ArXiV ML/AI/CV papers summary

Theme 1: Efficient Inference and Resource-Aware Adaptation

The “compute-budget” problem—running massive models on constrained hardware—is a defining challenge of modern AI. We are moving away from brute-force scaling toward intelligent resource management.

• Inference Optimization: To overcome networking bottlenecks, Topology-Aware Data Movement for Disaggregated GPU Inference introduces a topology-aware orchestrator that slashes latency by up to 18x. Complementing this, DeltaServe: Host-Agnostic Co-Serving of Inference and Fine-Tuning for LLMs maximizes hardware utility by utilizing idle GPU capacity for fine-tuning without compromising inference SLOs. For edge deployment, WitCert: Sound Runtime Risk Observability and Gating for KV-Cache Quantization and BLADE: Boundary-Expanded and Layer-Adaptive Dynamic Exit for Efficient LLM Reasoning provide methods to prune or quantize models while maintaining rigorous soundness guarantees.
• Efficient Adaptation: LARA: Lightweight Adapters in the Residual Stream for Composable Adaptation and Alignment allows multiple behaviors to reside on a single device without modifying base weights, while Versatile On-device Adaptation at the Edge by Unifying Few-shot, Zero-shot, Continual, and In-context Learning enables on-the-fly learning within micro-watt power budgets.

Theme 2: Scientific Machine Learning and Physics-Informed Modeling

We are transitioning from “curve fitting” to models that respect the governing equations of the universe. By embedding physical laws into neural architectures, we achieve models that are not only more accurate but physically plausible.

• Physics-Informed Discovery: Physics-Informed Differentiable Degradation Modeling for Lithium-Ion Battery State-of-Health Prediction and Predicting Steel Fatigue Life from Micrographs Using Physics-Informed Deep Learning demonstrate the power of integrating metallurgical features and Arrhenius kinetics.
• Operator Learning: HERO: History-Enriched Rollout Training for Long-Horizon Autoregressive Neural Operators and Feature Interaction Modeling for Physics-Informed Neural Networks and Neural Operators provide new ways to model complex PDEs, particularly those with sharp gradients or shocks, by focusing on feature interactions and history-enriched supervision.

Theme 3: Agentic Reasoning, Self-Correction, and Recursive Evolution

The field is shifting from static models to “agentic harnesses”—persistent systems that plan, use tools, and actively manage their own improvement.

• Self-Correction and Verification: SVR: Self-Verifying Refinement via Joint Verdict-Confidence Reinforcement Learning for Adaptive Test-Time Compute and AMTFV: Agentic Mathematical Tool-Flow Verification for LLM Self-Correction create “slow-thinking” loops where models verify their own outputs.
• Recursive Improvement: Code Is the Body: Agent-Owned Software Bodies for Recursive Evolution and Descent proposes an architecture where agents own their code to version and evolve themselves. This is supported by RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems and HarnessBank: Semantic Gene-Bank Search with Gated Verification for Agent-Harness Self-Evolution, which automate trial-and-error optimization. Self-Supervised Skill Optimization and LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents further demonstrate learning through execution trajectories without ground-truth labels.
• Benchmarking: MerchantBench: Benchmarking LLM Agents for Long-Term Coherence in E-Commerce Operations and AgentHPOBench: A Benchmark For Evaluating LLM Agents as Sequential Hyperparameter Optimizers emphasize the need to measure long-term coherence over single-turn accuracy.

Theme 4: Grounding, Safety, and the Evaluation Crisis

As AI enters high-stakes domains, we face a “crisis of evaluation” where current benchmarks often conflate capability with safety.

• The Evaluation Trap: The Formalism Trap: Are LLM-as-a-Judge Evaluators Blinded by Consensus Mimicry under Social Load?, Agreement Is Not Quality: Blind Expert Verification of Human and LLM Qualitative Coding When Human Consensus Is Not Ground Truth, and Agreement Metrics for LLM-as-a-Judge Evaluation: What to Report and Why challenge the reliance on human consensus as a proxy for quality. Validation Evidence in LLM Repair Agents: How Much of What Passes Actually Tests the Bug? warns that passing tests does not always imply functional correctness.
• Safety and Grounding: EvalSafetyGap: A Hybrid Survey and Conceptual Framework for LLM Evaluation-Safety Failures and Safety, or Just Capability? A Validity Audit of Agent-Safety Benchmarks call for more rigorous auditing. To ensure safety in practice, ActFovea: Runtime Safeguarding for VLA Policies via Spatiotemporal Visual-Action Consistency and LabGuard: Grounding Natural-Language Laboratory Rules into Runtime Guards for Embodied Laboratory Agents implement real-time physical constraints.
• Robustness: TextCloak: Thwarting Unauthorized LLM Exploitation via RL-Driven Unlearnable Text and TriShield: Zero-Utility-Loss Defense Against Privacy Backdoors in Federated Language Model Fine-Tuning protect against data exploitation and backdoors.

Theme 5: Geometric Foundations and Statistical Inference

By leveraging the rich mathematical structures of topology and geometry, we can build more expressive, stable, and interpretable models.

• Geometric Representation: TAGTorch: A PyTorch Library for Geometry, Topology, and Symmetry-Aware Machine Learning provides the infrastructure for this movement. Encoding the Euler Characteristic Transform and 1-Lipschitz Neural Networks on Hadamard Manifolds demonstrate robust representations for 3D analysis. Exponential Capacity in Multilayer Hetero-Associative Neural Networks explores how the geometry of encoding allows for exponential memory capacity.
• Statistical Inference: Distance Profile Embedding for Independence and Conditional Independence Testing of Random Objects maps complex objects into Hilbert spaces for rigorous testing. Longitudinal Adaptive Experimental Design for Learning Multiple Target Estimands with Semiparametric Efficient Inference and Bayesian fusion forests for heterogeneous treatment effects on survival from randomised and real-world data provide frameworks for causal discovery and data fusion.
• Multi-View Integration: Spectral Joint Subspace Estimation for Heterogeneous Multi-View Data: Geometry and Reweighting introduces “HeteroJIVE” to account for statistical heterogeneity in multi-modal datasets, ensuring that our perspectives are weighted by their true signal-to-noise ratio.