Do Language Models Know What Words Mean? Evaluating Semantic Knowledge of Adjective-Noun Compounds

Does Bielik Know What It Doesn't Know? Activation Dispersion Separates Entity Familiarity from Factual Reliability Across Model Scale

Large language models hallucinate most about entities they have never seen. We ask whether a model's activations betray entity familiarity before a single answer token is generated, and whether that signal predicts the factual reliability of the answers. On four Polish Bielik models (1.5B-11B parameters), we probe four entity domains (athletes, cities, writers, musicians), each with 42 well-known, 42 obscure-but-real, and 42 fabricated entities addressed by a one-sentence question (504 prompts per model). Two unsupervised, single-forward-pass dispersion measures over post-SwiGLU MLP activations, inverse participation ratio and spectral entropy, separate known from fabricated entities at AUROC 0.95-1.00 across all domains and scales; a supervised linear probe reaches 0.99-1.00. Both clear selection-aware permutation floors of about 0.70-0.74 (empirical p<=1e-3), survive held-out layer selection (0.93-0.99), and persist on real names (known vs. obscure-but-real: 0.96-1.00). The signal transfers across entity types (mean off-diagonal AUROC 0.92-0.99); a matched-template counterfactual shows the only large drops are template-caused, not entity-type effects, and the signal is diffuse across heads. This representational signal is already at ceiling at 1.5B, whereas behavioral factual reliability scales sharply: 0, 2, 10, and 19 of 42 known athletes are answered fully correctly by the 1.5B, 4.5B, 7B, and 11B models under a strict judge. Within known entities, separating correct from hallucinated answers is much harder (probe 0.93; dispersion no better than a first-token-entropy baseline). A five-sample semantic-entropy baseline reaches only 0.71-0.83 at 5x the inference cost. Despite this internal awareness, the models almost never abstain: an audit of 2,520 answers finds 2 refusals and 1 hedge. Entity familiarity and factual reliability are distinct phenomena on different scaling curves.

How Much Human Label Variation Does Formal Semantic Structure Explain?: Group-Level Effects and Item-Level Ceilings in NLI

Human label variation in natural language inference is increasingly treated as signal rather than noise, but how much of it formal semantic structure explains has not been measured directly. We measure it on the 3,113 SNLI and MNLI items of ChaosNLI, using a rule-based operator and monotonicity tagger validated against MED (0.883 agreement at the edit site, 0.807 on the sentence-level summary our analyses consume), three preregistered analysis blocks, and full reporting of negative results. Three bounds emerge. First, a group-level boundary: hypotheses that are not purely upward monotone show reliably higher label entropy (Cliff's delta = -0.284), and rank-based tests defend the effect against operator-presence and length reductions, though a bounded-outcome sensitivity check weakens the regression form of the length defense. Second, an item-level ceiling: the same formal profiles explain only 3.3 to 3.6 percent of entropy variance and reach a median-split AUC of 0.606, too weak to identify high-disagreement items. Third, composition invariance: across the boundary, three high-powered preregistered contrasts on validated error shares and explanation-type shares (VariErr, LiTEx) all return null results. In this sample, formal semantic structure shifts how much annotators disagree by a small amount and does not detectably change what they disagree about. ChaosNLI-S/M consists of items selected for low original agreement, and every claim is conditioned on that scope. All analyses were preregistered in a version-controlled research log, whose audit trail, including one corrected interpretation rule, the paper discloses.

Reversing Arrows in Large Language Models

Large language models (LLMs) have achieved strong performance on text-to-knowledge graph generation and related tasks. Nevertheless, it is still unclear whether they accurately model the direction-dependent semantics of inverse relations, in which reversing the order of the arguments alters the meaning of a relation (e.g., \textit{mother} versus \textit{child}). To the best of our knowledge, this work presents the first systematic study of inverse relation directionality in LLMs, using a benchmark consisting of 5,457 instances spanning 27 distinct inverse relation labels. We evaluate five open-source LLMs under a multiple-choice prompting framework and further examine the influence of relation descriptions and entity representations by substituting the original entities with synthetic and masked entities. Our findings reveal systematic asymmetries in inverse relation classification across LLMs, indicate that relation descriptions do not consistently improve performance, and show that model performance can be sensitive to variations in entity representations.

Predicting Human Semantic Priming Using Multilingual Large Language Models

Human-Like Anaphor Resolution in Large Language Models

Anaphors are expressions that refer to other expressions, called antecedents. The process of connecting the two is called resolution. Cognitive science has identified multiple factors that affect the speed and success of anaphor resolution, including discourse structure, situation-model properties, and semantic factors. Here, we investigate whether these factors also affect anaphor resolution in five Large Language Models (LLMs) with open weights: GPT-2-XL, Llama-3.1-8B, Pythia-12B, Mistral-7B, and Mistral-24B. To model processing difficulty, we adopt the standard linking hypothesis that relates human reading times to model surprisal at the anaphor. As a second behavioral measure, we compare model accuracy to human accuracy on comprehension questions probing the antecedents of anaphors. The results show selective cognitive alignment: some LLMs exhibit human-like sensitivity to discourse prominence and distance-based factors in anaphor resolution, while showing weaker or absent sensitivity to semantic interference effects. These findings delimit the conditions under which LLMs approximate human anaphor resolution.

Large Language Models Under Evaluation: An Acceptability, Complexity And Coherence Assessment In Italian

This paper discusses the results of various experiments assessing the morphosyntactic and semantic competence in Italian of four very large language models (vLLMs): davinci (GPT-3/ChatGPT), davinci-002, davinci-003 (both GPT-3.5 models) and gpt-4-1106-preview (GPT-4). We evaluated these models on (i) acceptability, (ii) complexity, and (iii) coherence judgments using 7-point Likert scales and on (iv) syntactic development through a forced choice task. The test sets were drawn from shared NLP tasks and standard linguistic assessments. The results suggest that, although fine-tuned transformers outperform all GPT models, GPT-4 represents a significant improvement over third-generation GPT models. According to our tests, even if GPT-4 and fine-tuned transformers cannot be considered descriptively or explanatorily adequate, they nonetheless pose a challenge to the poverty of the stimulus hypothesis. The "theory" expressed by GPT models is not linguistically intelligible in any relevant sense, and their training data is orders of magnitude larger than the primary linguistic input available to children. Nevertheless, GPT-4 captures certain generalizations, such as the constraints blocking the insertion of an overt resumptive clitic in specific gap positions, that are arguably unlearnable from just primary positive data.