Stanford HAI

A simple reasoning recipe unlocks massive zero-shot gains in robot foundation models, surpassing the benefits of larger models or more data.

Fine-grained 3D object generation is now possible by directly controlling local geometry and appearance with text and shape primitives, bypassing the need for model retraining.

AI agents are now building climate models that outperform published research and offer novel hypotheses for long-standing scientific debates.

LLM agents achieve dramatic efficiency gains in long-horizon research by offloading conversational history to a central harness, enabling focused, role-specific contexts for each task.

Bosonic Gaussian states can be identified with unprecedented accuracy using a novel warm-start strategy combined with non-Gaussian measurements, outperforming classical Gaussian estimation.

Fusing volumetric and surface brain MRI data in a self-supervised autoencoder uncovers substantially more heritable genetic loci than unimodal approaches.

Post-training LLMs for agentic tasks can drastically reduce solution coverage, a "Sharpening Tax" that a novel Bayesian sampler can mitigate.

Howard's policy iteration for deterministic MDPs is not strongly polynomial, exhibiting an exponential iteration complexity that separates it from the simplex method.

State-of-the-art VLMs struggle to execute scientific research workflows, revealing a critical gap in their ability to interact with specialized software and produce verifiable results.

A single time-series language model now masters forecasting, contextual prediction, and temporal reasoning, outperforming specialized models on key benchmarks.

Splitting models into fully decoupled, communication-free shards cuts zero-order gradient variance by roughly $1/N$, enabling backprop-free pretraining to beat standard monolithic baselines using significantly fewer perturbations.

A single egocentric human video can yield real-world dexterous manipulation policies that beat existing baselines by 3.5x, unlocked by treating video-to-sim translation as a closed-loop, agentic tool-synthesis problem rather than a rigid heuristic pipeline.

FAMOS, a feed-forward model that predicts movable-part segmentation and joint parameters from a sparse, unordered set of partial point clouds, is presented and a procedural data generator that synthesizes self-annotated assets during training is introduced to overcome the limited scale and diversity of existing datasets.

This work demonstrates a compact mobile manipulator that reuses its fingers for locomotion and interaction, without a separate locomotion mechanism.

Upper-bounding the number of distribution shifts in sequential data is provably impossible without parametric assumptions, but conformal inference can uniquely deliver valid, non-trivial lower bounds.

Standard motion benchmarks fail inside the cabin because drivers mostly sit still—until right before a maneuver, when arm movement spikes 3.4x.

Moving beyond passive prompt-and-generate workflows to genuine human-AI collaboration requires solving open technical challenges in mutual agency, real-time latent steering, and dynamic evaluation.

ToPO achieves superior image generation performance by optimizing token-conditioned preferences, outperforming existing methods across multiple evaluation metrics.

Iterative problem-solving reveals that LLMs, while less accurate, can generate insightful solutions that illuminate failure modes in student coding attempts.

Sycophantic agreement in language models is not just a flaw; it’s a systemic issue rooted in the very training objectives we use.

Static sample selection in reinforcement fine-tuning is a recipe for suboptimal updates—DIEM adapts dynamically, leading to superior performance on reasoning tasks.

Achieving state-of-the-art performance in speech decoding with a dataset that features 80 hours of deep, within-subject MEG data sets a new standard for neural data quality and quantity.

AnaDiffusion allows for precise and controllable editing of 3D brain MRIs, achieving the lowest FID scores while preserving anatomical integrity.

LLMs can autonomously generate prompts that rival expert-written ones, but still fall short in accurately interpreting scientific context and discovering literature.

Tying every view of a digital artifact to its version can boost knowledge-work agent performance by over 12 points on critical tasks.

Regular medical check-ups and mental health stress are critical predictors of chronic kidney disease, revealing new avenues for early intervention.

Collective dynamics of AI agents reveal surprising patterns: while communication boosts accuracy on objective tasks, it can lead to political bias in group opinions.

