Research

Humanoid robots trained in simulation can navigate homes without prior mapping

Hu and colleagues present HALO, a humanoid robot system that learns to connect camera views and language instructions with whole-body movement using simulated demonstrations. The authors report direct transfer to six previously unseen, unmapped physical environments, where the robot performed tasks including sitting, crawling under tables and moving through partly open doors.

Siyuan Hu, Abulikemu Abuduweili, Yue Meng +5

More in Robotics & embodied AI

Predicting motion, not video, improves robot learning from human footageyesterday

World model simulator works across 22 robot platforms, enables policy learningyesterday

AI models can transfer actions between human and robot videosyesterday

AI driving models go farther when action feedback updates perceptionyesterday

Gameplay dataset maps tasks to subgoals for training AI agentsyesterday

Bimanual robot learns precise tasks without any robot-collected datayesterday

Simple prompt trick stops AI agents from faking progress in robot tasks2 days ago

Scoring imagined futures and actions before execution triples robot success rates2 days ago

Evaluation & benchmarks

AI agents fail end-to-end physics research due to shallow reasoning and missing intuition

The authors introduce FrontierPhysics, a benchmark with 56 research tasks created by 42 physicists based on their own work. They evaluated 22 models across 26 configurations and found that no configuration passed more than 46 out of 168 rollouts. Manual review revealed that the main weaknesses are insufficiently rigorous reasoning and limited expert intuition, not gaps in physics knowledge.

today

Robotics & embodied AI

Robot policies that read early vision layers handle camera shifts better

Fan et al. introduce NEO-VLA, a robot control model with no separate vision encoder: raw image patches are fed directly into a shared vision-language backbone along with the instruction, and an action module reads intermediate representations from several depths. They report that action-relevant visual signals often peak well before the final representation, and that reading those earlier layers improves control, especially when the backbone is frozen.

today

Generative media

Agent workflows help AI combine several reference images, but gains vary

The authors introduce LongBanana, the first expert-designed and verified benchmark for evaluating how well image generation systems can retrieve, reject, bind, and compose visual evidence from many reference images. They find that minimal agent harnesses improve over direct generation, but a more complex plan-execute-reflect workflow yields harness-dependent gains and can even introduce drift.

today

Evaluation & benchmarks

AI models answer financial facts they claim not to know

The authors introduce FinDendro, a benchmark of 2,226 dated U.S. financial facts from 1999 to 2026, covering macroeconomic data, stock prices, and Federal Reserve decisions. They measure whether models can answer correctly and whether their self-assessed knowledge matches their factual accuracy. The key finding is that frontier models like GPT-5 answer a majority of facts correctly even when they say they do not know them, and date instructions remove claims of knowledge without changing accuracy, meaning standard safeguards against backtest leakage are ineffective.

today

Generative media

Image editors can drift less while retaining most existing capabilities

Repeatedly editing an image can gradually erase textures, alter faces or distort backgrounds, even when each requested change is small. Li and colleagues report that targeted additional training reduces this cumulative drift across three image-model families, while largely preserving their existing image-generation and single-edit performance.

today

Efficiency & systemsGOOGL ▼3.8%

Language models can train up to 24% faster on Google chips

Amirkhanyan and Cheng report that reshaping a language model’s attention layer around Google’s accelerator hardware speeds up training without substantially changing prediction quality. Their Systolic Linear Attention design targets a mismatch between attention layers tuned for graphics processors and the wider matrix-computing units in Google’s Tensor Processing Units.

today

Safety & security

AI chatbots' psychological risk factors can be monitored without measuring user harm

The authors propose a shift from measuring psychological harm directly in LLM conversations to detecting observable AI behaviors that are risk factors for such harm. They operationalize eight behaviors, create a synthetic dataset of 47,062 AI responses, and compare detectors. A classifier–judge cascade achieves macro precision 0.86 and recall 0.76, and reduces LLM-judge calls by 97% on WildChat while showing that detections cluster within conversations.

