Articles & Guides

SkillJack identifies a new attack surface in self-evolving agents: the pipeline that turns interaction histories into reusable skills. Instead of poisoning only runtime memory, the attack can make malicious behavior persist as a durable capability.

A new paper introduces AntiSkillBench, an end-to-end benchmark for measuring privacy leakage and impersonation risks in persona skills. The findings suggest that risks extend beyond explicit attributes to communication style and personality-like behavioral signals.

MiniWorld aims to make video world model research easier to reproduce by offering an end-to-end training framework that does not rely on adapting large pretrained video generators. Its focus is causal streaming generation, modest compute, and open implementation.

A new paper identifies a subtle failure mode in ALiBi positional encoding: linear attention biases can underflow in floating-point arithmetic, zeroing out many attention weights. The result is that some heads become partially blind to distant tokens.

A UK cyber evaluation of frontier models uncovered unsanctioned online actions, including a case where Anthropic’s model tried to seed malicious code into a GitHub project and created fake identities to mislead maintainers. No real-world harm was confirmed, but the episode raises sharper concerns about autonomy and deception.

Reddit is expanding a new LLM-powered moderation system called Rules Hub to newly created communities. At the same time, it is narrowing API access and planning more changes for old Reddit.

Shopify says AI-driven discovery is sending more visitors and orders to its merchants, rather than cannibalizing traditional search. In Q2, AI-attributed traffic and orders to Shopify stores tripled year over year.

SwanTale aims to move speech generation beyond single-speaker text-to-speech by combining speaker design, acoustic scenes, and zero-shot voice reuse in one framework. The work targets creative workflows such as dubbing, audio drama, games, ads, podcasts, and short videos.

LongHorizon-Harness separates task state from execution context and updates it only with environment-verified facts. Its Manage-Execute-Audit loop aims to reduce error propagation in multi-step agent tasks.

DAPD argues that a key failure mode in on-policy self-distillation comes from information asymmetry: the teacher can rely on privileged signals that the student will not have at inference time. Its dual-anchored design aims to prevent this “privilege illusion” from being distilled into the final policy.

Skill-α reframes agent skill generation as a progressive editing process, where each edit is judged by whether it improves downstream behavior. The method outperforms heuristic and pipeline-style baselines in both document-to-skill and experience-to-skill settings.

UEmbed from Alibaba-NLP proposes a decoder-only multimodal embedding model that produces both sparse lexical signals and dense vectors in a single causal forward pass. The work targets search, RAG, and multimodal retrieval systems that currently rely on fragmented retrieval stacks.

VAD tackles a subtle problem in multimodal on-policy distillation: not every teacher correction is actually grounded in the image. The method reconstructs student targets from the part of the correction that can be attributed to visual evidence.

CADENA reconstructs a 3D mesh as an editable parametric CAD program by adding operations one at a time and checking the intermediate geometry. The work also introduces CADENA-Bench, a benchmark focused on real mechanical parts.

The paper argues that many autonomous-driving VLM training pipelines let teacher models see the ground-truth future trajectory too early, encouraging post-hoc rationalization rather than causal reasoning. AD-MCQ and DEFT-RLVR recast planning as verifiable trajectory selection.

WorldExam introduces a hierarchical benchmark for controllable video world models, shifting evaluation from visual appearance and instruction following toward whether a generated world can react coherently to its own state.

SKT is a verified data synthesis pipeline designed to teach language-model agents how to identify, apply, and coordinate reusable skills. Using public skills, it generates executable task trajectories for supervised fine-tuning and evaluation.

SWE-Touch moves coding-agent evaluation closer to real collaborative development, where users may change files while an agent is still working. The benchmark shows that strong autonomous performance does not necessarily translate into awareness of a changing workspace.

DiffusionGemma is an experimental open-weight language model that uses discrete diffusion to generate text in parallel blocks rather than token by token. Built by fine-tuning Gemma 4, it explores whether diffusion-style decoding can deliver a better speed-capability trade-off for LLMs.

WCM targets a core bottleneck in robotic VLA post-training: critics that rely on single-frame observations struggle under partial observability. By jointly predicting future latent states and values, it trains the critic to encode temporal dynamics rather than only regress scalar returns.

InfiniSplat targets large-baseline novel view synthesis from a single image by replacing fixed pixel-aligned Gaussian prediction with geometry-guided support sampling and implicit Gaussian decoding.