ArXiV ML/AI/CV papers summary

We are currently witnessing a profound transition in artificial intelligence. We are moving beyond the era of “static” models—those that merely predict the next token or classify an image—and into the age of Agentic Intelligence. This shift represents a move from the “lab” to the “factory floor,” where systems are expected to reason, plan, interact with the physical world, and self-correct.

As we push these boundaries, the research community is coalescing around five critical pillars of progress.

Theme 1: Agentic Reasoning, Planning, and Self-Evolution

The frontier of AI is shifting from “prompt-and-pray” to “reason-and-revise.” Agents are no longer just executing single instructions; they are engaging in multi-turn, long-horizon workflows where they must manage their own beliefs and correct their own mistakes.

• Self-Correction & Reflection: DREAM: Forgetting is Not Always Bad: A Neuro-Inspired Memory Repair Mechanism for Poisoned LLM Agents, MENTOR: Learning from Failure with On-Policy Pedagogical Intervention, and SynAct: Synergistic Active Feedback Learning for LLM Agents demonstrate that agents can learn from failure and intelligently query human experts only when necessary.
• Long-Horizon Planning: TreeHCA: Hindsight-Guided Exploration and Credit Assignment for Multi-Turn LLM Agents, SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire, LongHarness: Advancing Agentic Reasoning over 100M-Token Corpora via Stateful Evidence Management, and Travel Back To Move Forward: Retrospective Exploration and Evolving for Long Horizon Agents address the “dead-end” problem, allowing agents to manage state and evidence to avoid cascading failures.
• Self-Improvement: EvoBelief: Falsifiable World Models for Self-Improving LLM Agents, Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution, Scale and Selection: What Makes Automatic Harness Evolution Work for Visual-Interface Robot Agents, and RSI-Anything: Unlocking Recursive Self-Improvement with World Models explore how agents can rewrite their own operational logic and update their world models to improve recursively.

Theme 2: Efficiency, Quantization, and Hardware-Aware Design

As models grow, physical constraints—memory bandwidth, energy, and latency—become the primary bottlenecks. We are rethinking numerical formats and hardware interaction to make AI sustainable.

• Quantization & Compression: Algebraic Integer Quantization for Machine Learning Applications, DualQuant: Dual-scale Calibration-free Weight-and-Activation Quantization for LLMs, TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models, and Deflating the Hessian: Rank-4 W4A4 Quantization for Multimodal Diffusion Transformers push the boundaries of low-precision training and inference.
• Architectural Efficiency: DuoRate: Two-Rate Per-Expert Precision Allocation for Mixture-of-Experts Serving, ReplaySSM: Cache SSM Inputs, Not State, Fast LapSum: Exact Differentiable Top-k at Million Scale, AudioTETRIS: Token Elimination via Transitional Importance Signals for Efficient Audio LLMs, and NeuronSpark: Selective Spiking State Space Models with Adaptive Computation for Language Modeling optimize how we store, access, and process information.

Theme 3: Embodied Intelligence and Scientific Discovery

AI is increasingly bridging the gap between digital reasoning and physical reality, whether in robotics or fundamental science.

• Embodied Interaction: HumanX: Toward Agile and Generalizable Humanoid Interaction Skills from Human Videos, Workhorse: Learning Robust Whole-Body Humanoid Loco-Manipulation from Human Data, Sampling-Based Safe Reinforcement Learning, RESCUE: Is This Alarm Worth Acting On? Counterfactual Evaluation of VLA Runtime Monitors, UniConFlow: A Unified Constrained Flow-Matching Framework for Certified Motion Planning, and VIGOR: Verifiable Industrial Geometry Optimization and Repair in Scene Programs focus on safety, robustness, and constraint-aware planning.
• Scientific Modeling: Learning Schwarzschild Geodesic Motion with Lagrangian Neural Networks, Recovering Irreducible Three-Body Interactions with Lagrangian Neural Networks, Enriched Holomorphic Neural Networks for 2D and 3D Laplace Problems, and Designing Protein Structures from Natural Language demonstrate how embedding physical laws into neural architectures allows models to “discover” the mechanics of the universe.

Theme 4: Interpretability, Trust, and Causal Auditing

In high-stakes environments, “black-box” performance is insufficient. We are moving toward verifiable reasoning and structural understanding of model internals.

• Interpretability: S-Space: Exploring Spatial Workspace in Multimodal Language Models, Playing the network backward: A Game Theoretic Attribution Framework, CORE-Seg: Reasoning-Driven Segmentation for Complex Lesions via Reinforcement Learning, VITAL: Visual-Semantic Dual Supervision for Enhanced and Interpretable Latent Reasoning in Medical MLLMs, and Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment provide tools to audit internal logic.
• Trust & Safety: AI Models Can Provably Hide Arbitrary Capabilities, PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs, SOAR: Source-Aware Reinforcement Learning with Precise Credit Assignment for Search Agents, and SEEDS: Preview-Guided Evidence Admission for Search Agents address the critical need for source attribution and the risks of hidden capabilities.

Theme 5: Evaluation, Adaptation, and Social Intelligence

As AI systems enter human environments, we must measure their reliability, economic impact, and ability to navigate social norms.

• Continual Learning: Monarch-MoE: Structured Experts for Multimodal Continual Instruction Tuning, Continually Post-training Small Models via Synthetic Self-distillation, and Specialists Learn Faster: Knowledge Organization in Pre-training Shapes Knowledge Update Efficiency address the stability-plasticity dilemma.
• Benchmarks & Governance: AgentSWE: Can Coding Agents Build the Agent You Actually Want?, Agents’ Last Exam, SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development Lifecycle, DRYLAB-Bench: Benchmarking the Biosecurity Risk of Agentic LLM Mutation Design, and API Benchmark Scores Do Not Reliably Transfer to Chatbot Interfaces highlight the need for rigorous, real-world evaluation.
• Social Intelligence: ZenGen: Social Mind for LLMs, The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs, Knowing Is Not Telling: Identity Ossification in LLM Representations of Trans-border Communities, and DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models explore the complexities of human-centric AI.