Guoxi Zhang

Guoxi Zhang, Hisashi Kashima

Offline reinforcement learning (RL) have received rising interest due to its appealing data efficiency. The present study addresses behavior estimation, a task that lays the foundation of many offline RL algorithms. Behavior estimation aims at estimating the policy with which training data are generated. In particular, this work considers a scenario where the data are collected from multiple sources. In this case, neglecting data heterogeneity, existing approaches for behavior estimation suffers from behavior misspecification. To overcome this drawback, the present study proposes a latent variable model to infer a set of policies from data, which allows an agent to use as behavior policy the policy that best describes a particular trajectory. This model provides with a agent fine-grained characterization for multi-source data and helps it overcome behavior misspecification. This work also proposes a learning algorithm for this model and illustrates its practical usage via extending an existing offline RL algorithm. Lastly, with extensive evaluation this work confirms the existence of behavior misspecification and the efficacy of the proposed model.

Guoxi Zhang, Jiawei Chen, Tianzhuo Yang et al.

As Large Language Models (LLMs) expand in capability and application scope, their trustworthiness becomes critical. A vital risk is intrinsic deception, wherein models strategically mislead users to achieve their own objectives. Existing alignment approaches based on chain-of-thought (CoT) monitoring supervise explicit reasoning traces. However, under optimization pressure, models are incentivized to conceal deceptive reasoning, rendering semantic supervision fundamentally unreliable. Grounded in cognitive psychology, we hypothesize that a deceptive LLM maintains a stable internal belief in its CoT while its external response remains fragile under perturbation. We term this phenomenon stability asymmetry and quantify it by measuring the contrast between internal CoT stability and external response stability under perturbation. Building on this structural signature, we propose the Stability Asymmetry Regularization (SAR), a novel alignment objective that penalizes this distributional asymmetry during reinforcement learning. Unlike CoT monitoring, SAR targets the statistical structure of model outputs, rendering it robust to semantic concealment. Extensive experiments confirm that stability asymmetry reliably identifies deceptive behavior, and that SAR effectively suppresses intrinsic deception without degrading general model capability.

Junpeng Zhang, Lei Cheng, Guoxi Zhang et al.

This paper explores a scientific question in supervised fine-tuning (SFT): why SFT is broadly effective for small-scale deep neural networks, yet can produce inconsistent or even detrimental effects when applied to large language models (LLMs). Recent advances in interaction-based explanations suggest that interactions between words/tokens provide a faithful metric for quantifying the inference patterns encoded by LLMs. We find that the evolution of interactions during SFT can effectively explain the inconsistent effectiveness of SFT for LLMs. Specifically, we find that (1) SFT primarily removes noise-like interactions, while rarely acquiring reliable new interactions. (2) This denoising stage is extremely brief, after which continued fine-tuning tends to introduce overfitted interactions. We validate these findings across multiple LLMs and datasets. Our findings provide new insights into early stopping and offer practical guidance for LLM training.

Xiaofeng Lin, Guoxi Zhang, Xiaotian Lu et al.

Treatment effect estimation can assist in effective decision-making in e-commerce, medicine, and education. One popular application of this estimation lies in the prediction of the impact of a treatment (e.g., a promotion) on an outcome (e.g., sales) of a particular unit (e.g., an item), known as the individual treatment effect (ITE). In many online applications, the outcome of a unit can be affected by the treatments of other units, as units are often associated, which is referred to as interference. For example, on an online shopping website, sales of an item will be influenced by an advertisement of its co-purchased item. Prior studies have attempted to model interference to estimate the ITE accurately, but they often assume a homogeneous interference, i.e., relationships between units only have a single view. However, in real-world applications, interference may be heterogeneous, with multi-view relationships. For instance, the sale of an item is usually affected by the treatment of its co-purchased and co-viewed items. We hypothesize that ITE estimation will be inaccurate if this heterogeneous interference is not properly modeled. Therefore, we propose a novel approach to model heterogeneous interference by developing a new architecture to aggregate information from diverse neighbors. Our proposed method contains graph neural networks that aggregate same-view information, a mechanism that aggregates information from different views, and attention mechanisms. In our experiments on multiple datasets with heterogeneous interference, the proposed method significantly outperforms existing methods for ITE estimation, confirming the importance of modeling heterogeneous interference.

Lirui Luo, Guoxi Zhang, Hongming Xu et al.

