Brief

Researchers have developed a new framework called UH-NAS, which uses LLMs to guide neural architecture search for physical neural networks. This approach co-optimizes task accuracy with hardware constraints like energy consumption and physical non-idealities. UH-NAS is designed to be hardware-agnostic, allowing for fair comparisons across different computing platforms and discovering more robust architectures than traditional methods. AI

Researchers have developed CoastlineVLM-7B, a vision-language model designed to directly predict coastlines as polylines rather than segmentation masks. This approach, built on the GeoChat-7B/LLaVA-1.5 architecture, focuses on geometric boundary localization using geomorphic proxies like vegetation lines or dune toes. Evaluations on the New Zealand Coastal Change Dataset showed improved geometric alignment, reducing Hausdorff distance and Earth Mover's Distance compared to traditional segmentation methods. AI

HiDream.ai has released its commercial image generation model, HiDream-O1-Image-1.5, which has achieved top rankings on the Artificial Analysis Text to Image Leaderboard. The model excels in complex tasks such as rendering text, detailed scene composition, and multi-subject consistency, surpassing many international competitors. This advancement is attributed to its novel native multi-modal architecture, Unified Transformer (UiT), which integrates various data types at a foundational level, moving beyond traditional modular approaches. AI

Researchers have developed new methods for open-vocabulary semantic segmentation, a task that involves assigning semantic labels to images using flexible category vocabularies without pixel-level training data. One approach, LASA, aggregates attention maps from different layers of Vision Transformers to capture both global structure and local details, improving segmentation accuracy and spatial coherence. Another method integrates differentiable fuzzy logic with foundation models like SAM to refine pseudo-labels and train segmentation models, achieving state-of-the-art results that surpass even densely supervised baselines. A third technique, Open-V, uses a training-free framework that coordinates frozen semantic priors from models like SAM and CLIP for generalized few-shot segmentation, demonstrating strong performance without parameter adaptation. AI

Two new research papers explore the internal geometry of transformer models, focusing on how representations evolve across layers. One paper investigates module-specific weight-space geometries for optimization, finding that assigning different manifold constraints to attention and MLP layers in GPT-2 improves performance and stability. The other paper analyzes the trajectory geometry of representations, using metrics like length, curvature, and convergence to understand how semantically related prompts evolve, revealing distinct phases of processing and correlating curvature with computational complexity across GPT-2, TinyLlama, and Qwen2.5. AI

Researchers have introduced IDEAL, an In-depth Alignment framework designed to improve discrete representation autoencoders (RAEs) for image generation. By combining both shallow and deep features from vision foundation models (VFMs), IDEAL enhances the preservation of fine-grained visual detail and semantic richness. This approach leads to superior reconstruction performance, achieving a new state-of-the-art rFID score of 0.61 on ImageNet and a gFID of 1.89 for autoregressive image generation. AI

Researchers have introduced "Next Forcing," a novel multi-chunk prediction framework designed to enhance autoregressive video generation. This method addresses limitations in current models by providing explicit signals about future dynamics, leading to faster training convergence and improved accuracy, particularly at high frame rates. The framework also accelerates inference and demonstrates better adherence to physical laws in generated videos. AI

Researchers have developed two novel deep learning approaches for improving Positron Emission Tomography (PET) image denoising. UniPET utilizes domain generalization and region-aware learning to create a universal model capable of denoising images across various dose reduction factors, addressing issues of style misalignment and over-smoothing. U-TTT employs test-time training with dual-domain adaptation (spatial and frequency) to dynamically adjust model parameters during inference, enabling robust generalization even with unseen dose levels or scanner types. AI

Researchers have developed POTATR, a new lightweight image-to-graph model for extracting tables from documents. This 29 million parameter model significantly outperforms existing methods on the PubTables-v2 benchmark, achieving a GriTS_Con score of 0.964. POTATR is also considerably faster and more cost-effective than current large language models, with its output being spatially grounded for verification and further integration. AI

Researchers have developed a novel data synthesis method to create neural machine translation (NMT) models for low-resource Indigenous languages, specifically Q'eqchi' Mayan. By transforming dictionaries into a synthetic corpus and using Parameter-Efficient Fine-Tuning (PEFT) with LoRA adapters on an mT5-base model, they achieved strong structural acquisition. However, the resulting model showed a significant gap in lexical grounding compared to organic language, indicating that while synthetic data is effective for learning grammar, authentic data is crucial for semantic refinement. AI

This article provides a practical guide to fine-tuning large language models like Llama 3 using Parameter-Efficient Fine-Tuning (PEFT) methods, specifically LoRA and QLoRA. It explains that while base LLMs are general, fine-tuning can adapt them for specific tasks, tones, or knowledge. LoRA achieves this by training only a small set of adapter weights instead of the entire model, significantly reducing computational cost. QLoRA further optimizes this by incorporating 4-bit quantization, enabling fine-tuning of very large models on limited hardware. AI

Kwai has released Keye-VL-2.0-30B-A3B, an open-source multimodal foundation model designed for long-video understanding and agentic intelligence. This model utilizes DeepSeek Sparse Attention to process up to 256K context, capturing essential frames and temporal dependencies in hour-long videos. It also incorporates Cross-Modal Multi-Teacher On-Policy Distillation to enhance multi-task alignment and agent collaboration across various scenarios. Evaluations show state-of-the-art performance on video understanding and temporal localization benchmarks. AI

Researchers have identified that Chain-of-Thought (CoT) fine-tuning, while improving reasoning, significantly degrades long-context recall in hybrid linear-attention models. This issue, termed "attention amnesia," causes performance drops on tasks like Needle-In-A-Haystack. A new training-free method called QK-Restore has been proposed to fix this by restoring specific query-key projection weights from a pre-fine-tuning checkpoint, successfully recovering long-context capabilities without sacrificing reasoning performance. AI

Researchers have developed a novel unsupervised learning algorithm for neural disentanglement using holographic reduced representations (HRR). This approach treats disentangled representations as symbolic structures, moving away from continuous representations common in prior work. The HRR unbinding operation demonstrates an inductive bias for separating factors, achieving competitive results on disentanglement metrics and showing robustness to noise. AI

Researchers have developed a new diagnostic theory and benchmark to understand how well local score models can extrapolate across different system sizes. They found that architectural locality alone is insufficient for stable size extrapolation, which is instead governed by the quasi-locality of the Gaussian-smoothed score. The study introduces the Finite-Depth Local Flow (FDLF) benchmark to empirically validate these findings, demonstrating that stable extrapolation depends on the interplay between spatial mixing, score quasi-locality, and model receptive fields. AI

A new research paper investigates how "thinking" mechanisms in large language models affect instruction following. The study found that while overall performance changes are minor, the "thinking" process alters error patterns, improving some instructions while worsening others. Specifically, "Planning" constraints benefit from thinking, whereas "Precision" constraints consistently degrade. Analysis of model traces revealed differing correlations between trace relevance and final answer compliance across these constraint types. AI

Researchers have developed new methods to improve supervised fine-tuning (SFT) for large language models. One approach, FisherAdapTune, uses the Fisher information geometry to dynamically select parameter groups for adaptation, enhancing in-distribution performance and zero-shot transfer. Another set of methods, including Target-SFT and PriFT, reinterprets SFT as target distribution design. These techniques aim to create more stable and effective training objectives by better aligning the fine-tuning process with the model's pretrained knowledge, leading to state-of-the-art results on various reasoning and code generation tasks. AI

Hugging Face has released TRL v1.0, a library for post-training reinforcement learning. A related announcement highlights RapidFire AI, a method that accelerates TRL fine-tuning by up to 20 times. These developments aim to improve the efficiency and adaptability of AI model training. AI

