Narrative World Model: Narratology-Grounded Writer Memory for Long-Form Fiction
The Narrative World Model (NWM) targets a specific gap in story-aware AI: general retrieval systems can’t reliably answer multi-hop questions that depend on narratology structure, like who learned what secret and when it was revealed. By combining a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval, NWM improves multi-hop narratological QA over established temporal-knowledge-graph memory baselines and also outpaces flat retrieval and GraphRAG-style approaches.
FirstResearch: Auditable Question Formation for LLM Scientific Discovery Agents
FirstResearch focuses on making LLM-generated scientific research questions inspectable rather than merely plausible by forcing them into a structured “Research Question Certificate” with assumptions, mechanism model, falsifiable hypothesis, decisive test, and failure update rules. Across multiple LLM-agent research topics, the certificate-centered approach outperforms prompt-only baselines under LLM-judge evaluation, with strong agreement across independent judge runs and evidence that removing certificates collapses quality.
Helping K–12 educators build practical AI skills
OpenAI Academy and the Walton Family Foundation are rolling out hands-on “AI Skills Jams” designed to help K–12 educators build practical AI capability in classroom settings. The emphasis is on immediately usable skills for teachers, bridging the gap between fast-moving AI research and day-to-day instructional needs.
Foundation Models for Automatic CAD Generation
This study evaluates how well today’s foundation models can generate mechanical CAD from natural-language specs using a unified pipeline and a 97-problem engineering benchmark. It also tests two critique regimes – lightweight analytic rendering metrics versus VLM-based visual-semantic critics – finding that strong compact instruction-tuned models can approach top performance while revealing systematic failure modes on geometries like rotationally symmetric cylinders.
Synthetic Consumer Insight Generation with Large Language Models
The paper examines whether LLMs can generate synthetic consumer responses for projective marketing-style techniques without simply copying human data patterns. Results show overlap with human responses in broad topics and associations, but also measurable differences in linguistic structure and how diversity is expressed, leading to practical guidance on when and how to use synthetic consumer insights responsibly.
From Hugging Face to Amazon SageMaker Studio in one click
This Hugging Face-focused workflow shows how to move from Hugging Face assets into Amazon SageMaker Studio with minimal friction, streamlining model experimentation and deployment paths. The “one click” promise is aimed at reducing glue-code work so teams can spend time evaluating models rather than rebuilding environments.
Meta’s new Muse Image model can pull other Instagram users into AI photos
Meta’s Muse Image model is being deployed across Meta AI, Instagram, and WhatsApp, with plans to expand further, and it’s framed as part of the Muse family where the system can reason, search, and plan before generating. The Verge’s reporting highlights that Muse Image can incorporate other Instagram users into AI photos, underscoring how quickly creator-facing AI features are turning into full product capabilities.
Akashic: A Low-Overhead LLM Inference Service with MemAttention
Akashic tackles the latency problem that prevents “memory every step” agent designs from working in practice by using MemAttention to chunk and relate context across steps efficiently. With hardware-software memory placement to reduce retrieval fragmentation, Akashic improves task accuracy up to 10.2 points and boosts throughput up to 1.21x across multiple workloads and model sizes.
From Graphs to Gradients: Physics-Inspired Structural Attribution for Cyber-Physical IoT Systems and Beyond
This work reframes explainability for cyber-physical systems by using an undirected, energy-based representation to produce dependency-aware attribution without needing to recover directed causal graphs. In simulations on an industrial-style hybrid IoT testbed, the method improves attribution accuracy and robustness while scaling better than graph-based alternatives, with potential value for diagnosis in high-stakes settings.
CSTutorBench: Benchmarking Small Language Models as Tutors for Block-Based Programming
CSTutorBench evaluates small language models as K–12 CS tutors in a block-based robotics environment using scenario questions and a tutoring/feedback rubric. The benchmark findings show that models often handle surface criteria (tone, vocabulary) well but struggle with deeper tutoring behaviors like avoiding answer leakage and engaging student debugging history, and prompt revisions guided by educational research significantly improve scores.
Prompt-to-Paper: Agentic AI System for Bioinformatics
Prompt-to-Paper aims to make AI manuscript generation more rigorous by grounding claims in verifiable literature, executing computational biology code instead of fabricating results, and scoring outputs with an eight-dimensional quality rubric. Across five bioinformatics case studies, the system generated submission-formatted PDFs with zero out-of-range citations and improved quality scores through a feedback-driven improvement loop.
ChatGPT’s upgraded voice mode is better at shutting up
OpenAI is overhauling ChatGPT Voice with GPT-Live-1, designed to interrupt users less and to naturally wait when someone pauses mid-thought. The update also routes user queries to stronger text models for reasoning or web search when needed, aiming to make voice conversations feel more like a continuous back-and-forth.
Native-speed vLLM transformers modeling backend
This NVIDIA-focused announcement looks at how to get closer to “native-speed” inference by optimizing the vLLM transformer backend for performance-oriented production use cases. The core pitch is lower overhead in serving transformer workloads, which is especially relevant for systems that need fast iteration and cost-efficient throughput.
Memory in the Loop: In-Process Retrieval as Extended Working Memory for Language Agents
Memory in the Loop argues that the bottleneck for in-loop retrieval isn’t the “pattern” but the networked store latency – so it studies what happens when memory is moved in-process. The results show that fast in-process memory can make read/write retrieval essentially continuous, improving recall and accuracy while collapsing the per-step latency tax compared to cloud retrieval, with an additional insight that embedding latency can dominate if not optimized.
ArtisanCAD: An Industrial-Level CAD Agent with Expert-Grounded Knowledge Distillation
ArtisanCAD introduces an industrial CAD agent that distills expert procedural knowledge into reusable “skills” using a CAD intermediate representation (CAD-IR) to carry parameters, operation order, tool bindings, dependencies, and verification rules. On CAD benchmarks and complex automotive component tests, CAD-IR helps bridge vague prompts into executable CATIA-native workflows, improving generation quality from intermediate descriptions.