Mengdi Li

Xufeng Zhao, Mengdi Li, Wenhao Lu et al.

Recent advancements in large language models have showcased their remarkable generalizability across various domains. However, their reasoning abilities still have significant room for improvement, especially when confronted with scenarios requiring multi-step reasoning. Although large language models possess extensive knowledge, their reasoning often fails to effectively utilize this knowledge to establish a coherent thinking paradigm. These models sometimes show hallucinations as their reasoning procedures are unconstrained by logical principles. Aiming at improving the zero-shot chain-of-thought reasoning ability of large language models, we propose LoT (Logical Thoughts), a self-improvement prompting framework that leverages principles rooted in symbolic logic, particularly Reductio ad Absurdum, to systematically verify and rectify the reasoning processes step by step. Experimental evaluations conducted on language tasks in diverse domains, including arithmetic, commonsense, symbolic, causal inference, and social problems, demonstrate the efficacy of enhanced reasoning by logic. The implementation code for LoT can be accessed at: https://github.com/xf-zhao/LoT.

Yuan Yao, Tianyu Yu, Ao Zhang et al.

Large-scale commonsense knowledge bases empower a broad range of AI applications, where the automatic extraction of commonsense knowledge (CKE) is a fundamental and challenging problem. CKE from text is known for suffering from the inherent sparsity and reporting bias of commonsense in text. Visual perception, on the other hand, contains rich commonsense knowledge about real-world entities, e.g., (person, can_hold, bottle), which can serve as promising sources for acquiring grounded commonsense knowledge. In this work, we present CLEVER, which formulates CKE as a distantly supervised multi-instance learning problem, where models learn to summarize commonsense relations from a bag of images about an entity pair without any human annotation on image instances. To address the problem, CLEVER leverages vision-language pre-training models for deep understanding of each image in the bag, and selects informative instances from the bag to summarize commonsense entity relations via a novel contrastive attention mechanism. Comprehensive experimental results in held-out and human evaluation show that CLEVER can extract commonsense knowledge in promising quality, outperforming pre-trained language model-based methods by 3.9 AUC and 6.4 mAUC points. The predicted commonsense scores show strong correlation with human judgment with a 0.78 Spearman coefficient. Moreover, the extracted commonsense can also be grounded into images with reasonable interpretability. The data and codes can be obtained at https://github.com/thunlp/CLEVER.

Xufeng Zhao, Mengdi Li, Cornelius Weber et al.

Programming robot behavior in a complex world faces challenges on multiple levels, from dextrous low-level skills to high-level planning and reasoning. Recent pre-trained Large Language Models (LLMs) have shown remarkable reasoning ability in few-shot robotic planning. However, it remains challenging to ground LLMs in multimodal sensory input and continuous action output, while enabling a robot to interact with its environment and acquire novel information as its policies unfold. We develop a robot interaction scenario with a partially observable state, which necessitates a robot to decide on a range of epistemic actions in order to sample sensory information among multiple modalities, before being able to execute the task correctly. Matcha (Multimodal environment chatting) agent, an interactive perception framework, is therefore proposed with an LLM as its backbone, whose ability is exploited to instruct epistemic actions and to reason over the resulting multimodal sensations (vision, sound, haptics, proprioception), as well as to plan an entire task execution based on the interactively acquired information. Our study demonstrates that LLMs can provide high-level planning and reasoning skills and control interactive robot behavior in a multimodal environment, while multimodal modules with the context of the environmental state help ground the LLMs and extend their processing ability. The project website can be found at https://matcha-agent.github.io.

Mengdi Li, Xufeng Zhao, Jae Hee Lee et al.

We study a class of reinforcement learning problems where the reward signals for policy learning are generated by an internal reward model that is dependent on and jointly optimized with the policy. This interdependence between the policy and the reward model leads to an unstable learning process because reward signals from an immature reward model are noisy and impede policy learning, and conversely, an under-optimized policy impedes reward estimation learning. We call this learning setting $\textit{Internally Rewarded Reinforcement Learning}$ (IRRL) as the reward is not provided directly by the environment but $\textit{internally}$ by a reward model. In this paper, we formally formulate IRRL and present a class of problems that belong to IRRL. We theoretically derive and empirically analyze the effect of the reward function in IRRL and based on these analyses propose the clipped linear reward function. Experimental results show that the proposed reward function can consistently stabilize the training process by reducing the impact of reward noise, which leads to faster convergence and higher performance compared with baselines in diverse tasks.

Wenhao Lu, Xufeng Zhao, Sven Magg et al.

