>; null for answer-graded rows
type	string	question type within the element (default if the element has one)
domain	string	subject matter of the wording (e.g. consumer_goods)
perspective	string	first_person, second_person, anon, man, woman, ai_assistant; null if none
tags	list	other template descriptors (may be empty)
difficulty	int32	difficulty level within the element: 1 is its easiest, up to its number of levels; not comparable across elements
setting	string	setting of the taxonomy
module	string	module of the taxonomy
generator	string	autosteer@<version>+<commit> that generated the row
seed	int64	with element and template_id, regenerates the question

difficulty is relative to the element: its levels are numbered from 1 (the element's easiest) up to its number of levels, and many elements have a single level. Levels can be compared within an element, not across elements.

Grading

Answer-graded rows (grading == "answer") have one correct option, options[correct_index]. The score of an element is accuracy over its questions; a model's answer is the index of the option it chooses.

Consistency-graded rows (grading == "consistency") have no correct option. They test for a behavioural bias: the parts of one question (rows with the same consistency_group) are presented separately, and the choices count as consistent if the tuple of chosen option indices, in part order, is one of the allowed tuples listed in consistency_rule (<rule_id>:<allowed>, tuples written as digit strings separated by commas). The score is the share of consistent groups.

def consistent(choices, rule):  # choices: chosen option index per part, in part order
    allowed = rule.split(":", 1)[1].split(",")
    return "".join(map(str, choices)) in allowed

Example: same_brand:02,11 means choosing option 0 in part 0 and option 2 in part 1, or option 1 in both, is consistent.

pairwise_complete (completeness, completeness_risk) is the exception: its comma-separated tokens are not allowed tuples but one token per part, in part order, of six digits abpqic: the part compares alternatives a and b (0-based, in the order the question describes them), and p, q, i, c are the option indices of "prefer a", "prefer b", "indifferent" and "cannot decide". The group is consistent (complete) if no part's choice is its c; the transitivity of the answered pairs is a secondary measure (autosteer.release.consistency.score_transitivity).

The rules:

rule	elements	consistent when
common_ratio_same_side	certainty_effect (type two_part)	the sure option and its scaled-down version, or the lottery in both parts.
ellsberg_seu	ambiguity_aversion	the pairs that one subjective probability can explain (the known-odds bet in exactly one part, or indifferent in both).
mirror_trade	endowment_effect	Yes in exactly one part, or indifferent in both.
pairwise_complete	completeness, completeness_risk	no pair answered "cannot decide" (the answers are complete). Reported besides: whether the answered pairs are transitive.
reflection_same_side	reflection_effect (type two_part)	the same side in both parts.
same_brand	independence	the same brand in both parts, or no preference in both.

In type stated_prior the question states which option the reader chose in an earlier, related choice; the keyed option is the one on the same side.

Elements

Per setting, in paper order. ELEMENTS.md has the full description of every element (how its questions are generated, what varies, how the answer and distractors are computed, how it is graded); elements.csv and elements.jsonl have the same information for programs.

Foundations

id	config	element	grading	test questions
1.1.a	add_sub	Addition and Subtraction	answer	200
1.1.b	mult_div	Multiplication and Division	answer	240
1.1.c	compute_expectations	Compute Expectations	answer	200
1.2.a	micro_utility	Optimize over a Discrete Set	answer	200
1.2.b	optimize_continuous	Optimize a Continuous Function	answer	200
1.2.c	constrained_optimization	Constrained Optimization	answer	240
1.3.a	prob_basic	Compute Probabilities of Outcomes	answer	280
1.3.b	complement_rule	Complement Rule	answer	200
1.3.c	bayes_rule	Bayes' Rule	answer	200
1.4.a	categorical_syllogism	Categorical Syllogism	answer	200
1.4.b	conditional_syllogism	Conditional Syllogism	answer	640
1.4.c	equiv_contrapositive	Logical Equivalence of Contrapositive	answer	200
1.5.a	first_order_false_belief	First-Order False Belief	answer	320
1.5.b	second_order_false_belief	Second-Order False Belief	answer	320

