NEW AI In-Context Reinforcement Learning for Agentic Tools (ICRL)

2026-03-12 · Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Large Language Models, Vision-Language Models · Depth: Expert, quick

Summary

Researchers from the National University of Singapore, Salesforce AI Research, UC Berkeley, and UC Santa Cruz have introduced In-Context Reinforcement Learning (ICRL), a novel method designed to enhance tool use in Large Language Models (LLMs) and Vision-Language Models (VLMs). ICRL integrates reinforcement learning principles directly into the in-context learning framework, allowing models to learn and adapt tool-use strategies without requiring explicit fine-tuning or gradient updates. This approach aims to improve the models' ability to select and apply external tools effectively, addressing limitations in current methods that often struggle with complex, multi-step tool interactions. The technique leverages demonstrations to guide the model's decision-making process for tool invocation.

Key takeaway

For research scientists developing advanced LLM and VLM applications, ICRL offers a promising avenue to improve tool integration without the overhead of traditional fine-tuning. You should explore ICRL for scenarios requiring dynamic tool selection and complex, multi-step interactions, as it could significantly enhance model adaptability and performance in real-world tasks. Consider its potential for agents needing to learn from demonstrations.

Key insights

ICRL enables LLMs and VLMs to learn tool use in-context via reinforcement learning, bypassing fine-tuning.

Principles

• Reinforcement learning can be integrated into in-context learning.
• Tool use can be improved without gradient updates.

Method

ICRL uses in-context demonstrations to guide LLMs/VLMs in selecting and applying external tools, adapting strategies through reinforcement learning principles.

In practice

• Enhance LLM/VLM tool selection.
• Improve multi-step tool interaction.

Topics

• In-Context Reinforcement Learning
• Tool Use
• Large Language Models
• Vision-Language Models
• Reinforcement Learning

Best for: Research Scientist, AI Researcher, AI Scientist, Deep Learning Engineer

Related on AIssential

• Paper Digest: ICLR 2026 Papers & Highlights — The ICLR 2026 highlights reveal a strong focus on enhancing LLM reasoning, multi-modal capabilities, and robust AI agent development.
• Google's Warning: ICL Context is Inert — LLMs encode complex topologies internally but cannot functionally utilize these in-context learned representations for multi-step reasoning.
• Are Tools All We Need? Unveiling the Tool-Use Tax in LLM Agents — Tool-use in LLM agents incurs a "tool-use tax" that can negate performance gains, especially with semantic noise.
• DeepTool: Scaling Interleaved Deliberation in Tool-Integrated Reasoning via Process-Supervised Reinforcement Learning — DeepTool uses process-supervised reinforcement learning and interleaved deliberation to significantly improve LLM tool-integrated reasoning performance and efficiency.
• Why Memory Pipelines Fail & ICL works for AI Agents — Complex memory pipelines often hinder, rather than enhance, continuous learning in AI agents; ICL proves more effective.
• Does RL Expand the Capability Boundary of LLM Agents? A PASS@(k,T) Analysis — RL expands LLM agent capabilities for compositional tool use, unlike its role in static reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.