Top AI Papers: Week of October 26, 2026
This week's papers highlight advancements in AI agents, particularly their ability to interact with complex environments and improve through self-correction. We also see progress in multimodal models, integrating diverse data for richer understanding and generation, and new tools for robust and efficient AI deployment.
10 picks from every AI/ML paper we summarized this week, each with the reason it made the cut. Selected and annotated by ArXiv TLDR with AI assistance.
๐ฌ Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week โ summarized, scored, and delivered to your inbox every Monday.
- #1cs.RO
Scaling Online Exploration for Robot Agents
Why it made the list: RAX is an open-source framework that enables robot agents to learn and improve through online exploration and code synthesis, a practical step towards autonomous robot learning.
TL;DR: RAX is an open-source framework enabling robot agents to learn and improve through online exploration and code synthesis, without human demonstrations.
- RAX agents generate robot control code and synthesize executable programs from online experience.
- Outperforms baselines by +37.4% in simulation and learns complex real-world tasks autonomously.
- Supports diverse robot embodiments (single-arm, bimanual, humanoid) using only their APIs.
- #2cs.RO, cs.CV
RoboDoctor: Training-Free Recursive Self-Improvement for Robot Policies
Why it made the list: RoboDoctor is a training-free framework that improves robot policies through recursive self-improvement via diagnosis, grounded action, and persistent experience, offering a new paradigm for robot learning.
TL;DR: RoboDoctor is a training-free framework that improves robot policies through recursive self-improvement via diagnosis, grounded action, and persistent experience.
- Connects diagnosis, grounded action, and persistent experience for robot policy adaptation.
- Allocates control between policy and agent, determines intervention, and refines physical methods.
- Retains linked experience records for decision reasons, methods, outcomes, and reuse conditions.
- #3stat.ME, stat.ML
Agentic Distributional Regression: Conditional Density Estimation through Agent-Written Programs
Why it made the list: Agentic Distributional Regression (ADR) uses an LLM agent to write programs for conditional density estimation, offering interpretable and performant models for complex data.
TL;DR: Agentic Distributional Regression (ADR) uses an LLM agent to write programs for conditional density estimation, offering interpretable and performant models.
- LLM agent writes short programs to estimate conditional densities from iid samples.
- Programs use a library of probability distributions, returning parametric, mixture, or factorized laws.
- Fitted conditional density is an inspectable formula with limited operations and constants.
- #4cs.CL
Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families
Why it made the list: A new continued-pretraining corpus, "coupled welfare," teaches LLMs ecological reasoning, reducing harmful decisions affecting living systems under pressure.
TL;DR: A new continued-pretraining corpus, "coupled welfare," teaches LLMs ecological reasoning, significantly reducing harmful decisions affecting living systems under pressure.
- Developed a real-data "coupled welfare" corpus for continued pre-training on human/biosphere/AI dependency.
- Applied corpus via QLoRA CPT, reducing harmful ecological decisions to near zero in 4/5 models under high pressure.
- Demonstrated robustness against adversarial use, with models declining harmful proposals even when instructed to greenwash.
- #5eess.IV, cs.CV
MedVTok: A General-Purpose Medical Visual Tokenizer
Why it made the list: MedVTok is a general-purpose medical visual tokenizer that unifies 2D images and 3D volumes from diverse modalities into a single token space, for medical AI.
TL;DR: MedVTok is a general-purpose medical visual tokenizer that unifies 2D images and 3D volumes from diverse modalities into a single token space.
- Unifies 2D images and 3D volumes from 9 diverse medical imaging modalities.
- Introduces slice differential regularization for anatomical coherence in 3D volumes.
- Uses multi-expert representation alignment to integrate modality-specific clinical knowledge.
- #6cs.LG, cs.CL, cs.CV
EmbeddingGemma 2: Unified and Efficient Multimodal Representations
Why it made the list: EmbeddingGemma 2 is an open, unified multimodal embedding model for text, images, audio, and video, setting new performance benchmarks for cross-modal understanding.
TL;DR: EmbeddingGemma 2 is an open, unified multimodal embedding model for text, images, audio, and video, setting new performance benchmarks.
- Unifies text, image, audio, and video into a single embedding space.
- Achieves state-of-the-art performance across unimodal, cross-modal, and mixed-modal benchmarks.
- Features a multi-stage curriculum and simplified per-layer embeddings for improved quality.
- #7cs.AI, cs.NE
Agentic Algorithm Design: LLMs that Propose, Formalise, and Prove Runtime Bounds for Evolutionary Algorithms in Lean
Why it made the list: Presents an LLM-powered pipeline that designs black-box optimization algorithms and proves their runtime bounds using Lean, bridging AI design with formal verification.
TL;DR: This paper presents an LLM-powered pipeline that designs black-box optimization algorithms and proves their runtime bounds using Lean.
- LLMs propose black-box optimization algorithms, filtered by Lean for correctness.
- The agent defines potential functions and proves drift bounds for runtime guarantees.
- Designs algorithms with provable runtime improvements (e.g., O(n^2) vs. e^ฮฉ(n) for CycleUncover).
- #8cs.AR, cs.SE
Automated Multi-Line Bug Repair Using Model Checker Aided LLMs: A Case Study on RISC-V Processors
Why it made the list: Introduces an automated multi-line bug repair methodology for RTL designs using LLMs guided by model checkers, improving hardware design verification.
TL;DR: This paper introduces an automated multi-line bug repair methodology for RTL designs using LLMs guided by model checkers.
- Proposes a novel methodology combining LLMs, assertions, and model checking for RTL bug repair.
- Leverages failing assertions to guide LLM patch generation for multi-line bugs.
- Formally verifies each candidate patch using a model checker (JasperGold) for correctness.
- #9cs.LG, cs.RO
Regularised RL improves sim-to-real transfer
Why it made the list: Introduces Soft MellowMax Actor-Critic (SMM-AC), a regularized RL algorithm that improves sim-to-real transfer and robot robustness, addressing a key challenge in robotics.
TL;DR: This paper introduces Soft MellowMax Actor-Critic (SMM-AC), a regularized RL algorithm that improves sim-to-real transfer and robot robustness.
- Introduces Soft MellowMax Actor-Critic (SMM-AC) for robust sim-to-real transfer.
- SMM-AC achieves higher returns and greater robustness on real robots than baselines.
- Demonstrates regularized RL's practical benefits for real-world policy deployment.
- #10cs.CV
Multimodal Generative AI Assists Comprehensive Echocardiographic Interpretation
Why it made the list: EchoChat, a multimodal generative AI, comprehensively interprets echocardiograms, emulating cardiologists to improve diagnostic accuracy and access in healthcare.
TL;DR: EchoChat, a multimodal generative AI, comprehensively interprets echocardiograms, emulating cardiologists to improve diagnostic accuracy and access.
- Introduces EchoChat, a multimodal vision-language model for comprehensive echocardiographic analysis.
- Emulates cardiologist workflow with chain-of-thought reasoning, integrating medical knowledge and visual evidence.
- Achieves state-of-the-art performance across 33 cardiac conditions, outperforming existing foundation models.
๐ฌ Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week โ summarized, scored, and delivered to your inbox every Monday.