0:00
/
0:00
Transcript

"Learn-by-interact: A Data-Centric Framework for Self-Adaptive Agents in Realistic Environments"

Below podcast is generated with Google's Illuminate.

LLMs struggle in realistic environments due to a lack of high-quality agent data. LEARN-BY-INTERACT synthesizes this data automatically by having LLMs interact with environments and refining instructions based on interaction histories.

-----

Paper - https://arxiv.org/abs/2501.10893

Original Problem 🤔:

→ Existing LLMs perform poorly in complex, real-world tasks.

→ Annotating agent data for training is expensive and challenging.

→ Current methods for adapting LLMs to new environments are limited.

-----

Solution in this Paper 💡:

→ LEARN-BY-INTERACT creates synthetic training data through LLM interaction with environments.

→ It generates instructions using self-instruct, guided by documentations.

→ LLMs attempt to complete tasks, generating interaction trajectories.

→ Backward construction creates refined instructions by summarizing or abstracting sub-trajectories, fixing misalignments between initial instructions and LLM-generated trajectories.

→ Agentic retrieval methods optimize data usage in both training and In-Context Learning (ICL).

-----

Key Insights from this Paper 😲:

→ Backward construction is crucial for generating high-quality training data.

→ Agentic retrieval improves performance and efficiency in ICL

→ Synthesized data improves LLM performance across diverse environments.

-----

Results 💯:

→ Improves baseline results by up to 12.2% for ICL with Claude-3.5 and 19.5% for training with Codestral-22B.

→ Backward construction provides up to 14.0% improvement in training.

→ Agentic retrieval outperforms conventional retrieval methods.

Discussion about this video