In deep reinforcement learning (DRL), an agent is trained from a stream of experience. In a continual learning setting, such agents can suffer from plasticity loss: their ability to learn new skills from new experiences diminishes over training. Recently, Mixture-of-Experts (MoE) networks have been reported to enable scaling laws and facilitate the learning of diverse skills. However, in continual reinforcement learning settings, their performance can degenerate as learning proceeds, indicating a loss of plasticity. To address this, building on Neural Tangent Kernel (NTK) theory, we formalize the plasticity loss in MoE policies as a loss of spectral plasticity. We then derive a tractable proxy for spectral plasticity, one expressible in terms of individual expert feature matrices. Leveraging this proxy, we introduce SPHERE, a practical Parseval penalty tailored for MoE-based policies that alleviates the loss of spectral plasticity. On MetaWorld and HumanoidBench, SPHERE improves average success under continual RL by 133% and 50% over an unregularized MoE baseline, while maintaining higher spectral plasticity throughout training.

Guoxi Zhang, Jiuding Duan

This paper addresses the cost-efficiency aspect of Reinforcement Learning from Human Feedback (RLHF). RLHF leverages datasets of human preferences over outputs of large language models (LLM)s to instill human expectations into LLMs. Although preference annotation comes with a monetized cost, the economic utility of a preference dataset has not been considered by far. What exacerbates this situation is that, given complex intransitive or cyclic relationships in preference datasets, existing algorithms for fine-tuning LLMs are still far from capturing comprehensive preferences. This raises severe cost-efficiency concerns in production environments, where preference data accumulate over time. In this paper, we discuss the fine-tuning of LLMs as a monetized economy and introduce an auction mechanism to improve the efficiency of preference data collection in dollar terms. We show that introducing an auction mechanism can play an essential role in enhancing the cost-efficiency of RLHF, while maintaining satisfactory model performance. Experimental results demonstrate that our proposed auction-based protocol is cost-effective for fine-tuning LLMs concentrating on high-quality feedback.

Lirui Luo, Guoxi Zhang, Hongming Xu et al.

Reward design is of great importance for solving complex tasks with reinforcement learning. Recent studies have explored using image-text similarity produced by vision-language models (VLMs) to augment rewards of a task with visual feedback. A common practice linearly adds VLM scores to task or success rewards without explicit shaping, potentially altering the optimal policy. Moreover, such approaches, often relying on single static images, struggle with tasks whose desired behavior involves complex, dynamic motions spanning multiple visually different states. Furthermore, single viewpoints can occlude critical aspects of an agent's behavior. To address these issues, this paper presents Multi-View Video Reward Shaping (MVR), a framework that models the relevance of states regarding the target task using videos captured from multiple viewpoints. MVR leverages video-text similarity from a frozen pre-trained VLM to learn a state relevance function that mitigates the bias towards specific static poses inherent in image-based methods. Additionally, we introduce a state-dependent reward shaping formulation that integrates task-specific rewards and VLM-based guidance, automatically reducing the influence of VLM guidance once the desired motion pattern is achieved. We confirm the efficacy of the proposed framework with extensive experiments on challenging humanoid locomotion tasks from HumanoidBench and manipulation tasks from MetaWorld, verifying the design choices through ablation studies.

Peixin Chen, Guoxi Zhang, Jianwei Ma et al.

Embodied agents must explore partially observed environments while maintaining reliable long-horizon memory. Existing graph-based navigation systems improve scalability, but they often treat unexplored regions as semantically unknown, leading to inefficient frontier search. Although vision-language models (VLMs) can predict frontier semantics, erroneous predictions may be embedded into memory and propagate through downstream inferences, causing structural error accumulation that confidence attenuation alone cannot resolve. These observations call for a framework that can leverage semantic predictions for directed exploration while systematically retracting errors once new evidence contradicts them. We propose Hypothesis Graph Refinement (HGR), a framework that represents frontier predictions as revisable hypothesis nodes in a dependency-aware graph memory. HGR introduces (1) semantic hypothesis module, which estimates context-conditioned semantic distributions over frontiers and ranks exploration targets by goal relevance, travel cost, and uncertainty, and (2) verification-driven cascade correction, which compares on-site observations against predicted semantics and, upon mismatch, retracts the refuted node together with all its downstream dependents. Unlike additive map-building, this allows the graph to contract by pruning erroneous subgraphs, keeping memory reliable throughout long episodes. We evaluate HGR on multimodal lifelong navigation (GOAT-Bench) and embodied question answering (A-EQA, EM-EQA). HGR achieves 72.41% success rate and 56.22% SPL on GOAT-Bench, and shows consistent improvements on both QA benchmarks. Diagnostic analysis reveals that cascade correction eliminates approximately 20% of structurally redundant hypothesis nodes and reduces revisits to erroneous regions by 4.5x, with specular and transparent surfaces accounting for 67% of corrected prediction errors.

Jiawei Chen, Tianzhuo Yang, Guoxi Zhang et al.

