Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers propose MESA, a task-adaptive multi-structure evidence selection framework for long-horizon agent memory, which dynamically selects and fuses a query-specific subset from five complementary structure views of…
Researchers introduced SBCO (Self-supervised Block Coordinate Optimizer), a verifier-grounded harness optimizer for planning agents that improves agent outputs from its own graded feedback without human labels or self-mo…
Researchers found that latent-space communication between vision-language model agents can be compressed 128x with minimal accuracy loss. Fitting a post-hoc sparse autoencoder to frozen Vision Wormhole activations, a uin…
A new arXiv study (arXiv:2608.10207v1) introduces an LLM-assisted semantic stop representation for reinforcement-learning-based bus holding control, reducing headway variability by 32.0%, bunching events by 69.2%, and pa…
Researchers introduced ELVAE, an evidential learning-based variational autoencoder that assigns each latent coordinate an input-dependent normal-inverse-gamma posterior, enabling explicit latent-location uncertainty for …
Researchers introduced Exploration-Driven Personalized Federated Reinforcement Learning via Intrinsic Motivation (EDPFRL-IM), a framework that adds curiosity-driven exploration to federated reinforcement learning to impr…
Researchers propose IADD-TR, a framework that decouples environment dynamics into action-intervention and action-free stages to reduce policy-induced data bias in model-based reinforcement learning. Experiments on five M…
Researchers from Huawei have developed proxy models that reduce accelerator requirements by 50%-87.5% and cut per-step NPU-hour cost by up to 33.3x for reproducing failures in large language model reinforcement learning …
Researchers introduced TACTICL, an automated task-aware compression framework for tabular in-context learning models that prunes transformer layers and replaces them with lightweight adapters, reducing inference costs wh…
Researchers introduced 4D-WAM, a world-action model that uses geometric foundation models during training to enforce 4D consistency in predicted driving scenes, improving trajectory planning. The model achieved state-of-…
A new arXiv paper (2608.10203v1) introduces a convolutional layer activation dimensionality reduction method that improves out-of-distribution and adversarial attack detection in convolutional neural networks. The author…
Researchers from an unnamed institution introduced Preference-Aware RLHF (PA-RLHF), a method that separates optimization across preference modes during reward learning, to address procedural fairness failures in Reinforc…
Researchers propose REATS, an LLM reasoning-based ensemble learning framework for adaptive time series forecasting, which uses chain-of-thought reasoning to generate interpretable, sample-adaptive ensemble weights from h…
Researchers introduced a generator-guided inverse sampling method for Lévy-driven generative models, addressing the challenge of nonlocal reverse processes caused by infinite jump activities. The method decomposes dynami…
A new arXiv paper (2608.10288v1) introduces Power Law Graph Attention (PLGA), a learned attention mechanism that exactly generalizes scaled dot-product attention (SDPA) and is used in the Power Law Decoder Representation…
A new study from arXiv (arXiv:2608.10235v1) finds that Hamiltonian Neural Networks (HNNs) reduce mean energy drift by 42-fold and mean trajectory MSE by 15.8-fold compared to parameter-matched feedforward networks on the…
Researchers introduced Fisher8, an output-layer gradient correction that uses Fisher geometry to stabilize neural heteroscedastic regression, improving likelihood-error tradeoffs and uncertainty calibration without data-…
Researchers proposed ReCBM, an uncertainty-gated relational reasoning framework for Concept Bottleneck Models (CBMs), to improve robustness under unreliable concept states. ReCBM models co-occurrence, implication, and ex…
Researchers introduced SINKFLEX-RL, a modular training system for reinforcement learning in long-horizon tool-use agentic tasks, which integrates a Gymnasium-compatible environment wrapper, VERL-style rollout dataflow, g…
LinkedIn's self-evolving agentic customer support system, integrating retrieval-augmented generation with evolutionary auto-prompting, increased QA self-serve by 9.0 percentage points, cancellation self-serve by 4.8 poin…