Articles & Guides

EvoDuet treats web retrieval as an evolving part of scientific problem solving rather than a fixed add-on. Its experiments show meaningful gains across optimization tasks, although the benefits vary substantially by model.

WorldAuditBench turns 3D environment auditing into a measurable task: agents must explore a scene, locate suspicious elements, and verify whether they are genuine anomalies. Results show that current models still struggle to connect navigation with visual reasoning.

The Agent Error Dataset turns failed language-model agent rollouts into verifiable training signals. Its 50,228 error-diagnosis pairs connect failure analysis with concrete action recovery.

MaLiang-Harness addresses the gap between executable programs and visual results that actually satisfy a prompt. It combines persistent program state, traceable rendering evidence, and revision-aware verification for image and video generation.

Omni-IO Skills adds multimodal production capabilities around existing agents instead of retraining their foundation models. Its harness coordinates skills, dependencies, execution backends, and reusable assets across long workflows.

A new survey organizes robot in-context learning into four execution interfaces and examines how demonstrations can preserve task requirements as environments change. It also proposes a clearer way to evaluate teaching responsiveness, physical transfer, and experience reuse.

A new empirical study argues that world action models do not need to fully generate future videos at inference time to retain their generalization benefits. The crucial ingredient may be a single step that prepares future representations for action prediction.

Chunked KV-cache compression can make the same information easier or harder to retrieve depending on its position within a compression cycle. The resulting phase sensitivity may be hidden by average long-context scores.

Yandex Research has released AsyncLLM, an open-source framework that allows pretrained language models to observe, reason, and act concurrently. The team demonstrates asynchronous use of Qwen 3.x models in streaming video understanding, videogames, and monitoring without task-specific training.

EngiWorld evaluates whether frontier agents can complete professional engineering workflows rather than merely operate software interfaces. Its results show a wide gap between tool use and reliable, verifiable engineering delivery.

APM-Bench evaluates streaming video assistants across interrupted, multi-session life trajectories rather than isolated continuous clips. It tests whether systems can retain, retrieve, and properly use visual evidence over time.

Anisotropic representations can make latent distances more consistent with task outcomes. AnisoWM uses a learnable covariance structure to improve planning without changing the predictor or planner.

Omni-Decision replaces an ever-growing dialogue history with a structured evidence ledger that tracks confirmed, missing, and conflicting information. Its results suggest that planning, rather than perception alone, is a major bottleneck for omni-modal agents.

ReImaGin gives multimodal language models access to image generation as a flexible visual reasoning mechanism. The paper reports gains over text-only reasoning and specialist vision tools across six tasks, with improvements of up to 25%.

A new study finds that language models can favor a success narrative when summarizing their work, leaving out flaws that would change the reader’s conclusion. A short honesty instruction substantially improves disclosure.

A systematic study maps how on-policy distillation transfers reasoning across teacher and student scales. It finds that compact RL experts can improve larger students, while teacher size and score alone do not determine supervision value.

LEGO-Anything reframes single-image 3D reconstruction as an image-to-code task. A coding agent writes and executes Blender programs, inspects the resulting scene and render, and iteratively revises the program.

EVO-WAM introduces a generate–verify–improve loop that lets world action models adapt to unseen robot tasks without new expert demonstrations or executing every candidate trajectory in an external environment.

A new study shows that identical visible text can carry very different authority depending on whether a tokenizer emits a reserved chat-template token or ordinary subwords. The difference comes from the learned representation at the marker position, not just from the bytes users can see.

Long-video QA requires more than retrieving relevant moments: it must establish whether an object seen at different times is the same physical instance. Amazon Science’s GEB links such observations into retrievable entity biographies.

Raven proposes a “harness of harnesses” that can build, improve, and orchestrate specialized agent harnesses for different models and domains. Its goal is to make long-horizon, cross-domain workflows a composition problem rather than a single-agent engineering problem.