ArXiV ML/AI/CV papers summary

We are witnessing a profound transformation in the architecture of intelligence. Much like the transition from the geocentric model of the universe to the heliocentric, our field is moving away from the “black-box” era of brute-force scaling toward a more elegant, diagnostic, and physically grounded understanding of machine cognition. We are no longer merely building larger statistical engines; we are engineering autonomous, verifiable agents capable of navigating the complexities of the physical and social world.

Theme 1: Mechanistic Interpretability and Diagnostic Frameworks

The “black box” is finally being pried open. We are moving toward a “white-box” era where we audit the internal reasoning circuits of models to ensure they are not just producing correct outputs for the wrong reasons.

• Attribution and Circuit Discovery: Researchers are mapping the internal “geography” of models. IG-Lens: Exact Additive Probability Attribution Across Transformer Layers via Telescoping Integrated Gradients and Symbolic Mechanistic Data Attribution: Tracing Training Influence to Learned Behavioral Policies allow us to trace outputs back to specific training data or layers. Meanwhile, MechRL: Reinforcement Learning Agents Perform Circuit Discovery for Mechanistic Interpretability, Explaining Attention with Program Synthesis, and Emergence of Minimal Circuits for Indirect Object Identification in Attention-Only Transformers provide methods to reverse-engineer the sparse subgraphs responsible for specific behaviors.
• Latent Reasoning and Safety: We are uncovering how models represent the world internally. Invariant Reasoning Directions in Latent Trajectories of Language Models and Do Models Read What They Write? Causal Registers in Scratchpad Reasoning show that models develop stable, manipulatable reasoning structures. Furthermore, Fuzzing Large Language Models to Elicit Hidden Behaviours, MemLeak: Diagnosing Information Leaks in Multimodal Agent Memory, and Attribution Graphs and Causal Probing for Mechanistic Discovery and Bias Repair in Multimodal Generative Learning offer rigorous forensic tools to detect backdoors, data leakage, and demographic biases.

Theme 2: Embodied Intelligence and World Foundation Models

Intelligence is not merely the processing of text; it is the ability to simulate and interact with the physical world. We are moving toward “world foundation models” that understand causal dynamics.

• World Models: Orca: The World is in Your Mind and OWMDrive: Causality-Aware End-to-End Autonomous Driving via 4D Occupancy World Model treat the world as a latent space to be navigated. Walking in the Implicit: Interactive World Exploration via Neural Scene Representation allows for consistent, long-horizon exploration by rendering neural scenes.
• Embodied Interaction: NormAct: A Benchmark for Hidden Social Norm Compliance in Embodied Planning reminds us that social constraints are as vital as physical ones. Robotic manipulation is being revolutionized by SceneBot: Contact-Prompted General Humanoid Whole Body Tracking with Scene-Interaction, S$^2$-VLA: State-Space Guided Vision-Language-Action Models for Long-Horizon Manipulation, and Training Vision-Language-Action Models with Dense Embodied Chain-of-Thought Supervision, which ground models in tactile dynamics and cross-embodiment reasoning.
• Spatial Intelligence: SpatialUAV: Benchmarking Spatial Intelligence for Low-Altitude UAV Perception, Collaboration, and Motion, HAT-4D: Lifting Monocular Video for 4D Multi-Object Interactions via Human-Agent Collaboration, and Towards Spatial Trace with Reasoning in Vision-Language Models for Robotics push the boundaries of metric-grounded reasoning in 3D space.

Theme 3: Agentic Reasoning and Reliability

As models transition into autonomous agents, we are shifting from simple accuracy metrics to “closed-loop” diagnostic frameworks that evaluate agents in complex, multi-step environments.

• Agentic Evaluation: CLQT: A Closed-Loop, Cost-Aware, Strategy-Consistent Benchmark for Diagnostic Evaluation of LLM Portfolio-Management Agents, MemDelta: Controlled Baselines and Hidden Confounds in Agent Memory Evaluation, and EvalSafetyGap: A Hybrid Survey and Conceptual Framework for LLM Evaluation-Safety Failures address the “measurement problem” in AI safety.
• Reasoning Horizons: Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent and The Complexity Ceiling Benchmark: A Multi-Domain Evaluation of Sequential Reasoning Under Depth Scaling highlight that depth of reasoning is the new frontier. Self-Evolving World Models for LLM Agent Planning and Mixture of Debaters: Learn to Debate at Architectural Level in Multi-Agent Reasoning demonstrate how agents can improve through self-evolution and internal consensus.

Theme 4: Physics-Informed and Scientific AI

By embedding physical laws into neural architectures, we ensure that AI remains a reliable partner in scientific discovery rather than a source of “hallucinated” results.

• Scientific Simulation: A Trainable-by-Parts Operator Learning Framework: Bridging DeepONet and Karhunen-Loeve Expansions for Large-Scale Applications, MALOQ: Massively Accelerated Learning of Operators for Quantum Transport, and Implementation of Hyperelastic Physics-Augmented Neural Networks in the Explicit Finite Element Codes Simcenter Radioss and OpenRadioss with Applications to Impact Events demonstrate the power of physics-constrained learning.
• Verifiable Discovery: Exploring the Cryptographic Limits of Transformer Networks, A Machine-Verified Proof of a Quantum-Optimization Conjecture, and Verifiable Geometry Problem Solving: Solver-Driven Autoformalization and Theorem Proposing show AI acting as a partner in formal mathematics and physics.

Theme 5: Efficiency, Scaling, and Inference Optimization

We are moving beyond the “more compute” mantra toward resource-aware, frugal intelligence that allocates computation only where it is truly needed.

• Dynamic Inference: HARD-KV: Head-Adaptive Regularization for Decoding-time KV Compression, Entropy-Gated Latent Recursion, and End-to-End Dynamic Sparsity for Resource-Adaptive LLM Inference allow models to adapt their computational footprint in real-time.
• Memory Management: RedKnot: Efficient Long-Context LLM Serving with Head-Aware KV Reuse and SegPagedAttention and Mandol: An Agglomerative Agent Memory System for Long-Term Conversations treat memory as a structured, dynamic object, while The Speedup Paradox: Rethinking Inference Speed-Quality Trade-off in Embodied Tasks provides a framework for balancing latency and performance in dynamic systems.

Theme 6: Safety, Governance, and Trustworthiness

As agents gain the ability to interact with the web and use tools, we must govern the entire ecosystem, not just the model.

• Agentic Security: Just Ask: Curious Code Agents Reveal System Prompts in Frontier LLMs, AI Snitches Get Glitches: Towards Evading Agentic Surveillance, and Agent-Native Immune System: Architecture, Taxonomy, and Engineering propose radical new defenses for the agentic loop.
• Forensics: VIGIL: Part-Grounded Structured Reasoning for Generalizable Deepfake Detection and Neural Gate: Mitigating Privacy Risks in LVLMs via Neuron-Level Gradient Gating provide surgical, evidence-based approaches to privacy and deepfake detection, ensuring that our digital reality remains verifiable.