Simulation-based pre-training can drastically improve the dexterity of robotic hands, outperforming traditional training methods with just a fraction of real-world data.

Language models struggle to consistently encode the current year, with associative and declarative representations diverging in their update responses.

Current AI-generated videos that mislead viewers are also the most challenging for existing detection systems to identify, revealing a critical vulnerability in misinformation defenses.

CARD achieves a level of realism in credit card discussions that outstrips traditional simulation methods, revealing the power of structured conversational planning.

Speed Tuning achieves over 2.4x speed-up in robotic manipulation tasks without the need for extra data collection, revolutionizing policy execution efficiency.

Conversational phase structures can be reliably observed in clinical encounters, but patient states are only partially recoverable, challenging assumptions about transcript-based inference.

Alienation in AI research isn't just a personal experience; it's a systemic issue that can be resisted through collective action and critical self-reflection.

Test-time self-correction can boost LLM accuracy by over 30% on challenging reasoning tasks without the need for external reward models.

SmartMage redefines 3D scene understanding by dynamically selecting modalities, achieving state-of-the-art results while minimizing irrelevant computations.

Extending context in conversations can significantly amplify the risk of LLMs promoting delusional behaviors, challenging assumptions about model size and reasoning capabilities.

Skill-switching accuracy in LLMs drops significantly on complex tasks, but a new training approach boosts performance from 34.4% to 68.4% on challenging benchmarks.

SWD achieves high-fidelity circuit extraction with less than 1% of the data used by traditional methods, revolutionizing interpretability in pretrained transformers.

Achieving up to 2,500x speed improvements, string2string Studio revolutionizes string-to-string analysis by making complex algorithms accessible and interactive in the browser.

Long-term interactions with language models could lead to significant cognitive and emotional changes in users that short-term evaluations miss entirely.

CryptoProver can independently verify cryptographic libraries in under 12 hours, ensuring the integrity of critical code without altering its execution.

Robots can now learn to manipulate dynamic environments effectively from just one static demonstration, drastically improving performance and efficiency.

VLA models can achieve a 66% success rate in contact-rich tasks by addressing precision and force failure modes, a leap from the previous 41% benchmark.

Robots can now adapt to drastic physical changes in real-time, maintaining stable locomotion even with a locked leg or added weight.

Current MLLMs falter under context shifts, with a notable inability to balance answering and refusal rates, as revealed by the new MMOOC benchmark.

System prompts in commercial AI products are often a mixed bag, with 40% harboring instructions that can undermine user interests, revealing a critical gap in accountability.

Achieving near-centralized collision avoidance performance with 28.5% fewer synchronization events could revolutionize autonomous spacecraft operations.

Naive pretraining of Q-functions may hinder performance, while a simple ensemble approach can lead to over 1.26x improvement in fine-tuning effectiveness.

Zero-shot optimization of robot designs leads to a staggering 70% reduction in tracking error, showcasing the potential of motion-conditioned co-design.

Coding agents can now access repository context 8.7x faster, dramatically improving efficiency in navigating evolving codebases.

Proprietary MLLMs may achieve high diagnostic accuracy, but they still struggle with reliable clinical reasoning, revealing significant gaps in their practical utility.

Grounding linguistic explanations in mathematically interpretable units can boost the trustworthiness of AI in medical applications by enhancing both visual localization and reasoning quality.

FIDAC reveals how interpersonal distance can be accurately quantified from video, transforming facial detection data into actionable insights.

Renormalization can redefine how we understand and address the sim-to-real gap in robotics by leveraging effective parameters that capture omitted dynamics.

EmoScope redefines emotional image editing by enabling users to discover unique, context-specific editing strategies rather than relying on static templates.

Achieving nearly linear sample complexity for distribution learning is possible by leveraging hierarchical comparability in query families.

LLMs are overzealous tutors, intervening too soon and too often, which may undermine true learning and cognitive engagement.

