[Paper] Reverse-engineering NLI: A study of the meta-inferential properties of Natural Language Inference

Source: arXiv - 2601.05170v1

Overview

The paper Reverse‑engineering NLI: A study of the meta‑inferential properties of Natural Language Inference digs into what the classic Natural Language Inference (NLI) benchmarks (especially SNLI) are actually teaching models about logical reasoning. By teasing apart three plausible interpretations of the “entailment / neutral / contradiction” labels, the authors reveal which logical reading the data really encodes – a crucial step for anyone building or evaluating language models on reasoning tasks.

Key Contributions

• Three formal readings of NLI labels – the authors define semantic entailment, pragmatic inference, and meta‑inferential interpretations and map each to concrete logical properties.
• Meta‑inferential consistency tests – they construct two novel probe sets: (1) shared‑premise pairs that should obey transitivity/consistency constraints, and (2) LLM‑generated NLI items designed to stress‑test the model’s logical behavior.
• Empirical analysis on SNLI‑trained models – a suite of BERT, RoBERTa, and DeBERTa models are evaluated on the probes, exposing systematic biases toward one of the three readings.
• Insightful diagnostic framework – the study provides a reusable methodology for auditing any NLI dataset or model for hidden logical assumptions.

Methodology

1. Define label semantics

   • Semantic entailment: classic truth‑preserving inference (if premise true, hypothesis must be true).
   • Pragmatic inference: inference based on typical world knowledge or speaker intent.
   • Meta‑inferential: inference about the relationship between premise and hypothesis (e.g., “the premise does not rule out the hypothesis”).

2. Create probe sets

   • Shared‑premise probes: group multiple hypotheses under the same premise and check whether model predictions respect logical constraints such as transitivity (if A entails B and B entails C, then A should entail C).
   • LLM‑generated probes: prompt a strong language model (e.g., GPT‑4) to produce NLI triples that deliberately violate one reading while satisfying another, yielding “adversarial” examples.

3. Train & evaluate – standard NLI models are fine‑tuned on the original SNLI training split, then tested on the probe sets. Accuracy, consistency scores, and confusion patterns are recorded.

4. Analysis – compare model behavior against the expected patterns of each reading, quantifying which logical view the dataset implicitly enforces.

Results & Findings

• Dominant meta‑inferential reading – Models trained on SNLI consistently obey the meta‑inferential constraints (e.g., they treat “neutral” as “the premise does not rule out the hypothesis”) while often violating pure semantic entailment expectations.
• Transitivity violations – On shared‑premise probes, >30 % of entailment chains break transitivity, indicating the dataset does not enforce strict logical closure.
• LLM‑generated stress tests – When presented with examples that are semantically entailed but labeled “neutral,” models follow the label rather than the underlying truth, confirming they learn the dataset’s idiosyncratic labeling scheme.
• Model‑agnostic pattern – The observed bias holds across architectures (BERT, RoBERTa, DeBERTa), suggesting it’s a property of the data rather than the model.

Practical Implications

• Benchmark interpretation – Developers should treat SNLI‑style scores as measuring compatibility with the dataset’s pragmatic/meta‑inferential conventions rather than pure logical reasoning ability.
• Model selection for downstream tasks – If an application needs strict entailment (e.g., legal document verification), relying on SNLI‑trained models may be risky; additional fine‑tuning on logically rigorous data is advisable.
• Dataset design – The diagnostic probes can be incorporated into new NLI corpora to enforce consistency, leading to higher‑quality training data for reasoning‑heavy applications such as question answering, fact‑checking, and dialogue systems.
• Evaluation pipelines – Adding the shared‑premise and LLM‑generated probe suites to a CI/CD test suite can catch regressions where a model inadvertently learns the wrong inference pattern after further fine‑tuning.

Limitations & Future Work

• Scope limited to SNLI – The analysis focuses on a single benchmark; other NLI datasets (e.g., MNLI, ANLI) may exhibit different meta‑inferential biases.
• Probe coverage – While the shared‑premise and LLM‑generated probes capture many logical constraints, they do not exhaust all possible inference patterns (e.g., modal or counterfactual reasoning).
• LLM generation bias – The adversarial examples rely on a strong LLM, which itself may embed its own biases, potentially influencing the probe difficulty.
• Future directions – Extending the framework to multilingual NLI, integrating formal logic verification tools, and designing training objectives that explicitly encourage semantic entailment are promising next steps.

Authors

• Rasmus Blanck
• Bill Noble
• Stergios Chatzikyriakidis

Paper Information

• arXiv ID: 2601.05170v1
• Categories: cs.CL
• Published: January 8, 2026
• PDF: Download PDF