Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2605.29655v4 Announce Type: replace Abstract: Autoregressive multimodal large language models (MLLMs) enable 3D generation but struggle to scale to high-resolution shapes due to inadequate 3D t...
The paper investigates the symmetric and antisymmetric components of bilinear forms that arise in the attention heads of large language models. It introduces an orthogonally invariant profile map that projects real bilinear forms onto a three‑dimensional simplex, revealing that the profiles of trained bilinear forms tend to cluster near those of rank‑one bilinear forms. Additionally, the authors prove that for a Zariski‑dense subset of query‑key matrices, the symmetric part of a bilinear form in an attention head can be decomposed into a hyperbolic form plus a zero form.
The paper presents a discrete generative model for neuronal spiking activity recorded on microelectrode arrays. It uses a shared vocabulary of spatiotemporal motifs learned by a residual vector‑quantized autoencoder and predicts motif occurrence with a factorized masked transformer. Evaluated on 31 assays from human brain organoids and ex vivo hippocampal tissue, the model achieves superior reconstruction and generation performance compared to baselines and shows that motifs are largely reused across assays.
The study benchmarks Vision Transformers (ViTs) against convolutional neural networks (CNNs) for fine‑grained orchid genus identification in New Guinea’s species‑rich, data‑poor flora. Using a two‑stage system that first predicts genus and then retrieves similar species images, the authors fine‑tuned four pretrained backbones on 16,701 photographs from 120 genera and 1,350 species. The self‑supervised ViT DINOv2 achieved the highest genus accuracy (macro top‑1 66.9 %) and outperformed both CNNs and a domain‑matched pretrained ViT, demonstrating strong species retrieval and open‑set detection capabilities.
The paper investigates whether the layer that yields the highest probing accuracy in omni‑modal large language models is also the most effective for steering interventions. Across three independently developed models, the authors find that the best probing layers differ widely, whereas the most steerable layers consistently lie in a narrow mid‑to‑late range of the network. Using emotion as a testbed, they demonstrate a significant causal gap between probing and steering, and propose a two‑factor account linking readability and downstream plasticity to steering effectiveness.
This paper conducts a systematic empirical study of multi‑object tracking (MOT) algorithms, focusing on how detection and association components affect overall performance. By evaluating state‑of‑the‑art methods on benchmarks such as MOT16/17/20, SportsMOT, DanceTrack, and CrowdTrack, the authors find that detection quality has a far greater impact than association strategies, and that transformer‑based end‑to‑end models are more robust to detection variations but computationally expensive. The study provides a unified pipeline diagram and practical guidance for researchers and practitioners in selecting and designing MOT systems.
ShearFuse-UNet is a lightweight deep learning model designed for next‑day wildfire spread prediction using multi‑modal satellite data. It incorporates three transform‑domain branches—Fast Walsh‑Hadamard Transform, Discrete Cosine Transform, and a cone‑adapted digital Shearlet residual—within each encoder block of a U‑Net backbone, fusing spectral representations via a learned SpectralFusion gate and adding Shearlet reconstruction as a residual. On the WildfireSpreadTS dataset, the model achieves an F1 score of 0.596 with only 267k parameters, outperforming a larger ResNet18‑based U‑Net and demonstrating a favorable accuracy‑efficiency trade‑off, with additional validation on the Google Next‑Day Wildfire Spread dataset.
arXiv:2609. 22415v1 Announce Type: new Abstract: Complex-valued Transformers have inherited softmax attention over the raw complex inner product.
The paper introduces BRIDGE, a method for efficient instruction tuning of large language models that jointly selects prompt context and response supervision. BRIDGE uses a shared validation-directed interaction surrogate to evaluate token utility conditioned on the other side’s retained state, and employs budgeted alternating selection and structure-aware projection to produce coherent supervision spans. Experiments across three model families show BRIDGE outperforms independent selection methods in mathematical reasoning, code generation, and instruction following, with its advantage increasing as compression grows.
arXiv:2609.22161v1 Announce Type: cross Abstract: Medical large language models are commonly trained on mixtures of didactic data (e.g., textbooks) and clinical data (e.g., patient records), yet how...
arXiv:2609.22478v1 Announce Type: cross Abstract: Behavioural evaluations of hosted language models can vary because the evaluated service, the measurement instrument, or both differ across runs. We...
arXiv:2609.22592v1 Announce Type: cross Abstract: Training agents with reinforcement learning requires a gym, comprising a task, an executable environment in which the task can be attempted, and a ve...
arXiv:2609.22939v1 Announce Type: cross Abstract: Long-context models read a novel the way a person reads a printout: one token after another, in narrative order, with the whole history competing for...
arXiv:2609.23043v1 Announce Type: cross Abstract: Large Language Models (LLMs) enable open-ended dialogue in interactive games, but their non-deterministic outputs make it difficult to preserve autho...
arXiv:2609.24057v1 Announce Type: cross Abstract: Medical image interpretation is central to diagnosis and care, yet adapting general-purpose multimodal large language models (MLLMs) often requires r...
arXiv:2609.24198v1 Announce Type: cross Abstract: This paper describes the SKstars submission to SHROOM-Visions 2026, a shared task on fine-grained hallucination detection in large vision-language mo...
arXiv:2609.24480v1 Announce Type: cross Abstract: Deploying Large Language Models (LLMs) in healthcare requires robust performance across two complementary dimensions - diagnostic reasoning: the conv...
arXiv:2609.24855v1 Announce Type: cross Abstract: Argumentative component detection (ACD) is a core subtask of Argument(ation) Mining (AM) and one of its most challenging aspects, as it requires join...
arXiv:2609.24967v1 Announce Type: cross Abstract: LLM agents are increasingly deployed in collaborative settings, yet long-term interaction may give rise to undesirable coordination. We study the eme...
arXiv:2609.22101v1 Announce Type: cross Abstract: Large language models can process increasingly long prompts, yet their ability to locate and use decisive evidence may degrade as irrelevant or confu...