Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A developer argues that AI agents, which complete tasks autonomously, represent a more significant shift than chatbots. The post claims agents can research, analyze, and deliver results without step-by-step instructions,…
Researchers introduced ForecastBench-Sim, a simulated-world forecasting benchmark built on game rollouts from Freeciv, to evaluate AI systems' probabilistic reasoning under dynamic world states. The benchmark generates c…
A new paper from arXiv formalizes what generalist agents must store in memory to act near-optimally across multiple environments. It proves a separation theorem showing that uniformly near-optimal policies require distin…
Researchers introduced R2D-RL, a reinforcement learning environment that bridges the RoboCup 2D Soccer Simulation platform with Python-based multi-agent reinforcement learning workflows. The environment supports full-fie…
Researchers from DiDi introduced ProfiLLM, an agentic LLM data pipeline that generates utility-aligned user profiles for industrial ride-hailing dispatch. Deployed on DiDi's production system, ProfiLLM achieved up to +6.…
Researchers introduced BeliefDiffusion, a framework combining diffusion models and model predictive control for navigation in partially observable environments. The method outperformed reinforcement learning and other ge…
Researchers introduced RTSGameBench, a benchmark built on the real-time strategy game Beyond All Reason, to evaluate strategic reasoning in Vision-Language Models. The benchmark includes diverse matchups, mini-games for …
Researchers propose ThinkDeception, a novel framework that uses multimodal large language models and chain-of-thought reasoning to detect deception with interpretability. It introduces a progressive training strategy and…
Researchers propose RODS (Reward-driven Online Data Synthesis) to address the depletion of informative samples in multi-turn tool-use reinforcement learning. RODS uses progress reward variance as a boundary detector to c…
Researchers introduced ARIADNE, a training-free routing framework for dynamic adapter selection in parameter-efficient fine-tuning ecosystems. It selects task-specialized adapters at inference time by measuring input pro…
Researchers introduced VISUALSKILL, a multimodal skill library for computer-use agents that combines text and visual figures to improve performance on long-horizon tasks. In tests, the system achieved a 0.456 average sco…
The Montreal Forced Aligner (MFA) 3.0 achieves state-of-the-art speech-to-text alignment with mean boundary errors below 15 ms across English, Japanese, and Korean. Released in 2016, MFA has become the most widely used f…
Researchers found that unlearning knowledge from large language models causes collateral damage that decays with semantic distance but persists across domains. They developed a method to audit this damage before unlearni…
Researchers introduced MCompassRAG, a metadata-guided retrieval framework that uses topic-level signals to improve paragraph-level retrieval in RAG systems. The method achieves 8.24% higher information efficiency with ov…
Researchers have developed a speech-driven end-to-end model for discriminating among Chinese dialects, combining MFCC features with word-level embeddings via a CNN. The approach outperforms state-of-the-art methods on tw…
Researchers developed a Chinese dialects discrimination framework using transfer learning and data augmentation to address low-resource challenges. The model outperformed state-of-the-art methods on two benchmark corpora…
Researchers introduced Steerable Cultural Preference Optimization (SCPO), a reward model training algorithm that balances diverse cultural preferences in large language models. SCPO improved minority reward model perform…
Researchers introduced Gaussian Mixture Attention (GMA), a probabilistic attention mechanism that replaces pairwise query-key comparisons with routing through K learned Gaussian mixture components, achieving linear-time …
Researchers introduced CAOA, a method that integrates point cloud completion with symmetry-aware pose estimation to align CAD models to objects in indoor RGB-D scans. On the Scan2CAD benchmark, CAOA achieved a 17% accura…
Researchers introduced PROPEL, a framework that trains task generators for reinforcement learning by using a lightweight activation probe to predict solver pass rates, avoiding costly solver rollouts. In tests on math, c…