Preprints and papers, rewritten in plain English.
A new arXiv study tests whether stack-based language models mirror universal linguistic preferences, such as SOV word order, when trained on mildly context-sensitive artificial languages.
A new arXiv paper argues that the scale of the reconstruction loss is a critical, overlooked factor in post-training quantization of large language models.
A new arXiv paper proposes replacing hand-crafted coding rules for AI agents with a self-evolving system, aiming to reduce manual effort and improve performance.
A multi-task QLoRA system estimates suicide-risk severity while pointing to the language and factors that support its assessment.
A new theory paper explains why some in-context learning behaviors can be compressed into a static vector while others cannot.
New research explores whether language models' recurring stylistic patterns can be identified and controlled without labeled data.
A new arXiv paper introduces quantum protocols that achieve information-theoretic privacy for two-party Hamming distance, a task where classical protocols require more communication.
A new arXiv preprint reports that the confidence large language models express in words is coupled with the probabilities they internally assign to their outputs.
A new arXiv study finds that large language models sharply lose accuracy in numerical claim verification when values are only slightly changed.
A new arXiv paper argues that alignment training, not just capability limits, can cause language models to quietly abandon the task they were given.
A new arXiv paper argues that agentic small language model ensembles should be judged on operational reliability, cost, latency, and token efficiency, not just accuracy.
A new arXiv paper notes that existing recipes for training looped language models require multi-stage training over trillions of tokens.
A new checker and benchmark tests whether LLM explanations stay aligned with structured evidence across four schema-level gates.
An arXiv paper proposes role-aware heuristic episodic attention to help conversational LLMs retain instructions and relevant information over long exchanges.
A new arXiv paper argues that curating synthetic training data for LLMs requires signals at the group level, not just individual examples.
A new inference-only technique applies receding-horizon planning to improve reasoning without any fine-tuning.
A new arXiv paper argues that standard divergence metrics may not tell deployers how many model decisions actually change after compression.
A new arXiv paper proposes using contextual trajectories and incremental displacement to study how LLMs construct meaning dynamically.