Articles & Guides

VoxMem evaluates whether large audio language models can remember information across spoken sessions, including who said something, how it sounded, and what was audible in the background. Its results expose a major gap between textual recall and genuine spoken memory.

RLVR training can waste rollout capacity when every sampled answer in a group fails. GRAFT addresses this gap by exchanging complementary trajectories between heterogeneous models while explicitly controlling off-policy mismatch.

A Berkeley-led study examines when image-to-image generation training transfers to image-to-text understanding. Its findings show that the gains are selective: the right generation tasks can strengthen specific perceptual and reasoning abilities.

LongCat-DeepResearch combines an enhanced LongCat model with a multi-agent pipeline for evidence-grounded reports. Its results are promising, while its analysis also suggests that adding more planning does not always improve research quality.

HybridCUA trains computer-use agents to combine graphical interfaces with command-line operations instead of relying on either mode alone. Its 9B model reaches 53.6% accuracy on OSWorld, 14.8 percentage points above the base model.

Marathoner is a post-training framework designed to help AI agents keep working toward difficult goals over extended execution horizons. It combines challenging task synthesis, rejection-sampling fine-tuning, sandboxed reinforcement learning, and a reward for useful actions late in a trajectory.

Researchers at the Hong Kong University of Science and Technology introduce WorldLine, an action-driven visual simulator for robotic manipulation. The system separates transferable robot–object dynamics learning from the grounding of heterogeneous robot controls.

Meta researchers introduce Context Language Models, which represent context as a writable file that the model can update without fixed external rules. The paper reports stronger results and lower compute costs across several long-horizon agent tasks.

A new study explains when dense retrieval should use shared or separate query-document projections. Its bias-variance analysis shows that dual projections become more attractive as data grows or query and document geometries diverge.

Google Research introduces TabFM, a 400-million-parameter foundation model for tabular prediction. It treats supervised table learning as in-context learning and produces predictions without task-specific fine-tuning.

CompoWorld builds complex agent tasks by composing reusable, verified services into dependency graphs. The approach targets workflows that require information to move across multiple tools and services.

FlashForward reuses the key–value caches already computed during denoising instead of rebuilding clean history with extra model passes. Sparse clean anchors then help control the visual and motion drift that can arise from using noisy, stage-matched history.

A new study examines why training and inference engines can assign different probabilities to the same token during RLVR. Its Calibrated Importance Sampling method uses confidence-aware clipping to control the resulting policy-update error.

VQS replaces fragile answer voting with structured image records and fixed programs that generate and compute visual questions. The approach reaches 94% answer accuracy in human evaluation and improves Qwen3-VL across ten benchmarks.

Looped language models gain effective depth by repeatedly applying a shared block, but this also creates a sequential decoding bottleneck. Wavefront Decoding uses intermediate states as drafts and schedules states from different positions and recurrence depths in a diagonal, batched wavefront without additional training.

SentZero is a sentence-centric vision-language pretraining framework for multi-task, zero-shot chest X-ray analysis. It addresses limited positive-pair diversity, false negatives in contrastive learning, and the mismatch between long radiology reports and simple prompts.

MassAlloc Attention keeps access to every legal causal QK interaction, then uses normalized attention contribution to reduce post-score work in low-impact regions. The result is adaptive compute allocation with accuracy close to dense FullAttn.

Skill2Env introduces a capability-oriented pipeline for converting reusable skills into challenging tasks and executable environments. It also uses solver traces to iteratively strengthen tasks that fail to demand enough from an agent.

EditWorld is a video world model designed to support continuous editing in interactive environments. It accepts editing instructions and reference images during autoregressive generation rather than limiting control to the initial prompt.

A broad evaluation suggests that general-purpose AI systems are absorbing more capabilities once reserved for dedicated vision models. Yet precise geometry, faithful reconstruction, temporal consistency, and specialist knowledge remain difficult.

VideoPhysEdit moves video editing beyond appearance changes by reconstructing a rigid-body scene, simulating an intervention, and using the resulting trajectories to generate a counterfactual video.