ArXiV ML/AI/CV papers summary

Theme 1: Mechanistic Interpretability and Verifiable Reasoning

We are witnessing a paradigm shift from treating neural networks as opaque “black boxes” to treating them as complex, auditable circuits. The field is moving toward “mechanistic oversight,” where we demand that a model’s reasoning is not just plausible, but causally linked to its output.

• Mechanistic Insights: Researchers are using causal interventions to map internal “circuits.” Hidden not Deleted: How Networks Suppress Entangled Features reveals that unlearning often merely suppresses features rather than deleting them, while Judge Circuits Explain Format-Induced Inconsistency in LLM-as-a-Judge uses edge attribution patching to show how judgment logic is decoupled from output formatting.
• Faithfulness and Auditing: To ensure models aren’t “silently bypassing” reasoning, From Concept Alignment to Causal Grounding: An Intervention Test of Chain-of-Thought Faithfulness and From Decorative to Load-Bearing: Task Difficulty Shapes the Causal Role of Chain-of-Thought provide frameworks to verify if reasoning actually drives predictions. Similarly, Discover, Falsify, Revise: Auditing Input-Use Claims from Source Code to Predictive Contribution in Agent-Discovered Cell Models and What Makes a Terminal-Bench Task Hard? Separating Genuine Hardness from Fake-Hardness on an Adjudicated Agentic Corpus offer rigorous screens to validate agentic claims and benchmark difficulty.
• Formal Verification: Moving beyond natural language, Faithful Autoformalization via Roundtrip Verification and Repair and FormalTCS: Benchmarking End-to-End Frontier Formal Theoretical Computer Science Research of Large Language Models emphasize that in high-stakes logic and mathematics, we require formal proof rather than mere “plausible” text.

Theme 2: Agentic Governance and Long-Horizon Reliability

As agents transition from simple chatbots to autonomous actors, the focus has shifted to “agentic sovereignty”—the ability of a system to maintain independent, accurate judgment over long horizons without succumbing to social pressures or early-stage errors.

• Planning and State: Clarification Is Not Correction: LLMs Fail to Let Go warns of “early posterior collapse,” where agents commit to incorrect initial interpretations. To counter this, X-Planner: Event-Structured Task Planning for Embodied Intelligence and The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks propose explicit, event-structured planning and “taste” metrics to guide long-term decision-making.
• Social and Multi-Agent Dynamics: Recovering Agentic Sovereignty: Mitigating the Consensus Paradox via Contrastive Epistemic Decoding and Unanimity Without Persuasion: A Single Round of Debate Erases the Disagreement That Verification Needs highlight the dangers of groupthink in multi-agent systems, while The Bystander Effect in Multi-Agent Reasoning: Quantifying Cognitive Loafing in Collaborative Interactions quantifies how collaboration can paradoxically degrade performance.
• Governance: ActGov: Governing LLM Agent Actions via Policy-Constrained Validation and LabourCrew: A Multi-Agent RAG Framework for Trustworthy Adversarial Deliberation and Statutory Reasoning over Labour Law demonstrate how to enforce policy constraints at runtime using SMT-based checkers.

Theme 3: Physics-Informed and Embodied AI

The physical world serves as the ultimate “ground truth” for AI. This theme explores the integration of physical laws into machine learning to create models that are stable, geometrically aware, and capable of real-world interaction.

• Physics-Grounded Architectures: PhyMo: A Physical-Field Modality for Multimodal AI4Physics and TinyUDE: Solver-Free Universal Differential Equations on Microcontrollers via Lie-Taylor Jet Matching integrate PDE-associated operators and derivative matching to respect conservation laws. Stability is further enforced in Data-driven discrete-time deep recurrent neural network-based modeling for dissipative systems and KATOsuper: Surrogate-accelerated neural topology optimization with sensitivity-consistent Fourier neural operators.
• Visual Geometry and Robotics: VGM-VS: Rethinking Visual Geometry Model for High-Precision Visual Servoing and DAVIO: Dense Monocular-Inertial SLAM with Feed-Forward Initialization and Pose-Conditioned Mapping bridge the “geometry gap” in vision. For manipulation, Task-Prototype Guided Flow Matching for Few-Shot Generalization in Vision-Language Robot Manipulation and MemBodied: Recurrent Associative Memory for Vision-Language-Action Models enable robots to learn from few examples while maintaining history-dependent state.

Theme 4: Memory, Efficiency, and System-Level Optimization

To scale agents effectively, we must move beyond “more compute” toward structural efficiency and intelligent memory management.

• Memory Systems: ChipMEM: Verification-Grounded Memory for EDA Agents, Learning from Failures: Heterogeneous Graph Memory for Small Language Model Tool-Using Agents, and EnSIMem: Entity-Structured Indexing for Long-Term Agent Memory demonstrate that structured, retrievable memory outperforms reliance on internal weights. Security concerns are addressed in A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle.
• Architectural Efficiency: KITE: KV-Invariant Transformer Expansion for Efficient Agentic LLM Scaling and Crossflow: Prefill-Decode Elasticity for Agentic LLM Serving optimize the serving stack, while GTR: Gated Token Recurrence for Efficient Dense Prediction and AdaGScale: Viewpoint-Adaptive Gaussian Scaling in 3D Gaussian Splatting to Reduce Gaussian-Tile Pairs reduce the quadratic costs of high-resolution tasks.

Theme 5: Trustworthy Deployment and Domain-Specific Intelligence

The final frontier is the deployment of AI in high-stakes domains like medicine and law, where the cost of error is absolute and the need for provenance is critical.

• Robustness and Calibration: LWCal: Loss-Weighted Calibration for Tabular Classifiers with Noisy Calibration Labels and CS-WCP: Robust Conformal Sets for LLM-Judge Traffic Shifts with Uncertain Group Proportions ensure models remain calibrated under distribution shift. Security is bolstered by Backdoors Leave Structural Traces: FedMAST for Backdoor Detection and Containment in Federated Learning.
• Domain-Specific Applications: EndoCogniAgent: Closed-Loop Agentic Reasoning with Self-Consistency Validation for Endoscopic Diagnosis, Rachel: A general-purpose language model directs and revises retrosynthetic routes, and CORTEX: A Structured Reasoning Benchmark for Trustworthy 3D Chest CT MLLMs showcase how agentic frameworks can provide traceable, evidence-based reasoning in specialized fields.
• Forgery Detection: ASAP: Visual Analytics for Identifying and Analyzing Image Patterns in AI-generated Images and UniShield: An Adaptive Multi-Agent Framework for Unified Forgery Image Detection and Localization provide the necessary tools to verify authenticity in an era of increasingly sophisticated AI-generated content.