arXiv subjects

Zhehao Zhou

Publications and source records attributed to Zhehao Zhou.

5 recordsLinked to original sources

arXiv2026

Learning When to Think Aloud in Latent Reasoning

Latent Chain-of-Thought (CoT) reasoning improves efficiency by shifting intermediate computation into continuous hidden states. However, fully latent methods often struggle on precision-critical tasks such as mathematical reasoning, while existing hybrid approaches retain substantial token-generation overhead. We introduce VITAL, an importance-guided hybrid reasoning framework that selectively verbalizes informative tokens as reasoning anchors while keeping most computation latent. A lightweight gating head dynamically selects between explicit token generation and direct hidden-state feedback. VITAL constructs hybrid trajectories from explicit CoT using token- importance estimation and trains the model through curriculum-based supervised training, followed by reinforcement learning with a reward that balances correctness and conciseness. Experiments show that VITAL consistently out- performs existing fully latent reasoning methods, achieving relative average accuracy gains of 9.6% on GPT-2 and 5.1% on Llama-3.2-1B at comparable reasoning lengths. Compared with hybrid baselines, VITAL achieves higher reasoning accuracy with substantially shorter reasoning trajectories on Qwen3-1.7B.

cs.LG↗

arXiv2026

Tools Need Not Speak in Text: Latent Context Compression for Agents

Large language model (LLM) agents typically append raw tool outputs to their context. Over multi-step interactions, long and noisy observations incur substantial context overhead and dilute the agent's attention to task-relevant information. Existing token pruning and summarization methods remain constrained by discrete text representations, limiting the achievable compression ratios. In this work, we challenge the prevailing assumption that agents must consume tool observations as explicit text tokens. To this end, we propose Coca-Cola, a continuous latent context compression model for LLM agents. Coca-Cola compresses tool observations into latent tokens and embeds them into the input embedding space of the agent. We further introduce Action-Preserving Training, which aligns the agent's next-action distributions under full and compressed observations. This trains the compression model to learn action-sufficient latent tokens that retain decision-critical information. Experiments with web-search agents on four QA benchmarks and two model families show that Coca-Cola reduces the agent's token consumption by 70.01%-89.24% and end-to-end runtime by 24.45%-55.97%, while consistently improving task performance. These results suggest that Coca-Cola offers a promising approach for context compression in agents. Our code is available at https://github.com/THU-Agent/Coca-cola.

cs.AI↗

arXiv2026

Learning Domain- and Class-Disentangled Prototypes for Domain-Generalized EEG Emotion Recognition

Electroencephalography (EEG)-based emotion recognition plays a critical role in affective Brain-Computer Interfaces (aBCIs), yet its practical deployment remains limited by inter-subject variability, reliance on target-domain data, and unavoidable label noise. To address these challenges, we propose a Multi-domain Aggregation Transfer learning framework with domain-class prototypes (MAT) for emotion recognition under completely unseen target domains. MAT introduces a feature decoupling module to disentangle class-invariant domain features from domain-invariant class features, enabling more robust and interpretable EEG representations. A Hierarchical-Domain Aggregation (HDA) mechanism based on Maximum Mean Discrepancy (MMD) constructs superdomains to model shared distributional structures across subjects, while adaptive prototype updating refines domain and class prototypes to capture stable intrinsic representations. Moreover, a pairwise learning strategy reformulates classification as similarity estimation between sample pairs, effectively mitigating the effect of label noise. Extensive experiments on three public EEG emotion datasets (SEED, SEED-IV, and SEED-V) show that the accuracy of MAT is improved by 2.87%, 3.84%, and 2.05% compared with the state-of-the-art (SOTA) model for unseen target domains. Our results provide a promising direction for emotion recognition under real-world unseen-subject scenarios.The source code is available at https://github.com/WuCB-BCI/MAT.

cs.LG↗

arXiv2025

PL-DCP: A Pairwise Learning framework with Domain and Class Prototypes for EEG emotion recognition under unseen target conditions

Electroencephalogram (EEG) signals serve as a powerful tool in affective Brain-Computer Interfaces (aBCIs) and play a crucial role in affective computing. In recent years, the introduction of deep learning techniques has significantly advanced the development of aBCIs. However, the current emotion recognition methods based on deep transfer learning face the challenge of the dual dependence of the model on source domain and target domain, As well as being affected by label noise, which seriously affects the performance and generalization ability of the model. To overcome this limitation, we proposes a Pairwise Learning framework with Domain and Category Prototypes for EEG emotion recognition under unseen target conditions (PL-DCP), and integrating concepts of feature disentanglement and prototype inference. Here, the feature disentanglement module extracts and decouples the emotional EEG features to form domain features and class features, and further calculates the dual prototype representation. The Domain-pprototype captures the individual variations across subjects, while the class-prototype captures the cross-individual commonality of emotion categories. In addition, the pairwise learning strategy effectively reduces the noise effect caused by wrong labels. The PL-DCP framework conducts a systematic experimental evaluation on the published datasets SEED, SEED-IV and SEED-V, and the accuracy are 82.88\%, 65.15\% and 61.29\%, respectively. The results show that compared with other State-of-the-Art(SOTA) Methods, the PL-DCP model still achieves slightly better performance than the deep transfer learning method that requires both source and target data, although the target domain is completely unseen during the training. This work provides an effective and robust potential solution for emotion recognition. The source code is available at https://github.com/WuCB-BCI/PL_DCP.

cs.LG↗

arXiv2025

Learning Diverse Natural Behaviors for Enhancing the Agility of Quadrupedal Robots

Achieving animal-like agility is a longstanding goal in quadrupedal robotics. While recent studies have successfully demonstrated imitation of specific behaviors, enabling robots to replicate a broader range of natural behaviors in real-world environments remains an open challenge. Here we propose an integrated controller comprising a Basic Behavior Controller (BBC) and a Task-Specific Controller (TSC) which can effectively learn diverse natural quadrupedal behaviors in an enhanced simulator and efficiently transfer them to the real world. Specifically, the BBC is trained using a novel semi-supervised generative adversarial imitation learning algorithm to extract diverse behavioral styles from raw motion capture data of real dogs, enabling smooth behavior transitions by adjusting discrete and continuous latent variable inputs. The TSC, trained via privileged learning with depth images as input, coordinates the BBC to efficiently perform various tasks. Additionally, we employ evolutionary adversarial simulator identification to optimize the simulator, aligning it closely with reality. After training, the robot exhibits diverse natural behaviors, successfully completing the quadrupedal agility challenge at an average speed of 1.1 m/s and achieving a peak speed of 3.2 m/s during hurdling. This work represents a substantial step toward animal-like agility in quadrupedal robots, opening avenues for their deployment in increasingly complex real-world environments.

cs.RO↗