Personalizing agents through cross-session interaction data boosts task success

In 600 tasks across 30 individuals, the authors' TAHI method raises solo task success by 4.5–20.9% and yields evaluation rubrics that catch 16.0–22.3% more failures than LM- or human-generated rubrics.

PaperIndependentcs.AIarXiv:2609.04141v1 ↗

Zora Zhiruo Wang · Apurva Gandhi · Rulin Shao · Aspen Chen · Jonas Mueller · Zhiqi Liang · +19 more

Research Digest·7 September 2026·1 min read

Read the paper →

Threads:Agent Self-Improvement Memory Management for Agents

Wang et al. propose TAHI (test-time adaptation through human-agent interaction), which leverages a user's cross-session interaction history to update agent context and weights while maintaining an evolving rubric that captures personal evaluation criteria. Applied to writing and visual creation across 30 individuals, the adapted agents outperformed non-adapted baselines on individual tasks, and part of the personalization gains transferred to other users.

What they did

The authors developed TAHI, a framework that integrates a user's repeated feedback across human-agent sessions into the agent's context and model weights without retraining on the full population. At the core is an evolving rubric module, which progressively crystallizes the user's training and evaluation criteria — many of which cannot be fully specified up front — and acts as a scalable annotation guide.

They evaluated TAHI on 30 individuals across two high-utility domains (writing and visual creation), totaling 600 open-ended tasks with heterogeneous success criteria. The method was compared against non-personalized agents and against rubrics generated by LMs or humans alone, measuring both per-user task success and rubric failure-detection rate.

Analysis

TAHI sits at the intersection of test-time adaptation and human-in-the-loop agent design, echoing a growing shift from static prompting to context engineering that evolves with the user. The rubric module is especially notable: it converts implicit, repeatedly applied criteria into an explicit document that can be inspected, reused, and potentially improved, which may help bridge the gap between 'average' model behavior and personal expertise. An open question is whether such adaptation can scale beyond individual interaction histories to institutional or team-level standards, and whether the gains persist as models themselves improve.