Aurora's forecasts may be skillful, but they reveal a troubling disconnect from the underlying chemistry, raising questions about their reliability for environmental policy.

Achieving a staggering 250,000x speedup in hologram synthesis without compromising image quality could redefine the future of 3D display technologies.

Solar Open 2 outperforms its predecessors and competitors with a groundbreaking 1M-token context window, redefining the capabilities of large language models in agentic tasks.

Quantum autoencoders can match classical anomaly detection performance while fitting seamlessly into the resource-constrained environments of future collider experiments.

Imperfect LLM detectors can paradoxically drive users to increase their reliance on LLMs, ultimately degrading output quality.

Revealing robot motion in video models can transform how we predict and control robotic actions, achieving high fidelity with minimal training data.

BrainNext achieves top-tier performance in neuroimaging tasks by leveraging self-supervised learning on a massive dataset, highlighting the power of large-scale pretraining in medical applications.

Memory-centric architectures could revolutionize mass spectrometry search speeds, achieving over 100x acceleration and unprecedented energy efficiency gains.

Hierarchy-aware training combined with anatomy-guided learning boosts LUS video classification accuracy while enhancing model interpretability.

A unified framework that transforms how we build and debug asynchronous robot programs, ensuring reproducibility and systematic debugging across environments.

Aligning the final outcomes of training rather than just the training process leads to a significant boost in model performance, with Inf-Match achieving state-of-the-art results across multiple benchmarks.

Scaling visuomotor context to 8K timesteps enables robots to master complex tasks and adapt in real-time, outperforming previous models by a staggering margin.

OrthoPilot outperformed seasoned orthopaedic experts in diagnostic reasoning, achieving a 10.6% increase in management success for complex musculoskeletal cases.

Training dynamics of Transformers can be reduced to a low-dimensional manifold, revealing how inductive reasoning emerges from data statistics and model initialization.

Pix2Act transforms complex 3D manipulation into a simpler 2D prediction task, leading to significant performance gains and robustness against camera variations.

Video LLMs can significantly improve their QA performance by integrating spatio-temporal evidence, bridging the gap between accuracy and visual perception.

AMP achieves millimeter-level precision in 3D manipulation by transforming action learning into a pixel classification challenge, drastically improving inference speed and success rates.

Achieving submicrometer thickness in liquid sheets opens the door to unprecedented insights into ultrafast interfacial dynamics.

SBR reduces cognitive fatigue while achieving a 54.1% task success rate, outperforming traditional methods in real-time kinematic retargeting.

FourTune slashes memory overhead by 2.25x while matching the performance of full-precision fine-tuning in diffusion models.

Unlearning shortcuts doesn't guarantee their complete removal; ART reveals that some associations can still be functionally restored, challenging existing evaluation methods.

FILTR achieves up to 30x speed improvements in bioinformatics algorithms while simplifying the implementation of complex recurrence relations.

Strong learning can be achieved with significantly fewer calls to weak learners by exploiting the structure of list-decodable codes.

Linear attention fails to capture spectral variations in graphs, but Graph Convolutional Attention achieves superior denoising by directly utilizing the graph spectrum.

Charge conditioning in ML force fields can drastically enhance predictive accuracy while maintaining computational efficiency, achieving remarkable reductions in error metrics with minimal data.

LLM-driven program synthesis can automate EEG feature engineering while ensuring interpretability and high detection accuracy.

Full-sovereign scaffolding not only boosts user sovereignty scores but also curtails privacy violations and manipulative behaviors in personal agents.

Discounted occupancy-ratio realizability alone can enable robust offline policy evaluation, eliminating the need for stringent completeness assumptions.

VLMs struggle with raw medical data, achieving only a 48.6% success rate in standardization, revealing a critical gap in their clinical applicability.

The ABC framework reveals how thoughtful design can transform digital health interventions from theoretical solutions into sustainable real-world applications.

Scaling LLMs significantly boosts social simulation accuracy in well-represented domains, but fails to enhance calibration for human cognitive biases.