The Archive

Evaluation of LLM performance on classical Arabic maqama generation, a stylistically constrained literary form with rhymed prose and dense rhetoric.

FakeContextBench evaluates whether LLMs can distinguish when contextual information should override learned patterns, revealing gaps in in-context learning robustness.

Framework identifies semantic abstraction gaps in LLM natural language inference via constructed higher-order semantic knowledge and reasoning diagnostics.

Face/Off evaluation framework reveals LLMs overweight lexical cues in code comprehension tasks despite access to semantic program structure.

EADC benchmark evaluates LLM compliance with AI laws and regulations, detecting implicit covert risks beyond static explicit compliance checks.

Informed Masking prioritizes high-leverage token positions in diffusion LLM RL by identifying upstream/downstream structure rather than uniform random masking.

Study analyzes reverse reasoning ability in GPT-o1, GPT-o3, and DeepSeek-R1, identifying five error categories in complex problem-solving.

Novel membership inference detector using entropy-corrected loss boundaries to distinguish pretraining data exposure from learned generalization in LLMs.

LLM-Falsifier applies large language models as optimizers to falsify cyber-physical system specifications written in Signal Temporal Logic.

Paper proposes 'summarization bias'—LLMs collapse narrative structure into abstract summaries rather than reconstructing inferential meaning—with formal testing protocol.

Study of 67,200 responses from GPT, Claude, and Gemini across 112 languages shows geopolitical bias variance in LLM answers about Ukraine war.

Cunning questions training enhances LLM safety vigilance by teaching detection of latent risks hidden in benign contexts beyond surface compliance.

LLMs form linear belief-state geometry for HMM hidden states during in-context learning; probed across 6 models on 40 HMMs.

LLM-based extraction of ontology-compliant knowledge graphs from scientific materials literature enables structured reuse.

K-Bench evaluates 33 base models from 14 providers on 200 multi-turn mental health vignettes with suicide, self-harm, and domestic violence scenarios—clinician-calibrated safety benchmark.

Multi-signal hallucination detection pipeline achieves F1=0.915 on HaluEval using DeBERTa-v3, MC Dropout, and temperature calibration.

Φ-Bench evaluates LLM capability on open-ended LLM infrastructure engineering tasks—kernel optimization, operator design, compiler tuning—beyond isolated benchmarks.

YallaMorph benchmark evaluates Arabic morphological generation in LLMs across 600K examples with diacritized and undiacritized settings.

Formal framework for assembling context in LLM-based systems engineering design with modular units and modelling-as-code evaluation.

Knowledge graph-based evaluation framework to assess contextual understanding in LLMs beyond surface-level metrics.

Systematic review of 66 studies on LLM applications for HVAC building automation, assessing deployment readiness and physical system safety.

Mixture-of-Experts (MoE) architectures provide an efficient paradigm for scaling large language models (LLMs), yet fixed top-k routing activates the same number of expert slots for every token, causing substantial redundant computation. Existing expert-skipping methods often rely on router confidence, calibration data, or additional training, and therefore cannot reliably estimate the actual contribution of routed experts. To this end, we propose ACE, a training-free, calibration-free, and checkpoint-preserving framework for token-adaptive expert skipping in MoE-based LLMs. ACE contains two c...

FlexPension-LLM domain-specialized model with DKI-RDistill predicts pension enrollment behavior for Chinese flexible workers.

arXiv proposes time-to-first-spike SNNs with reference-based encoding for energy-efficient large language models.

Controlled pre-training experiments show auxiliary views (knowledge reformulations) improve LLM factual recall more than repetition at fixed token budget.

On-policy distillation of LLMs recovers full-data performance gains with single-query training via state coverage and alignment dynamics.

FLY-EVAL++: safety-focused evaluation protocol for LLMs in physics-governed domains; measures constraint violations and physical inconsistency beyond accuracy metrics.

InSituMeasure benchmark evaluates MLLMs on continuous-valued measurement tasks in realistic industrial settings with gauge reading and instrument-specific context.

Zero-shot/few-shot LLM framework for chronic kidney disease screening using structured prompts on tabular clinical features; compares with traditional ML/DL baselines.

Structured reasoning framework for LLM-powered telecom root cause analysis with evidence grounding to reduce hallucination.