Articles & Guides

CANOPY builds hierarchies over retrieved multimodal items and keeps evidence at different levels of detail according to the query. A critic can also request targeted retrieval when the current evidence is incomplete.

A new study finds that the first tokens of a model’s response can steer the reasoning behavior that follows. Carefully fixing those tokens can bring base models surprisingly close to reinforcement-learning-trained counterparts.

HLA-WM combines geometry-guided retrieval with recurrent linear attention to recover relevant history during long video rollouts. It improves scene revisits and camera control without additional training while sharply reducing memory use.

Periscope is a training-free inference method that replaces one costly long read with a set of local and document-spanning probes. It identifies answers and supporting evidence while using far less KV-cache memory.

Kandinsky 6.0 Video is a family of models that generates five-second clips with synchronized 44 kHz audio and lip sync from text or images. Its dual-stream CrossDiT design jointly aligns visual and audio content instead of adding sound after video generation.

OSWorld-Pro evaluates computer-use agents at the level of sequential subgoals rather than checking only the final task state. It exposes process failures such as inaccurate clicks, keyboard errors, and irrelevant actions.

A new study finds that conventional deepfake detectors degrade sharply against newer generators and adversarial perturbations. Its alternative is to test whether known generators can faithfully reconstruct an image, then calibrate the risk of certifying it as real.

OmniConfess is a training-free inference method that probes a multimodal model’s token-level dependence on text, image, audio, and video evidence before correcting unreliable commitments.

RealtimeWAM tackles two major latency sources in world action models: iterative action denoising and sequential execution between video and action experts. The paper reports less than 1% accuracy degradation and up to 25× end-to-end speedup on H100.

RealCompanion introduces a longitudinal benchmark for AI companions, evaluating not only whether systems can retrieve past messages but whether they know when the past matters. Its results suggest that memory is needed less often than standard benchmarks imply—and is still difficult to trigger correctly.

VeriHarness is an agentic verification framework designed to improve long-horizon task results without reference answers or test-time grading rubrics. It combines rollout disagreement, consensus challenging, environmental evidence, and evidence-backed revision.

SciUtopia is a persistent closed-loop simulation in which LLM agents choose research topics, collaborate, submit papers, review, seek funding, and evolve across simulated years. Its experiments suggest that more submissions can raise output while also increasing review pressure and reducing researcher retention.

Queen combines a silent chess expert encoder with an instruction-tuned language model, aiming to unite strong play with understandable explanations. The paper reports an Elo rise from 1782 to 2697 after seven rounds of iterative distillation.

WEFT treats environments, tasks, agent harnesses, and evaluators as one scalable system for training general-purpose tool-using agents. It combines execution-driven evolution with methods for more reliable long-horizon post-training.

A study of long-form medical factuality evaluation finds that retrieval-then-verify systems can score well on closed benchmarks while failing badly on open-ended clinical answers. It separates the problem into evidence quality failures and verifier reasoning failures.

CacheBack introduces receiver-conditioned latent communication: the receiving agent describes its information needs, and the sender filters its KV cache accordingly. The reported FanOutQA results show less state transfer alongside gains in accuracy and latency.

JEPA-TTT enables a pretrained JEPA world model to keep adapting when test-time dynamics differ from training. It updates only the latent dynamics predictor through persistent self-supervision, while preserving the visual representation and task objective.

Long-horizon video generation needs more than a larger memory: it needs to decide which past frames will matter later. FrameMorrow predicts future information needs and uses them to select useful frames from the generation history.

ProWAM extends world action models with an ordered sequence of sparse visual subgoals, giving robots intermediate guidance without generating dense future video. The reported results show stronger robustness on several simulation benchmarks.

SimuVerity introduces a benchmark for evaluating whether agent-generated Simulink models satisfy engineering requirements, rather than merely compiling or resembling a reference. Across 101 tasks, the strongest of six evaluated systems scored only 42.86 overall.

Video world models rely on growing KV caches to maintain long-range consistency, but aggressive quantization can cause flicker. QuantWM addresses the problem by preserving attention behavior rather than minimizing reconstruction error alone.