Stanford HAI

LLM agents can learn to proactively build optimized, hierarchical file system workspaces that anticipate future queries, outperforming reactive retrieval by 33 points on BrowseComp-Plus.

Even the best AI agents fail to deliver substantial business value in realistic customer delivery scenarios, struggling to uncover unstated constraints and achieve high impact.

Frontier LLM agents fail to recognize superior human-corrected scientific research plans, performing worse than chance even with full project context.

Unexplained observations, not just supporting evidence, are the key to LLMs discovering truly novel hypotheses.

Autonomous agents can now discover their own evaluation metrics, bypassing the costly bottleneck of real-world engineering benchmarks.

Future LLM agents can be prevented from reward hacking by using today's models to automatically harden their evaluation environments.

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

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

Robust optimal policies and values for RMDPs are now computable in strongly polynomial time via a single LP, unifying and extending prior results across diverse uncertainty models.

VideoGen-Agent is presented, a multimodal agent trained through multitask agentic reinforcement learning to use external tools for video generation to support learning tool use across video-generation tasks and show that the trained agent can benefit from subsequent advances in generation tools.

A framework for evaluating vulnerability reports via probes, executable checks of security properties, which surfaced 23 previously unreported vulnerabilities, the majority of which have been confirmed by maintainers.

This work instantiates Real-Time EXPO-FT, an RL framework for finetuning real-time VLA policies that meets the real-time control requirements of dynamic real-world manipulation, demonstrating rapid, sample-efficient adaptation to challenging real-world dynamics.

FlashVector is an agentic system that optimizes performance across all layers of the model serving stack, and the key contribution is an extensible framework to generalize the single kernel optimization agent paradigm to heterogeneous technical stacks, and to deliver performance improvements holistically.

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

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

AI agents can handle the engineering aspects of research but fall short in addressing the core research questions, leading to outright rejections from experts.

Value-Aware MARO prevents performance collapse in multi-agent systems during communication failures, achieving over 20% improvement in returns under high-attrition scenarios.

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

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

Memory management emerges as a high-leverage skill that can double or quadruple the performance of LLMs in complex tasks without altering their core action behaviors.

PaperPilot transforms scientific literature search by enabling users to iteratively refine their search strategies through an interactive workflow, achieving a remarkable reduction in execution errors.

Learning from failures can boost agent success rates by over 6% without extra training, reshaping how we approach agent improvement.

Traditional recommendation algorithms falter when faced with LLM agents, revealing a surprising shift from personalization to mere structural pattern matching.

Terminal-use agents are still far from achieving reliable general-purpose performance, with top models only scoring 65.8% on a new benchmark that spans diverse real-world tasks.

BiPACE transforms credit assignment in LLM training, boosting validation success rates by over 6% without the need for critics or extra rollouts.

Agents can boost their task completion rates by over 20% simply by grounding their actions in observed context rather than assumptions.

Zero-shot sim-to-real transfer for articulated tool manipulation is now achievable with just a few clicks, revolutionizing how robots interact with complex objects.

Current AI agents excel in structured tasks but falter at generating novel insights and tackling open-ended scientific challenges.

Data2Story not only automates data journalism but also ensures every claim is traceable back to its source, revolutionizing trust in automated reporting.

Agents collaborating on EinsteinArena achieved breakthroughs that surpassed previous human and AI solutions, showcasing the power of collective intelligence in scientific discovery.

Decentralized coordination in multi-agent systems can boost reasoning performance and cut costs by 50% in large language model applications.

Design choices in agent memory systems can significantly shift operational costs, revealing critical trade-offs that impact long-horizon task performance.

Current agents are alarmingly susceptible to skill-based attacks, with success rates reaching over 86%, exposing a critical vulnerability in AI safety.

Over a quarter of tasks in popular AI benchmarks contain critical flaws that distort model evaluations, and this automated auditing framework can catch them.

AI-driven scientific discovery is closer than you think, but current systems still struggle with reproducibility, cross-domain robustness, and accountable scientific closure.

People systematically overestimate the efficiency gains from using AI for simple tasks, even when it wastes their time.

AI agents are shockingly easy to manipulate into leaking API keys, deleting user data, and initiating unauthorized transactions across a wide range of real-world applications.

Current LLM agents are woefully inadequate for real-world clinical tasks, achieving only 46% success on a new benchmark that demands long-horizon reasoning and verifiable execution within electronic health records.

LLMs can now automatically design and execute experiments to resolve debates between cognitive science theories, even discovering the models and experiments themselves.

Looping language models isn't just for single agents anymore: Recursive Multi-Agent Systems (RecursiveMAS) show that agent collaboration itself can be scaled through recursion, yielding faster and more efficient problem-solving.

Turns out, coding agents in the wild are only writing useful code 44% of the time, and are introducing more security vulnerabilities than human developers.

Achieve real-time video understanding with transparent reasoning: \model{} aligns response timing with visual evidence, offering a breakthrough for online video LLMs.

RadAgent doesn't just give you the answer; it shows its work, offering clinicians a transparent, step-by-step reasoning trace for AI-generated CT reports.

A lightweight, RL-trained context curator can match GPT-4o's context management abilities, slashing token consumption by 8x and opening the door to efficient long-horizon LLM agents.

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

LLMs can decide when they need more "thinking time" – and boost their accuracy while slashing compute costs by up to 65% – simply by checking if they agree with themselves.

Scaling prompt learning by 17x without sacrificing accuracy is now possible, unlocking efficient self-improvement for LLM agents.

LLM agents can autonomously outperform fixed evolutionary search by 3-10x on open-ended discovery tasks when given persistent memory, asynchronous collaboration, and heartbeat-based interventions.

Unlock richer, more realistic agent simulations by moving beyond individual personas to unified group representations that capture collective behavior.

Medical AI Scientist leapfrogs generic LLMs in clinical research, generating higher-quality, evidence-backed hypotheses and manuscripts that rival top-tier medical publications.