Large Language Model News

A training-free, CPU-only lexical prompt-compression pipeline reduced prompt tokens by a mean of 40.3% (sigma = 9.2) while holding BERTScore-F1 at 0.876 against original-prompt output, according to an arXiv paper (2609.1…

Researchers introduced TestHallVQA, a multi-image visual question answering benchmark built from scientific exams that combines document-level scale with human-examination difficulty, according to the arXiv paper 2609.13…

A new arXiv paper, RFCLLM, evaluates how well large language models reason about network protocol state machines, testing 16 protocols across 4 tasks and 1482 task queries. The study compares each LLM's implicit finite-s…

A maxillofacial cone beam CT report generation system scored 0.4122 against a composite objective that weights 80% on large language model factual entailment and 20% on lexical overlap, versus 0.2909 for a report selecte…

A new arXiv paper (2609.13454v1) introduces a paired benchmark of 171 PubMed Central Open Access case reports — 40 sepsis and 131 GLP-1/diabetes cases — to measure hindsight bias in clinical temporal reasoning by large l…

Researchers identified Harmfulness Propagation Dynamics (HPD), a cross-layer signature in which the projection of the last-token hidden state onto a learned harm direction rises monotonically with transformer depth for h…

Researchers proposed ModelLog, a declarative probabilistic framework for pre-training evaluation of large language models that specifies evaluation targets as symbolic constraints over token-level predictions, according …

Researchers built pseudo-references for the WMT26 General MT task's 10 reference-free language pairs by translating 3,277 official documents with seven models across up to five prompt conditions, producing 26 system-prom…

A new arXiv paper (2609.13581v1) presents an evidence-driven alignment framework for hospital discharge summarization that treats provenance as a first-class constraint, using semantic graphs and deep learning models to …

The University of Melbourne's submission to the WMT 2026 Creole Language Translation Shared Task reports that its machine translation models for Pacific creoles Tok Pisin, Bislama, and Solomon Pijin beat open model basel…

Researchers introduced LayerRoute, a parameter-efficient method for adaptive transformer layer-skipping that combines per-layer hard-gated routing with joint LoRA fine-tuning, according to an arXiv paper (2609.13682v1). …

A new arXiv paper introduces Continual Search, an iterative framework for automated root-cause attribution (RCA) in long-horizon AI agent failures, reporting that it improves GPT-5.5's F1 score by more than 40%, from 0.3…

A multi-agent supply chain analytics system described in arXiv paper 2609.13561v1 achieved 90% accuracy while cutting input token usage roughly fourfold compared with a single-agent baseline, according to the paper's aut…

Researchers introduced Asclepius, an adaptive agent scaffolding with a self-evolving harness, externalized clinical skills library, and three isolated subagents, which improved critical-action correctness by 22% (p = 0.0…

Researchers proposed FLoKD, an adaptive knowledge-distillation framework for federated LoRA fine-tuning of large language models over wireless networks that transmits intermediate LoRA activations instead of logits or fu…

A new arXiv paper (2609.13637v1) introduces PAI-Bench, a provider-neutral benchmark for measuring persistent identity in deployed AI agents across recall, composition, behavioral enactment, resistance, persistence, linea…

Researchers introduced MACE, a Multi-Agent Candidate Event acquisition method that uses evidence-specialized LLM agents to refine event structure before linking, according to a new arXiv paper (arXiv:2609.13670v1). MACE …

ArtSociety, a multi-agent framework that coordinates heterogeneous multimodal experts with two training-free controllers, achieved an Overall Score of 0.8870 on the official AffectiveArt Multidimensional Art Emotion Unde…

A structured temporal video reasoning pipeline combining a discrete EventGraph, a continuous EventField, and a human-readable EventGlyph view achieved 0.98 overall accuracy on a calibrated EPIC-KITCHENS subset of 10 vide…

A new arXiv paper (2609.13422v1) introduces Vibe Patenting, an end-to-end patent-drafting testbed showing that a separately-invoked LLM judge providing structured feedback consistently improves judge-assessed patent draf…