CounterBench: LLM Counterfactual Benchmark

• CounterBench is a structured benchmark designed to evaluate counterfactual reasoning in LLMs using deterministic structural causal models.
• It includes 1,000 formally specified queries across basic, joint, nested, and conditional types to assess algorithmic what-if reasoning.
• The introduced CoIn algorithm dramatically improves LLM performance by guiding systematic search and backtracking for complex symbolic inferences.

CounterBench is a structured benchmark specifically designed to evaluate the counterfactual reasoning capabilities of LLMs within the framework of deterministic Structural Causal Models (SCMs). It offers a formally specified battery of 1,000 counterfactual queries with diverse causal-graph structures and controlled complexity, enabling precise assessment of LLM performance on algorithmic what-if reasoning, independent of world knowledge memorization. The benchmark demonstrates that leading LLMs perform near random on these formally specified tasks under standard and causal chain-of-thought prompts, and introduces CoIn, an explicit counterfactual reasoning algorithm, which dramatically improves LLM accuracy by guiding systematic search and backtracking (Chen et al., 16 Feb 2025).

1. Formal Task Definition and Structural Causal Models

CounterBench is grounded in the SCM formalism as described by Pearl (2009). An SCM is represented as $M = \langle U, V, f \rangle$ with three components:

• $U = \{U_1, \ldots, U_m\}$: exogenous (noise) variables,
• $V = \{V_1, \ldots, V_n\}$: endogenous variables,
• $f = \{f_1, \ldots, f_n\}$: deterministic structural equations, $V_i = f_i(\mathrm{Pa}(V_i), U_i)$ for parents $\mathrm{Pa}(V_i) \subseteq V$.

An intervention $do(X=x)$ fixes variable(s) $X \subseteq V$ to value(s) $x$, yielding a modified model $M_x$. The fundamental query format is $Y_x(u)$: given exogenous context $u$, what would the outcome $Y$ have been under the intervention $X = x$? CounterBench includes more complex query templates such as:

• Joint: $Y_{x,z}(u)$ (multiple simultaneous interventions),
• Nested: $Y_{Z_x}(u)$ (sequential interventions, with downstream effects),
• Conditional: $Y_x(u)\mid Z_x(u)=z$ (interventional outcome, conditioned on a side-constraint).

Each counterfactual question is paired with the binary correct answer ($\text{yes}$/$\text{no}$) determined by the specified SCM and context.

2. Dataset Construction and Characteristics

The CounterBench dataset comprises 1,000 questions, partitioned for diagnostic granularity:

Query Type	Number	Example Structure
Basic	250	Single-step chain $X \to V_1 \to \ldots \to Y$
Joint	250	Multiple parents ($V_1$, $V_2$) to $V_3$
Nested	250	Stepwise, e.g., $Y_{V_3_{X=0}}$
Conditional	250	Chain with hurdle/observation on $V_1$

Difficulty levels are set by number of variables (5–9; 200 questions per level, balanced yes/no), inducing complexity scaling in inference. Causal graphs feature AND/OR/NOT combinatorics and chain/multi-parent/nested/conditional dependencies. All variable names use randomly generated tokens (e.g., "Ziklo", "Blaf") to prevent aliasing to memorized knowledge and enforce structural, context-independent reasoning. Query templates are designed to explicitly formalize the causal structure and the intervention or observation at issue.

3. Experimental Evaluation and Metrics

Evaluation runs multiple state-of-the-art LLMs, including GPT-3 (Davinci-002, Babbage-002), GPT-3.5 Turbo, GPT-4o (full and mini), Claude-3 (Sonnet, 3.5 Haiku), Gemini-1.5-flash/8B, and Deepseek-V3. Each model is assessed under several prompting strategies:

1. Standard: default prompt, no specific causal instructions.
2. CausalCoT: explicit chain-of-thought (CoT) with causal reasoning, as per Jin et al. (2023).
3. Solver: problem-solving formalism, from Hua et al. (2024).
4. CoIn: explicit search-and-backtrack (details below).

Responses are temperature 0 (deterministic), classified as "yes", "no", or "incomprehensible". The principal metric is overall binary accuracy (percentage correct yes/no). Intermediate error analysis tracks whether mistakes stem from parsing/representation or from the actual inference chain (the latter dominating).