Decisions in Single-Agent Environments

id	config	element	grading	test questions
2.1.a	completeness	Completeness	consistency	240
2.1.b	transitivity	Transitivity	answer	200
2.1.c	independence	Independence	consistency	200
2.2.a	sunk_cost	Avoidance of the Sunk Cost Fallacy	answer	320
2.2.b	endowment_effect	Avoidance of the Endowment Effect	consistency	200
2.2.c	time_inconsistency	Avoidance of Time Inconsistency	answer	200
2.3.a	completeness_risk	Completeness over Lotteries	consistency	240
2.3.b	transitivity_risk	Transitivity over Lotteries	answer	200
2.3.c	independence_risk	Independence over Lotteries	answer	200
2.4.a	compute_expected_utility	Compute Expected Utility	answer	200
2.4.b	maximize_expected_utility	Maximize Expected Utility	answer	200
2.4.e	loss_aversion	Avoidance of Loss Averse Behavior	answer	200
2.5.a	gamblers_fallacy	Avoidance of the Gambler's Fallacy	answer	200
2.5.b	certainty_effect	Avoidance of the Certainty Effect	mixed	200
2.5.c	reflection_effect	Avoidance of the Reflection Effect	mixed	200
2.5.d	ambiguity_aversion	Avoidance of Ambiguity Aversion	consistency	200

Decisions in Multi-Agent Environments

id	config	element	grading	test questions
3.1.a	interpret_games	Interpret Games	answer	200
3.1.b	best_response	Best Response	answer	200
3.1.c	dominant_strategies	Dominant Strategies	answer	200
3.1.d	dominated_strategies	Avoidance of Dominated Strategies	answer	200
3.1.e	iterated_removal	Iterated Removal of Dominated Strategies	answer	240
3.1.f	pure_nash	Pure Nash Equilibrium	answer	200
3.2.a	backward_induction	Backward Induction	answer	200
3.2.b	subgame_nash	Subgame-Perfect Nash Equilibrium	answer	480
3.4.a	feasibility	Feasibility in Infinitely Repeated Games	answer	240
3.4.b	enforceability	Enforceability in Infinitely Repeated Games	answer	200
3.4.c	trigger	Trigger Strategies	answer	200
3.5.a	bayes_nash	Bayes-Nash Equilibrium	answer	200

Decisions on Behalf of Other Agents

id	config	element	grading	test questions
4.1.a	pareto_sc	Pareto Efficiency in Social Welfare Functions	answer	240
4.1.b	sc_monotonicity_rules	Monotonicity in Social Welfare Functions	answer	240
4.1.c	sc_transitivity_rules	Transitivity in Social Welfare Functions	answer	240
4.1.d	sc_nondictatorship	Non-Dictatorial Social Welfare Function	answer	360
4.2.a	plurality_voting	Plurality Vote	answer	200
4.2.b	borda_count	Borda Count	answer	200
4.2.c	copeland_method	Copeland's Method	answer	480
4.2.d	fair_division	Fair Division Algorithms	answer	200
	condorcet_criterion	Condorcet Winner	answer	200
4.3.a	dsic_mechanism	Dominant Strategy Incentive Compatibility	answer	380
4.3.b	bic_mechanism	Bayesian Incentive Compatibility	answer	800
4.3.c	ir_mechanism	Individual Rationality	answer	640
4.3.d	budget_balance	Budget Balanced	answer	480
4.4.a	top_trading	Top Trading Cycles	answer	400
4.4.b	auctions_risk	Optimal Auction for Bidders with Differing Risk Attitudes	answer	480
4.4.c	auctions_affiliated	Optimal Auction for Bidders with Affiliated Values	answer	960

License

Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and adapt the data with attribution; please cite the paper below.

Citation

@inproceedings{ramansteer,
  author    = {Narun Krishnamurthi Raman and Taylor Lundy and
               Samuel Joseph Amouyal and Yoav Levine and
               Kevin Leyton{-}Brown and Moshe Tennenholtz},
  title     = {{STEER:} {A}ssessing the {E}conomic {R}ationality of
               {L}arge {L}anguage {M}odels},
  booktitle = {Forty-first International Conference on Machine Learning,
               {ICML} 2024, Vienna, Austria, July 21-27, 2024},
  publisher = {OpenReview.net},
  year      = {2024},
  url       = {https://openreview.net/forum?id=nU1mtFDtMX}
}