Aligning Large Language Models (LLMs) with nuanced human values remains a critical challenge, as existing methods like Reinforcement Learning from Human Feedback (RLHF) often handle only coarse-grained attributes. In practice, fine-tuning LLMs on task-specific datasets to optimize value alignment inevitably incurs an alignment tax: the model's pre-calibrated value system drifts significantly due to latent bias absorption from training data, while the fine-tuning process also causes severe hallucinations and semantic information loss in generated responses. To address this, we propose VISA (Value Injection via Shielded Adaptation), a closed-loop framework designed to navigate this trade-off. VISA's architecture features a high-precision value detector, a semantic-to-value translator, and a core value-rewriter. The value-rewriter is trained via Group Relative Policy Optimization (GRPO) with a composite reward function that simultaneously optimizes for fine-grained value precision, and the preservation of semantic integrity. By learning an optimal policy to balance these competing objectives, VISA effectively mitigates the alignment tax while staying loyal to the original knowledge. Our experiments demonstrate that this approach enables precise control over a model's value expression while maintaining its factual consistency and general capabilities, significantly outperforming both standard fine-tuning methods and prompting-based baselines, including GPT-4o.

Guoxi Zhang, Jiawei Chen, Tianzhuo Yang et al.

The increasing prevalence of large language models (LLMs) is influencing global value systems. However, these models frequently exhibit a pronounced WEIRD (Western, Educated, Industrialized, Rich, Democratic) cultural bias due to lack of attention to minority values. This monocultural perspective may reinforce dominant values and marginalize diverse cultural viewpoints, posing challenges for the development of equitable and inclusive AI systems. In this work, we introduce a systematic framework designed to boost fair and robust cross-cultural consensus among LLMs. We model consensus as a Nash Equilibrium and employ a game-theoretic negotiation method based on Policy-Space Response Oracles (PSRO) to simulate an organized cross-cultural negotiation process. To evaluate this approach, we construct regional cultural agents using data transformed from the World Values Survey (WVS). Beyond the conventional model-level evaluation method, We further propose two quantitative metrics, Perplexity-based Acceptence and Values Self-Consistency, to assess consensus outcomes. Experimental results indicate that our approach generates consensus of higher quality while ensuring more balanced compromise compared to baselines. Overall, it mitigates WEIRD bias by guiding agents toward convergence through fair and gradual negotiation steps.

Lirui Luo, Guoxi Zhang, Hongming Xu et al.

Neuro-symbolic reinforcement learning (NS-RL) has emerged as a promising paradigm for explainable decision-making, characterized by the interpretability of symbolic policies. NS-RL entails structured state representations for tasks with visual observations, but previous methods cannot refine the structured states with rewards due to a lack of efficiency. Accessibility also remains an issue, as extensive domain knowledge is required to interpret symbolic policies. In this paper, we present a neuro-symbolic framework for jointly learning structured states and symbolic policies, whose key idea is to distill the vision foundation model into an efficient perception module and refine it during policy learning. Moreover, we design a pipeline to prompt GPT-4 to generate textual explanations for the learned policies and decisions, significantly reducing users' cognitive load to understand the symbolic policies. We verify the efficacy of our approach on nine Atari tasks and present GPT-generated explanations for policies and decisions.

Guoxi Zhang, Han Bao, Hisashi Kashima

In preference-based reinforcement learning (PbRL), a reward function is learned from a type of human feedback called preference. To expedite preference collection, recent works have leveraged \emph{offline preferences}, which are preferences collected for some offline data. In this scenario, the learned reward function is fitted on the offline data. If a learning agent exhibits behaviors that do not overlap with the offline data, the learned reward function may encounter generalizability issues. To address this problem, the present study introduces a framework that consolidates offline preferences and \emph{virtual preferences} for PbRL, which are comparisons between the agent's behaviors and the offline data. Critically, the reward function can track the agent's behaviors using the virtual preferences, thereby offering well-aligned guidance to the agent. Through experiments on continuous control tasks, this study demonstrates the effectiveness of incorporating the virtual preferences in PbRL.

Guoxi Zhang, Hisashi Kashima

A shortcoming of batch reinforcement learning is its requirement for rewards in data, thus not applicable to tasks without reward functions. Existing settings for lack of reward, such as behavioral cloning, rely on optimal demonstrations collected from humans. Unfortunately, extensive expertise is required for ensuring optimality, which hinder the acquisition of large-scale data for complex tasks. This paper addresses the lack of reward in a batch reinforcement learning setting by learning a reward function from preferences. Generating preferences only requires a basic understanding of a task. Being a mental process, generating preferences is faster than performing demonstrations. So preferences can be collected at scale from non-expert humans using crowdsourcing. This paper tackles a critical challenge that emerged when collecting data from non-expert humans: the noise in preferences. A novel probabilistic model is proposed for modelling the reliability of labels, which utilizes labels collaboratively. Moreover, the proposed model smooths the estimation with a learned reward function. Evaluation on Atari datasets demonstrates the effectiveness of the proposed model, followed by an ablation study to analyze the relative importance of the proposed ideas.