HuggingFace

This paper decomposes transformer representation updates into parallel and perpendicular components to study evolution geometry, linking it to editing robustness, compression diagnosis, and training improvements.

ModaLens is a paired image-swap audit that measures how report availability reduces image sensitivity in medical vision-language models, demonstrated using MedGemma-27B on the MIMIC-CXR dataset.

This paper investigates the losslessness of Orthrus, a hybrid autoregressive-diffusion model for inference acceleration, finding that it requires high numerical precision (FP32) for exact trajectory matching, while BF16 divergence does not impair downstream performance.

LynnReal-Omni is a unified multimodal video diffusion framework that integrates agentic visual controls with high-fidelity generation and real-time acceleration for stable, controllable video creation.

This paper proposes an exploration-guided prompt scaffolding framework that dynamically adjusts training prompts for multimodal reinforcement learning, achieving up to 9.7% relative improvement in performance on benchmarks.

PhysBrain 1.5 is a unified model that integrates physical environment understanding, action generation, and future state prediction via autoregressive training, achieving state-of-the-art open-source performance on 28 embodied benchmarks.

KaiNinja extends a native 3D generator to produce part-level outputs using a dual-volume representation, improving both part and whole-object fidelity without requiring segmentation.

Atria Dawn Preview is a foundation agentic language model designed for scientific research, achieving competitive benchmark results and demonstrating a shift toward human-AI project-level collaboration.

This paper introduces Elo-per-token analysis to study how LLM agents allocate test-time compute, revealing that agents initially outperform independent sampling but slow down over time, with parallel sessions offering performance gains.

This paper introduces a method for vibe design agents to explore diverse UI alternatives by separating exploration from implementation through structured design specifications, as evaluated on 168 prompts and a large online experiment with over 300,000 tasks.

This paper introduces SAILS, a method for selecting optimal poison sets in backdoor attacks against large language models, improving worst-case attack success by 30 percentage points over baselines.

HazardAuditor introduces an execution-grounded framework for supervising safety in computer-use agents, with Guard Policy Optimization improving safety outcomes by up to 16.5% accuracy over prior methods.

RSIAgent is a training-free multi-agent framework that enables digital agents to adapt to new environments through recursive self-improvement, autonomous memory construction, and broad-then-deep exploration, outperforming closed-source models on benchmarks.

Omni-Streaming Thinking improves streaming omni-modal reasoning by deferring claims until cross-modal verification, reducing premature commitment and auditory hallucinations.

Dream-RSI is a framework for scalable recursive self-improvement in AI agents that uses historical discovery trees to create a replay simulator for offline policy evaluation, reducing online costs and improving discovery efficiency.

The paper introduces BVB, a benchmark for agentic video understanding via programmatic reconstruction in Blender, evaluating models on perceptual similarity and spatiotemporal fact retention.

Discovery Foundation Models are proposed as general-purpose systems for enabling open-ended scientific discovery through iterative problem formulation, hypothesis testing, and evidence-based revision across dry and wet lab settings. The paper introduces a framework with capabilities like problem discovery and continual improvement, instantiated with systems like Zetema and GALILEO.

This is an audio tokenizer and LoRA tool for the YuE2-3B model, enabling users to tokenize real audio recordings and generate new songs or covers.

This paper introduces techniques to manage memory peaks in training large Mixture-of-Experts models with long context lengths, including Pipelined LLEP, Ring-DTP, SCO, and OffloadStreamAdamW, which enable fixed GPU working sets and improve throughput up to 10.4x.

SpectralShift introduces a spectral reparameterization approach to effectively extend the context window of Gated DeltaNet models by reshaping the decay spectrum, improving long-context capabilities through continual pretraining.