Daily AI Research Pulse: 2025-11-28

Top 10 Trending AI Papers

Here are the most discussed AI papers from the last 24-48 hours, categorized by their primary contribution.

I. AI Agents

1. Kosmos: An AI Scientist for Autonomous Discovery

• Publication Date: Nov 04, 2025 (Revised Nov 2025)

• Problem Solved: Existing "AI Scientist" agents lose coherence after a few steps and cannot conduct deep, long-horizon research (e.g., executing code, reading literature, and hypothesizing simultaneously).

• Why it Solves it: Kosmos uses a "World Model" to maintain coherence over 200+ steps. It cycles through parallel data analysis, literature search, and hypothesis generation, scaling findings linearly with compute time.

• Key Takeaways:

  1. Automates the entire scientific discovery loop (data -> hypothesis -> report).

  2. Read 1,500 papers and wrote 42,000 lines of code in a single run.

  3. Independent scientists verified 79.4% of its statements as accurate.

  4. Made 7 novel discoveries in materials science and metabolomics.

  5. Demonstrates "linear scaling" of scientific output with agent cycles.

• Source: arXiv:2511.02824

2. The Landscape of Agentic Reinforcement Learning

• Publication Date: Nov 08, 2025 (Revised)

• Problem Solved: The field of "Agentic RL" is fragmented, with confusion between simple LLM prompting and true reinforcement learning for autonomous decision-making.

• Why it Solves it: Provides a unifying taxonomy that distinguishes "LLM RL" (optimizing the model) from "Agentic RL" (optimizing the agent's interaction with a dynamic world).

• Key Takeaways:

  1. Synthesizes over 500 recent papers into a single framework.

  2. Identifies key gaps in agent self-improvement and perception.

  3. Proposes a new definition for "Agentic Capabilities" (Planning, Tool Use, Memory).

  4. Highlights the shift from "passive sequence generators" to "active decision makers."

  5. Essential reading for defining the roadmap of autonomous agents.

• Source: arXiv:2509.02547

3. How Do AI Agents Do Human Work?

• Publication Date: Nov 06, 2025 (Revised)

• Problem Solved: We don't know how agents differ from humans when executing white-collar work. Do they think like us, or do they "cheat" via brute force?

• Why it Solves it: Uses a new toolkit to induce "structured workflows" from agent activity. It compares human vs. agent execution traces across design, data analysis, and coding.

• Key Takeaways:

  1. Agents are 88.3% faster and ~95% cheaper than humans.

  2. Agents take an "overwhelmingly programmatic" approach, even for visual design tasks.

  3. Humans rely on intuition; agents rely on code execution loops.

  4. Identifies specific "human-agent alignment" gaps in open-ended tasks.

  5. Suggests a future of "delegating programmable sub-tasks" rather than full job replacement.

• Source: arXiv:2510.22780

4. Browser-Use: Browser Automation Benchmark

• Publication Date: Trending Nov 28, 2025 (Reddit/GitHub)

• Problem Solved: Existing browser agents are brittle, slow, and expensive to run, often failing on complex modern websites.

• Why it Solves it: A new library/benchmark "Browser-Use" that wraps agents in a high-performance environment, achieving 100% success on tasks where previous agents had 30%.

• Key Takeaways:

  1. Achieved 100% success rate on benchmark tasks (vs 30% baseline).

  2. Reduced token costs by 65% by optimizing the context sent to the agent.

  3. 82% fewer steps required to complete tasks.

  4. Viral GitHub repository trending on Reddit r/MachineLearning.

  5. Allows wrapping any LangChain agent in ~10 lines of code.

• Source: Reddit Discussion

---

II. AI Foundation / Large Models

5. Microsoft Research Asia: Spatial & Embodied Foundation Models

• Publication Date: Nov 27, 2025

• Problem Solved: Current foundation models are "disembodied"—they understand text/images but lack 3D spatial awareness and physics required for robotics.

• Why it Solves it: Defines a new class of "Spatial AI" models that integrate 3D visual understanding directly into the LLM, enabling "synergy between AI and the brain."

