Researchers have developed a sparse cache approach for sequence models that stores only unique entries, dramatically improving memory efficiency while matching full-attention recall levels. The method, which uses DP-mean…
Researchers developed a CNN-LSTM-RL framework that uses reinforcement learning to refine EEG-based brain-computer interface predictions, achieving up to 41.5% improvement in movement decoding accuracy in 2D and VR scenar…
AgentFootprint, a new benchmark, reveals that AI agents' storage footprint varies by up to 6.7x across frameworks, with configurations achieving 100% accuracy differing by 15.7x in retained bytes. The benchmark highlight…
Researchers introduced Self-Verified Reasoner (SVR-R1), a framework enabling AI models to self-verify answers through internal dialogue, significantly boosting accuracy on vision-language reasoning tasks. The approach us…
A new dataset called CRiT-QA exposes flaws in large language models' multi-hop reasoning, revealing their reliance on memorized knowledge and superficial shortcuts rather than genuine reasoning. The benchmark forces mode…
A new study on distributional reinforcement learning agents reveals that 40-95% of their risk trade-off claims are statistically invalid, with risk perception emerging as a training artifact rather than a genuine assessm…
Researchers have developed a method using case-based decision theory to decompose neural network decisions into weighted influences of past training cases, enabling audits without retraining. The approach achieved high c…
A new framework integrates contemplative traditions like mindfulness and compassion into large language models to improve ethical alignment, initially targeting mental health applications. The modular, domain-agnostic sy…
Researchers introduced Agentic-DPO, a new method for training language models that uses state-conditioned preference supervision instead of costly environment rollouts. On the tau-bench, Agentic-DPO improved accuracy fro…
Researchers tested large language models in Texas Hold'em poker to study their risk-taking behaviors, revealing distinct risk profiles ranging from conservative to aggressive. The findings highlight how AI adapts to unce…
6G networks are integrating artificial intelligence deeply into their infrastructure, using self-supervised learning methods like joint-embedding predictive architecture (JEPA) to manage data and optimize operations such…
Companies including Duolingo, Salesforce, Klarna, Harvey, and Adobe are integrating artificial intelligence into their products to personalize learning, refine customer relationship management, streamline payments, assis…
RFC BE ships the History tool in v1.20.0, providing ten operations for accessing past chats including list, get, search, rename, annotate, pin, archive, recap, resume, and related. The tool introduces session-level metad…
Mathematician Terence Tao developed an interactive web app that visualizes the logical flow of mathematical papers using JSON dependency files, leveraging large language models to generate diagrams for demonstration. The…
A developer introduces 'mechanical coding,' a workflow where ChatGPT proposes small, bounded steps and the human verifies each step via terminal output, preventing the model from inventing project state. The method treat…
Researchers found that deep learning models for click-through rate predictions suffer performance degradation after the first training epoch due to feature sparsity. By removing highly sparse features and merging infrequ…
New AI models, including transformer-based BERT and graph attention networks, are transforming puzzle game design by predicting player behavior more accurately than traditional methods. Tested on Candy Crush Saga, these …
The GUARD Act, a federal bill, would require mandatory identity verification for all chatbot users, ban minors from AI companions, and impose criminal liability on developers under a vague standard, raising privacy and F…
OpenAI's ad business is projected to miss its own five-year revenue forecast by 90%, according to Emarketer. The AI lab forecast $2.5 billion in ad revenue this year and $100 billion by 2030, but Emarketer estimates stan…
Samsung Electronics won eight Red Dot Design Award honors in the design concept category, including two 'Best of the Best' prizes for a home appliance consumables concept and a kids' robot concept called Dremo & Minimo, …