Teaching robots to handle real-world chores—no special training required This research shows a new way for robots to complete long, multi-step tasks by combining off‑the‑shelf foundation models (the same kind powering today’s AI) with a constantly updated “scene graph”—a smart map of objects and their
AI
From doodle to design Imagine sketching a rough poster layout and instantly getting a polished version. That’s the promise of Sketch‑to‑Layout, a new approach that uses your sketch as the design guide. * How it works: A multimodal transformer takes your quick sketch plus content (text/images) and
ExplainableAI
Meet “Modules of Influence”: groups of features that drive AI decisions Most explainers like SHAP or LIME tell you which single features mattered for one prediction. This paper goes a step further: it builds an explanation graph across many predictions and uses community detection to reveal feature groups that act
AI
What's new Spiking Neural Networks (SNNs) mimic brain cells that communicate with brief spikes, not continuous signals. That makes them naturally energy-efficient and great at timing. * Surrogate-gradient training gets SNNs within 1-2% of standard ANN accuracy, converging by about 20 epochs with latency near 10 ms. * ANN-to-SNN conversion
AgenticAI
What is "Agentic AI" really? Sebastian Benthall and Andrew Clark argue it’s best seen as a software delivery mechanism—like SaaS—that lets an application work autonomously inside complex enterprises. * Applications first: Agentic AI systems should be judged as products, not as foundation models. * Validate with users:
From 300 guidelines to precise answers—grounded, not guessed. This RAG system uses a hybrid retriever over 10,195 chunks (from 300 NICE guidelines). Retrieval shines: MRR 0.814; 81% recall at first chunk and 99.1% within top 10 (7,901 queries). In generation (70 QA pairs), RAG lifted
New attack objective, bigger search space—higher risk. This work proposes the first gradient-based untargeted jailbreak attack (UJA): instead of matching a fixed unsafe target, it maximizes the probability of an unsafe response using a judge model and a differentiable decomposition. In evaluations, UJA achieved over 80% success against recent
Descriptive vs. expressive semantics—same sentence, different signals. Participants watched emotional movie clips, then described their experiences. The study shows descriptive semantics align with intended emotions, while expressive semantics track the emotions actually evoked (valence/arousal and self-ratings). Why it matters: Separating content from expression can make Speech Emotion Recognition
Different labels, same pitfalls. This position paper argues that reward models (for RL-based LLM training) and evaluation metrics face overlapping challenges—spurious correlations, reward hacking, data quality, and meta-evaluation. In some tasks, metrics even outperform reward models. Why it matters: Aligning these research communities could improve preference elicitation, robustness to
LLM agents work better together—with the right prompts. This work enhances the CoELA framework for collaborative embodied agents, testing multiple LLMs and prompt strategies to boost cooperation and decision-making in shared virtual spaces. The best combo improved efficiency with Gemma3 by 22% over the original setup. They also added
When chains get long, LLMs lose the thread. Self-Anchor pins it down. Self-Anchor structures a plan from the reasoning trajectory and automatically aligns the model’s attention to the most relevant steps while generating. Across six benchmarks, it outperforms state-of-the-art prompting and narrows the gap between general and specialized reasoning
Counterintuitive, but it works: add tiny shifts to boost robustness. This training-free, architecture-agnostic method uses stochastic resonance: apply small image translations, align features, aggregate, and map back—no extra modules, no attack-specific tuning. It recovers up to 68.1% of accuracy loss in classification, 71.9% in stereo, and 29.