ArXiV ML/AI/CV papers summary

Theme 1: Agentic Reasoning and Self-Evolution

The frontier of AI has shifted from passive text generation to “agentic” workflows—systems that actively plan, verify, and improve their own reasoning. We are moving toward autonomous agents capable of self-correction and iterative refinement.

• Self-Evolution & Co-Evolution: Models are increasingly trained to improve through iterative cycles. Questioning the Questions: Sustaining Self-Evolution in Reasoning Models addresses “performance collapse” in self-training, while DUET: Co-Evolving Solver and Grader Agents and Trinity: Self-Evolving Vision-Language Models with a Self-Verifier propose frameworks where agents co-evolve with their own verifiers to curate high-quality training data. To prevent “co-cheating” (where agents converge on shared errors), False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents introduces “CrossFit” to ensure feedback remains objective.
• Test-Time Adaptation & Scaling: Agents are learning to adapt during inference. ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience and Dynamic Harness Search: Building Multi-Agent Systems Per-Query via Prediction demonstrate how agents can optimize their own “harness” (roles, tools, instructions) per query. Furthermore, Towards Better Exploration in Sequential Test-Time Scaling and Divide-and-Conquer CoT: RL for Reducing Latency via Parallel Reasoning explore how to scale compute at inference time to improve reasoning depth and reduce latency.
• Agentic Autonomy: Systems like AgentDiscover: Autonomous Discovery with Minimal Search Scaffolding and EvoCast: Reliable Autonomous Research Agents for Iterative Forecasting Architecture Evolution allow AI to act as the researcher, planning experiments and iterating on its own architecture. GitSwarm: Decentralized Compounding Inference introduces “compounding inference,” where agent work persists and builds upon itself, mirroring human research teams.

Theme 2: Reliability, Verification, and Safety

As agents gain the ability to act in the world, the focus has shifted from simple output filtering to deep behavioral oversight and formal verification.

• Formal Verification & Auditing: From Transformers to Weighted Automata: Towards the Verification of Large Language Models and FORGE: Verification-Gated Behavioral Repair for Generative Language Models provide mathematical foundations for rigorous behavioral repair. Tools like Correct Verdicts, Flawed Reasoning: Structured Auditing of LLM-based Vulnerability Reasoning and Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate’s Log-Probabilities allow us to audit the process of reasoning, not just the final output.
• Safety & Jailbreak Defense: Don’t Judge an LLM Only by Its Activations: Discovering Suppressed Safety Features via Counterfactual Activation Potential and Reactivating Alignment: Defending LLMs from Jailbreaks via Intention-Aware Input-Output Matching reveal that jailbreaks often work by suppressing safety features, providing new metrics to identify and block these attacks. Reflections and Fragments: Securing LLMs Against Sequential Mosaic Attacks develops a theory of “mosaic defense” against multi-turn attacks.
• Machine Unlearning: UnAct: Gradient-Free Unlearning via Targeted Activation Intervention and No Concept Escapes the Audit: Auditing-Aware Unlearning for Verifiable Concept Erasure in Diffusion Models provide efficient, verifiable ways to remove sensitive data or concepts from models without full retraining. Lethe: Principled Dual-Stream Update for Persistent Knowledge Erasure in Federated Unlearning addresses “knowledge resurfacing” as a failure mode.

Theme 3: Mechanistic Interpretability and Representation Geometry

We are moving beyond “black-box” evaluations toward understanding the geometric structure of model internal states, treating concepts like “correctness” or “bias” as recoverable geometric directions.

• Circuit & Feature Discovery: Slaying the Hydra: Interaction-Aware Circuit Discovery in Language Models and ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection map the internal “circuits” that drive behavior.
• Geometric Steering: Correctness Is a Direction: Geometric Answer Selection in Language Models and Extracting Persona Subspaces Through Iterative Nullspace Projection For Modulation show that we can steer model behavior by intervening in specific latent subspaces. The Answer-Basin Representation Hypothesis: We Are Not Probing or Steering Concepts challenges current assumptions, suggesting that linear directions are organized by the model’s “answer measure.”
• Supply Chain Security: GrayShield: Bit-Level Sanitization for Transformer Model Supply-Chain Security and Backdooring Sparse Autoencoders warn that even our interpretability tools can be weaponized, necessitating a new discipline of AI supply-chain security.

