State of Simulation for Physical AI: An Overview
The study maps how today’s simulation pipelines tackle the “physical mismatch” problem for embodied AI – from sensor fidelity and domain randomization to validation loops that connect virtual performance to real-world behavior. It positions simulation as a practical accelerator for physical training, but emphasizes that measurable benchmarks and reproducibility are still uneven across research groups.
Design and Validation of a Lightweight 1D CNN for Affective Touch Classification in Soft Plush Companions
This work delivers an open-source MATLAB framework for affective touch recognition using compact 1D CNNs, backed by a FAIR-compliant dataset of 1,326 labeled gesture sequences from 25 participants. It finds a 13.2k-parameter dilated 1D CNN can hit 75% test accuracy and 85% leave-one-subject-out cross-validation, with quantized inference estimated at 3.2 MMAC per window for real-time embedded deployment.
Google launches Gemini 3.5 Flash Cyber for cheaper security scanning
Google is adding Gemini 3.5 Flash Cyber, a security-focused model designed for fast, cost-efficient vulnerability discovery and patching. It’s built on Gemini 3.5 Flash, comes through Google’s CodeMender coding agent first to governments and trusted partners, and targets “multiple calls at high speed and low cost” to make security workflows more scalable than heavyweight models.
A Survey on GNN-based Link Prediction: Techniques, Applications, and Challenges
This survey lays out link prediction research through a dedicated GNN lens, organizing advances by encoder architectures such as GCN, GAE, GAT, and GFormer, and by downstream settings like knowledge graphs and recommendation systems. It also calls out open challenges – including how different graph structures and modeling choices affect reliability and evaluation – to help researchers avoid comparing “apples to oranges.”
PlanFlip: Attacking Multi-Agent LLM Systems via Planning-Phase Prompt Injection
PlanFlip shows how injecting malicious instructions into a multi-agent system’s planner can cascade into corrupted sub-tasks across an entire run, even when downstream components use filtering or critics. Across 3,479 episodes and nine frontier LLMs, it reports high attack success – with GPT-5 reaching ASR = 0.68 – and argues that reasoning-heavy models may be harder to exploit, while heterogeneous model diversity improves defenses.
Here are the 30,000 songs Sony is suing Udio’s AI music generator over
Sony has sued Udio again, alleging the AI music generator infringes copyrighted recordings spanning more than 30,000 songs, from classic rock staples to modern pop catalogs. The filing in New York claims Udio’s use represents only “a small portion” of affected works, escalating the broader copyright fight over how generative music systems learn and reproduce styles and content.
Anthropic’s $1.5 billion book piracy settlement approved by judge
A federal judge has approved Anthropic’s $1.5 billion class action settlement tied to accusations that the company trained on copyrighted books. The order says authors receive about $3,000 per allegedly infringed book, marking what Reuters describes as the largest known copyright recovery in history.
Deterministic Replay for AI Agent Systems
agrepl tackles a core debugging pain for tool-using LLM agents by making executions reproducible, even when APIs, sampling variance, and network state normally break determinism. The CLI framework MITM-proxies external interactions, records structured traces, and replays them offline in an isolated environment with a determinism invariant and reported replay fidelity of F = 1.0 across 250 instances.
Rater State Bias in RLHF Preference Data: An Audit Framework
This paper argues RLHF preference datasets can carry a structured confound: a rater’s emotional or stressful state during annotation can shift preferences independently of output quality. It proposes an audit framework and survival-level emotional authenticity measures, laying out falsifiable predictions about how state-dependent bias can survive aggregation and distort learned reward models.
Substack adds an AI detector to help spot blogs written by no one
Substack is rolling out a “Scan for AI text” tool that estimates how much of a post may be AI-generated or AI-assisted, powered by Pangram. The feature analyzes posts, notes, replies, and comments, with longer content support and expansion beyond iOS planned for Android.
Advancing next-gen AI with materials science innovation
The report connects AI progress to the often-overlooked hardware side – materials that enable the next jump in compute efficiency, memory density, and power constraints. It frames materials science as a parallel “infrastructure layer” for emerging AI systems, where breakthroughs in physical components can unlock new model capabilities without proportional energy cost.
Democratizing AI with Small Language Models: Structured Benchmarking and Parameter-Efficient Fine-Tuning for Local Deployment
This work makes a case for “AI democratization” via small, locally deployable models trained for structured tasks, not just broad generality. Using a 16-topic benchmark with strict output protocols and parameter-efficient fine-tuning under a budget, it reports that sub-3B models can become viable local experts – with Qwen Coder 3B reaching 75.67% strict accuracy and fine-tuning yielding large gains on held-out adaptation.
Generative Ontology Induction: Domain-Agnostic Schema Discovery from Document Corpora Using Large Language Models
GOI focuses on the ontology engineering bottleneck by inducing a typed “generative blueprint” – entities, properties, relationships, and constraints – from document corpora and exporting it as a structured graph in YAML/JSON. Across four diverse ontologies, it reports 95-100% node coverage for generated structural backbones, with coverage performance measured via a new Node Coverage Score.
Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL
This paper argues masked diffusion language models (MDLMs) address key weaknesses in autoregressive text world models by enabling bidirectional, anchor-aware denoising. It reports better coherence, groundedness, and rollout diversity than larger AR baselines, plus a plug-and-play GRPO training approach and up to 47% absolute gains on out-of-distribution environments in zero-shot transfer tests.
It Takes 8 Tokens: Weak-to-Strong Off-Policy RL via Auxiliary Branches
W2SPO proposes a weak-to-strong RL strategy for verifiable-reward training when models get stuck in redundant reasoning “basins.” By inserting short auxiliary segments into trajectories (often as few as 8 tokens) from a weaker model and restricting updates to those segments, it reports improved Pass@1 (62.3% to 64.2%) and a training speedup of 3.55x under the same sampling budget.