• Key Takeaways:

  1. Part of the "StarTrack Scholars 2026" program.

  2. Focuses on "Embodied Foundation Models" that seamlessly integrate vision, language, and action.

  3. Addresses data scarcity in 3D/spatial domains.

  4. Aims to "redefine robot intelligence" by 2026.

  5. Establishes a roadmap for "World Simulators" beyond video generation.

• Source: Microsoft Research

6. DeepSeek R1 & Global AI Innovation

• Publication Date: Nov 2025 (Policy Brief)

• Problem Solved: The dominance of closed Western models vs. the rise of efficient open-weights models from Asia.

• Why it Solves it: Analyzes "DeepSeek R1," an MIT-licensed model geared toward advanced reasoning that rivals ChatGPT, showing how open-weight models are closing the gap despite hardware export controls.

• Key Takeaways:

  1. DeepSeek R1 achieves parity with ChatGPT in reasoning tasks.

  2. Demonstrates high performance despite "stringent export controls" on chips.

  3. Changes the global landscape of "Foundational Model Sovereignty."

  4. Highlights the "high-end NVIDIA cluster" strategy used by Chinese labs.

  5. MIT license allows for massive downstream innovation.

• Source: Observer Research Foundation

7. AdaLite: Efficient Knowledge Distillation

• Publication Date: Nov 26, 2025

• Problem Solved: Running depth estimation and vision models on edge devices (like Raspberry Pi) is too slow using standard Transformers.

• Why it Solves it: "AdaLite" uses a dual-supervision distillation scheme to train a compact student model that learns from a large teacher without quantization.

• Key Takeaways:

  1. 94% reduction in model size.

  2. 11x faster inference on CPU devices (Raspberry Pi).

  3. Preserves 96.8% of the teacher model's accuracy.

  4. Does not rely on pruning or quantization (which degrade quality).

  5. Enables "foundation model" capabilities on cheap hardware.

• Source: ResearchGate/MDPI

8. Automated Hierarchy Restructuring with LLMs

• Publication Date: Nov 26, 2025

• Problem Solved: Knowledge graphs and hierarchies often have suboptimal structures (imbalanced, too deep) that make them hard to use for RAG or embedding.

• Why it Solves it: Proposes a prompt-based approach where LLMs automatically analyze and "refactor" knowledge hierarchies to optimize them for hyperbolic embeddings.

• Key Takeaways:

  1. LLMs can act as "Knowledge Engineers" to fix ontology structures.

  2. Optimizes "branching factor" and "inheritance" for better embeddings.

  3. Robust to imbalance in the original data.

  4. Improves downstream performance in knowledge retrieval tasks.

  5. Bridge between classical Knowledge Representation and modern LLMs.

• Source: arXiv:2511.21444

9. AssurAI: Korean Multimodal Safety Dataset

• Publication Date: Nov 26, 2025

• Problem Solved: AI Safety benchmarks are too English-centric and miss "cultural hallucinations" or specific risks relevant to non-Western contexts.

• Why it Solves it: Introduces "AssurAI," a quality-controlled multimodal dataset specifically for evaluating Generative AI safety in the Korean socio-cultural context.

• Key Takeaways:

  1. Defines 35 distinct AI risk factors (universal + cultural).

  2. Addresses the "non-English safety gap" in foundation models.

  3. Evaluates multimodal (Image+Text) safety, not just text.

  4. Crucial for deploying global models in specific regions.

  5. Highlights how "safe" US models can be "unsafe" elsewhere.

• Source: arXiv:cs.AI/new

10. Dual Structure-Aware Image Filtering (DSAIF)

• Publication Date: Nov 28, 2025

• Problem Solved: Semi-supervised learning for medical images fails because it treats medical scans like random internet images, ignoring anatomical structure.

• Why it Solves it: Uses "Dual Structure-Aware" filtering that forces the model to respect the prior biological structure of organs during the unsupervised training phase.

• Key Takeaways:

  1. New State-of-the-Art (SOTA) for medical image segmentation.

  2. Leverages unlabeled data more effectively than previous methods.

  3. Crucial for "Healthcare Foundation Models" where labels are expensive.

  4. Integrates structural priors into the learning loop.

  5. Validates on complex datasets (likely organ segmentation).

• Source: Hugging Face Daily