alphaXiv connects papers, researchers, and organizations, grounding the answer in the underlying work.
GLM-5.3 is built on the same base model as GLM-5.2, with every reported gain coming from scaled post-training on long-horizon task environments. It posts open-source SOTA on Terminal Bench 3.0 (28.3 vs 4.6) and Agents' Last Exam (28.5), and a 50% improvement on Z.ai's in-house Code Bench while spending fewer output tokens. Cyber capability grew fastest of all: 84.5% on CyberGym is the best result on that benchmark, and exploitation scores more than doubled. Weights are slated for release two weeks after launch, once safety hardening completes.
DreamX-Phi 1.0 introduces an action-conditioned video world model for robotic manipulation, incorporating geometry-aware SE(3) action representation and comprehensive physical-consistency supervision. This model achieved first place on the WorldArena 2.0 Track 1 leaderboard for video prediction with an EWMScore-P of 60.65 and tied for second on Track 2 for policy training, demonstrating a 67.19% success rate on the "Adjust Bottle" task.
V-RAE constructs generative latent spaces for video by integrating frozen Vision Foundation Models (VFMs) with a learnable temporal pooling module and a spatiotemporal decoder. This method improves video generation quality, reduces diffusion model training time by up to 6x, and maintains strong semantic information in the latent representations.
Meituan and MBZUAI researchers developed AutoDesign, a meta-harness optimization framework enabling design systems to recursively improve their operational components based on human-aligned evaluations. Applied to academic paper-to-poster generation, it achieved a PosterBench Score of 78.32, outperforming other systems by over 7 points, and autonomously generated posters in 40 minutes at under $3 each.
The Intern-S2-Preview Team at Shanghai AI Laboratory developed Intern-S2-Preview-397B, a scientific agentic foundation model capable of multimodal scientific understanding, reasoning, and long-horizon tasks through iterative, tool-grounded problem-solving. This model demonstrated competitive or leading performance across diverse scientific, multimodal, and agentic benchmarks, and features specialized modules for time series processing and modular domain adaptation.
Researchers uncovered an architectural vulnerability in major proprietary large language model APIs where encrypted internal reasoning traces can be extracted in plaintext by weaker models within the same provider's ecosystem. This enables unauthorized model distillation, secret extraction, jailbreaking, and invisible prompt injection attacks.
Researchers at Inherent developed Faraday, an AI Scientist agent designed to replicate research papers by reproducing experimental figures using a "Coding Agent as a Tool" (CAT) paradigm. Trained on the Replica task space with a novel rubric-based reward system, Faraday demonstrated superior scientific rigor and replication accuracy compared to leading frontier models, including Claude Opus 4.8 and GPT-5.5.
Synthetic Persona Pretraining (SPP) introduces a method to instill desired assistant values and identity into large language models directly from "token zero" during pretraining. This foundational approach yields improved constitution following, enhanced jailbreak robustness, and better alignment to out-of-distribution moral dilemmas, demonstrating a more deeply embedded and robust alignment compared to post-training interventions.
Agent instruction files like CLAUDE.md exhibit unbounded growth due to "catastrophic remembering," a process where the original rationale for instructions is lost, making safe deletion difficult. Implementing prompt comments that record outcome-grounded reasoning effectively halts this growth, reducing excess instruction size from 211.3% to 1.4% and improving agent instruction-following correctness by 11.6 percentage points.
STREAMTTT introduces a streaming Video-Language Model that overcomes the perception-memory trade-off in real-time video understanding through a dual-memory architecture and a dedicated real-time QA corpus. Its 4B parameter model achieved a 68.59 two-track average on OVO-Bench, outperforming HERMES-7B by 9.39 points and improving real-time perception by 1.4 points and backward tracing by 3.7 points compared to a matched-scale baseline.
OmniScientist presents an end-to-end AI scientist capable of conducting multidisciplinary research directly from raw, heterogeneous scientific evidence, integrating perception throughout the research lifecycle. It successfully completes full research workflows across diverse scientific domains and modalities, demonstrating improved research quality and multimodal grounding compared to systems relying on pre-processed data.
A "strong-to-weak scaffolding" method enables a powerful builder language model to automatically create inference-time harnesses, significantly enhancing a weaker target model's performance on complex tasks without any parameter modification. This approach yielded up to an 87% relative improvement in GPT-5.4-mini's macro-average accuracy on Theory-of-Mind benchmarks, achieving 0.912 and approaching human-engineered solutions.
Alaya-EVOKE introduces an interactive world model that enables continuous, open-ended virtual world generation by decoupling persistent state from the denoiser and utilizing a long-horizon, dynamically conditioned teacher. The system achieves hour-scale coherent generation with bounded computational costs and supports responsive mid-session interaction and geometric recall.
This research from FAIR at MSL Meta and New York University establishes that reliable scaling laws are observable even in very small foundation models, provided hyperparameters are rigorously tuned. The study presents a methodology that enables cost-effective, model-centric research by revealing how hyperparameter sensitivity at smaller scales can obscure these laws.
The CAKE framework, a co-design approach from NVIDIA and Carnegie Mellon University, integrates GPU kernel optimization agents with an evolving compiler harness to overcome limitations of traditional black-box compilation. It enables agents to author high-performance kernels by providing localized diagnostics and a typed, hardware-explicit intermediate representation, achieving significant speedups over baselines and synthesizing frontier kernels for modern GPU architectures.
Morpho, Inc. developed a fidelity-constrained anchoring framework that post-processes black-box denoiser outputs by linearly blending them with the noisy input. The approach effectively controls output fidelity to the input image, balancing denoising performance and naturalness, with SSIM-based anchoring demonstrating robustness across varying noise levels.
A systematic evaluation framework was developed for autonomous AI R&D agents, moving beyond single final scores to diagnose performance using process-level metrics, controlled comparisons for experience reuse, and harness impact analysis. The study revealed that current frontier models reliably optimize artifacts but exhibit rare genuine innovation (1.2% novel solutions) and inconsistent performance heavily influenced by bottlenecks in solution framing and feedback control, as well as the effectiveness of experience reuse.
PlayWorld introduces a benchmark for interactive video world models, employing an adaptive "Agent Player" to achieve long-horizon objectives and enable consistent cross-model comparison. This evaluation reveals that current models frequently struggle with maintaining sustained world evolution and global spatial consistency during complex interactions.
Latent On-Policy Self-Distillation (LOPD) introduces a framework where the privileged context for self-distillation is a learnable latent representation, moving beyond human-engineered artifacts. This approach consistently enhanced agent performance on tool-use and code generation benchmarks, yielding superior aggregate results and improved sample efficiency across multiple LLM backbones.
This research systematically investigates Massive Activations (MAs) in Hybrid Linear Attention (HLA) Large Language Models, identifying two novel, architecture-aligned morphologies: Pre-attention Spikes (PAS) and Inter-spike Plateaus (ISP). The work demonstrates these patterns' robust recurrence across model scales up to 397B parameters and diverse settings, providing a mechanistic explanation for their emergence and evolution.