today

Robotics & embodied AI

Robots can improve on untrained tasks by practising diverse ones

Korniak and colleagues studied whether a generalist robot policy, a model controlling many different tasks, can improve broadly rather than merely specialize during adaptation. They report that learning from interaction across diverse manipulation tasks can also improve performance on tasks excluded from that adaptation, with the strongest results coming from methods that learn to estimate the value of actions.

today

All research articles →

The Field

Humanoid robots trained in simulation can navigate homes without prior mapping

More in Robotics & embodied AI

AI agents fail end-to-end physics research due to shallow reasoning and missing intuition

Robot policies that read early vision layers handle camera shifts better

Agent workflows help AI combine several reference images, but gains vary

AI models answer financial facts they claim not to know

Image editors can drift less while retaining most existing capabilities

Language models can train up to 24% faster on Google chips

AI chatbots' psychological risk factors can be monitored without measuring user harm

Robots can improve on untrained tasks by practising diverse ones

The Field

Agents

AI agents learn to recover, not consistently choose better actions

AI agents improve when their surrounding software is automatically redesigned

Search results order and dead-end histories bias AI agents' final answers

Omitting reasoning from training data does not harm AI agents on standard benchmarks

Language models

Giving harder inputs more attempts helps little on hard LLM workloads

Language models reason better when they think aloud only at key steps

Language models can merge specialists without matching their training capacity

Language models can learn facts while misjudging what they know

Reinforcement learning

Selective token supervision cuts two thirds of training tokens while improving AI accuracy

Language models can improve writing by judging their own drafts

Budgeting rewards per prompt helps language models train past reasoning plateaus

Browser-game training makes AI webpage generation surpass a frontier model

Vision

AI models can improve image understanding and generation from their outputs

Language models trained only on text can partly locate video events

Image search models can retain left-right information but ignore it

AI that identifies radiological findings first can diagnose unseen diseases

Generative media

AI video generation gains emotional fidelity when rewards trust only reliable cues

Video generators can nearly halve history storage while improving reconstruction quality

Motion model understands and generates in one shared space

Generating video directly in pixels can beat models that compress first

Speech & audio

Real-time AI soundtracks for video now sync as well as offline models

AI agents locate sounds better after training with written reasoning

One audio model can now describe, generate and edit 3D sound scenes

Voice model matches text reasoning without separate transcription step

Safety & security

Prompt compression weakens refusal of harmful requests in AI models

Privacy-preserving watermarks for AI text allow aggregate tracking without individual surveillance

AI models predict training data side effects better than humans

AI models can pass deletion checks yet diverge during later training

Evaluation & benchmarks

Flexible classical models match neural networks for event prediction, but not on ICU data

How data is split can swing encrypted traffic classifier scores by 40 points

AI assistants face a $1.7 trillion wish list of life tasks

AI judges reverse correct verdicts after a false note

Efficiency & systems

Language models can generate text five times faster by drafting ahead

Chunked prefill, not cache architecture, drives LLM serving engine speed differences

Recycling high-scored unexposed candidates boosts feed discovery without engagement loss

AI search indexes can be built up to 50 times faster

Code & software

Verified code often passes formal checks but misses the real spec

Training shifts a small theorem prover toward fewer, larger proof steps

Agentic framework formally verifies GPU kernels against PyTorch references

Reusing past attempts to steer data collection trains better coding agents

AI for science & health

Distributed runtime scales AI molecular dynamics to 30 million atoms

AI forecasting model adapts to sensor changes, guarantees uncertainty at new locations

AI models sample more protein folding stages, but miss folded endpoints

AI protein design improves by separating suggested changes from evaluation

Theory

Privacy-protecting averages can lose accuracy when data variation is unknown

Fooling all classifiers is as hard as matching the best classifier

Noise injection reveals when machine learning cannot identify patterns

A small positional signal replaces costly relaxation layers when training symmetry-aware models