AI Research Digest

Agentic Systems and Autonomous Research

• 🔥 AutoScientists: Introduces decentralized AI agent teams that self-organize around hypotheses,...

No significant updates today.

三篇新作勾勒出智能体开发全栈路径：

• QUEST 用统一评分树合成8K任务，通过中训练+SFT+RL训练2B-35B开源模型，在八个深度研究基准接近闭源前沿水平。
• Shepherd 提供函数式运行时基座，将执行轨迹转为Git式可分叉对象，支持元代理实时监控、回滚与并行干预，性能提升显著。
• SkillEvolBench 构建180任务基准，揭示当前代理多停留在情景复用，难以稳定形成可迁移程序化技能。
  全栈闭环正加速形成。

Two complementary methods advance digital-twin creation from limited inputs.

• Pantheon360 employs 3D-aware 360° video diffusion plus an explicit 3D...

• 模型合并的限制：单语预训练模型直接合并会因表示差异过大导致性能崩溃，混合数据训练更稳定。
• 原生多模态路线图：定义架构原生性，划分Multi-to-Text等三类模型，并系统梳理从数据到部署的完整工业流程。
• 趋势启示：预训练阶段需早期对齐，未来多模态模型或需从头设计原生架构而非后期合并。

视频世界模型正从生成迈向交互评估，三篇新作展现并行进展。

• 并行工具调用：ParaVT首创多代理RL框架，一次性调度多时间窗裁剪，解决Tool Prior Paradox，六基准平均提升7.9%。
• 交互式基准：WBench推出289案例、1058回合的多轮测试，覆盖五维度与四交互类型，无模型全优。
• 自回归蒸馏：AFD提出on-policy对抗流蒸馏，无需教师分数即可实现黑箱提炼，提升运动与物理一致性。

• DeepMind's AlphaProof Nexus solves 9 Erdős problems and 44 OEIS conjectures using LLM proposals checked by Lean, at a few hundred dollars each.
  -...

Google has launched Gemini for Science, a suite of experimental tools on Google Labs—including Hypothesis Generation with Co-Scientist, Computational...

A new arXiv audit of ChatGPT, Copilot, Gemini and Perplexity found ~16% of cited sources across 712 real-world queries were AI-generated, raising risks that users may treat synthetic content as authoritative.

Ai2's MolmoAct2 open robotics model surpasses π0.5 on real-world and simulation benchmarks, runs up to 37x faster, and ships with the largest open bimanual dataset (720 hours, 34,500 demos). All weights, code, and tokenizer are fully public.

Video and Geometry Model Advances

• VGenST-Bench: A benchmark for spatio-temporal reasoning via active video synthesis.
• Good Token Hunting: A...

Unified world models are emerging that turn raw data into sim-ready, interactive environments.

• PhysX-Omni delivers the first unified framework for...

scpFormer introduces a transformer foundation model that unifies single-cell proteomics across technologies via amino acid sequence tokenization and...

Two fresh papers reveal a clear pattern: targeted decoupling and smart pruning tackle core VLM limits in perception, reasoning, and compute.

-...

StepAudio 2.5 shows a single audio-language foundation model can match or exceed specialized systems across ASR, TTS, and realtime spoken interaction....

• Rocket introduces a recurrent self-play RL framework that tunes mixed-type hyperparameters with no manual priors or domain knowledge required
• A...

PhotoFlow introduces a Director-Reviewer-Reflector agent that combines 3D scene understanding with closed-loop camera search to execute...

Two new papers signal a shift toward treating model-generated skills as optimizable, reusable artifacts rather than ad-hoc prompts.

• SkillOpt...

Three concurrent papers signal video AI's shift toward integrated perception-generation-evaluation systems.

• SWIM aligns vision-language attention...

African AI faces an evaluation crisis where model-building capacity vastly exceeds available linguistic expertise and community infrastructure.

-...