BEIJING INSTITUTE OF TECHNOLOGY

Human Language Technology Group

HLT 语言技术研究组

Institute of Language Intelligence and Social Computing · School of Computer Science and Technology · Beijing Institute of Technology

Understanding language.
Connecting context and action.

We study how language connects perception and action: from machine translation and multimodal interaction to agents that understand tasks, learn from experience and use tools reliably.

RESEARCH DIRECTIONS

What we study

Starting with language, grounded in real-world understanding and interaction.

01 / LANGUAGE

Language understanding & translation

How can we recover meaning across languages and forms of expression? We explore information extraction, machine translation, in-image translation and translation evaluation.

RATE · PRIM
02 / INTERACTION

Multimodal interaction & agents

How can vision, audio and context turn task understanding into action? We study smart-home interaction, embodied planning, GUI agents and agent memory.

PEAP · HomeBench · LifeMem
03 / RELIABILITY

Model reliability & safety

Can models follow requirements, recognize risks and use tools safely? We study reliability from generated outputs to execution through capability improvement and evaluation.

ReFF · SafeToolBench

SELECTED WORK & OPEN SOURCE

Publications & open source

All repositories
2026EMNLP 2026 · Main
LifeMem Perception & agents

LifeMem: Enabling Lifelong Experience Reuse for LLM Agents

Extracting, organizing and retrieving skills from task trajectories so agents can reuse experience across environments.

2026EMNLP 2026 · Main
JarvisGUI Perception & agents

JarvisGUI: Towards Cross-Device GUI Agents with Dynamic Task Composition

Evaluating cross-device GUI agents through dynamically composed tasks that require coordination and state transfer across operating systems.

2026EMNLP 2026 · Findings
VGEBench Perception & agents

Towards Generalizable Visually Grounded Exploration of Household Devices

Benchmarking visually grounded exploration of household devices and generalization to new devices and contexts.

2026EMNLP 2026 · Findings
Behavior2Trip Perception & agents

Behavior2Trip: Towards Personalized Travel Planning via User Behavior Trajectory

Inferring preferences from user behavior trajectories to create personalized travel plans with tool retrieval and memory.

2026EMNLP 2026 · Findings
SC-GRPO Reliability & safety

Learning from Own Solutions: Self-Conditioned Credit Assignment for Reinforcement Learning with Verifiable Rewards

Using the model’s own verified solutions to assign token-level learning weights for reinforcement learning with verifiable rewards.

2026ACL 2026 · Main
RATE Language & translation

Beyond Literal Mapping: Benchmarking and Improving Non-Literal Translation Evaluation

Evaluating non-literal translation through a human-annotated benchmark and an agentic framework for reflection, retrieval and comparison.

2026ACL 2026 · Main
PEAP Perception & agents

PEAP: Proactive Embodied Action Sequence Planning with Joint Understanding of Vision and Audio Perception

Jointly understanding visual and audio context for proactive embodied action-sequence planning.

2026ACL 2026 · Findings
PEC-Home Perception & agents

PEC-Home: Interpretation of Progressively Elliptical Commands in Smart Homes

Interpreting progressively elliptical instructions in smart-home interactions using conversational context.

2026arXiv · 2026
LivingScreen Perception & agents

Benchmarking Living-Screen-Native GUI Agents on Short-Video Platforms

Evaluating GUI agents on the continuously changing interfaces of short-video platforms.

2025EMNLP 2025 · Main
PRIM Language & translation

PRIM: Towards Practical In-Image Multilingual Machine Translation

Exploring practical multilingual translation within images, connecting textual meaning with visual presentation.

2025ACL 2025 · Main
HomeBench Perception & agents

HomeBench: Evaluating LLMs in Smart Homes with Valid and Invalid Instructions Across Single and Multiple Devices

Evaluating smart-home instruction understanding across single and multiple devices, including valid and invalid instructions.

2025EMNLP 2025 · Findings
SafeToolBench Reliability & safety

SafeToolBench: Pioneering a Prospective Benchmark to Evaluating Tool Utilization Safety in LLMs

Evaluating the safety of tool use by large language models through a dedicated benchmark.

2025AAAI 2025
ReFF Reliability & safety

ReFF: Reinforcing Format Faithfulness in Language Models across Varied Tasks

Reinforcing output-format faithfulness across tasks to make language-model outputs more reliable for downstream applications.

Find more publications on the full publication list and ACL Anthology.

TEACHING

Teaching & learning

Build understanding from core ideas and interactive examples. Courses & apps ↗

DS2026

DATA STRUCTURES

Data Structures

Course information, chapter summaries and interactive materials released after class.

NLP2026

NATURAL LANGUAGE PROCESSING

Natural Language Processing Practice

Course schedule, practical guides and student exercise packages.

PEOPLE & COLLABORATION

People & collaboration

Institute of Language Intelligence and Social Computing · School of Computer Science and Technology · Beijing Institute of Technology

Group lead

Yuhang Guo 郭宇航

We welcome discussions on language intelligence, multimodal interaction and reliable agents.

Doctoral students

  • 姚嘉树
  • 田炎智
  • 曾理
  • 刘艳平

Master’s students

  • 邱昱力
  • 单赢宇
  • 郑林昊
  • 文浩宇
  • 苏炜
  • 张红博
  • 艾华喜
  • 陈旺科
  • 钱博傲