Ask questions across all of research

Samuel SchmidgallXiaokai ZhuMarian Shaw

Google DeepMind's Co-Scientist, a Gemini-based multi-agent system, transitions from *in silico* hypothesis generation to an execution-grounded research partner, demonstrating closed-loop scientific workflows in real-world materials science, biology, and computer science. The system significantly reduces result hallucination to 4% and eliminates complete data fabrication in autonomous paper generation, while also improving research safety.

Aozhe WangZhengxi LuJianze Wang

Researchers from Zhejiang University and Alibaba Group introduced Test-Time Policy Optimization (TTPO), a method for improving large language models' mathematical reasoning without ground-truth labels by applying an asymmetric objective for dense, token-level supervision. TTPO achieved a 7.2 percentage point absolute gain over the base Qwen3-1.7B model in a label-free test-time training setting, consistently surpassing existing label-free baselines and enabling smaller models to match larger untuned models.

Chen Yueh-HanJiaxin WenJan Hendrik Kirchner

Automated Alignment Researchers (AARs) reliably mitigate ten common AI alignment failures, preserving model capabilities, generalizing across diverse evaluations, and outperforming methods proposed by human researchers on established benchmarks. AARs successfully post-trained an early Claude Opus 4.8 checkpoint to production-level alignment using significantly less data.

Lukas KuhnLucas MaesGiuseppe Serra

LeVJEPA presents an efficient and scalable self-supervised video pretraining method that eliminates complex architectural heuristics by employing a single encoder with statistically principled regularization. This approach yields competitive or superior performance on downstream tasks while requiring significantly less computational cost than existing methods.

Google Research and Virginia Tech researchers developed WikiSkill, a framework that compiles agent experience into a persistent knowledge base (wiki) to guide the continuous evolution of agent skills. The framework consistently improved agent performance across diverse models and tasks, demonstrating average performance gains of 3.3 to 12.0 points over existing methods and enabling effective skill transfer.

Maximilian DuZhanyi SunChen Xu

This research introduces "Memory Anchors," a concept for identifying a critical subset of past experiences essential for preventing catastrophic forgetting in continual robot learning. It proposes the ANCHORER algorithm, which strategically populates experience replay buffers with these anchors, leading to substantial reductions in forgetting on robotic manipulation benchmarks and achieving a 1.7x higher final success rate on a real-world garment folding task.

Sarthak KamatAdam RashidSatvik Sharma

Researchers from Stanford University, MIT, and Scale AI developed Simulation Pre-training for Dexterity (SPD), a framework enabling scalable data collection in simulation for dexterous robot hands. Policies pre-trained with SPD on 75 hours of simulation data achieved up to an 18 percentage point improvement in real-world task progress compared to policies trained from scratch with limited real-world data, particularly when using a combination of long history conditioning and short action chunks.

Niklas MuennighoffZhengyang WangZeyi Chen

A novel context management strategy called Prefix Sliding enables large language models to perform extremely long reasoning tasks with constant computational and memory costs, achieving approximately a 3x speedup for inference and facilitating more effective reinforcement learning on extended rollouts. This method allows models to process sequences hundreds of thousands of tokens long without prohibitive expense or memory exhaustion.

This research formally characterizes the latency of cascade voice agents, establishing an irreducible lower bound for response times and defining conditions for achieving human-competitive conversational speeds. The work quantifies how architectural choices and multi-step reasoning impact the 'silence gap' perceived by users, providing a framework for optimizing system design.

Yiwen ChenGuosheng LinChi Zhang

The Code World Model integrates large language models as coding agents with video models to construct open-ended virtual worlds. This approach functionally separates logical world evolution from high-fidelity visual realization, enabling consistent, causally complex environments and precise spatiotemporal visual control via an intermediate "proxy" representation.

Zekai LiJiaming TangZhijian Liu

FlashVLA introduces a streaming action decoding framework for flow-matching Vision-Language-Action (VLA) models, reducing action decoding latency by up to 2.81x and enhancing asynchronous execution stability. This approach improves task success rates, particularly for long-horizon tasks (e.g., a 36.6 percentage point increase on RoboTwin 2.0 long-horizon tasks), enabling real-time robotic control at sustained frequencies of 30-50 Hz on a single GPU.

