Theme 1: Efficient Inference & Model Compression

The rapid deployment of Large Language Models (LLMs) has shifted from brute-force scaling to “surgical” optimization. We are moving beyond static pruning toward dynamic, input-dependent execution. Accelerating GPU Inference of Large Language Models with Moderately Unstructured Sparse Weight Matrices utilizes a three-layer storage format to outperform dense tensor cores, while COBS: Cumulant Order Block Sparse Attention mitigates the KV-cache bottleneck by using second-order statistics to select attention blocks.

Further efficiency gains are realized through adaptive computation: Sensitivity-Aware Thresholding and Token Routing for Activation Sparsification in Large Language Models and HALO: Hybrid Adaptive Latent Reasoning for Language Models optimize the quality-throughput trade-off by selectively applying computation. These are complemented by inference engineering breakthroughs like DominoTree: Conditional Tree-Structured Drafting with Domino for Speculative Decoding and AugServe: Adaptive Request Scheduling for Augmented Large Language Model Inference Serving, which maximize hardware utilization through smarter token drafting and request scheduling.

Theme 2: Agentic Reasoning & Planning

AI is evolving from passive text generation into active, multi-step problem-solving. This transition requires robust frameworks for planning, memory, and tool use. GATS: Graph-Augmented Tree Search with Layered World Models for Efficient Agent Planning and Hierarchical Chain-of-Thought: Enhancing LLM Reasoning Performance and Efficiency emphasize systematic search and planning over simple exploration. To manage long-horizon tasks, ECHO: Prune To Act, Trace To Learn With Selective Turn Memory In Agentic RL, AgentKGV: Agentic LLM-RAG Framework with Two-Stage Training for the Fact Verification of Knowledge Graphs, and Remember Your Trace: Memory-Guided Long-Horizon Agentic Framework for Consistent and Hierarchical Repository-Level Code Documentation highlight the necessity of maintaining provenance and iterative refinement.

Innovation in this space also demands new representational primitives, as argued in Beyond Fixed Representations: The Vocabulary and Verifier Gaps in Open-Ended AI, and the ability to move reasoning into continuous latent spaces, as seen in Latent Thoughts Tuning: Bridging Context and Reasoning with Fused Information in Latent Tokens. Finally, Toward Auditable AI Scientists: A Hypothesis Evolution Protocol for LLM Agents provides a formal protocol for verifiable scientific reasoning.

Theme 3: Alignment, Safety, & Unlearning

As agents become more complex, safety can no longer be treated as a static property. Hair-Trigger Alignment: Black-Box Evaluation Cannot Guarantee Post-Update Alignment and Refused in Chat, Written in Code: Workflow-Level Jailbreak Construction in IDE Coding Agents demonstrate that single-turn benchmarks are insufficient to capture risks in multi-turn, agentic workflows.

Addressing these vulnerabilities requires sophisticated internal probing and surgical removal of knowledge. Forget Narrowly, Retain Broadly: Unlearning as an Asymmetric Generalization Problem proposes frameworks for targeted unlearning, while Optimizing Against Safety Representations: Activation-Guided Adversarial Suffixes and the Geometry of Refusal reveals that refusal behavior is distributed across the model’s forward pass, necessitating activation-guided defense strategies.

Theme 4: Scientific Discovery & Domain-Specific AI

AI is increasingly acting as a surrogate for complex physical simulations, grounding its reasoning in the laws of nature. A Machine Learning Surrogate for Component Criticality Ranking in Interdependent Power-Communication Networks, Learning Physics-Informed Surrogate Model of Linear Elastic Displacement Fields from Geometry, and GReFEM: Multimodal LLMs as Zero-Shot Semantic Assistants for Physics-Guided 3D Mesh Refinement demonstrate the power of physics-informed surrogates.

In the biological and astronomical sciences, TheBioCollection: Unified Pre-Training Scale LLM Corpus for Biology and Variable-Length Generative Protein Design via Generalized Poisson Flow push the boundaries of generative design. Meanwhile, DKCD: Domain Knowledge-Enhanced Causal Discovery from Unstructured Data, Beyond LLMs: A Linguistic Approach to Causal Graph Generation from Narrative Texts, and AS-Bridge: A Bidirectional Generative Framework Bridging Next-Generation Astronomical Surveys showcase how AI can synthesize disparate data sources to uncover causal structures and bridge scientific surveys.

Theme 5: Interpretability, Representation, & The Gradient Bottleneck

We are moving toward a rigorous, dynamical understanding of neural representations. How are linear representations learned? Exact solutions to the dynamics of abstraction and Riemannian Geometry for Pre-trained Language Model Embeddings explore how concepts emerge as geometric structures. Neural Collapse Is Forbidden: Information Floors in Language Models challenges existing hypotheses by identifying the necessity of within-class variance, while All Explanations are Wrong, But Many Are Useful: Exploring the Rashomon Explanation Set with Large Language Models suggests that coupling explanation with prediction improves model accuracy.

Underpinning these representations are the gradients themselves. Lost in Backpropagation: The LM Head is a Gradient Bottleneck identifies a fundamental inefficiency in current output architectures, while Smooth Scaling Laws Hide Stepwise Token Learning reveals that learning occurs in localized events, offering a roadmap for more efficient training. Finally, VTaMo: Video-Text Alignment Model for Sign Language Translation, Weaving Light and Time: Unified Harmonic-Geometric Representation Learning for Dense RGB-Event Parsing, StereoSplat+: Feed-Forward Stereo Gaussian Splatting with Diffusion-Assisted Progressive Inference, and Glob3R: Global Structure-from-Motion with 3D Foundation Models demonstrate how we can navigate these latent spaces to achieve high-fidelity multimodal perception and 3D reconstruction.