Meta has laid off approximately 8,000 employees, with a significant portion impacting development and management roles as the company pivots towards AI. This move, part of a broader strategy to streamline its culture, saw over 1,400 management positions and nearly 1,000 software engineering roles affected. In contrast, marketing and sales roles experienced fewer layoffs. Separately, DeepSeek founder Liang Wenfeng's academic achievements have resurfaced, highlighting his decision to pursue AI at Zhejiang University over a prestigious offer from Tsinghua University. AI

Anthropic's Claude Fable model exhibited overly strict safety guardrails, refusing to answer basic questions like "What is DNA?" In a separate demonstration, the same model successfully completed a playthrough of "Pokémon FireRed" without human intervention, taking 50 hours. Meanwhile, Google's NotebookLM is being highlighted for its ability to help users quickly understand multiple documents. AI

OpenAI CEO Sam Altman informed employees that the company plans to go public within the next year, having submitted a draft S-1 filing to the SEC. Separately, a new, powerful, and expensive model from Anthropic called Fable 5 has been released, with a caution for general users due to its capabilities. AI

A new approach called Dexterity-BEV is being introduced to address the data challenges in embodied intelligence by adapting the Bird's-Eye View (BEV) methodology from autonomous driving. This method aims to unify heterogeneous robot data, including visual inputs, sensor readings, and action commands, into a common spatial reference frame. This unified representation is intended to enable more scalable and transferable training for robots, moving beyond simple data aggregation to establishing a foundational data infrastructure for embodied AI. AI

Researchers have introduced a new framework to unify and clarify concept-based representational similarity in machine learning models. The framework decomposes alignment into representation vs. concept and instance-wise vs. distributional levels, identifying four key properties. They also developed an intervention-based benchmark called \InterVenchA to measure these properties and proposed the Coupled Sparse Autoencoder (CoSAE) method, which demonstrates that strong alignment emerges when multiple objectives are jointly enforced, even with minimal paired data. AI

A new research paper investigates whether video foundation models possess an understanding of intuitive physics. The study probes frozen representations of models like V-JEPA, VideoMAE, and LTX-Video using benchmarks such as IntPhys2 and Minimal Video Pairs. Results indicate that V-JEPA performs best, particularly with temporal dynamics probes, while VideoMAE is competitive, and LTX-Video shows weaker but present signals. The research also found that physics knowledge is more accessible in intermediate to late layers of these models. AI

Researchers have developed Hypnos, a new foundation model for sleep physiology that utilizes next-token prediction for representation learning. Trained on eight different sensing modalities from over 20,000 polysomnography recordings, Hypnos tokenizes physiological signals and uses an auto-regressive RQ-Transformer to predict future data points. This approach significantly outperforms existing models on various benchmarks, including sleep stage classification and atrial fibrillation detection, while requiring substantially less labeled data. AI

Researchers have developed a novel method for automatically generating Individualized Education Programs (IEPs) in Traditional Chinese, addressing a significant gap in special-education NLP. The proposed Corpus-Grounded Feature Diffusion (CGFD) pipeline utilizes a low-resource fine-tuning approach with a modified Breeze-7B model. This system achieves state-of-the-art results on a held-out test set, outperforming several leading LLMs in zero-shot performance while ensuring privacy-preserving, local inference. AI

Researchers have developed a new method to evaluate the quality of generated samples from conditional models, particularly when exploring novel or unobserved conditions. This approach uses a post-hoc trust score that combines global realism and attribute faithfulness, requiring only the original training distribution for assessment. The score can effectively filter, rank, and abstain from generations, demonstrating improvements in downstream predictive performance in biological imaging and vision benchmarks. AI

Researchers have developed scTransformer, a novel approach that integrates gene regulatory information into Transformer models for analyzing single-cell RNA sequencing data. This method enhances interpretability and robustness by incorporating prior biological knowledge into the model's attention mechanisms. Evaluations show scTransformer improves cell-type classification accuracy and produces more biologically meaningful representations compared to standard Transformers. AI