01Physical design
BrickWright gives coding agents a simpler way to design with real brick parts, then checks whether pieces collide, float or disconnect. The authors report valid builds on all 512 prompts, compared with 9% to 69% for the same agents without the framework.
Why it mattersA picture or sentence can become an assembly that is meant to work in the physical world, not only look convincing on screen.
Keep in mind: The tests cover LEGO-style construction inside the authors framework, not general mechanical engineering.
Read the source · BrickWright authors ↗02Creative AI
Latent Space Jam steers a video model with rough boxes that say where objects should appear and move. A first pass makes a coarse guide; a second pass turns it into a cleaner video using the same pretrained model. No extra training is required.
Why it mattersCreators could choreograph motion and placement without collecting new data or rebuilding the video model.
Keep in mind: The paper reports short benchmark examples; long scenes and production reliability remain unproven.
Read the source · Latent Space Jam authors ↗03Agent memory
Researchers compared a plain memory recipe with systems that rewrite and reorganize what an agent remembers. Dense retrieval over the raw conversation ranked first in six of seven benchmark-and-model settings, beating the strongest agentic alternative by 0.13 to 5.53 points.
Why it mattersKeeping the original conversation may preserve useful links that get lost when an agent repeatedly summarizes its past.
Keep in mind: The result covers four conversation-memory benchmarks, not every long-running agent task.
Read the source · Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought authors ↗04AI safety
Researchers tested 38 activation readers across 15 instruction-tuned models on harmful requests rewritten with persuasive and jailbreak-style language. The internal signals weakened under that shift, and a simple word-based classifier ranked harm better than the original Latent Guard in every tested model.
Why it mattersLooking inside a model does not automatically make a safety check robust to new wording.
Keep in mind: This evaluation covers selected readers and prompt shifts, not every possible form of latent monitoring.
Read the source · Faithful Latent Harm Monitors Work No Better Than the Models Own Behaviour at a Matched False Alarm Rate authors ↗05Edge AI
AgentDP turns several fiddly deployment steps into one agent-led process. It clarifies the target requirements, splits a model across compatible runtimes and reuses earlier tuning settings. The authors report up to 80% less adaptation time and higher average task accuracy than their baselines.
Why it mattersMoving an AI model onto unfamiliar cameras, boards or embedded computers could demand much less manual trial and error.
Keep in mind: The speed and accuracy gains are author-reported on the evaluated devices and tasks.
Read the source · AgentDP authors ↗06Reasoning
A new training framework gives a language model a separate decision about how much reasoning to spend. It rewards short correct answers on easy prompts and reserves longer work for harder ones. Across the authors tests, that improved the balance between accuracy and computation.
Why it mattersAI systems may stop spending the same effort on a simple lookup and a difficult puzzle.
Keep in mind: The gains come from the authors benchmark suite and do not yet establish behavior in production.
Read the source · Route-and-Generate authors ↗07Model behavior
A new test separates two reasons a multilingual model accepts a false premise: it may not know the answer, or it may abandon an answer it already knew under pressure. The authors found that similar headline scores can hide very different mixes of these failures.
Why it mattersTeams can choose a better fix once they know whether the problem is missing knowledge or pressure to agree.
Keep in mind: The diagnosis covers the authors selected models, languages and pressure prompts.
Read the source · Disentangling Robustness and Sycophancy in Multilingual LLMs authors ↗