AlphaEarth Foundations (AEF) introduces a task-agnostic geospatial foundation model that learns a compact, time-continuous embedding field of Earth’s surface.
- Agentic-R1
- Context Rot
- Scaling up RL
- A Survey of AIOps
- Chain-of-Thought Monitorability
- One Token to Fool LLM-as-a-Judge
- A Survey of Context Engineering for LLMs
Read on for more:
1. One Token to Fool LLM-as-a-Judge
Investigates the surprising fragility of LLM-based reward models used in Reinforcement Learning with Verifiable Rewards (RLVR).
- H-Net
- HIRAG
- Kimi K2
- MemAgent
- Adaptive Branching MCTS
- A Survey on Latent Reasoning
- What Has a Foundation Model Found?
Read on for more:
1. Kimi K2
Moonshot AI introduces Kimi K2, a 1T parameter Mixture-of-Experts model (32B active) optimized not just for knowledge tasks but for agentic capabilities, models that act, not just respond.
- xLSTMAD
- AI4Research
- Deep Research Agents
- SLMs are the Future of Agentic AI
- Chain-of-Thought Is Not Explainability
- Survey on Evaluation of LLM-based Agents
Read on for more:
1. Small Language Models are the Future of Agentic AI
This position paper argues that small language models (SLMs), defined as those runnable on consumer-grade hardware, are not only sufficient but superior for many agentic AI applications.
Here are the top AI Papers of The Week (June 23 - 29):
- MEM1
- AlphaGenome
- Diffusion Steering via RL
- Towards AI Search Paradigm
- AI Agent Communication Protocols
- Ultra-Fast Diffusion-based Language Models
Full list below:
1. Ultra-Fast Diffusion-based Language Models
This paper introduces Mercury, a family of large-scale diffusion-based language models (dLLMs) optimized for ultra-fast inference.
Here are the top AI Papers of The Week (June 16-22):
- RAG+
- ALE-Agent
- From Bytes to Ideas
- Agentic Misalignment
- Future of Work with AI Agents
- Eliciting Reasoning with Cognitive Tools
...
1. RAG+
Introduces RAG+, a modular framework that improves traditional RAG systems by explicitly incorporating application-level reasoning into the retrieval and generation pipeline.
- Open Thoughts
- RewardBench 2
- The Illusion of Thinking
- Knowledge or Reasoning
- From Tokens to Thoughts
- Self-Challenging Language Model Agents
Read on for more:
1. The Illusion of Thinking
Investigates the capabilities & limitations of frontier Large Reasoning Models like Claude 3.7, DeepSeek-R1, and OpenAI’s o-series by analyzing their performance on reasoning tasks as a function of problem complexity.
- QwenLong-L1
- Spurios Rewards
- New Lens on RAG
- Self-Improving Agents
- Scalable Agentic Reasoning
- Learn to Reason without External Rewards
Here is the full list:
1. New Lens on RAG Systems
Introduces a new conceptual and empirical framework for analyzing RAG systems through the lens of sufficient context, whether the retrieved content alone enables answering a query.
Read on for more:
1. Does RL Incentivize Reasoning in LLMs Beyond the Base Model?
By analyzing models across tasks (math, code, vision), the authors find that RLVR improves sample efficiency but does not expand reasoning capacity beyond the base model.
Here are the top AI Papers of the Week (Mar 24-30):
- AgentRxiv
- Play2Prompt
- Chain-of-Tools
- Qwen2.5-Omni
- Tracing the Thoughts of LLMs
- Synthetic Data Generation Using LLMs
Read on for more:
1). Tracing the Thoughts of LLMs
Anthropic researchers unveil new interpretability tools for peering inside LLMs, using Claude 3.5 Haiku. Their two new papers show how to trace model internals like circuits, plans, and conceptual thinking in real time.
Here are the top AI Papers of the Week (Mar 10-16):
- Gemma 3
- Search-R1
- Gemini Robotics
- Post Training of LLMs
- Improving Planning of Agents
- Transformers without Normalization
Read on for more:
1). Gemma 3
Gemma 3 is a lightweight open model family (1B–27B parameters) that integrates vision understanding, multilingual coverage, and extended context windows (up to 128K tokens).
Many developers are still writing long prompts or post-processing LLM outputs to get consistent results.
Using structured outputs is a better option in many LLM applications.
Let's discuss a few important points:
What are structured outputs useful for?
Structured outputs enable LLMs to generate responses in a specific schema like JSON, often leveraging data validation tools like Pydantic. This ensures consistent formatting, predictable data structures and type safety.
Mar 9 • 11 tweets • 4 min read
Here are the top AI Papers of the Week (Mar 3-9):
- Agentic Reward Modeling
- Fractal Generative Models
- Cognitive Behaviors in LRMs
- Overview of Reasoning LLMs
- Conversational Speech Model
- A Few Tokens Are All You Need
Read on for more:
1). A Few Tokens Are All You Need
Proposes a new approach to boost reasoning in LLMs by only fine-tuning on the first few tokens of generated solutions.
Here are the top AI Papers of the Week (Feb 24 - Mar 2):
- GPT-4.5
- PlanGEN
- Protein LLMs
- Chain-of-Draft
- Claude 3.7 Sonnet
- Emergent Misalignment
Read on for more:
1). Claude 3.7 Sonnet
Anthropic releases a system card for its latest hybrid reasoning model, Claude 3.7 Sonnet, detailing safety measures, evaluations, and a new "extended thinking" mode.
Here are the top AI Papers of the Week (Feb 10-16):
- AI Co-Scientist
- Open-Reasoner-Zero
- The AI CUDA Engineer
- Native Sparse Attention
- The Danger of Overthinking
- Large Language Diffusion Model
Read on for more:
1). AI Co-Scientist - Google introduces AI co-scientist, a multi-agent AI system built with Gemini 2.0 to help accelerate scientific breakthroughs.
- s1
- OmniHuman-1
- Less Is More for Reasoning
- Advancing Reasoning in LLMs
- Rethinking Mixture-of-Agents
- Chain-of-Associated-Thoughts
Read on for more:
1). s1: Simple test-time scaling
Introduces s1, a method to boost LLM performance by using extra compute at inference (“test-time scaling”). A new decoding trick appends the token “Wait” when the model tries to stop, forcing it to think longer.
