Research

New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.

2566 articles · page 1 of 27 · The Research front page

All Agents Language models Reinforcement learning Vision Generative media Speech & audio Robotics & embodied AI Safety & security Evaluation & benchmarks Efficiency & systems Code & software AI for science & health Theory

Viewgrid list compact

Robots can change task strategies when their controllers change

One-camera humanoids learn to catch objects they have never seen

Quadrotors learn to fly through moving gaps using differentiable simulation

New memory architecture lets legged robots remember maze layouts while exploring

Future 3D motion prediction boosts robot learning with fewer demonstrations

Audio model composes timestamped captions for queryable timelines

Contact-preserving retargeting improves robot task success from 13% to 83%

Language model that indexes 200 million web pages outperforms embedding search for navigation

Image generation models can be repurposed for speech synthesis while preserving voice

Dexterous hands match grippers on many tasks, but precision tasks remain difficult

Active exploration lets humanoid monitors spot unrecoverable falls before they happen

AI speech translators can improve quality with little extra initial delay

AI agents find better evidence when search considers their original question

Robots can learn better when humans correct deliberately introduced deviations

Robots can use human videos before learning their own controls

Robots learn tasks better with demonstrations generated by a video model

AI models can reason better by predicting videos of outcomes

Filtering atypical records lets differential privacy noise match typical data sensitivity

Training data attribution for AI language models degrades with redundancy and distractors

World models learn true dynamics with fewer parameters

Training chatbots to state a safety decision before answering defeats multi-turn attacks

Teacher corrections with dense feedback double AI success on hard logic problems

AI agents that succeed at tasks may still fail their mid-way steps

Diffusion models can be composed and steered at inference time without retraining

Language models benefit from repeated layers, but more can hurt

AI agents can improve by tracing how shared evidence changes answers

Batching makes AI grading verdicts depend on the other items

AI web agents satisfy only part of revised goals mid-task

New watermark for language models detects AI text without distorting output or slowing decoding

Several routers pick experts in language models, boosting accuracy without extra compute

AI models can be trained end to end with one algorithm

AI mathematical reasoning improves more with textbook structure than with isolated problems

Fake content and structure in tool outputs combine to fool all tested AI agents

AI agents can improve accuracy by discarding cluttered task histories

AI model controls characters in unfamiliar game worlds in real time

AI diffusion models can follow deterministic paths through discrete data

AI agents’ false completion claims are caught more often with follow-up checks

AI models count visible appearances more reliably than distinct physical objects

Language models solve more problems when training accounts for confidence

Language models improve psychiatric interviews after training with simulated patients

Language models gain from varied answers, not necessarily altered weights

AI agents can match the theoretical limit for exploration losses

Video generators can run twice as fast without retraining

Long AI prompts use disproportionately more energy than output length predicts.

Image generators can outperform their multi-step originals on image statistics

Coding up definite operations improves AI agent consistency

AI agents that satisfy more search constraints recommend fewer places

AI health overviews cite fewer top sources and often lack evidence

Replacing language-style decoding cuts robot trajectory errors by 60%

Textual personas fail to make vision-language models answer like real people

Document parser reads hundreds of pages without growing per-step compute

Repurposing pretrained image generators for any-order editing costs a fraction of retraining

Language models gain no advantage from synthetic warmups at equal compute

Joint video and 3D motion prediction reduces trajectory errors by over half

AI video narrators fit silences well, but pacing still trails human professionals

Single-image 3D generation creates assets that move and deform realistically

New hinge losses ensure optimal routing in adaptive AI classification

Transformer-like networks predict electron densities accurately enough to skip self-consistent calculations

Visual scaffolds improve AI reasoning by bridging the reference gap, not perception

AI image generators and robot controllers can follow guidance more reliably

AI image editors get better results comparing alternatives than repeatedly refining

AI models miss image details when attention drifts during reasoning

AI chatbots can answer sooner without permanently discarding visual context

AI models can align language and vision using 42-fold fewer samples

Language models can get optimizer updates about 13 times faster

AI agents recall conversations better when memories preserve events and relationships

Targeting uncertain local decisions cuts GUI agent training cost by 56-61%

Online reinforcement learning on live iOS apps yields UI agents with 7x better cross-platform task success

Language models struggle to turn social understanding into appropriate responses

On-the-fly GPU mesh preprocessing adds just 2% overhead to physics AI training

Extra attempts help small language models solve terminal tasks, but not verify solutions

Multimodal model tops specialist speech systems for low-resource African languages

AI reasoning models learn from reference examples alone, no reward needed

Self-teaching language models improve when the teacher gets feedback too

Feedback criteria that track a model's growing ability halve its training steps

AI assistants that learn separate style modules from response feedback outperform fixed-prompt personalization

Game engine expertise and verification feedback boost AI 3D scene completion from 18% to 48%

Malicious project files can push code agents toward vulnerabilities, often without developer awareness

One-line fix makes language models learn better and decode faster for free

Human egocentric video lifts robot grasp success from 16% to 58%

AI models struggle to distinguish objects surrounded by similar ones

Multi-agent AI systems improve by learning from execution traces

AI image models fail to combine multiple views of the same scene

AI agents follow attacks in documentation 2.6 times more than in code

Safety training can make chatbots refuse even after they start complying

Video super-resolution sharpens and runs 10x faster via sub-pixel frame fusion

Proper coordinate alignment reshuffles novel-view AI rankings

AI agents often delay checks most closely linked to success

Language models can accelerate more when handling one request alone

Language models can improve by learning from their weakened copies

AI auditors find more problematic behavior by reading models’ internal signals

AI weather forecasts can capture extreme rain without blurring storm patterns

Teacher rewrites of drifting student reasoning stabilize training and improve accuracy

Structured tags help AI chatbots ignore hidden instructions in data

Magic-trick videos expose the reasoning gap between AI models and humans

Coding agents teach humanoid robots new whole-body skills without retraining