Research map / Multi-agent systems and games / LLM agents

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

ICLR 2025 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Learn-by-interact, a fully autonomous pipeline that synthesizes environment-specific agent trajectories via self-instruct, LLM rollouts, and backward instruction construction, then uses the data for agentic retrieval in ICL and for LoRA fine-tuning.

Problem

LLM agents underperform on realistic digital environments (SWE-bench, WebArena, OSWorld, Spider2-V) due to a lack of high-quality environment-specific agent data, and human annotation of long agent trajectories is costly.

Approach

Generate tasks from environment documentation, collect interaction trajectories, then relabel every sub-trajectory with LLM-written summary or abstraction instructions (backward construction) to fix instruction-trajectory misalignment and multiply data; use model-based plus observation-based agentic retrieval for ICL and SFT on the filtered data.

Key results

• On OSWorld with Claude-3.5-sonnet, Learn-by-interact ICL raises success from 12.4% to 22.5%, nearly doubling the baseline.
• On WebArena, Codestral-22B improves from 4.7% to 24.2% after LoRA tuning on the synthesized data, versus 10.2% for data distillation.
• Training-free Learn-by-interact outperforms Reflexion and LATS across all four benchmarks for both Gemini-1.5-pro and Claude-3.5-sonnet (e.g., SWE-bench 60.0 vs 55.2 for Claude).

Measured effects reported in the paper

Surprising findings

• Data distillation without backward construction gives at most about 2% gain over baseline with Gemini, suggesting raw distilled trajectories are noisy.
• Short-trajectory data is more useful than medium or long trajectories for both ICL and training.

Limitations

• Requires many LLM calls for generation and committee-based filtering.
• Depends on environment-related documentation, which may be incomplete or unavailable.

Experimental setup

Datasets: SWE-bench Verified documentation (python files) WebArena documentation sources OSWorld documentation sources Spider2-V documentation sources

Benchmarks: SWE-bench Verified WebArena OSWorld Spider2-V

Metrics: resolved rate pass@1

Baselines: RAG Data distill Reflexion LATS

Builds on: self-instruct Reflexion LATS ReAct CodeAct

Closest papers by meaning

• Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.90)
• WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.89)
• How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.89)
• Self-Generated In-Context Examples Improve LLM Agents for Sequential Decision-Making Tasks (NeurIPS 2025 · similarity 0.89)
• AutoManual: Constructing Instruction Manuals by LLM Agents via Interactive Environmental Learning (NeurIPS 2024 · similarity 0.89)
• GenEnv: Difficulty-Aligned Co-Evolution Between LLM Agents and Environment Simulators (NeurIPS 2026 · similarity 0.89)