Google DeepMind

Frozen video diffusion models can effectively serve as competitive encoders for a wide range of tasks, merging generation and understanding seamlessly.

VLMs are prone to critical failures that vary significantly across cultures, exposing the inadequacy of Western-centric safety benchmarks.

Gemma 4's unified architecture and reasoning mode enable it to outperform larger models in human-rated tasks while maintaining high efficiency.

RMMD not only accelerates model inference by 7.5x but also outperforms its teacher model on nearly all target weather variables, showcasing a breakthrough in distillation techniques.

Claw-like agents are vulnerable to severe security breaches, with malicious plugins achieving a 100% success rate in attacks.

TempoWave reveals that rethinking numerical embeddings can unlock significant improvements in LLM forecasting performance.

Achieving efficient uncertainty quantification in multi-modal regression could redefine the landscape of trustworthy large-scale learning.

TokenMinds reveals that combining discrete SID-based user tokens with dense embeddings can significantly enhance user modeling in recommender systems at scale.

LLM critiques can be systematically evaluated for alignment with human judgment, revealing that better models significantly enhance evaluation reliability.

Achieving sound probabilistic verification for AI agents could redefine how we secure complex systems against policy violations in uncertain environments.

The Physics-IQ Verified benchmark reveals that over half of the evaluated samples can be significantly refined, leading to notable shifts in model performance rankings.

Language adherence in ASR can be significantly improved using soft prompting techniques, leading to better transcription quality in multilingual contexts.

A mere 1% of poisoned samples can flip classifier labels, leading to catastrophic false positives and negatives in jailbreak detection systems.

Real-time LLM-generated user personas can dramatically enhance viewer engagement by dynamically balancing existing interests with new content recommendations.

ATLAS achieves a staggering 5-10x increase in sample efficiency for discovering interpretable behavioral models, revolutionizing experimental design in cognitive science.

The journey from AGI to ASI could unfold through multiple pathways, challenging the notion of a single transformative leap in AI progress.

Hierarchical VLA agents can outperform traditional flat control systems by leveraging unified design principles that enhance task performance across various complexities.

Control interventions are often detected by LLMs, with awareness levels varying significantly across models and tasks, revealing vulnerabilities in AI safety protocols.

Reusing training data during inference can boost imitation learning performance by up to 46%, reshaping how we approach generalization in AI systems.

Calibrated safety flags in medical summaries can reduce unflagged omissions by up to 5 times compared to existing methods, enhancing clinician confidence in LLM outputs.

RLNS turns a classic heuristic into a powerful MCMC sampler, enabling efficient combinatorial optimization without the need for exact solutions.

VLMs struggle with procedural 3D modeling, often producing flawed outputs due to API mismatches and geometric disconnections, but performance can be significantly boosted through iterative refinement.

Weather models can do climate, too: ArchesWeather and ArchesWeatherGen, originally built for short-term forecasting, show surprisingly strong performance in multi-decadal climate simulations when forced with SST and SIC.

Larger models learn more not just because of increased capacity, but because they experience less interference during training, allowing them to retain rare and complex tasks that smaller models forget.

LLM-powered honeypots can trick even frontier models into longer interactions than rule-based systems, all while costing less to run.

Splitting attention and feedforward networks onto separate GPUs can unlock 4x higher MoE LLM throughput, but only if you carefully tune the GPU partitioning strategy based on the workload.

Forget vague AGI claims – this cognitive taxonomy provides a concrete, measurable framework to map AI capabilities against human cognitive abilities.

AI can now autonomously solve open math problems, cracking 9 Erdős problems and 44 OEIS conjectures at a reasonable cost.

Forget hand-engineered safety constraints: MARL agents trained for high-speed quadrotor racing not only beat human champions, but also cut collision rates in half.

Training a foundation model on a trillion minutes of wearable sensor data unlocks surprisingly accurate predictions across a wide range of health conditions, even with limited labeled data.

Dynamic quantization, a widely adopted optimization for efficient ML serving, can leak your data to adversaries sharing the same batch.

Generative AI evaluation can be sped up by 8-65x without sacrificing accuracy by proactively focusing on the most informative test cases using a pre-trained Gaussian Process surrogate model.

Entropy regularization makes planning provably easy: SmoothCruiser achieves polynomial sample complexity in MDPs where standard methods fail.

DPP-based Monte Carlo integration can offer variance reduction, but choosing the right DPP—fixed vs. tailored to the integrand—determines whether you get a biased but faster converging estimator or an unbiased but standard-rate estimator.

Human-inspired context sensitivity boosts visual reasoning in machines, closing the gap between AI and human perception.

Forget sub-Gaussian assumptions: this semi-bandit algorithm adapts to the true covariance structure of outcomes, leading to tighter regret bounds and better performance.

Ethics interventions in AI development often fail because practitioners don't trust them – here's a breakdown of why, and how to fix it.

Unpacking Google's AI literacy partnerships reveals the surprising complexities of aligning research, industry, and public needs.

LLMs' struggle to grasp subtext—even generating literal clues 60% of the time—reveals a critical gap in their ability to understand nuanced human communication.

LLMs can now generate more relevant and factual movie recommendations by dynamically bridging retrieval and generation with a novel reinforcement learning approach.