Theme 4: Scientific Machine Learning and Physics-Informed Models

Scientific AI is shifting toward “Neural Operators” and physics-informed surrogates that generalize across parametric regimes, moving from predicting tokens to predicting physical states.

• Physics-Informed Learning: Poisson-GENERIC Neural Operators: Exact Metriplectic Structure in Function Space via Casimir Entropies and Physics-informed neural operator for parametric phase-field modelling of interfacial degradation and microstructural evolution embed physical laws directly into the model architecture. Generative World Models Enable Predictive Control of Laser Melt Pool Dynamics and Xaurora: Generative Weather Forecasting with Denoising Stochastic Interpolants from a Foundation Model Prior demonstrate AI acting as a “digital twin” for complex physical systems.
• Efficiency in Scientific AI: FTD-GNO: Memory-Efficient Graph Neural Operators through Functional Tensor Decomposition of the Kernel and Green-Routed Neural Operators: Physics Determines Where the Network Reads optimize the computational cost of solving PDEs.
• Inverse Problems: GeoFunFlow: Geometric function flow matching for joint probabilistic inference of physical fields and complex geometries and Inverse Cross-spectral Neural Networks for Multivariate Time Series demonstrate that embedding physical constraints allows for stable, data-efficient solutions to inverse problems.

Theme 5: Efficiency, Optimization, and Architecture

As models scale, the focus has turned to compute-optimal training, hardware-aware sparsity, and efficient inference.

• Advanced Optimizers: Clean: Second-order LLM Training at Linear Memory Cost via Nystr"om Sketching and Sven: Singular Value Descent as a Computationally Efficient Natural Gradient Method offer robust alternatives to AdamW. A Solvable Model of Adaptive Learning Rate Rescaling: Acceleration, Stability & Scaling provides a theoretical framework for “edge-of-stability” training regimes.
• Quantization & Sparsity: StagQ: Constraint-Driven Multi-Precision Weight Quantization for LLMs and PATCH: Learnable Tile-level Hybrid Sparsity for LLMs enable flexible, hardware-aware model compression. Dissecting Quantization Error: A Concentration-Alignment Perspective introduces the “Concentration-Alignment Transform” (CAT) to improve quantization performance.
• Inference Acceleration: ManifoldCache: Training-Free Diffusion Acceleration via Constraint Manifold Caching and SharpDraft: Accelerating Long-Context Speculative Decoding with Cardinality-Aware Query Scaling represent the cutting edge of making generative models run in real-time.

Theme 6: Embodied Intelligence and World Modeling

The “physical turn” in AI focuses on teaching machines to understand the laws of physics and spatial reasoning, moving beyond static image-text matching.

• Embodied Agents: ForeAct3D: Policy-Grounded Future World Modeling for VLA Policies and EvoMem-VLA: State-Evolution Memory for Long-Horizon Robot Manipulation allow robots to anticipate the consequences of their actions. RocketAgent: A Long-Horizon Engineering Agent for Multidisciplinary Design of Liquid-Rocket Thrust Chambers showcases agents coordinating heterogeneous engineering tools.
• Spatial & 4D Reasoning: Kepler4D: Controllable Future Video Generation via 4D Scene State Evolution and 3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models enable models to perform mental rotation and 3D reconstruction. Mobile-4DGS: Unified Static-Dynamic Real-time Mobile Gaussian Splatting enables high-fidelity rendering of dynamic environments.
• Memory for Embodied Tasks: MemTrace: State-Consistent Memory for Long-Horizon Coding Agents and Spatially Grounded Conversational Memory for Complex Queries in Egocentric Assistants ensure that memory is anchored to physical objects and execution history, preventing agents from acting on stale information.