GPT-4.1 is a language model from OpenAI that recent coverage often uses as the base model for agents. Hugging Face's ALTK-Evolve used GPT-4.1 to distill behavioral rules, reaching 67.5% Pass^4 on τ²-bench retail tasks, while its consistency analysis found a GPT-4.1 ReAct agent averaged 77.4% success on AppWorld but only 53.0% success across all five repeats. A researcher is also running 576 pipelines with GPT-4.1 and Claude Sonnet 4.6.
ALTK-Evolve tops customer-service benchmark via learned memoryHugging Face introduced ALTK-Evolve, which uses GPT-4.1 to distill short behavioral rules from an agent's training-task successes and failures and append them to the system prompt, without updating model weights. On Sierra's τ²-bench retail customer-service tasks, a basic ReAct agent reached 67.5% Pass^4, beating every retail entry on the leaderboard snapshot.
Forum: isolating back-and-forth LLM interaction vs one-way sharingA researcher seeks controlled studies comparing back-and-forth interaction between two different LLMs against independent drafts, one-way sharing and self-refinement under an equal compute budget, currently running 576 pipelines with GPT-4.1 and Claude Sonnet 4.6 on 12 tasks.
Agent Consistency: Hugging Face's New DiagnosticHugging Face introduces the Consistency Analyzer and consistency guidelines to diagnose and reduce agent variability. They show a GPT‑4.1 ReAct agent scores 77.4% average success on AppWorld but only 53.0% success across all five repeats, a 24.4‑point consistency gap; the new approach improves repeatability.
The A.I. Daily