Xiaomi's 8 EMNLP 2026 Papers: Mobile Agents, Efficient LLMs, Proactive AI

Mobile GUI Agents

GSAR: Goal-State-Anchor Rewards for Mobile GUI Agents with Self-Evolving Data Synthesis

Collaborators: Wuhan University. Authors: Zhang Long, Chen Yuhan, Zhang Chaoran, Cao Wanxia, Huang Kun, Gao Pengzhi, Liu Wei, Luan Jian, Li Chenliang, Zou Lixin. Acceptance: Main conference long paper. arXiv: https://arxiv.org/abs/2608.22847

Training phone AI assistants via reinforcement learning requires massive tasks and reliable reward signals. GSAR automates both: the agent generates increasingly difficult tasks and execution paths through self-interaction with apps, and uses the goal-state screen as an anchor to automatically identify task-relevant UI elements for reward computation, eliminating hand-crafted rules. GSAR achieves over 90% accuracy in task-completion judgment and significantly improves agent performance on AndroidWorld and other benchmarks.

GUI-PRA: Process Reward Agents for GUI Tasks

Collaborators: Zhejiang University. Authors: Xiong Tao, Hu Xueyu, Chen Yurun, Liu Yuhang, Wu Changqiao, Gao Pengzhi, Liu Wei, Luan Jian, Zhang Shengyu. Acceptance: Main conference long paper. arXiv: https://arxiv.org/abs/2509.23263

Multi-step GUI tasks suffer from error propagation when only final outcomes are verified. GUI-PRA upgrades the reward model into an active verifier: it distills principles from past success/failure trajectories, generates step-specific criteria, and proactively calls visual tools to verify UI details when evidence is insufficient, forming a principle–criterion–evidence–conclusion reasoning chain. On AndroidWorld and Mobile-MiniWoB++, GUI-PRA improves over prior process reward methods by ~5.0 and ~6.5 percentage points respectively, and achieves competitive results on OS-Critic Bench without training a dedicated critic.

Towards Omni-dimensional GUI Agent Navigation with Masked Trajectory Prediction (MaP)

Collaborators: Institute of Information Engineering, Chinese Academy of Sciences; Nankai University. Authors: Zhang Yan, Fu Pei, Wu Daiqing, Shen Huawen, Zhang Ruozeng, Zhang Shaojie, Yang Jiahui, Zhou Yu, Ma Can, Luo Zhenbo, Luan Jian. Acceptance: Findings long paper.

GUI navigation requires single-step action prediction, state-action matching, and multi-step planning. MaP unifies these by packing instruction, screenshot, reasoning, and action into an interaction trajectory and randomly masking segments during training, forcing the model to reconstruct them. A role-separation mechanism prevents interference. On five GUI navigation benchmarks, MaP yields zero-shot gains of 2.8%, 5.4%, and 2.1% on the three core abilities over mixed training. On AndroidControl-High and AndroidControl-Low, success rates rise from 72.9% to 74.2% and from 89.0% to 90.9%.

Large Model Efficiency

Informed Masking: Structure-Aware Perturbation for Reinforcement Learning in Diffusion Large Language Models

Collaborators: Renmin University of China; University of Modena and Reggio Emilia. Authors: Yu Xiaoyi, Enver Sangineto, Fu Pei, Fiorenzo Parascandolo, Tan Wenhui, Zhang Ruikang, Rita Cucchiara, Song Ruihua, Luan Jian. Acceptance: Findings long paper.

Diffusion LLMs (dLLMs) generate text by iteratively denoising masked tokens. Random masking mixes easy and hard tokens, lowering RL training efficiency. The authors find that masking certain critical tokens causes widespread prediction errors (high difficulty), while others affect only local context (suitable for practice). Informed Masking biases training toward appropriately difficult tokens without extra compute. Integrated into three recent dLLM RL methods and applied to LLaDA-8B-Instruct, it delivers relative average improvements up to 2.01%, 8.52%, and 5.77% on math and planning benchmarks, with more stable training.