Subhadeep PalFiona Y. WangMarkus J. Buehler

Researchers from MIT introduce SwarmWorld, an environment where initially homogeneous language-model agents spontaneously self-organize and collectively evolve functional technologies through stigmergy. The study found that interacting societies achieved higher technological diversity and resilience compared to independent agents, with performance benefits dependent on interaction mechanisms and timescale.

A collaborative research effort from Tsinghua University, Zhejiang University, and Alibaba Group introduced Self-OPD, a teacher-free on-policy distillation framework for Flow Matching models that generates dense, step-wise supervision via self-exploration. This method achieves high performance across single and mixed-reward visual generation tasks, surpassing existing teacher-based and reinforcement learning approaches while considerably reducing training time.

Researchers at Tsinghua University developed PuRo-2B, a collection of 2-billion parameter language models trained on 1.4 trillion tokens using consumer-grade RTX 5090 GPUs for under $6.9K. The canonical PuRo-2B model achieved performance comparable to Qwen2.5-1.5B, while a cost-optimized variant surpassed Qwen2-1.5B for approximately $4.4K.

Jiaming ZhouQihang ZhangGangwei Xu

Zero-WAM, a causal video-action model developed by researchers at Robbyant and HKUST, enables zero-shot cross-task robotic generalization by interpreting human demonstration videos as in-context task specifications. The approach leverages a scalable pipeline for generating human-robot paired data and a novel training objective, leading to improved success rates on unseen manipulation tasks in both simulation and real-world settings.

Yijia ShaoDora ZhaoVishakh Padmakumar

Stanford University researchers analyzed nearly 250,000 real-world human-AI conversations from Claude.ai to understand collaborative dynamics, finding that over half of interactions involve consequential work and are human-led, with users actively engaging in adaptation and learning despite frequent friction.

Researchers from Siemens, UC Berkeley, Microsoft, and ETH Zurich developed ARLI, a framework that enables reinforcement learning finetuning of generalist robot policies, effectively overcoming the significant inference latency inherent in large models. This approach allowed a robot policy to achieve nearly 100% success in real-world manipulation tasks after 100-125 episodes, a substantial improvement over baselines that struggled with latency.

Google Research developed the Planetary Prediction Engine (PPE), an autonomous AI architecture that translates natural-language queries into executed geospatial predictive models for global challenges. This system streamlines complex workflows by autonomously discovering and integrating diverse geospatial data, including foundation model embeddings, and consistently improves predictive accuracy across various tasks like epidemiological nowcasting and food security mapping.

Lehong WuYuxiao QuZheyuan Hu

Carnegie Mellon University researchers developed ℛ3, a two-stage post-training method to enable Vision-Language Models (VLMs) to generate free-form natural language reasoning for guiding robotic manipulation policies. This approach significantly enhanced robotic task generalization, particularly for out-of-distribution scenarios, and demonstrated that explicit inference-time reasoning acts as a dynamic computational mechanism.

Recent AI systems promise autonomous scientific discovery, claiming to discover algorithms and produce research papers, yet understanding whether they exhibit creativity, the capacity to produce solutions that are both novel and useful, remains an open question. We present a framework for evaluating multi-turn LLM research agents' creativity using ML engineering tasks as a testbed, through three dimensions: P-Creativity (psychological novelty: novel relative to the agent's own prior solutions within a run), H-Creativity (historical novelty: novel relative to the corpus of human solutions), and Usefulness (task performance). Evaluating two agent frameworks, AIDE and AIRA-Dojo, on 10 Kaggle-style machine learning tasks from MLE-Bench, we develop an LLM-as-a-Judge pipeline and verify its strong correlation with human creativity judgments, providing a reliable automated metric for P-Creativity evaluation at scale. Applying this pipeline to agent trajectories, we find: (1) all agents exhibit declining P-Creativity as they transition from exploration to exploitation; (2) LLMs exhibit greater H-Creativity than medal-winning humans, yet achieve lower performance. Our findings reveal that current agents can explore novel regions of the solution space but lack the capacity to convert this novelty into improved task performance.