Today for AI
HOT RADAR
agentHEAT 19.7°

AI CURATED HOT STORY · 10/20/2026

IBM ALTK-Evolve: Learned Memory Tops Tau2-Bench Without Frontier Models

1 reports archived1 independent sourcesupdated 10/20/2026, 23:19:48
Synthesis & Latest Updates
1 Sources Cross-Validated

IBM Research introduces ALTK-Evolve, a framework that uses 'learned memory'—prompt-level guidelines automatically extracted from agent runs—to enable stock models to top the τ²-bench leaderboard. This approach matches or beats fine-tuned systems in retail tasks without human annotation or test-set leakage.

LATEST/IBM Research introduces ALTK-Evolve, a framework that uses 'learned memory'—prompt-level guidelines automatically extracted from agent runs—to enable stock models to top the τ²-bench leaderboard. This approach matches or beats fine-tuned systems in retail tasks without human annotation or test-set leakage.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. Hugging Face BlogT1·78 pts
    • ALTK-Evolve builds 'learned memory' using prompt-level guidelines automatically extracted from the agent's own execution traces.
    • The mechanism enables stock models to match or outperform fine-tuned systems on the τ²-bench customer service benchmark.
    • The process requires no human annotation and strictly avoids test-set leakage, indicating high generalization potential.