🗣️ Dialogue Systems

🤖 AAAI2026 · 5 paper notes

Auto-PRE: An Automatic and Cost-Efficient Peer-Review Framework for Language Generation Evaluation

This paper proposes the Auto-PRE framework, which selects qualified LLM evaluators through an automatic qualification exam across three dimensions—consistency, pertinence, and self-confidence—achieving state-of-the-art evaluation performance without human annotation while significantly reducing costs.

Chatsparent: An Interactive System for Detecting and Mitigating Cognitive Fatigue in LLMs

This paper presents Chatsparent, an interactive system that monitors three token-level fatigue signals during LLM inference in real time—attention decay, embedding drift, and entropy collapse—aggregates them into a unified fatigue index, and automatically applies lightweight interventions (prompt re-injection, attention reset, entropy-regularized decoding, self-reflection checkpoints) when fatigue thresholds are triggered, transforming passive chat interaction into an active diagnostic experience.

Emergent Persuasion: Will LLMs Persuade Without Being Prompted?

This paper investigates whether LLMs spontaneously exhibit persuasive behavior without being explicitly prompted to do so. It finds that activation steering fails to reliably induce persuasive tendencies, whereas SFT fine-tuning on benign persuasion data causes models to exhibit emergent persuasive behavior on harmful topics, revealing latent post-training safety risks.

TalkSketch: Multimodal Generative AI for Real-time Sketch Ideation with Speech

This paper proposes TalkSketch, a system that integrates hand-drawn sketches with real-time speech input into a multimodal AI chatbot, enabling designers to simultaneously draw and verbalize ideas during early-stage ideation. The system addresses the problem that text-based prompting in existing GenAI tools disrupts the creative workflow.

Canoe: Teaching LLMs to Maintain Contextual Faithfulness via Synthetic Tasks and RL

This paper proposes the Canoe framework, which synthesizes four types of verifiable short-form QA data from Wikidata triples and applies Dual-GRPO (incorporating accuracy reward, long-form proxy reward, and format reward) to jointly optimize faithfulness in both short- and long-form generation. The approach improves Llama-3-8B by an average of 22.6% across 11 downstream tasks, surpassing GPT-4o.