# ALTK‑Evolve pushes retail customer‑service agent to top of τ²‑bench leaderboard

> Using GPT‑4.1 and learned prompt guidelines, IBM Research’s agent hits 67.5% Pass⁴ on the retail benchmark, beating all frontier models.

Oossa · 2026-10-20 · https://oossa.com/en/altk-evolve-pushes-retail-customer-service-agent-to-top-of-2-bench-leaderboard

IBM Research released ALTK‑Evolve, a method that adds a short playbook of learned guidelines to a basic ReAct customer‑service agent that runs on GPT‑4.1. The added rules raise the agent’s Pass⁴ score on the τ²‑bench retail task to 67.5%, the highest reported in the October 12 2026 leaderboard snapshot. No model weights are changed and no human labeling is required.

## The facts

- Pass⁴ = 67.5% on τ²‑bench retail (October 12 2026 leaderboard)
- Guidelines extracted with GPT‑4.1, 292 rules in the lossless configuration

## Why it matters

Retail companies can achieve top‑ranking, reliable automated support using only prompt tweaks rather than expensive model fine‑tuning.

## Sources & references

1. [ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model](https://huggingface.co/blog/ibm-research/altk-evolve-bc) – Hugging Face, 2026-10-20

Last updated: 2026-10-20