Explaining the behaviour of intelligent agents learned by reinforcement learning (RL) to humans is challenging yet crucial due to their incomprehensible proprioceptive states, variational intermediate goals, and resultant unpredictability. Moreover, one-step explanations for RL agents can be ambiguous as they fail to account for the agent's future behaviour at each transition, adding to the complexity of explaining robot actions. By leveraging abstracted actions that map to task-specific primitives, we avoid explanations on the movement level. To further improve the transparency and explainability of robotic systems, we propose an explainable Q-Map learning framework that combines reward decomposition (RD) with abstracted action spaces, allowing for non-ambiguous and high-level explanations based on object properties in the task. We demonstrate the effectiveness of our framework through quantitative and qualitative analysis of two robotic scenarios, showcasing visual and textual explanations, from output artefacts of RD explanations, that are easy for humans to comprehend. Additionally, we demonstrate the versatility of integrating these artefacts with large language models (LLMs) for reasoning and interactive querying.

Kun Chu, Xufeng Zhao, Cornelius Weber et al.

Reinforcement Learning (RL) plays an important role in the robotic manipulation domain since it allows self-learning from trial-and-error interactions with the environment. Still, sample efficiency and reward specification seriously limit its potential. One possible solution involves learning from expert guidance. However, obtaining a human expert is impractical due to the high cost of supervising an RL agent, and developing an automatic supervisor is a challenging endeavor. Large Language Models (LLMs) demonstrate remarkable abilities to provide human-like feedback on user inputs in natural language. Nevertheless, they are not designed to directly control low-level robotic motions, as their pretraining is based on vast internet data rather than specific robotics data. In this paper, we introduce the Lafite-RL (Language agent feedback interactive Reinforcement Learning) framework, which enables RL agents to learn robotic tasks efficiently by taking advantage of LLMs' timely feedback. Our experiments conducted on RLBench tasks illustrate that, with simple prompt design in natural language, the Lafite-RL agent exhibits improved learning capabilities when guided by an LLM. It outperforms the baseline in terms of both learning efficiency and success rate, underscoring the efficacy of the rewards provided by an LLM.

Jiaye Lin, Mengdi Li, Xufeng Zhao et al.

Reward models trained through Reinforcement Learning from AI Feedback (RLAIF) methods frequently suffer from limited generalizability, which hinders the alignment performance of policy models. This challenge stems from various issues, including distribution shift, preference label noise, and mismatch of overly challenging samples with model capacity. In this paper, we aim to enhance the generalizability of reward models through a data-centric approach, driven by the insight that these issues are inherently intertwined from a uniform perspective of data difficulty. Accordingly, we propose a novel framework, Curriculum-RLAIF, which constructs preference pairs with varying difficulty levels and then produces a specific curriculum for reward model training. Comprehensive experimental results suggest that reward models trained with Curriculum-RLAIF achieve improved generalizability, boosting the alignment performance of policy models by a significant margin without incurring additional inference costs compared to various existing non-curriculum baselines. Further analysis and comparison with alternative strategies highlight the superiority of Curriculum-RLAIF in simplicity, efficiency, and effectiveness.

Xiaowen Sun, Matthias Kerzel, Mengdi Li et al.

Vision-language models (VLMs) have shown remarkable performance in various robotic tasks, as they can perceive visual information and understand natural language instructions. However, when applied to robotics, VLMs remain subject to a fundamental limitation inherent in large language models (LLMs): they struggle with numerical reasoning, particularly in object detection and object-state localization. To explore numerical reasoning as a regression task in VLMs, we propose a novel training strategy to adapt VLMs for object detection and object-state localization. This approach leverages box decoder outputs to compute an Auxiliary Regression Loss (ARL) during fine-tuning, while preserving standard sequence prediction at inference. We leverage this training strategy to develop StateVLM (State-aware Vision-Language Model), a novel model designed to perceive and learn fine-grained object representations, including precise localization of objects and their states, as well as graspable regions. Due to the lack of a benchmark for object-state affordance reasoning, we introduce an open-source benchmark, Object State Affordance Reasoning (OSAR), which contains 1,172 scenes with 7,746 individual objects and corresponding bounding boxes. Comparative experiments on adapted benchmarks (RefCOCO, RefCOCO+, and \mbox{RefCOCOg}) demonstrate that ARL improves model performance by an average of 1.6\% compared to models without ARL. Experiments on the OSAR benchmark further support this finding, showing that StateVLM with ARL achieves an average of 5.2\% higher performance than models without ARL. In particular, ARL is also important for the complex task of affordance reasoning in OSAR, where it enhances the consistency of model outputs.

Junchi Yao, Shu Yang, Jianhua Xu et al.