4. Baseline LLM Performance on CounterBench

Baseline segmentation reveals that most LLMs perform at or near random guessing under Standard or CausalCoT prompts:

Model	Basic	Conditional	Joint	Nested	Avg. (Standard)
GPT-3 (Davinci-002)	56.8	50.2	48.8	51.6	51.8
GPT-3.5	49.6	51.2	50.4	50.0	50.3
GPT-4o	50.4	54.4	50.4	54.8	52.5
Gemini-1.5-flash	75.2	65.6	67.2	76.0	71.0
Deepseek-V3	50.4	50.4	50.0	50.0	50.0

Applying CausalCoT improves some model scores (e.g., GPT-4o to 78.8%, Deepseek-V3 to 76.3%), but most models cluster near 50% accuracy—the performance expected under coin-flip guessing. Error decomposition attributes 86% of failures under CausalCoT to deficiencies within the chain-of-thought inference process, not the initial causal graph extraction, indicating fundamental limits in uncontrolled symbolic computation for these queries.

5. The CoIn Reasoning Paradigm

CoIn (Counterfactual Inference) is an explicit algorithmic paradigm engineered to scaffold LLM reasoning for systematic counterfactual inference using stepwise search and backtracking.

Phases of CoIn:

• Counterfactual Information Extraction: The model parses the prompt to construct a causal graph $\mathbb{R}$ (set of directed edges) and collects all observed/intervened variable values into a set $\mathbb{N}$.
• Counterfactual Reasoning Algorithm: (As specified by Alg. 1 in the source)
  1. Iteratively select a variable $K$ not yet assigned and attempt to infer its value using known causal relations and assigned values.
  2. If inferral fails, backtrack and try alternative paths.
  3. Continue until the target outcome variable $Y$ is resolved.

This paradigm is explicitly prompted so that the LLM is directed to execute search, symbolic inference, and backtracking akin to algorithmic causality solvers rather than relying on heuristic or memory-based CoT.

6. Comparative Empirical Results of CoIn

CoIn enables substantial performance gains across all models tested. The average accuracy boost ranges from 15 to 40 percentage points over prior strategies, with state-of-the-art LLMs surpassing 90%:

Model	Standard	CausalCoT	Solver	CoIn
GPT-4o	52.5%	78.8%	52.1%	92.0%
GPT-4o mini	50.0%	61.7%	47.5%	80.6%
Claude-3 Sonnet	50.0%	59.0%	51.9%	91.6%
Gemini-1.5-flash	71.0%	73.5%	53.8%	93.0%
Deepseek-V3	50.2%	76.3%	49.5%	93.5%

By problem difficulty (number of variables, shown for GPT-4o mini):

Variables	Standard	CausalCoT	CoIn
5	50.0%	67.0%	92.0%
6	50.0%	61.0%	82.0%
7	50.0%	63.5%	82.0%
8	50.0%	61.5%	73.5%
9	50.0%	55.5%	73.5%

Performance gradually declines with increased structural complexity, but CoIn retains a dominant lead at all levels.

7. Insights, Limitations, and Trajectories

CounterBench reveals that unassisted or even classic chain-of-thought prompting leaves LLMs inept at formally specified counterfactual reasoning, with accuracy approximating the random baseline even on the latest, largest model generations. The core bottleneck is not in parsing or representing causal structures, but in reliably performing the multi-step symbolic inferences required for rigorous counterfactual judgments. CoIn's explicit algorithmic guidance, mirroring formal search and backtracking, equips LLMs to execute these inferences, producing $>$90% accuracy in most cases.

Limitations include:

• Exclusively deterministic SCMs; real-world scenarios feature probabilistic dependencies and hidden confounders.

• Short, abstract variable names; richer, more contextually embedded tasks may present new challenges.
• Absence of statistical significance testing (though the effect sizes reported are far beyond typical sampling variance).

Planned extensions involve probabilistic SCMs (soft interventions and stochastic outcomes), incorporation of hidden confounders and instrumental variable queries, domain-rich causal narratives (e.g., in medicine, policy), and automated prompt optimization to reduce manual overhead. CounterBench thus provides a necessary foundation and a reproducible, transparent metric for advancing LLM capabilities in formal, algorithmic counterfactual reasoning (Chen et al., 16 Feb 2025).