CoBe: A Benchmark For Conversational Counterfactual Text Editing

AI summary

Overview

Research area: Natural Language Processing / evaluation benchmarks, sitting at the intersection of causal inference (Pearl's causal hierarchy) and LLM text generation.

Technical level: Intermediate. The paper is written to be readable without deep causal-inference background (it includes its own primers in appendices), but it assumes familiarity with standard LLM benchmarking vocabulary.

Scope: CoBe is a 2,000-sample, expert-verified benchmark that tests whether language models can perform conversational counterfactual text editing — revising a short narrative in response to a colloquially phrased hypothetical change — in a zero-shot setting.

What This Paper Is About

The paper argues that when a model is asked "what would have happened had X been different?", it must do genuine causal reasoning: hold fixed the background facts of the given scenario and change only the events causally downstream of the hypothetical change. The problem is that models frequently slide into "associational generation" — producing a plausible alternative scenario correlated with the altered condition, even if that means rewriting facts that should have stayed the same. CoBe was built to separate these two modes of reasoning and measure how well current models manage the counterfactual one.

The motivating illustration: if a doctor had not recommended surgery, the model should keep the patient's medical history intact and change only the consequences of the decision — not make the symptoms milder just because milder cases correlate with no surgery.

Key Contributions

1. CoBe, a conversational counterfactual editing benchmark: 2,000 samples (200 public), each pairing a short narrative scenario with single-turn prompts written in everyday subjunctive language ("would have…had"), spanning 7 domains and 6 graph structures.

2. Expert-verified, causally grounded evaluation criteria: Rather than reference-matching, each sample carries criteria specifying which facts must remain unchanged and which must change, plus a system-level check on causal connectors. Multiple wordings and outcomes can pass.

3. A nine-model evaluation across scales and training paradigms: three frontier closed models (GPT-5.4-Pro, Claude-Opus-4.7, Gemini-3.1-Pro) and six open-source models (Llama-3.1-8B, Llama-3.1-70B, Mistral-7B, Qwen-7B, Phi-4, Gemma-27B).

4. Failure-mode taxonomy and model-specific error profiles, plus prompt-ablation experiments showing how much (and how little) additional procedural guidance helps.

Main Findings

• Frontier models average 58.05% accuracy. Averaged across three query phrasings, the frontier average is 58.05% as of July 2026. The highest single accuracy is 61.34% (GPT-5.4-Pro).

• Full per-model accuracies (Table 5, averaged across three queries): GPT 61.34 ± 0.45; Gemini 56.68 ± 0.44; Claude 56.14 ± 2.50; Llama-8B 33.77 ± 2.50; Mistral 32.90 ± 3.36; Qwen 36.83 ± 3.41; Phi 44.73 ± 2.02; Gemma 48.26 ± 1.45; Llama-70B 49.64 ± 1.42.

• Scale does not reliably buy counterfactual ability. Llama-70B lands within a 6.5% gap of Claude despite far fewer parameters. The authors read this as evidence the failure reflects training objectives rather than insufficient model capacity. Performance also does not scale consistently with output token length.

• The dominant error is editing facts that should have stayed fixed (E1). This pattern is consistent across all models tested. Errors from failing to make required downstream changes (E2) and from wrong causal connectors (E3) are generally lower.

• Two finer-grained failure types were identified: E4, reasoning wrongly about the direction of a downstream edit (especially involving numerical changes), and E5, failing to hold latent factors invariant across the factual and counterfactual worlds. GPT is the weakest frontier model on E4; Claude does well on E4 but struggles on E5; large frontier models outperform open-source models on E5.

• Graph structure matters. Accuracy is lower on Diamond-like and Fork-like graphs, indicating difficulty propagating intervention effects across multiple downstream branches. Models also tend to misread correlated graphs as cause-and-effect graphs.

• Query phrasing barely matters. Three paraphrased versions of each query produce negligible performance differences, which the authors take as evidence that failures stem from a lack of native conversational counterfactual reasoning rather than prompt ambiguity.

• Model-specific biases: Gemini is described as strong on grammatical coherence but makes slightly more upstream changes; GPT is comparatively good at refraining from changing non-descendant variables; Claude is best on numerical changes but weaker at preserving latent state.

• Prompt ablations (500 random samples): additional conversational prompts that more strongly suggest the counterfactual interpretation improve frontier models by around 5% on average and lower performance for smaller models. A simple clarification ("Keep prior and unaffected facts unchanged, and update the consequences of the requested change.") improves scores by around 3%. Even a fully detailed step-by-step causal recipe leaves the benchmark unsaturated, with frontier models scoring 81.3% and 76.1%.

• Judge reliability: Claude-Opus-4.7 serves as the LLM judge across three checks. It agrees with three independent expert human raters at Cohen's κ = 0.812. Expert evaluation covered 250 sampled responses.

• Dataset composition: 2,000 scenarios across 7 domains; average 39 words, 2.35 sentences, 5.43 clauses, and 2.44 clauses per sentence. Domain shares: Science & Research 16.0%, Engineering 15.7%, Business & Services 14.5%, Everyday 14.2%, Healthcare 6.0%, Legal/Policy & HR 4.7%, Others 28.9%. Graph types: Chain-like 51.0%, Collider-like 21.9%, Hybrid 11.9%, Correlated 5.7%, Fork-like 4.8%, Diamond-like 4.5%.

