Researchers introduced Occupancy-based Policy Compression (OPC), a new approach to deep reinforcement learning that improves efficiency by focusing on long-term behavior rather than immediate actions.…
Researchers introduced SpreadsheetArena, a platform that evaluates language models on spreadsheet generation through blind pairwise preference votes. The system reveals that even top-ranked models oft…
A new data-driven pipeline reduces GPU requirements for LLM adapters by 60% by predicting optimal resource allocation. The system uses a digital twin, a distilled machine learning model, and a greedy …
ARCQuant, a new framework for Large Language Model inference, uses the NVFP4 numerical format to achieve up to 3x speedup on GPUs while maintaining accuracy comparable to full-precision baselines. The…
Researchers are advancing autonomous coordination by using self-supervised goal-reaching, where AI agents learn to reach target states without complex reward systems. This method outperforms tradition…
Researchers have developed a new VideoQA approach that separates causal reasoning from answer generation using natural language causal chains, improving explainability and trust. The model, comprising…
Recent studies highlight that AI agents face efficiency challenges in memory, tool learning, and planning, which are critical for real-world deployment. Researchers emphasize balancing effectiveness w…
Researchers have developed DrugAgent, a multi-agent system using large language models to assess drug-target interactions, achieving 98.8% faithfulness to input evidence on kinase screening data. The …
Researchers introduced TREK (Teacher-Routed Exploration via Forward KL), a new AI training method that enhances learning through unconventional exploration strategies. TREK significantly improved perf…
Researchers introduced MARS, a training-free method that enhances safety in multimodal large language models by leveraging textual refusal directions to steer model outputs away from unsafe content. T…
Researchers have developed STREAM, a new AI model that uses a diffusion transformer architecture to generate dance motions from both text and music cues, overcoming the modality collapse problem where…
Researchers have developed WhisperBench, a benchmark of 108 test cases, and MemGhost, a framework that crafts one-shot email payloads, to demonstrate how persistent personal agents can be compromised …
AI agents are becoming more capable, but efficiency in memory usage, tool learning, and planning is often overlooked, hindering real-world deployment. Experts argue that speed and cost-effectiveness a…
Local large language models like Qwen2.5-32B are now capable of constructing fuzzy cognitive maps from text data, such as hotel reviews, by extracting quantitative insights and mapping them to user sa…
Researchers introduced TREK, a method that improves AI problem-solving by using verified output trajectories to extend model learning. TREK boosted Qwen3-8B's performance on AIME 2024 from 36.9 to 40.…
AgenticAI-Supervisor, a new RL Gym environment for evaluating AI agents, decouples environment creation from execution and uses multi-dimensional reward shaping to combat reward hacking. Its first cas…
New research challenges assumptions about infinite-width neural networks, introducing a width-solid learnability theorem that shows infinite width does not guarantee an edge over polynomial width for …
Researchers introduced DT-Guard, a new AI safety model that combines reasoning-focused training with efficient inference to improve safety judgments while reducing latency. The model achieves average …
Researchers have introduced a new approach to Markov decision processes (MDPs) called minimum action distance (MAD), which measures the minimum number of actions needed to transition between states. T…
Researchers have developed physics-informed neural networks (PINNs) that embed the axisymmetric equations of linear elasticity to model wave propagation in bimaterial systems like steel and aluminum. …