ArXiV ML/AI/CV papers summary

Theme 1: Mechanistic Interpretability and Model Behavior

The “black box” era of AI is drawing to a close. Researchers are increasingly treating models as physical systems that can be disassembled, audited, and steered. By moving beyond simple input-output observation, we are uncovering the internal “gears” of intelligence.

• Steering and Control: As a Language Model…: Chat Template Switches LLM Self-Referential Voice and Activation Steering Reproduces It reveals that a model’s “personality” is often a configurable byproduct of its chat template rather than its core weights. Similarly, Rewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMs shows that instruction tuning acts as a “gate” for existing circuits rather than a total rewrite, allowing us to apply interpretability tools across different model versions.
• Transparency and Auditing: To make models reliable, we must see how they think. Attention as a Routing Graph: Live Circuit Extraction from a Single Forward Pass provides a “cheap sketch” of causal reasoning, while The Delegation Blind Spot: Auditing Product Decisions from Agent Choices and RankCert: When Can Simulated Learners Safely Select an AI Tutor? establish rigorous protocols to ensure that “task completion” is not mistaken for “correct decision-making.”

Theme 2: Agentic Reasoning and Governance

We are transitioning from static, single-turn question answering to long-horizon, multi-agent workflows. The challenge is no longer just generating text, but coordinating, verifying, and constraining autonomous behavior.

• Coordination and Scale: Agensh: Scaling Organizational Intelligence to 1,024 Agents proves that complex cooperation emerges as agent populations grow, while CONCAT: Consensus- and Confidence-Driven Ad Hoc Teaming for Efficient LLM-Based Multi-Agent Systems optimizes these interactions without the need for expensive retraining.
• Reliability and Safety: Safety is increasingly viewed as an external, policy-driven constraint. ActGov: Governing LLM Agent Actions via Policy-Constrained Validation and Distributed Legal Infrastructure for a Trustworthy Agentic Web propose formalizing authorization as a runtime gate. Furthermore, LLM Ghostbusters: Surgical Package Hallucination Suppression via Adaptive Unlearning and FIRE: Failure-Informed Runtime Engineering for Reliable Language-Model Agents offer surgical interventions to suppress failure modes like hallucinations post-deployment.
• Evidence-Based Evaluation: Quantifying Overclaiming Propensity in Frontier LLM Agents and ShowTellArena: Evaluating Business Workflow Understanding from Demonstrations emphasize that we must prioritize verifiable outcomes over model-generated claims.

Theme 3: Embodied Intelligence and Physical Grounding

True intelligence requires a bridge between high-level reasoning and the physical world. This theme explores how models can move beyond pixels to understand causal dynamics, tactile feedback, and spatial geometry.

• World-Action Models (WAMs): Modern robotics is shifting toward models that predict and act simultaneously. DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation and CSWAM: Better Causal Semantic Representations for Out-of-Distribution Generalization in World Action Models integrate tactile and causal data to ground actions. 4DGS-JEPA: Temporally Compositional Joint-Embedding Prediction for Dynamic Gaussian Splatting and Skytopia: Monocular Drone Navigation with Action-Conditioned Latent World Models represent the next generation of world models that learn the underlying physics of their environment.
• Tactile and Geometric Intelligence: With datasets like N0-Foundation: Towards the Age of Tactile Intelligence and ME-Dex 1.0: Bringing Heterogeneous Tactile Sensing into World Action Modeling, robots are learning to “feel.” Meanwhile, X-GS: An Extensible Framework for Perceiving and Thinking with 3D Gaussian Splatting and GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation establish 3D geometry as a first-class citizen in AI.

Theme 4: Efficient Inference and Memory Management

As models grow, the “context window” and computational cost become primary bottlenecks. The field is moving toward smarter memory architectures and hardware-aware compression.

• Dynamic Memory: DTOC: Dynamic Tool Output Compression for Adaptive Context Management in AI Agents and RPMem: Learning Long-Term Recurrent Parametric Memory Across Sessions for LLM Agents allow agents to manage their own memory, while LatentPort: Beyond KV Cache - Cross-Model Transfer of Recurrent Memory in Hybrid Language Models explores “memory handoff” between models.
• Hardware-Aware Compression: Disaggregated Quantization: Specializing LLM Prefill and Decode, FuncCode: Compressing Kolmogorov–Arnold Networks in Function Space with Hardware-Aware Quantization, and GeoPair: Geometry-Preserving Cross-Layer Factorization for Training-Free Transformer Compression demonstrate that we can achieve massive speedups by exploiting the underlying geometry of weights and functions.
• Token Efficiency: Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs and PACE: A Unified Condense-and-Extract Paradigm for Fast VLM Inference show that we can maintain high performance while drastically reducing the number of tokens processed.

Theme 5: Scientific Discovery and Domain-Specific AI

AI is increasingly applied to high-stakes scientific domains where the cost of error is high. These models are moving from simple retrieval to active scientific inquiry and evidence-linked reasoning.

• Scientific Orchestration: Rachel: A general-purpose language model directs and revises retrosynthetic routes and Multi-Agent Design Assistant for the Simulation of Inertial Fusion Energy demonstrate LLMs acting as scientific orchestrators. ScholarStack: Layered Research Asset Orchestration and Cross-Task Reuse for Scientific Agents emphasizes the need for “cumulative” AI that builds upon previous research.
• Specialized Foundation Models: In fields like medicine and engineering, general-purpose models are being replaced by specialized architectures. ORION-CMR: On-scanner Reporting with Integrated Foundation Model for End-to-End Cardiac MRI Analysis and Interpretation and SurgMotion: A Video-Native Foundation Model for Universal Understanding of Surgical Videos show that foundation models can handle entire pipelines—from raw data to clinical insight—with expert-level precision.
• Causal and Counterfactual Reasoning: CLARITY: Medical World Model for Guiding Treatment Decisions by Simulating Context-Aware Disease Trajectories and Generating Chest X-Ray Counterfactuals by Specialising Foundation Image Models move beyond prediction to “what-if” analysis, allowing for safer, evidence-based decision-making.