• Construction and QA: Four stages — 425 candidate examples contributed by student researchers who had completed a graduate causal inference course; refinement into 106 core examples; LLM-assisted augmentation producing 25 variations per core example; and final filtering that removed 650 of 2,650 augmented examples to yield 2,000 retained samples. Three causal inference experts uninvolved in construction independently assessed 100 random samples for coherence, grammar, and counterfactual validity, surfacing a systematic inconsistency that prompted further exclusions.

• A stated evaluation limitation: the E2 check accepts removal of a listed fact as well as its modification, so an incomplete rewrite can pass by omitting a consequence that should have been updated. The authors did not observe systematic use of this shortcut in manually inspected responses but do not claim the criteria are robust to deliberate omission.

Methodology in Plain English

The researchers built the benchmark by having people with formal causal-inference training write short narratives containing causal structure, then refining the strongest into 106 core examples. An LLM generated variations of each core example to add domain and linguistic diversity while keeping the intended causal relationships, and the authors then filtered out 650 of the resulting 2,650 samples, leaving 2,000.

Each sample contains a story, a domain tag, a conversational edit request, a non-scored illustrative answer, and evaluation criteria. The request is phrased three ways per scenario using subjunctive "would have…had" grammar, and scores are averaged over all three to reduce prompt-related ambiguity.

For evaluation, models answer each story–query combination, and Claude-Opus-4.7 judges each response through three binary checks: whether facts that should remain true were preserved (E1), whether original facts that should change did change (E2), and whether causal connectors are consistent (E0). A response counts as correct only if all three pass. The authors deliberately avoid reference-similarity metrics like BLEU, ROUGE, and MoverScore, arguing these do not test which facts should remain invariant or change under an intervention. To validate this pipeline, three independent expert human raters assessed a subset, yielding κ = 0.812 agreement with the judge.

Why This Matters

Impact on research. The paper argues that prior counterfactual-text datasets (TimeTravel, PASTA, CRASS, CLadder, CounterBench and others) do not cleanly separate Rung 1 associational generation from Rung 3 counterfactual editing — either by pre-specifying which sentences must stay fixed, by using multiple-choice formats, or by relying on crowd-sourced similarity evaluation. CoBe is presented as the first testbed for zero-shot conversational counterfactual edits grounded in formal causal semantics over narrative stories, and the first to characterize distinct error profiles across models on this task. Because scoring permits multiple valid rewrites, the authors suggest CoBe can be plugged into fine-tuning and inference-time pipelines as a training signal.

Real-world applications:

• Healthcare: generating case studies for the same patient under different treatments without corrupting the patient's history — the paper's Figure 3 shows a frontier model (Gemini 3.1) incorrectly altering patient symptoms and cascading errors through the workflow.
• Law: reasoning about what would have happened under a different legal or procedural decision while holding the established facts of the case fixed.
• Business and policy analysis: counterfactual questions about decisions such as oil prices had a country quit OPEC earlier.
• Fairness and debiasing: the paper notes that naive counterfactual data augmentation can introduce new biases through spurious correlations, which a causally grounded benchmark helps surface.

Industry relevance. AI assistants are being embedded in everyday and professional workflows, including personal agents designed for conversational interaction. Users increasingly expect to state a hypothetical change in ordinary language without supplying a causal task recipe or tuning prompts per case. The paper's prompt ablations suggest this is a real gap: even explicit procedural instructions only move frontier scores by roughly 3–5%.

Future Directions

1. Extending the benchmark to other modalities and domains, such as vision, which the authors name as a natural next step given that counterfactual text editing is only one specific manifestation of counterfactual reasoning.

2. Closing the gap between conversational and procedural performance, since even a fully detailed causal recipe following structural causal semantics leaves frontier models at 81.3% and 76.1% rather than saturating the benchmark.

3. Improving the evaluation criteria's robustness, particularly the E2 check's acceptance of omitted facts, so that incomplete rewrites cannot pass by deletion.

4. Building systems designed natively for causal reasoning rather than relying on scale, given that Llama-70B sits within 6.5% of Claude and performance does not track model size or output length consistently.

Target Audience

Researchers and practitioners working on LLM evaluation, causal reasoning, and safety-critical text generation — especially those in clinical documentation, legal drafting, and counterfactual data augmentation. It is also useful for model developers who need a benchmark that distinguishes associational plausibility from causal validity, and for anyone studying how model training objectives, rather than capacity, shape reasoning failures.

Authors’ abstract

Counterfactual text editing is the ability to revise a given text in light of a hypothetical change while preserving facts not causally affected by that change. When asked what would have happened if a doctor had not recommended surgery, the model output should retain the patient's medical history and modify only the consequences of the decision. This requires causal understanding: a model must determine, without being explicitly instructed to, what events are affected by the hypothetical change and what should remain fixed. This skill is central to human reasoning, especially in safety-critical domains like healthcare or law. By contrast, the task of associational text generation merely generates text statistically associated with the altered condition. In the medical example, an associational response might make the patient's symptoms less severe because milder cases correlate with no surgery. Such responses produce a plausible scenario, but not the counterfactual version of the given one. Prior counterfactual-text datasets do not clearly separate these modes of reasoning. We introduce the Counterfactual Benchmark (CoBe), a testbed for models' zero-shot ability to counterfactually edit a given textual scenario when prompted to do so in a conversationally natural way. CoBe contains 2000 questions spanning healthcare, science and engineering, business, and everyday scenarios. We test a suite of frontier and smaller models, with frontier systems achieving an average score of $58.05\%$. Our findings indicate that models frequently default to associational generation rather than preserving the causal structure of the original scenario. Moreover, performance does not scale consistently with either model size or output token length, suggesting that the failure may reflect model training objectives rather than insufficient model capacity.