alphaXiv connects papers, researchers, and organizations, grounding the answer in the underlying work.
HarnessEval-W introduces an agentic evaluation framework for interactive visual world models, generating transparent "evidence trees" to explain model performance beyond scalar scores. The system aligns strongly with human preferences and provides fine-grained diagnoses for eight distinct evaluation axes, outperforming existing benchmarks in discriminative power.
An AI agent, Claude, autonomously conducts *de novo* protein binder design campaigns by orchestrating open-source tools and managing all design decisions from target research to final selection. The system achieved a 26.8% overall binder hit rate across 15 targets, with 90 designs exhibiting sub-10 nM affinity, and its performance matched or exceeded human expert teams in open competitions.
Terence TaoTerence Tao's "Mathematics in the Age of AI" explores how the mathematical community should respond to advanced AI capabilities, arguing this era presents a "crisis of values and practices" analogous to the early 20th-century foundational crisis. The paper examines implicit mathematical goals such as verification, exposition, community acceptance, and canonicalization, positing they risk diverging from mere solution generation due to AI-driven optimization, and asserts that human understanding remains paramount.
Ali BehrouzProteus introduces incremental memory activation for long-context sequence modeling, which adaptively expands effective memory capacity during processing. This mechanism consistently improved average downstream accuracy and lowered perplexity across four memory-based model architectures, exhibiting enhanced robustness and length extrapolation, particularly at extended context lengths up to 16K.
A "latent-to-pixel" adaptation strategy from Alibaba Token Hub enables the training of pixel-space text-to-image diffusion models that achieve quality comparable to or surpassing latent-space models. This approach demonstrated 3.18x to 4.75x end-to-end inference speedups over latent-space counterparts and a 100.6x speedup over the original 100 NFE latent-space pipeline.
Researchers from Google DeepMind, Carnegie Mellon, MIT, and Columbia University reduced the upper bound on the matrix multiplication exponent from 2.371339 to 2.371177 by employing modern gradient-based optimization and AI-driven algorithmic refinement with AlphaEvolve, enabling computation at an unprecedented scale.
τ0-VLA is a hierarchical robot foundation model that employs world-model-guided test-time computation for high-level subtask generation. This approach dynamically allocates computational resources to evaluate hypothetical subtask outcomes, leading to improved task success rates on long-horizon manipulation tasks and enhanced robustness in out-of-domain scenarios.
GenRouter introduces a dynamic workflow routing framework for agentic image generation, leveraging a unified workflow space called GenCanvas to adaptively match heterogeneous prompts with optimal generative plans. This system achieved superior visual alignment compared to existing static pipelines, concurrently reducing execution cost by over 95% and latency by 65%.
Zetta introduces a closed-loop embodied harness that enables online self-evolution of code-based runtime critics and recovery skills for robotic agents, significantly enhancing task success rates without modifying base policy models. It integrates a dedicated rollout infrastructure (Z-Infra) to accelerate the learning process, demonstrating robust and transferable physical intelligence.
ClawGym II introduces a unified black-box reinforcement learning framework for optimizing general autonomous agents through complex and opaque agent harnesses. The framework consistently improved agent performance, achieving Pass@1 gains of 9.98 to 17.28 points across different harnesses and benchmarks, while maintaining stable training dynamics and supporting joint optimization across heterogeneous harness environments.
Agentic ESOpt introduces an Evolution Strategies (ES) framework for fine-tuning long-horizon LLM agents, enabling full-parameter optimization with inference-level GPU memory requirements. This approach consistently outperforms reinforcement learning baselines on various agentic tasks, including multi-turn Sudoku, ReAct-style tool use, and WebArena, and integrates flexibly with prompt-space optimization techniques.
A neurosymbolic framework for embodied agents combines task-directed visual grounding with constrained symbolic planning to ensure plan executability by design. This approach achieves 94.5–99.5% task success in VirtualHome and 90.3–97.8% in ALFWorld, substantially outperforming direct VLM policies while reducing token generation and visual queries.
HiPHI, a large-scale benchmark, provides 617.5 hours of high-precision optical motion capture data, including 245.7 hours of human-object interaction with synchronized object trajectories and meshes. This dataset, developed by researchers primarily from Noitom Robotics, offers superior physical consistency and broader motion-space coverage than previous datasets, leading to enhanced performance in humanoid robot learning tasks and successful real-world deployment.
Havoc Explorer is a semantic knowledge graph of 611 real vulnerabilities drawn from 11 source classes — public audits, CTF writeups, coordinated disclosures, fuzzing case studies, web3 audits, and academic work — grounded in 3,508 raw artifacts across 59 ecosystems. Rather than a flat CVE list or a single CWE label, each record is represented as a directed mechanism path: source (what the attacker controls) → capability (what the primitive achieves) → impact, with mitigations hanging off capabilities. The graph is built by a multi-agent discovery, curation, and evidence-extraction pipeline, and every node traces back to the evidence record that supports it. The author also tested the graph as a retrieval tool for LLM agents on CVE-Bench, where planning-phase access helped and late-turn access hurt.
This survey reviews how mathematical research in democracy is evolving from idealized theories to data-driven, practical approaches, addressing contemporary challenges in voting, participatory budgeting, and deliberative processes. It highlights the successful application of diverse mathematical tools to design fairer electoral systems and representative citizen assemblies, while also framing mathematics as a critical scaffold for integrating AI into democratic functions.
GigaBrain-0.7 introduces a three-system architecture for embodied foundation models, coordinating understanding, prediction, and action with extensive heterogeneous data pretraining. It achieves improved generalization across diverse robot embodiments and tasks, demonstrating enhanced zero-shot capabilities, language-conditioned instruction following, and task success rates.
The Humanoid Adaptation Framework (HAF) enables generalist Vision-Language-Action (VLA) models to perform complex humanoid whole-body loco-manipulation tasks by introducing a hierarchical action-flow generation and a spectral latent reinforcement learning approach. This framework improved overall average task performance from 53.3% to 70.5% on physical humanoid robots across seven real-world tasks.
HiFi-BRep, from Zhejiang University and Hangzhou Dianzi University, presents a generative framework for B-Rep models that employs a topology-aware dual-stream encoder and a single-stage, validity-constrained decoder. This approach yields the highest structural validity rates on DeepCAD (72.20%) and ABC (32.66%) datasets, significantly narrowing the gap between compilability and true validity while generating models up to 6 times faster than previous methods.
This dissertation rigorously reconciles Adam's theoretical convergence with its widespread empirical success, demonstrating that vanilla Adam converges under problem-dependent hyperparameter choices. It also uncovers intrinsic near-block-diagonal and block-heterogeneous Hessian structures in neural networks, using these insights to develop Adam-mini, a memory-efficient optimizer that reduces optimizer state memory by 50% while matching AdamW's performance on Llama 2 architectures.
On-Policy Distillation (OPD) in Large Language Models exhibits a dual nature of generalization: same-origin teacher-student pairs facilitate broad skill transfer across diverse domains, while cross-origin pairs lead to narrower, distribution-specific fitting. This research found that in multi-teacher OPD (MOPD), prompt routing does not isolate teacher influences, instead causing a "seesaw effect" where increasing one teacher's prompt share dynamically shifts overall student capabilities across multiple domains towards that teacher's profile.
