Product-of-Experts Training Reduces Dataset Artifacts in Natural Language Inference

As many benchmarks have become saturated, it has become increasingly important to create new datasets that evaluate the generalization capacity of current state-of-the-art models in reasoning. However, designing high-quality reasoning datasets is challenging, as their manual construction is costly, and their automatic generation is unreliable, often leading to synthetic data with limited scope. In this paper, we propose the Minimal Expression-Replacement GEneralization (MERGE) test that evaluates the robustness of reasoning models against non-adversarial variants of existing evaluation datasets. We automatically obtain high-quality variants from the original instances with Minimal Expression REplacement (MERE) generation, which uses Masked Language Models (MLMs) and safeguarding filters. We apply the MERGE test to Natural Language Inference (NLI), a popular task of reasoning. We generate new NLI datasets from two widely used existing ones with the MERE generation and use them to evaluate multiple strong NLI models. The results indicate that both LLMs and fine-tuned NLI models generalize poorly: they struggle to consistently and correctly classify variants minimally different in form and reasoning from the original ones. Further, we also analyze how certain aspects in variant generation, such as the word class and the source MLMs, affect model performance.

We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.

While Large Language Model (LLM)-based Natural Language Inference (NLI) systems achieve high accuracy, their decision-making processes lack auditable structures. This paper explores whether NLI can be performed using only interpretable, graph-based representations of evidence. We introduce a fully graph-based pipeline where the classifier never directly processes the input text. Instead, sentences are decomposed into atomic propositions, converted into ConceptNet triples via constrained decoding, and represented as three graphs per pair: premise, hypothesis, and a retrieved ConceptNet subgraph. These graphs are then fed into a fine-tuned 0.8-billion-parameter language model. On the SNLI dataset, our pipeline achieves 89.7% accuracy, just 1.9 points below an identically trained text-based model. On ANLI, it matches the published performance of RoBERTa-large on rounds R2 and R3 (48.0% vs. 48.9% and 44.9% vs. 44.4%) but trails by 16 points on R1, resulting in an overall gap of 9 to 14 points compared to its text counterpart. We term this gap the price of interpretability and demonstrate that it stems from representational limitations rather than data constraints. Ablation studies further reveal that graphs and text are complementary: combining both modalities achieves 92.1% accuracy on SNLI.