Jianfei Ma

Zhaoxin Feng, Zheng Chen, Jianfei Ma et al.

Alignment techniques often inadvertently induce sycophancy in LLMs. While prior studies studied this behaviour in direct-answer settings, the role of Chain-of-Thought (CoT) reasoning remains under-explored: does it serve as a logical constraint that mitigates sycophancy, or a tool for post-hoc rationalization that masks it? We evaluate a range of models across objective and subjective tasks to investigate the issue. Results show that reasoning generally reduces sycophancy in final decisions but also masks sycophancy in some samples, where models construct deceptive justifications through logical inconsistencies, calculation errors, and one-sided arguments etc. Furthermore, LLMs are more prone to sycophancy in subjective tasks and under authority-bias. Our mechanistic analysis on three open-source models reveals that the tendency of sycophancy is dynamic during the reasoning process rather than being pre-determined at the input stage.

Jianfei Ma

Temporal difference learning (TD) is a foundational concept in reinforcement learning (RL), aimed at efficiently assessing a policy's value function. TD($λ$), a potent variant, incorporates a memory trace to distribute the prediction error into the historical context. However, this approach often neglects the significance of historical states and the relative importance of propagating the TD error, influenced by challenges such as visitation imbalance or outcome noise. To address this, we propose a novel TD algorithm named discerning TD learning (DTD), which allows flexible emphasis functions$-$predetermined or adapted during training$-$to allocate efforts effectively across states. We establish the convergence properties of our method within a specific class of emphasis functions and showcase its promising potential for adaptation to deep RL contexts. Empirical results underscore that employing a judicious emphasis function not only improves value estimation but also expedites learning across diverse scenarios.

Jianfei Ma

Soft actor-critic is a successful successor over soft Q-learning. While lived under maximum entropy framework, their relationship is still unclear. In this paper, we prove that in the limit they converge to the same solution. This is appealing since it translates the optimization from an arduous to an easier way. The same justification can also be applied to other regularizers such as KL divergence.

Jianfei Ma

While on-policy algorithms are known for their stability, they often demand a substantial number of samples. In contrast, off-policy algorithms, which leverage past experiences, are considered sample-efficient but tend to exhibit instability. Can we develop an algorithm that harnesses the benefits of off-policy data while maintaining stable learning? In this paper, we introduce an actor-critic learning framework that harmonizes two data sources for both evaluation and control, facilitating rapid learning and adaptable integration with on-policy algorithms. This framework incorporates variance reduction mechanisms, including a unified advantage estimator (UAE) and a residual baseline, improving the efficacy of both on- and off-policy learning. Our empirical results showcase substantial enhancements in sample efficiency for on-policy algorithms, effectively bridging the gap to the off-policy approaches. It demonstrates the promise of our approach as a novel learning paradigm.

Jianfei Ma

Deep reinforcement learning was instigated with the presence of trust region methods, being scalable and efficient. However, the pessimism of such algorithms, among which it forces to constrain in a trust region by all means, has been proven to suppress the exploration and harm the performance. Exploratory algorithm such as SAC, while utilizes the entropy to encourage exploration, implicitly optimizing another objective yet. We first observed this inconsistency, and therefore put forward an analogous augmentation technique, which combines well with the on-policy algorithms, when a value critic is involved. Surprisingly, the proposed method consistently satisfies the soft policy improvement theorem, while being more extensible. As the analysis advises, it is crucial to control the temperature coefficient to balance the exploration and exploitation. Empirical tests on MuJoCo benchmark tasks show that the agent is heartened towards higher reward regions, and enjoys a finer performance. Furthermore, we verify the exploration bonus of our method on a set of custom environments.

Haoxi Li, Qinglin Hou, Jianfei Ma et al.

To navigate partially observable visual environments, recent VLM agents increasingly internalize world modeling capabilities into their policies via explicit CoT reasoning, enabling them to mentally simulate futures before acting. However, relying solely on passive reasoning over visited states is insufficient for sparse-reward tasks, as it lacks the epistemic drive to actively uncover the ``known unknown'' required for robust generalization. We ask: Can VLM agents actively find signals that challenge and refine their internal world model through curiosity-driven exploration? In this work, we propose GLANCE, a unified framework that bridges reasoning and exploration by grounding the agent's linguistic world model into the stable visual representations of an evolving target network. Crucially, GLANCE leverages the discrepancy between linguistic prediction and visual reality as an intrinsic curiosity signal within reinforcement learning, steering the agent to actively explore areas where its internal model is uncertain. Extensive experiments across a series of agentic tasks show the effectiveness of GLANCE, and demonstrate that aligning ``what the agent thinks'' with ``what the agent sees'' is key to solving complex or sparse agentic tasks.

Jianfei Ma

Future sequence represents the outcome after executing the action into the environment (i.e. the trajectory onwards). When driven by the information-theoretic concept of mutual information, it seeks maximally informative consequences. Explicit outcomes may vary across state, return, or trajectory serving different purposes such as credit assignment or imitation learning. However, the inherent nature of incorporating intrinsic motivation with reward maximization is often neglected. In this work, we propose a policy iteration scheme that seamlessly incorporates the mutual information, ensuring convergence to the optimal policy. Concurrently, a variational approach is introduced, which jointly learns the necessary quantity for estimating the mutual information and the dynamics model, providing a general framework for incorporating different forms of outcomes of interest. While we mainly focus on theoretical analysis, our approach opens the possibilities of leveraging intrinsic control with model learning to enhance sample efficiency and incorporate uncertainty of the environment into decision-making.

Pengxiang Ding, Jianfei Ma, Xinyang Tong et al.

This paper addresses the limitations of current humanoid robot control frameworks, which primarily rely on reactive mechanisms and lack autonomous interaction capabilities due to data scarcity. We propose Humanoid-VLA, a novel framework that integrates language understanding, egocentric scene perception, and motion control, enabling universal humanoid control. Humanoid-VLA begins with language-motion pre-alignment using non-egocentric human motion datasets paired with textual descriptions, allowing the model to learn universal motion patterns and action semantics. We then incorporate egocentric visual context through a parameter efficient video-conditioned fine-tuning, enabling context-aware motion generation. Furthermore, we introduce a self-supervised data augmentation strategy that automatically generates pseudoannotations directly derived from motion data. This process converts raw motion sequences into informative question-answer pairs, facilitating the effective use of large-scale unlabeled video data. Built upon whole-body control architectures, extensive experiments show that Humanoid-VLA achieves object interaction and environment exploration tasks with enhanced contextual awareness, demonstrating a more human-like capacity for adaptive and intelligent engagement.

Zhaoxin Feng, Jianfei Ma, Emmanuele Chersoni et al.

Autoregressive Large Language Models (LLMs) demonstrate exceptional performance in language understanding and generation. However, their application in text embedding tasks has been relatively slow, along with the analysis of their semantic representation in probing tasks, due to the constraints of the unidirectional attention mechanism. This paper aims to explore whether such constraints can be overcome by enabling bidirectional attention in LLMs. We tested different variants of the Llama architecture through additional training steps, progressively enabling bidirectional attention and unsupervised/supervised contrastive learning.

Jianfei Ma, Wee Sun Lee

At the boundary between the known and the unknown, an agent inevitably confronts the dilemma of whether to explore or to exploit. Epistemic uncertainty reflects such boundaries, representing systematic uncertainty due to limited knowledge. In this paper, we propose a Bayesian reinforcement learning (RL) algorithm, $\texttt{EUBRL}$, which leverages epistemic guidance to achieve principled exploration. This guidance adaptively reduces per-step regret arising from estimation errors. We establish nearly minimax-optimal regret and sample complexity guarantees for a class of sufficiently expressive priors in infinite-horizon discounted MDPs. Empirically, we evaluate $\texttt{EUBRL}$ on tasks characterized by sparse rewards, long horizons, and stochasticity. Results demonstrate that $\texttt{EUBRL}$ achieves superior sample efficiency, scalability, and consistency.

Ziqi Zhang, Jianfei Ma, Emmanuele Chersoni et al.

Classifiers are an important and defining feature of the Chinese language, and their correct prediction is key to numerous educational applications. Yet, whether the most popular Large Language Models (LLMs) possess proper knowledge the Chinese classifiers is an issue that has largely remain unexplored in the Natural Language Processing (NLP) literature. To address such a question, we employ various masking strategies to evaluate the LLMs' intrinsic ability, the contribution of different sentence elements, and the working of the attention mechanisms during prediction. Besides, we explore fine-tuning for LLMs to enhance the classifier performance. Our findings reveal that LLMs perform worse than BERT, even with fine-tuning. The prediction, as expected, greatly benefits from the information about the following noun, which also explains the advantage of models with a bidirectional attention mechanism such as BERT.