Articles & Guides

An AI reviewer may judge the same science differently when its wording changes. A new study defines rhetorical robustness and introduces SciCore, which combines full-manuscript review with a structured science core.

HUST Vision Lab introduces Multimodal Flow, a fully continuous approach for jointly modeling language and images. It replaces discrete visual tokens with continuous hyperchunks while retaining structure from both modalities.

AutoDataBench introduces a controlled benchmark for measuring how well LLM research agents diagnose, organize, and construct training data. It also tests whether agents can predict the effects of data interventions and whether their research traces can support further training.

Self-evolving search agents can learn to agree on shared mistakes, raising internal rewards without improving factual accuracy. CrossFit breaks this feedback loop by separating the evidence used for training and evaluation.

A new MIST benchmark shows that irrelevant images can destabilize vision-language model judges without reliably informing their decisions. The presence of an image matters more than what the image depicts.

OSWorld-Science brings scientific workflows into computer-use evaluation, covering molecular drawing, pathology analysis, statistics, and physical simulation. Instead of judging clicks alone, it examines whether agents produce verifiable scientific artifacts.

A Meta research team proposes Loop Scaling Laws, a framework that models recurrence, MoE sparsity, model size, and data together. The results suggest that repeated computation and sparse capacity can reinforce each other rather than compete.

A study of flow-based VLA models finds that reinforcement learning changes are strongly concentrated in a small action-expert component called Timestep Modules. Its shift-vector updates also encode task success and transfer relationships.

MILO turns Agent Harness design into an adaptive multi-agent evolution process, optimizing both the harness being searched and the strategy used to discover it. The framework reports strong gains across three long-horizon agent benchmarks.

CUA-SWE evaluates agents that must edit code, run commands, operate graphical interfaces, and use visual feedback to diagnose and verify software changes. It targets the gap between coding-agent benchmarks and computer-use evaluations.

A new study tests visual self-supervised learning under a more realistic input regime: frames arrive in temporal order, without global shuffling or repeated epochs. Its proposed StreamMAE shows how adapting the data pipeline can recover much of the performance lost by standard methods.

NVIDIA researchers introduce Physis-Lang, a framework that uses detailed physical language as a shared representation for video curation, training, and generation. Assertion-level evaluation and language-guided retrieval help models produce videos with more plausible physical behavior.

A2Z GameSpec-Bench evaluates games generated from 100 long-form game design documents, asking whether agents preserve the intended design rather than merely producing runnable code. It combines source inspection with replay and adaptive playtesting to test interacting requirements.

PatchHolmes combines hybrid retrieval with an agent that compares a candidate list as a whole. The approach substantially improves top-1 patch retrieval while using one agent conversation per vulnerability.

AREX-2 treats reflection and long-horizon execution as complementary foundations for self-improving agents. Training on verifiable machine-learning and algorithmic-programming trajectories, its agent transfers the learned behavior to deep-research benchmarks.

Automated research is not only about running more experiments; it also requires deciding which research directions deserve attention. AIM makes idea organization, selection, implementation auditing, and budget allocation explicit parts of the search loop.

Reward models can mistake longer, more confident, or more agreeable answers for higher-quality ones. BiasReducer reduces these distortions by editing only the linear reward head and selecting bias corrections for each dataset.

A new study shows that trainable latent links can weaken refusal behavior in aligned agents without changing the underlying models. Malicious link optimization can amplify harmful compliance, while safety-oriented rewards can help repair compromised links.

WUSH-KV brings data-aware transforms to KV-cache quantization, constructing separate transforms for keys and values from calibration statistics. The paper reports competitive or better results than the OSCAR transform at 2-bit precision across evaluated models and downstream tasks.

A new study shows that non-invasive brain-to-text systems may exploit word-duration cues rather than brain activity. Separating time windows makes decoding more dependent on neural information and improves the value of repeated observations and language-model priors.

A remembered object location can become wrong while a robot is on its way to inspect it. EvolvingNav combines time-indexed 3D memory, probabilistic belief updates, and visual evidence to support navigation in evolving environments.