Training LLMs to optimize for conflicting objectives between the final output and the reasoning process can significantly degrade the monitorability of Chain-of-Thought, making oversight more difficult.

Refining generative models with discriminator guidance provably improves generalization, offering a theoretical justification for techniques like score-based diffusion.

Forget finetuning: DynaEdit unlocks complex video edits like action modification and object insertion, all without training, using clever manipulation of pretrained text-to-video models.

Forget fine-tuning: surprisingly, single neuron activations in VLMs can be directly probed to create classifiers that outperform the full model, with 5x speedups.

LLMs get *more* honest when they have time to reason, defying human tendencies and revealing surprising insights about their internal representational geometry.

Mixture-of-Experts models might be hiding more of their reasoning than we thought, thanks to a newly quantified "opaque serial depth" metric.

Forget black-box policies: CSRO uses LLMs to generate human-readable code policies in multi-agent RL, achieving performance competitive with traditional methods.

LLM-powered diagnostic AI is ready for prime time: a real-world clinical trial shows it's safe, patients love it, and doctors find it useful.

Ditch the slow sampling dance of diffusion models: Variational Flow Maps let you condition image generation in a single pass by learning the right initial noise.

Achieve significantly better code generation and mathematical problem solving from diffusion language models with a simple, training-free sampling tweak that encourages diversity.

LLMs can drastically accelerate robot planning in cluttered environments by injecting common-sense priors about object locations and co-occurrences, slashing planning time by up to 72% in real-world experiments.

Cracking DNNs is now easier than ever: Kraken extracts parameters from GPU Tensor Cores via near-field EM attacks and even sniffs LLM weights from a meter away.

LLMs are becoming "epistemic agents" that shape our knowledge environment, so we need a new framework for evaluating and governing them based on trustworthiness, not just performance.

Frontier models are surprisingly good at taking actions at extremely low, calibrated probabilities, raising concerns about their ability to evade pre-deployment safety evaluations designed to catch malicious behavior.

DINOv2's impressive unimodal performance doesn't translate to cross-modal understanding, but a simple training tweak can align embeddings across RGB, depth, and segmentation without sacrificing feature quality.

Forget slow prefix trees: STATIC unlocks massive speedups (up to 1033x) for constrained LLM decoding on GPUs/TPUs by vectorizing trie traversals into sparse matrix operations.

Unlock asymptotically normal and semiparametrically efficient estimators in adaptive data collection by using a novel target-specific condition called "directional stability," which is weaker than previous target-agnostic conditions.

Tri-modal masked diffusion models can now be trained from scratch, achieving strong results in text generation, text-to-image, and text-to-speech, thanks to a systematic exploration of the design space and a novel SDE-based batch size reparameterization.

Existing deforestation monitoring maps misclassify smallholder agroforestry as "forest," risking unfair penalties under regulations like the EUDR.

LLMs can autonomously discover novel MARL algorithms that outperform hand-designed baselines, revealing untapped potential in automated algorithm design.

LLMs can be taught to proactively seek and effectively use conversational feedback, generalizing across tasks and improving their ability to handle ambiguity.

Forget scaling laws: teaching LLMs to learn from feedback lets smaller models rival giants and generalize to new tasks.

Language models organize concepts like months and years into surprisingly clean geometric structures because of hidden symmetries in language statistics, even when those statistics are heavily perturbed.

Forget scaling laws: the secret to AGI might be teaching AI to argue with itself through high-quality conversational scaffolds.

Boost macrocycle generation rates from 1% to 99% by guiding diffusion models with persistent homology, opening new avenues for drug discovery.

Robots can now learn long-horizon tasks far more effectively by distilling complex histories into a few key visual moments, outperforming standard imitation learning by 70% on real-world tasks.

Forget rigid heuristics: this adaptive AI delegation framework dynamically adjusts task allocation, authority transfer, and trust-building, promising more robust agentic systems.

People prefer AI advisors, but AI delegates that autonomously negotiate on their behalf actually lead to higher individual gains and improve overall group welfare in multi-party bargaining games.

A redesigned AlphaFold Protein Structure Database offers improved usability and expanded structural coverage, making high-accuracy protein structure predictions even more accessible.

Reasoning-based safety guardrails, once thought to be a strong defense against jailbreaks, crumble with just a few strategically placed tokens.

AlphaFold didn't just solve protein structure prediction; it unlocked a new era of biological discovery, making nearly the entire genome structurally accessible.

Ditch the high-fidelity simulator: IRL-VLA uses a lightweight reward world model trained with inverse reinforcement learning to enable efficient and effective closed-loop RL training for autonomous driving.

DPO's success isn't just clever engineering—it's deeply rooted in human choice theory, unlocking a surprisingly flexible framework for preference optimization and justifying many DPO extensions.

Ditch reward models: Nash Mirror Prox achieves fast, stable convergence to a Nash equilibrium directly from human preferences, sidestepping the limitations of traditional RLHF.

AlphaFold3 doesn't just predict single protein structures; it tackles the messy reality of biomolecular interactions, from protein-protein binding to protein-nucleic acid complexes, opening new doors for drug discovery and genomic research.