ECHO: Early-layer Collaborative Hierarchical Orchestration with Bonus Logits in Speculative Decoding

Collaborators: Wuhan University. Authors: Ma Ziyang, Zhang Zihong, Li Zuchao, Zhang Lefei, Qi Baoyuan, Li Siqi, Yu Siming. Acceptance: Main conference long paper.

Speculative decoding accelerates LLMs by drafting tokens with a small model and verifying with the large model. Prior work uses only the final layer for drafting, leading to slow draft updates and high verification cost (the verification wall). ECHO exploits layer-wise specialization: early layers draft quickly using contextual understanding, while the full model verifies, creating a fast-slow dual loop. Without adding parameters, ECHO achieves 2.4× to 3.01× speedup on Spec-Bench, HumanEval, and GSM8K, substantially improving draft quality and acceptance rate as a plug-and-play solution.

BORA-BUDGET: Learning When to Think Under a Token Budget

Collaborators: Shanghai Jiao Tong University. Authors: Wang Zilong, Pan Kang, Zhang Xiao, Meng Erli, Li Shuai. Acceptance: Main conference long paper.

Deep reasoning models are powerful but token-hungry. BORA-BUDGET allocates a token budget: a fast model gives an initial answer, then a controller predicts the gain, risk, and cost of deep reasoning, triggering it only when worthwhile. The decision to think and the decision to accept the reasoned answer are separated to avoid overturning correct fast answers. The controller is trained via RL on feedback from deep reasoning traces. On Qwen3-8B MATH500: fast mode 73.40%, full reasoning 93.33% (avg 4293 tokens), BORA 91.40% (3534 tokens), random 50% trigger 82.71% (same tokens). On Qwen3-14B: BORA 93.13% (3360 tokens) vs full 95.07% (3903 tokens). This concentrates compute on genuinely hard problems.

Intelligent R&D and Proactive Intelligence

Accelerating Neural Architecture Search with Code Semantics (ECO-SCE)

Collaborators: Xi'an Jiaotong University; North China University of Technology; Zhongguancun Academy. Authors: Liu Zhen, Zhou Wanqi, Bai Shuanghao, Song Wei, Zhang He, Liu Yuhan, Fu Jingwen. Acceptance: Main conference long paper.

As LLMs rewrite neural network code, evaluating every candidate by full training is prohibitively expensive. ECO-SCE analyzes code-level changes before evaluation, filtering out modifications unlikely to yield gains, preserving budget for promising candidates. Unlike mere compilability checks, it identifies semantically ineffective changes. Integrated with OpenEvolve, ECO-SCE saves 70.88% GPU time on average across seven benchmarks, with six tasks maintaining or improving performance. This provides Xiaomi with reusable low-cost automated model design for on-device AI, intelligent vehicles, and smart home scenarios.

Perceive Before Reasoning: A Pre-Reasoning Perception Framework for Efficient and Reliable Proactive Mobile Agents (PRPF)

Collaborators: Zhongnan University of Economics and Law; Jilin University; CUHK-Shenzhen. Authors: Ding Zhijie, Hong Weinan, Zhu Zicheng, Li Lei, Kong Dezhi, Wang Hao, Zhou Peng, Jiang Xuchu, Xu Jiaming. Acceptance: Main conference long paper. arXiv: https://arxiv.org/pdf/2606.03236

Proactive mobile agents must decide whether to intervene and what to do — two decisions with different risk profiles. PRPF adopts a two-stage perceive-then-reason design: a lightweight model first judges if the context warrants a recommendation; only if yes does the main reasoning model engage deeply. On the ProactiveMobile benchmark, PRPF lifts task success rate from 20.82% to 41.15%, cuts false trigger rate from 13.76% to 7.21%, and reduces reasoning compute by 69.3%.

Conclusion

The eight EMNLP 2026 papers reflect Xiaomi's sustained exploration across large models, agents, and proactive intelligence. From fundamental research to engineering capabilities and concrete applications in the phone-car-home ecosystem, Xiaomi is building a synergistic AI technology stack. Future work will continue to drive innovation grounded in real user needs and scenarios, making AI integrate more naturally into devices, services, and environments for a more continuous and intelligent experience.