Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers propose TurnOPD, a turn-level budgeting strategy for efficient on-policy distillation of long-horizon language agents. The method addresses inefficiencies in vanilla agent OPD by using adaptive rollout-depth …
Researchers introduced ArtisanCAD, an industrial-level CAD agent that uses expert-grounded knowledge distillation to convert ambiguous text prompts into executable CAD programs. The system leverages a CAD intermediate re…
Researchers introduced Design-CP, two context-parallel inference strategies for RFdiffusion 3 that distribute quadratic activations across multi-GPU meshes, enabling all-atom design of large protein nanoparticles. The me…
A study on the 'Granularity Paradox' in time-series forecasting finds that finer temporal disaggregation improves in-sample fit but degrades out-of-sample accuracy due to recursive error compounding. Benchmarking 10 mode…
Researchers propose Patch Knowledge Transfer (PKT), a knowledge distillation framework that enables efficient AI-generated image quality assessment by transferring knowledge from a complex teacher model to a lightweight …
Researchers introduced a Tri-Branch Modular Fusion Neural Network that combines 3D geometry, SMILES embeddings, and physicochemical descriptors for molecular property prediction. The model achieved a validation MAE of 0.…
Researchers formalize the execution harness of LLM agents as a learnable control layer using offline reinforcement learning, training a lightweight controller with advantage-weighted regression to improve verification be…
Researchers introduced Algorithmic Core Extraction (ACE) to isolate compact subspaces, or algorithmic cores, that are necessary and sufficient for a task and recur across independently trained transformers. Across synthe…
Researchers introduced a rendering-aware Bayesian 3D Gaussian splatting framework that provides native uncertainty estimates and adaptive complexity control, outperforming standard ensembles in active view selection task…
A controlled comparison of byte-pair encoding (BPE) and Unigram-LM tokenizers on chemistry SMILES reveals that the two algorithms build near-disjoint subword vocabularies across all 22 tested conditions, with cross-algor…
Researchers revisited the relation between language model perplexity and word error rate for modern end-to-end automatic speech recognition systems, finding that internal language modeling capacity and the use of neural …
Researchers introduced ResonatorLM, a new language model mechanism that replaces transformer self-attention with causal functions of damped resonators derived from physics. In a 6M parameter setting, ResonatorLM achieved…
Researchers introduced a new method for structural sequence analysis based on Algorithmic Information Theory, using the Ladderpath approach to extract nested and hierarchical repetitions. The method defines three distanc…
Researchers introduced InvWeaver, a neuro-symbolic framework that synthesizes loop invariants for programs with multiple interacting loops by exposing inter-loop dependencies and propagating proof obligations. The system…
Researchers introduced SPANUQ, a lightweight probe for span-level uncertainty quantification in large language models, achieving 10-20x faster inference than sampling-based methods while outperforming them in uncertainty…
Researchers introduced Light-Omni, a multimodal agent framework for reflexive video understanding that uses dual contextual states to avoid costly iterative reasoning. It achieved a 2.4% accuracy gain over M3-Agent with …
Researchers introduced Statistically Meaningful Geometry (SMG), a framework modeling over-parameterized learning systems as infinite-dimensional Orlicz fiber bundles, to differentiate genuine intelligence from statistica…
Researchers introduced FedPPO-PG, a federated multi-agent reinforcement learning framework for transient stability control in smart grids. The system achieved 100% stabilization in simulations on the IEEE 39-bus benchmar…
Researchers introduced FASR++, a diffusion-model-based super-resolution algorithm that leverages multiple low-quality images from surveillance cameras to enhance facial features and improve face recognition. The method a…
Researchers show that incorrect L0 hyperparameter settings in Sparse Autoencoders (SAEs) cause feature mixing and degenerate solutions, preventing proper disentanglement of LLM internal features. They propose a proxy met…