Large language models (LLMs) have made remarkable progress in various domains, yet they often suffer from repetitive text generation, a phenomenon we refer to as the "Repeat Curse". While previous studies have proposed decoding strategies to mitigate repetition, the underlying mechanism behind this issue remains insufficiently explored. In this work, we investigate the root causes of repetition in LLMs through the lens of mechanistic interpretability. Inspired by recent advances in Sparse Autoencoders (SAEs), which enable monosemantic feature extraction, we propose a novel approach, "Duplicatus Charm", to induce and analyze the Repeat Curse. Our method systematically identifies "Repetition Features" -the key model activations responsible for generating repetitive outputs. First, we locate the layers most involved in repetition through logit analysis. Next, we extract and stimulate relevant features using SAE-based activation manipulation. To validate our approach, we construct a repetition dataset covering token and paragraph level repetitions and introduce an evaluation pipeline to quantify the influence of identified repetition features. Furthermore, by deactivating these features, we have effectively mitigated the Repeat Curse. The source code of our work is publicly available at: https://github.com/kaustpradalab/repeat-curse-llm

Yuan Yao, Ao Zhang, Xu Han et al.

Scene graph generation aims to identify objects and their relations in images, providing structured image representations that can facilitate numerous applications in computer vision. However, scene graph models usually require supervised learning on large quantities of labeled data with intensive human annotation. In this work, we propose visual distant supervision, a novel paradigm of visual relation learning, which can train scene graph models without any human-labeled data. The intuition is that by aligning commonsense knowledge bases and images, we can automatically create large-scale labeled data to provide distant supervision for visual relation learning. To alleviate the noise in distantly labeled data, we further propose a framework that iteratively estimates the probabilistic relation labels and eliminates the noisy ones. Comprehensive experimental results show that our distantly supervised model outperforms strong weakly supervised and semi-supervised baselines. By further incorporating human-labeled data in a semi-supervised fashion, our model outperforms state-of-the-art fully supervised models by a large margin (e.g., 8.3 micro- and 7.8 macro-recall@50 improvements for predicate classification in Visual Genome evaluation). We make the data and code for this paper publicly available at https://github.com/thunlp/VisualDS.

Shu Yang, Shenzhe Zhu, Zeyu Wu et al.

We introduce Fraud-R1, a benchmark designed to evaluate LLMs' ability to defend against internet fraud and phishing in dynamic, real-world scenarios. Fraud-R1 comprises 8,564 fraud cases sourced from phishing scams, fake job postings, social media, and news, categorized into 5 major fraud types. Unlike previous benchmarks, Fraud-R1 introduces a multi-round evaluation pipeline to assess LLMs' resistance to fraud at different stages, including credibility building, urgency creation, and emotional manipulation. Furthermore, we evaluate 15 LLMs under two settings: 1. Helpful-Assistant, where the LLM provides general decision-making assistance, and 2. Role-play, where the model assumes a specific persona, widely used in real-world agent-based interactions. Our evaluation reveals the significant challenges in defending against fraud and phishing inducement, especially in role-play settings and fake job postings. Additionally, we observe a substantial performance gap between Chinese and English, underscoring the need for improved multilingual fraud detection capabilities.

Kun Chu, Xufeng Zhao, Cornelius Weber et al.

Although there has been rapid progress in endowing robots with the ability to solve complex manipulation tasks, generating control policies for bimanual robots to solve tasks involving two hands is still challenging because of the difficulties in effective temporal and spatial coordination. With emergent abilities in terms of step-by-step reasoning and in-context learning, Large Language Models (LLMs) have demonstrated promising potential in a variety of robotic tasks. However, the nature of language communication via a single sequence of discrete symbols makes LLM-based coordination in continuous space a particular challenge for bimanual tasks. To tackle this challenge, we present LAnguage-model-based Bimanual ORchestration (LABOR), an agent utilizing an LLM to analyze task configurations and devise coordination control policies for addressing long-horizon bimanual tasks. We evaluate our method through simulated experiments involving two classes of long-horizon tasks using the NICOL humanoid robot. Our results demonstrate that our method outperforms the baseline in terms of success rate. Additionally, we thoroughly analyze failure cases, offering insights into LLM-based approaches in bimanual robotic control and revealing future research trends. The project website can be found at http://labor-agent.github.io.

Shu Yang, Jiayuan Su, Han Jiang et al.

With the rise of large language models (LLMs), ensuring they embody the principles of being helpful, honest, and harmless (3H), known as Human Alignment, becomes crucial. While existing alignment methods like RLHF, DPO, etc., effectively fine-tune LLMs to match preferences in the preference dataset, they often lead LLMs to highly receptive human input and external evidence, even when this information is poisoned. This leads to a tendency for LLMs to be Adaptive Chameleons when external evidence conflicts with their parametric memory. This exacerbates the risk of LLM being attacked by external poisoned data, which poses a significant security risk to LLM system applications such as Retrieval-augmented generation (RAG). To address the challenge, we propose a novel framework: Dialectical Alignment (DA), which (1) utilizes AI feedback to identify optimal strategies for LLMs to navigate inter-context conflicts and context-memory conflicts with different external evidence in context window (i.e., different ratios of poisoned factual contexts); (2) constructs the SFT dataset as well as the preference dataset based on the AI feedback and strategies above; (3) uses the above datasets for LLM alignment to defense poisoned context attack while preserving the effectiveness of in-context knowledge editing. Our experiments show that the dialectical alignment model improves poisoned data attack defense by 20 and does not require any additional prompt engineering or prior declaration of ``you may be attacked`` to the LLMs' context window.

Mengdi Li, Yuming Liu, He Wang et al.

Code clone detection (CCD) supports software maintenance, refactoring, and security analysis. Although pre-trained models capture code semantics, most work reduces CCD to binary classification, overlooking the heterogeneity of clone types and the seven fine-grained categories in BigCloneBench. We present Full Model, a multimodal fusion framework that jointly integrates heuristic similarity priors from classical machine learning, structural signals from abstract syntax trees (ASTs), and deep semantic embeddings from CodeBERT into a single predictor. By fusing structural, statistical, and semantic representations, Full Model improves discrimination among fine-grained clone types while keeping inference cost practical. On the seven-class BigCloneBench benchmark, Full Model raises Macro-F1 from 0.695 to 0.875. Ablation studies show that using the primary model's probability distribution as a prior to guide selective arbitration by a large language model (LLM) substantially outperforms blind reclassification; arbitrating only ~0.2% of high-uncertainty samples yields an additional 0.3 absolute Macro-F1 gain. Overall, Full Model achieves an effective performance-cost trade-off for fine-grained CCD and offers a practical solution for large-scale industrial deployment.

Jiaming Zhang, Mingxi Lei, Meng Ding et al.

Reinforcement Learning with Human Feedback (RLHF) has emerged as an influential technique, enabling the alignment of large language models (LLMs) with human preferences. Despite the promising potential of RLHF, how to protect user preference privacy has become a crucial issue. Most previous work has focused on using differential privacy (DP) to protect the privacy of individual data. However, they have concentrated primarily on item-level privacy protection and have unsatisfactory performance for user-level privacy, which is more common in RLHF. This study proposes a novel framework, AUP-RLHF, which integrates user-level label DP into RLHF. We first show that the classical random response algorithm, which achieves an acceptable performance in item-level privacy, leads to suboptimal utility when in the user-level settings. We then establish a lower bound for the user-level label DP-RLHF and develop the AUP-RLHF algorithm, which guarantees $(\varepsilon, δ)$ user-level privacy and achieves an improved estimation error. Experimental results show that AUP-RLHF outperforms existing baseline methods in sentiment generation and summarization tasks, achieving a better privacy-utility trade-off.

Wenhao Lu, Xufeng Zhao, Thilo Fryen et al.

Reinforcement learning (RL) is a powerful technique for training intelligent agents, but understanding why these agents make specific decisions can be quite challenging. This lack of transparency in RL models has been a long-standing problem, making it difficult for users to grasp the reasons behind an agent's behaviour. Various approaches have been explored to address this problem, with one promising avenue being reward decomposition (RD). RD is appealing as it sidesteps some of the concerns associated with other methods that attempt to rationalize an agent's behaviour in a post-hoc manner. RD works by exposing various facets of the rewards that contribute to the agent's objectives during training. However, RD alone has limitations as it primarily offers insights based on sub-rewards and does not delve into the intricate cause-and-effect relationships that occur within an RL agent's neural model. In this paper, we present an extension of RD that goes beyond sub-rewards to provide more informative explanations. Our approach is centred on a causal learning framework that leverages information-theoretic measures for explanation objectives that encourage three crucial properties of causal factors: causal sufficiency, sparseness, and orthogonality. These properties help us distill the cause-and-effect relationships between the agent's states and actions or rewards, allowing for a deeper understanding of its decision-making processes. Our framework is designed to generate local explanations and can be applied to a wide range of RL tasks with multiple reward channels. Through a series of experiments, we demonstrate that our approach offers more meaningful and insightful explanations for the agent's action selections.

Liangyu Wang, Huanyi Xie, Xinhai Wang et al.

Group-based reinforcement learning algorithms such as Group Reward Policy Optimization (GRPO) have proven effective for fine-tuning large language models (LLMs) with human feedback. However, generating and storing multiple responses per prompt incurs substantial memory overhead, especially as the sample group size increases, limiting scalability under constrained hardware. We propose Infinite Sampling, a framework that enables efficient and stable GRPO training by decoupling group size from GPU memory usage. It consists of: (1) micro sampling groups that decompose large groups into memory-feasible rounds; (2) continuous sampling that interleaves generation across groups to improve utilization; and (3) a length-aware scheduler combining token-conditioned sequence length prediction with a two-stage plan: global grouping via FPTAS and runtime refill via SJF. Experiments show that our Micro Sampling Groups reduce peak memory usage by over 50% compared to full-group decoding (e.g., from 21.55 GB to 10.64 GB on Qwen3-1.7B). Building on this, Infinite Sampling improves throughput by over 25% compared to the naive micro sampling group method, reducing decoding steps while maintaining full-length completions and memory usage. Our hybrid scheduling ensures efficient and stable GRPO training with larger groups under realistic GPU memory constraints.

Tong Li, Shu Yang, Junchao Wu et al.

We present a comprehensive evaluation framework for assessing Large Language Models' (LLMs) capabilities in suicide prevention, focusing on two critical aspects: the Identification of Implicit Suicidal ideation (IIS) and the Provision of Appropriate Supportive responses (PAS). We introduce \ourdata, a novel dataset of 1,308 test cases built upon psychological frameworks including D/S-IAT and Negative Automatic Thinking, alongside real-world scenarios. Through extensive experiments with 8 widely used LLMs under different contextual settings, we find that current models struggle significantly with detecting implicit suicidal ideation and providing appropriate support, highlighting crucial limitations in applying LLMs to mental health contexts. Our findings underscore the need for more sophisticated approaches in developing and evaluating LLMs for sensitive psychological applications.

Mengdi Li, Guanqiao Chen, Xufeng Zhao et al.

Reward models (RMs), which are central to existing post-training methods, aim to align LLM outputs with human values by providing feedback signals during fine-tuning. However, existing RMs struggle to capture nuanced, user-specific preferences, especially under limited data and across diverse domains. Thus, we introduce PersRM-R1, the first reasoning-based reward modeling framework specifically designed to identify and represent personal factors from only one or a few personal exemplars. To address challenges including limited data availability and the requirement for robust generalization, our approach combines synthetic data generation with a two-stage training pipeline consisting of supervised fine-tuning followed by reinforcement fine-tuning. Experimental results demonstrate that PersRM-R1 outperforms existing models of similar size and matches the performance of much larger models in both accuracy and generalizability, paving the way for more effective personalized LLMs.

Lijie Hu, Liang Liu, Shu Yang et al.

Chain-of-Thought (CoT) holds a significant place in augmenting the reasoning performance for large language models (LLMs). While some studies focus on improving CoT accuracy through methods like retrieval enhancement, yet a rigorous explanation for why CoT achieves such success remains unclear. In this paper, we analyze CoT methods under two different settings by asking the following questions: (1) For zero-shot CoT, why does prompting the model with "let's think step by step" significantly impact its outputs? (2) For few-shot CoT, why does providing examples before questioning the model could substantially improve its reasoning ability? To answer these questions, we conduct a top-down explainable analysis from the Hopfieldian view and propose a Read-and-Control approach for controlling the accuracy of CoT. Through extensive experiments on seven datasets for three different tasks, we demonstrate that our framework can decipher the inner workings of CoT, provide reasoning error localization, and control to come up with the correct reasoning path.

Mengdi Li, Cornelius Weber, Matthias Kerzel et al.

Reasoning about potential occlusions is essential for robots to efficiently predict whether an object exists in an environment. Though existing work shows that a robot with active perception can achieve various tasks, it is still unclear if occlusion reasoning can be achieved. To answer this question, we introduce the task of robotic object existence prediction: when being asked about an object, a robot needs to move as few steps as possible around a table with randomly placed objects to predict whether the queried object exists. To address this problem, we propose a novel recurrent neural network model that can be jointly trained with supervised and reinforcement learning methods using a curriculum training strategy. Experimental results show that 1) both active perception and occlusion reasoning are necessary to successfully achieve the task; 2) the proposed model demonstrates a good occlusion reasoning ability by achieving a similar prediction accuracy to an exhaustive exploration baseline while requiring only about $10\%$ of the baseline's number of movement steps on average; and 3) the model generalizes to novel object combinations with a moderate loss of accuracy.