Large Language Model News

An open-source, bilingual guide explaining the inner workings of Transformers has been published, covering topics such as attention collapse and KV-cache compression with reproducible code. The guide is available in Engl…

A developer argues that agent runtimes lack a session state infrastructure, forcing users to act as the retry protocol when tool calls fail or context overflows. The proposed solution includes a typed, inspectable state …

Meta hit a compute wall when Google throttled its access to Gemini models due to insufficient GPU capacity, disrupting internal projects. The incident underscores that cloud AI resources are finite, forcing developers to…

Researchers propose LaViD, a framework that transfers semantic knowledge from a large language model to a vision-only student model using multiple-choice questions, outperforming existing distillation methods on fine-gra…

Researchers introduced Grounded Iterative Language Planning (GILP), a hybrid method combining a small parameterized world model with LLM-based agent reasoning, reducing hallucinated-state rates from 0.176 to 0.035 on gra…

Researchers propose a framework to expand human motion generation by using large-scale synthetic motion data and a redesigned VQ-VAE tokenizer, overcoming the limited diversity of motion capture datasets. The approach im…

A new position paper argues that the term 'machine unlearning' is overused in large language model research and should be reserved for dataset-defined deletion. The authors contend that many tasks labeled as unlearning, …

Researchers propose EntMTP, a training-free scheduler that dynamically adjusts multi-token prediction depth based on local entropy, achieving up to 1.36x speedup over Medusa baselines in LLM inference without quality los…

Researchers introduced Ko-WideSearch, a Korean breadth-search benchmark for web agents that evaluates exhaustive set enumeration across 228 tables spanning 190 entities. Testing on twenty web agents revealed consistent f…

Researchers introduced Masked Language Flow Models (MLFMs), combining masked diffusion and flow-based methods for efficient language generation. MLFMs enable conditional generation via continuous flows and support multi-…

Researchers introduced Smooth Maximum Mean Discrepancy (SMMD) to improve numerical prediction accuracy in large language models by aligning predicted numeric distributions with targets using distance-based kernels and gr…

Researchers introduce the Prism Transformer, a new architecture that progressively increases head counts across layers to create a hierarchical attention processing structure. This design improves performance on zero-sho…

Researchers propose ATOD, a hybrid online distillation algorithm that combines on-policy distillation and reinforcement learning to train small language-model agents for multi-turn tasks. ATOD uses an annealed schedule t…

Researchers fine-tuned Gemini 2.5 Pro on 400 clinician-rated home videos to score 30 behavioral features for autism assessment, improving inter-rater reliability with clinicians by 40% and boosting zero-shot ASD diagnosi…

Researchers propose SD-GPS, a solver-driven framework for geometry problem solving that integrates autoformalization and theorem proposing. The framework uses a symbolic solver as an execution oracle to improve formaliza…

Researchers introduced PEBS, a per-rater empirical-Bayes shrinkage estimator for calibrating reward models in RLHF, which reduces within-user held-out RMSE by 8.58% on PRISM and 9.66% on PluriHarms compared to the pooled…

Researchers propose COOPA, a modular LLM agent architecture for operations research problems that combines iterative confidence-based modeling, element-level provenance explanations, and multi-solver routing. In benchmar…

Researchers introduced an axiomatic evaluation framework for latent thought representations in LLMs, comprising four axioms (Causality, Minimality, Separability, Stability) that measure representation quality independent…

Team HSA_CORAL submitted to the FinCausal 2026 shared task on extracting cause-effect relations from financial narratives via extractive question answering in English and Spanish. Their best system, GPT-4.1 Mini fine-tun…

Researchers used a developmental approach to study how neural language models learn statistical patterns, finding that Transformers first acquire abstract global statistics before local dependencies, with early over-gene…