✨Scaling Agents via Continual Pre-training

📝 Summary:
Current agentic LLMs underperform due to training tensions. This paper proposes Agentic Continual Pre-training CPT to build powerful agentic foundation models. Their AgentFounder model achieves state-of-the-art performance on benchmarks with strong tool-use.

🔹 Publication Date: Published on Sep 16

🔹 Paper Links:
• arXiv Page: https://arxiv.org/pdf/2502.06589
• PDF: https://arxiv.org/pdf/2509.13310
• Project Page: https://tongyi-agent.github.io/blog/
• Github: https://tongyi-agent.github.io/blog/

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLMAgents #ContinualPretraining #FoundationModels #AIResearch #ToolUse

✨In-the-Flow Agentic System Optimization for Effective Planning and Tool Use

📝 Summary:
AgentFlow is a trainable agentic framework that optimizes its planner in-the-flow within multi-turn interactions. It uses Flow-GRPO to train its modules and significantly outperforms top baselines and GPT-4o on various reasoning and tool-use tasks.

🔹 Publication Date: Published on Oct 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.05592
• PDF: https://arxiv.org/pdf/2510.05592
• Project Page: https://agentflow.stanford.edu/
• Github: https://github.com/lupantech/AgentFlow

✨ Spaces citing this paper:
• https://huggingface.co/spaces/AgentFlow/agentflow
• https://huggingface.co/spaces/bioliveir4/agentflow2
• https://huggingface.co/spaces/bioliveir4/agentflow

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#AI #MachineLearning #AIagents #ToolUse #Planning

✨LoopTool: Closing the Data-Training Loop for Robust LLM Tool Calls

📝 Summary:
LoopTool is an automated framework that closes the data-training loop for LLMs. It iteratively refines data and models to improve tool-use capabilities, achieving state-of-the-art results and surpassing larger models cost-effectively.

🔹 Publication Date: Published on Nov 12

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.09148
• PDF: https://arxiv.org/pdf/2511.09148
• Github: https://github.com/Rednote-ExperienceAI-Lab/LoopTool

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #AI #MachineLearning #DataScience #ToolUse

✨M3-Bench: Multi-Modal, Multi-Hop, Multi-Threaded Tool-Using MLLM Agent Benchmark

📝 Summary:
M3-Bench is a new benchmark evaluating multimodal LLM agent tool use in complex, multi-hop workflows requiring visual grounding and tool dependencies. It introduces a similarity-driven alignment method and interpretable metrics. Evaluations show significant gaps in current MLLMs, especially in ar...

🔹 Publication Date: Published on Nov 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.17729
• PDF: https://arxiv.org/pdf/2511.17729
• Github: https://github.com/EtaYang10th/Open-M3-Bench

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#MLLM #LLMAgents #AI #Benchmarking #ToolUse

✨Budget-Aware Tool-Use Enables Effective Agent Scaling

📝 Summary:
Tool-augmented agents struggle to scale with more tool calls due to a lack of budget awareness. This paper introduces Budget Tracker for continuous budget awareness and BATS for adaptive planning, dynamically adjusting strategy based on remaining resources. These methods significantly improve cos...

🔹 Publication Date: Published on Nov 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.17006
• PDF: https://arxiv.org/pdf/2511.17006

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#AIAgents #ToolUse #ResourceManagement #AgentScaling #AIResearch

✨Unlocking Implicit Experience: Synthesizing Tool-Use Trajectories from Text

📝 Summary:
This paper introduces GEM, a text-based pipeline to synthesize multi-turn tool-use trajectories for LLMs from text corpora. It addresses data scarcity and reduces costs with a specialized Trajectory Synthesizer. GEM-32B significantly improves performance on multi-turn benchmarks, showing strong g...

🔹 Publication Date: Published on Jan 15

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.10355
• PDF: https://arxiv.org/pdf/2601.10355

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #AI #NLP #ToolUse #DataSynthesis

✨D-CORE: Incentivizing Task Decomposition in Large Reasoning Models for Complex Tool Use

📝 Summary:
D-CORE is a two-stage training framework improving large reasoning models' task decomposition and reasoning. It overcomes Lazy Reasoning using self-distillation and diversity-aware reinforcement learning. D-CORE achieves superior tool-use performance, setting new state-of-the-art results even wit...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02160
• PDF: https://arxiv.org/pdf/2602.02160
• Github: https://github.com/alibaba/EfficientAI

🔹 Models citing this paper:
• https://huggingface.co/bowiehsu/D-CORE-8B

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #TaskDecomposition #ToolUse #ReinforcementLearning #AIResearch

✨τ-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge

📝 Summary:
τ-Knowledge extends τ-Bench to evaluate conversational agents in fintech customer support, integrating external knowledge with tool use. Its τ-Banking domain involves navigating 700 documents and executing tool-mediated updates. Frontier models achieve only ~25.5% pass, struggling with document r...

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04370
• PDF: https://arxiv.org/pdf/2603.04370

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#ConversationalAI #Fintech #LLMEvaluation #KnowledgeIntegration #ToolUse

✨EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings

📝 Summary:
EnterpriseOps-Gym is a new benchmark for evaluating LLM agents in realistic enterprise settings, featuring a complex sandbox and curated tasks. It reveals current models struggle with strategic planning and task refusal, achieving low success rates, indicating they are not ready for autonomous de...

🔹 Publication Date: Published on Mar 13

🔹 Paper Links:
• arXiv Page: https://arxiv.org/pdf/2603.13594
• PDF: https://arxiv.org/pdf/2603.13594
• Project Page: https://enterpriseops-gym.github.io/
• Github: https://enterpriseops-gym.github.io

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLMAgents #EnterpriseAI #AIResearch #Benchmarking #ToolUse