CoSy: Conversational Synthesis for Grounded Question Answering

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

CoSy (Conversational Synthesis) is a new framework designed to generate diverse, steerable, multi-turn conversations at scale, addressing the scarcity of high-quality conversational datasets needed for training effective on-device language models (~1B parameters). The framework integrates three core mechanisms: conversational graphs for natural dialogue flow, turn-based prompt augmentations for diversity, and explicit linguistic phenomena for coherence. Evaluated on conversational grounded reasoning tasks, a key on-device use case, models trained with CoSy-synthesized data achieve competitive performance against human-annotated baselines. These models also outperform instruction-tuned models of up to 70B parameters in zero-shot settings.

Key takeaway

For Machine Learning Engineers developing on-device conversational AI, CoSy presents a critical solution to data scarcity. By leveraging synthetic data generation, you can train smaller, efficient models (~1B parameters) that achieve performance competitive with human-annotated baselines and even surpass larger instruction-tuned models (up to 70B parameters) in zero-shot scenarios. Consider integrating conversational synthesis techniques to accelerate your model development and deployment for resource-constrained environments.

Key insights

CoSy synthesizes diverse, multi-turn conversational data to train effective on-device language models.

Principles

• Conversational graphs ensure natural dialogue flow.
• Turn-based prompt augmentations enhance diversity.
• Explicit linguistic phenomena improve coherence.

Method

CoSy combines conversational graphs, turn-based prompt augmentations, and explicit linguistic phenomena to generate diverse, steerable, multi-turn conversations at scale for language model training.

In practice

• Train on-device language models (~1B parameters).
• Improve conversational grounded reasoning task performance.

Topics

• Conversational AI
• Data Synthesis
• On-device Language Models
• Grounded Question Answering
• Dialogue Systems
• Large Language Models

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Syn-TurnTurk: A Synthetic Dataset for Turn-Taking Prediction in Turkish Dialogues — Synthetic datasets can effectively address data scarcity for turn-taking prediction in under-resourced languages like Turkish.
• Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering — CorVer uses Wikipedia co-occurrence for lightweight, corpus-grounded process supervision to boost factual QA accuracy.
• DiscoTrace: Representing and Comparing Answering Strategies of Humans and LLMs in Information-Seeking Question Answering — DiscoTrace reveals LLMs lack rhetorical diversity and over-address question interpretations compared to human answerers.
• Build Small with Cohere Labs — Cohere focuses on compact, efficient, and multilingual AI models for diverse applications and deployment environments.
• We Evaluated ChatGPT vs. Google on 500 Search Queries — ChatGPT significantly outperforms Google on coding queries and ties it on general informational queries, despite not being optimized for search.
• Beyond Code Pairs: Dialogue-Based Data Generation for LLM Code Translation — A dual-LLM Questioner–Solver generates dialogue-based data, improving LLM code translation in low-resource domains.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.