arXiv:2606. 30815v1 Announce Type: cross Abstract: Recent work suggests that transformer language models show a bias towards human languages over unnatural ("impossible") languages argued to be unacquirable by humans.
Large language models increasingly produce and interpret verbal probability expressions, yet whether these expressions carry consistent meaning across models (or match human perceptions of uncertainty...
The study evaluates how large language models (LLMs) interpret verbal probability expressions by mapping words to numbers and testing consistency across 19 models. Results show that LLMs largely mirror human benchmarks—preserving word order, recovering key anchor points, and reflecting the high variance of the term "possible"—but they exhibit a systematic upward bias for negative expressions like "unlikely" and "improbable." Explanation elicitation reduces within‑model variance but increases divergence between models, while a bidirectional roundtrip test reveals that leading models maintain coherent internal representations.
As the chain-of-thought reasoning capabilities of large language models improve, evaluating and calibrating their reasoning confidence is becoming increasingly important for quantifying the uncertaint...
The paper investigates how large language models (LLMs) handle two usage‑based grammatical strategies: coercion, where a mismatched word is reinterpreted to fit its context, and preemption, where a plausible structure is avoided because it never appears in expected contexts. Experiments across models of varying sizes show that larger LLMs successfully recognize and produce constructional productivity in coercive situations, even with novel words, but they fail to exhibit preemptive behavior, lacking the negative evidence that would guide avoidance of unconventional but semantically viable constructions. The authors argue that coercion is detectable due to its surprising, semantically motivated nature, whereas preemption lacks such a signature, leading to an asymmetry where models coerce but do not preempt.
arXiv:2604. 14180v2 Announce Type: replace-cross Abstract: We train a 318M-parameter Transformer language model from scratch on a curated corpus of 1.
The paper introduces Divergent Token Confidence (DTC), a method that estimates large language model confidence by counting tokens where two models strongly disagree during decoding. DTC uses Jensen-Shannon divergence between next-token distributions along the same reasoning trajectory and shows a near-negative correlation with answer accuracy. Experiments on multiple model families and six mathematical benchmarks demonstrate that DTC improves calibration over traditional probability-based and verbalized baselines, achieving lower expected calibration errors in both white-box and black-box settings.
The paper investigates how adjectival modifiers affect the semantic plausibility of events, using the Adept benchmark of 16,000 English sentence pairs that differ by a single adjective. Experiments show that sentence transformers, despite being conceptually suited to the task, underperform compared to models like RoBERTa. The authors provide an error analysis and discuss the implications of their findings for future work on balancing training and test data.
arXiv:2609.37497v1 Announce Type: new Abstract: Modern transformer models excel at capturing semantic relationships through sentence embeddings, yet their ability to perform pragmatic reasoning remai...
The paper evaluates two large language models, Claude Sonnet 4.5 and Claude Opus 5, on the bidirectional English Resource Grammar (ERG) tasks of generating English from Minimal Recursion Semantics (MRS) and parsing English into MRS. In generation, Opus achieves 76.3 BLEU—surpassing a 72k‑pair trained system and matching a million‑pair system—while Sonnet scores 65.7 BLEU, rising to 69.6 when selecting from ACE’s candidates. In parsing, both models lag behind ACE, attaining only 57.2 and 65.5 F₁ respectively, with exact‑match on about 1 % of sentences, highlighting that high generation scores do not guarantee accurate semantic parsing.
FreqBLiMP is a frequency‑controlled extension of the BLiMP minimal‑pair benchmark that regenerates all 67 paradigms under explicit Zipf‑frequency regimes while preserving grammatical contrasts. The study evaluates multiple open‑weight LLM families and finds that lower lexical frequency consistently reduces sentence likelihood, yet overall contrastive acceptability accuracy drops only modestly. However, the stability in aggregate accuracy hides significant variability across linguistic phenomena, with models remaining robust on overt morphosyntactic generalization but degrading on lemma‑specific tasks.
arXiv:2608. 03803v1 Announce Type: cross Abstract: Multilingual language models are deployed across a hundred or more languages, yet most benchmarks test whether a model can perform a task _in_ a language rather than whether it commands the language itself, conflating fluency with proficiency.