AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

ClarifyVC: Clarifying Ambiguous Commands in Vehicle Control with a Hybrid Data Augmentation Pipeline

Hange Zhou, Zhonglin Jiang, yingjie cui, Mingzhe Zhang, Xiaotang Wang, Hengwei Dai, Qiyao Yu, Yong Chen, Yongqi Zhang

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ClarifyVC, a framework with a hybrid data-augmentation pipeline, reference models, and an evaluation protocol for resolving ambiguous natural-language vehicle-control commands.

Problem

Natural language interfaces for vehicle control must handle vague commands, evolving dialogue context, and strict protocol constraints.

Approach

A unified framework pairing an agent-orchestrated hybrid data-augmentation pipeline that generates diverse, ambiguity-rich dialogues with reference models trained on that data and a dedicated evaluation protocol.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
supervised fine-tune on ClarifyVC-Data vs zero-shotQwen2.5-7B · ClarifyVC-Data single-roundaccuracy: 74.3 -> 89.0Improved
supervised fine-tune on ClarifyVC-Data vs zero-shotLLaMA3-70B · ClarifyVC-Data single-roundaccuracy: 81.2 -> 94.1Improved
add 4 in-context examplessmall models (Qwen2.5-0.5B) · ClarifyVC-Data single-roundaccuracy: 59.2 -> 64.5Improved
add 4 in-context examplesLLaMA3-70B · ClarifyVC-Data single-roundaccuracy: 81.2 -> 80.1Hurt
use 7B SFT model instead of larger backbonesQwen2.5 · ClarifyVC-Dataaccuracy: 89.0 vs 95.8 (72B)No clear effect
scale model size under SFTQwen2.5 0.5B to 72B · ClarifyVC-Data single-roundaccuracy: 75.1 -> 95.8Improved
add fuzz injection (FIA) stage (ablation)data pipeline · ambiguity diversityambiguity diversity: 0.68 -> 0.85Improved
add adversarial generation (AGA) stage (ablation)data pipeline · ambiguity diversityambiguity diversity: 0.71 -> 0.85Improved
run AGA before FIA rather than reversed (ablation)data pipeline · ambiguity diversityambiguity diversity: 0.79 -> 0.85Improved
add multi-turn evolution (MEA) stage (ablation)data pipeline · avg scoreavg score drop when removed: -4.6%Improved
constrain FIA with drift control and expert-curated pairs (ablation)ambiguity injection · intent preservationintent preservation: 89-92% -> 98.9%Improved
use ClarifyVC-Data in place of existing datasets as a benchmarkzero-shot LLMs · benchmark difficultyintent accuracy: 92.5% -> 70.5% (GPT-4, APIGen vs ClarifyVC-Data)Hurt

Experimental setup

Benchmarks: ClarifyVC-Eval

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy