SkillRouter: Skill Routing for LLM Agents at Scale
Research introduces 'SkillRouter' for LLM agent skill routing, addressing scalability challenges in large skill ecosystems.
arXiv cs.LG — Machine Learning·21 Jul 2026
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Research introduces 'SkillRouter' for LLM agent skill routing, addressing scalability challenges in large skill ecosystems.
Research proposes Precision-Varying Prediction (PVP) to robustify ASR systems against adversarial attacks by randomly sampling precision during inference.
New neural approach for global optimization from noisy samples shows iterative refinement can find global minima, addressing multi-modal function challenges.
DIB-OD proposes a novel Graph Neural Network pretraining method to improve generalization across heterogeneous graph domains by disentangling invariant knowledge.
DORA is an asynchronous reinforcement learning system that speeds up LLM post-training by overlapping generation with model training, addressing the rollout phase bottleneck.
Research introduces StructGP, a continuous-time Gaussian process model that learns sparse, ordered directed acyclic graphs for interpretable forecasting.
Research systematically investigates Reinforcement Learning (RL) based jailbreaking techniques in LLMs, highlighting threats to safe model deployment.
New research introduces 'Unpack,' a backward recursion method to decompose credit through transformer sublayers for mechanistic interpretability.
Research proves a 'Behavioral Credibility Trilemma' for confidence-gated autonomous agents: maximum helpfulness, optimal calibration, and full autonomy are mutually exclusive.
Research introduces AOE, a method improving out-of-distribution detection for ML models by recalibrating outlier labels during training.
Research identifies commonalities between neural scaling laws and the Vendi Score, showing both are submodular and a special case of matrix spectral functions.
New research proposes "singularity-aware optimization" to address non-smooth loss landscapes in deep learning, aiming for stable convergence.
Research proposes ADWM, an autoregressive diffusion world model, to evaluate LLM agent policies off-policy using pre-collected trajectories.
FAIR-Calib proposes a Post-Training Quantization (PTQ) method to mitigate errors in Diffusion Large Language Models (dLLMs) caused by early, fragile token decisions.
FlashMemory-DeepSeek-V4 proposes Lookahead Sparse Attention, a new inference method to reduce GPU memory for ultra-long context LLMs.
RoVE (Rotary Value Embeddings) is a parameter-free modification to make attention values position-sensitive, building on Rotary Position Embeddings (RoPE).
Research paper details an interpretable machine learning pipeline using XGBoost and TreeSHAP to decompose equity return predictability in Chinese A-shares.
Researchers propose Overlapping Schwarz Attention, a hierarchical attention mechanism for LLMs inspired by domain decomposition methods, tested on operator learning.
Research explores Role-Typed Credit Assignment (RTCA) for agentic reinforcement learning, improving credit distribution beyond standard GRPO for actions.
Research demonstrates that biases in LLM-based pairwise judges are not fully identifiable or removable through standard statistical debiasing methods.
Research introduces Tensor-Train Joint Modeling for discrete diffusion models, promising faster generation by overcoming a structural limitation.
Research finds low-rank optimizers like GaLore, used for memory-efficient LLM training, lack a stable, trackable subspace beyond a small core.
Research explores using LLM-guided task-semantic field factorization for industrial process forecasting with scarce labeled data and changing regimes.
Research introduces EHM, an online mirror descent algorithm for generalized linear bandit problems under heavy-tailed noise.
Research proposes TSRouter, a system using dynamic modality-model selection for time series reasoning, combining LLMs and VLMs for data analysis.
Research finds Tabular Foundation Models (TFMs) struggle with discrete choice estimation due to their assumption of row-independent data, limiting direct application.
Research reveals gradient-free privacy leakage in federated language models (FLMs) through selective weight tampering, exposing sensitive client data.
Research demonstrates gradient span algorithms exhibit deterministic behavior in high-dimensional Gaussian functions, explaining consistent training outcomes.
Research benchmarks generative methods for zero-shot environmental sound classification, an underexplored area compared to computer vision.
Research explores regret minimization for optimizing unknown, stochastic piecewise linear rewards in microeconomic models like contract design and auctions.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion