Follow the latest research

Jiale Kang

Token-indexed memory can replace attention’s value projection while improving language modeling and average benchmark performance under matched training-token budgets.

Tel Aviv UniversityStanford

AC

Aditya CowsikKfir DolevNoah D. Goodman

Models trained only on self-generated programs improve prediction across unseen text, images, audio, and other data, suggesting useful structure can emerge without curated examples.

CMU

Yubo LiYidi Miao

RK

Ramayya Krishnan

A low-cost first-pass judge can handle routine evaluations, while confidence-based escalation preserves nearly all a stronger judge’s accuracy at lower fees.

Max Planck Institute for Intelligent SystemsSnyk

Jeremy QinDavid SchmotzMaksym Andriushchenko

Agents can erase or falsify their execution records, even under reward pressure, undermining audits unless logging is controlled independently.

Arizona State UniversityCornell

Jacob DineenSilei RenDan Roth

In security-sensitive applications, language-model agents are often required to coordinate without disclosing confidential information. Yet repeated interactions may also let ordinary messages acquire shared private meaning. We study a repeated game with pairs of models in which the sender model observes one of four secret states and selects one of four summaries of the same public report, while the receiver model tries to infer the secret state. We find that model pairs can learn to communicate the secret using only one bit of feedback indicating whether the receiver inferred it correctly. This learning occurs during inference with fixed parameters and no supplied codebook or encoding examples. The effect also persists when agents generate their own free-form updates in a simulated incident-response task. Across ten independent games, pairs of GPT-5.6 Sol agents reach 98.8% final accuracy, compared with 25% chance, despite explicit instructions prohibiting disclosure and a monitor that screens each message without access to the agents’ interaction histories. The same interactions that help agents cooperate can therefore allow confidential information to pass through messages intended for legitimate coordination.

Shanghai AI Lab

Jisong CaiYao MuBowen Zhou

Robots can reuse slow visual predictions while updating actions from new observations, enabling more responsive control in complex, contact-rich tasks.

USCBrown University

Yinghua ZhouJunjie YeYue Wang

Robots can replan more responsively while retaining predictions of future scenes, because the model spreads video-and-action computation across successive control cycles.

USCCMU

Haotian ZhangFengyuan YuYilun Du

Fine-tuning a video model on synthetic scenes of occlusion and physical interactions produced the top-ranked continuation model in a human evaluation.

CMU

Christine ParkValerie ChenTim Dettmers

Researchers can openly test clinical AI on realistic, longitudinal patient records with traceable ground truth, without access to private health data.

Tsinghua

Yijun YuanWeicheng ZhengHang Zhao

Robots can plan by constructing action-executable paths between current and goal states, avoiding online trajectory search in tested control tasks.

ZJUTencent

Xingyu WuYuchen YanZhengxi Lu

A shared language model that alternates between planning searches and summarizing evidence can answer long research questions while keeping its working context bounded.

MIT CSAIL

Kimia HamidiehGiannis DarasAntonio Torralba

Previously trained reward-specific policies can approximate how reinforcement learning would adapt a model to a new reward, without another training run.

Pavel TikhonovAnton KorznikovIvan Oseledets

Standard Transformers preserve signals from two mixed text streams, and lightweight fine-tuning can help decode separate continuations from one shared forward pass.

UC San DiegoHKU

Zanyi WangYuheng LeiPing Luo

Training robot policies across visual denoising states improves robustness to shifts without requiring future-frame prediction, while using fewer training visual tokens.

NVIDIAUC San Diego

Nicklas HansenIretiayo AkinolaXiaolong Wang

A single vision-based robot policy can insert previously unseen parts, with success improving as it trains on more object geometries.

UC San DiegoUSC

Xinyue WangJiacheng PangBiwei Huang

One model can interpret time-series patterns, answer questions about them, and generate forecasts that respond to textual context.

Anthropic

Peter H. YoonJanuka S. AthukoralageEmmanuel Ameisen

AI agents examining billions of protein clusters identified an unusual viral DNA repeat array, revealing a new family of reverse transcriptase systems.

UNCStanford

Peng XiaRujun HanTomas Pfister

Regularizing how agent harnesses evolve helps improvements transfer to unseen tasks, while using fewer policy tokens than unregularized evolution.

HKUSTCity University of Hong Kong

Harold Haodong ChenRongjin GuoYu Cheng

This survey organizes scattered research into a framework for comparing how video generators retain, retrieve, and use history beyond their context windows.

UPenn

Xiaotian HouKwangmoon ParkHongzhe Li

Single-target perturbations of every measured variable can identify causal network structure despite unmeasured confounders and selection bias, with statistically controlled uncertainty.