IBM Research says a customer-service agent improved its τ²-bench retail result after learning guidelines from its own training runs. The result reached 67.5% Pass^4 with a stock GPT-4.1 agent, placing it above every retail entry in the team’s October 12, 2026 leaderboard snapshot.
The key idea is simple: instead of writing every instruction by hand, the system learned guidelines automatically and added them to the agent’s system prompt. That change lifted the GPT-4.1 customer-service agent to a 67.5% Pass^4 rate on τ²-bench’s retail domain.
Training Runs Became a Source of New Instructions
IBM Research described the learned guidelines as rules drawn from the agent’s own training runs. Once retained, those rules were appended to the system prompt used by the customer-service agent. The approach gave the stock GPT-4.1 agent an updated set of instructions without changing the model name or replacing the agent with another system.
IBM Research summarized the result this way: “guidelines learned automatically from an agent’s own training runs and appended to its system prompt lifted a stock GPT-4.1 customer-service agent to 67.5% Pass^4 on τ²-bench’s retail domain, above every retail entry in the team’s October 12, 2026 leaderboard snapshot.”
That wording puts the focus on the instructions rather than a new model. The reported result connects the improvement to guidelines learned during training and then added to the system prompt. The number to watch is 67.5% Pass^4, measured in τ²-bench’s retail domain.
What the Retail Evaluation Included
The evaluation used four trials per task. The setup also identified gpt-5.5 as the customer-simulator model, while gpt-4.1 served as the customer-service agent. τ²-bench provided the simulation framework, and Sierra Research maintained τ²-bench.
The guideline counts show how the retained instructions changed under different support requirements. Retail kept 31 guidelines when each retained guideline had support of at least three. When all guidelines were kept, the retail set contained 292 guidelines.
Those two counts describe different levels of selection. The 31-guideline figure refers to rules that met the support threshold of at least three, while the 292-guideline figure includes all guidelines kept in retail. Both numbers sit alongside the same reported 67.5% Pass^4 result for the GPT-4.1 customer-service agent.
Why the Result Matters for AI Agents
AI agents depend on system prompts to guide how they handle tasks. This result tests a way to improve those instructions through the agent’s own training experience, rather than relying only on a prompt written before evaluation.
The reported gain also gives the method a clear comparison point. IBM Research placed the 67.5% Pass^4 result above every retail entry in its October 12, 2026 leaderboard snapshot. That comparison makes the retail result the central measure of the report.
The approach does not rely on a new customer-service model in the facts IBM Research reported. It uses a stock GPT-4.1 agent, learned guidelines, and an updated system prompt. The evaluation then measures the agent in τ²-bench’s retail domain with four trials per task.
IBM Research Shares the Findings
IBM Research reported the results on Hugging Face on October 20, 2026. Jonas Reeve was one of the 11 IBM Research authors connected with the report.
The report leaves a focused takeaway: guidelines learned from training runs can lift a stock GPT-4.1 customer-service agent to a 67.5% Pass^4 rate in τ²-bench retail testing. The result also shows the range of retained guidance, from 31 guidelines with support of at least three to 292 when all guidelines were kept.
