Article | Yiyang "Diana" Wang

CultivAgents: Cultivating Relationship-Centered Multi-Agent Systems for Personalized Gardening

Fri, 22 May 2026 00:00:00 +0000

Abstract

Gardening supports well-being, cultural continuity, and food autonomy, yet digital tools often give generic advice that overlooks gardeners’ skills, local ecologies, and cultural contexts. We introduce CultivAgents, a relationship-centered multi-agent system grounded in ethics of care that coordinates an Experience Agent, an Environmental Agent, and an Ethnobotanical Agent to deliver personalized, socio-culturally grounded support. A three-phase mixed-methods study with experts, HCI researchers, and community gardeners found that CultivAgents helped gardeners translate interest into situated action and increased their confidence, motivation, and trust in acting on AI advice.

TextReg: Mitigating Prompt Distributional Overfitting via Regularized Text-Space Optimization

Wed, 20 May 2026 00:00:00 +0000

Abstract

LLMs are highly sensitive to their prompts, and iterative prompt-optimization methods often overfit: prompts grow longer, accumulate sample-specific rules, and generalize poorly. We study this as prompt distributional overfitting and introduce TextReg, a regularization framework that realizes a soft-penalty objective through regularized textual gradients — combining Dual-Evidence Gradient Purification, Semantic Edit Regularization, and Regularization-Guided Prompt Update. Across reasoning benchmarks, TextReg improves out-of-distribution generalization by up to +11.8% over TextGrad and +16.5% over REVOLVE.

UniSD: Towards a Unified Self-Distillation Framework for Large Language Models

Thu, 07 May 2026 00:00:00 +0000

Abstract

Self-distillation offers a promising path for adapting large language models without stronger external teachers, but it remains hard to apply reliably in autoregressive LLMs. We propose UniSD, a unified framework that systematically studies self-distillation by integrating complementary mechanisms — multi-teacher agreement, EMA teacher stabilization, token-level contrastive learning, feature matching, and divergence clipping. Across six benchmarks and six models from three families, UniSD clarifies when and why self-distillation helps, and its integrated pipeline improves over the base model by +5.4 points and the strongest baseline by +2.8 points.

MASCOT: Towards Multi-Agent Socio-Collaborative Companion Systems

Tue, 20 Jan 2026 00:00:00 +0000

Abstract

Multi-agent systems are emerging as promising socio-collaborative companions for emotional and cognitive support. However, existing systems frequently suffer from persona collapse, where agents revert to generic assistant behaviors, and social sycophancy, where agents produce redundant, non-constructive dialogue. We propose MASCOT, a multi-agent framework for multi-perspective socio-collaborative companions, with a bi-level optimization strategy that harmonizes individual identities and collective discourse. MASCOT improves persona consistency by up to +14.1 and social contribution by up to +10.6 across in-domain and out-of-domain settings.