Tianqing Fang

Jiaxin Bai, Yue Guo, Yifei Dong et al.

Text-agent environments are typically modeled as partially observable Markov decision processes (POMDPs), assuming that the simulator's latent state and transition dynamics are hidden from the agent. Yet little work has examined whether executable code can be induced to serve as a world model for prediction and planning under partial observability. We introduce PatchWorld, a gradient-free framework that turns offline trajectories into executable Python world models through counterexample-guided code repair. Instead of predicting the next observation with a black-box model, PatchWorld induces symbolic belief-state programs whose action updates can be inspected, replayed, and locally patched. Across seven AgentGym environments, PatchWorld-Simple achieves the highest code-based planning score among evaluated methods, reaching 76.4\% macro success in live one-step lookahead while invoking no LLM calls inside the world-model prediction module itself. We further find that a human-specified residual-memory bias improves surface observation fidelity but weakens decision utility. This exposes a tradeoff in executable world models, since improving observation fidelity can come at the expense of action-discriminative dynamics, and vice versa. Code is available at https://github.com/HKBU-KnowComp/PatchWorld.

Tianqing Fang, Quyet V. Do, Hongming Zhang et al. · tencent-ai

Commonsense Knowledge Base (CSKB) Population aims at reasoning over unseen entities and assertions on CSKBs, and is an important yet hard commonsense reasoning task. One challenge is that it requires out-of-domain generalization ability as the source CSKB for training is of a relatively smaller scale (1M) while the whole candidate space for population is way larger (200M). We propose PseudoReasoner, a semi-supervised learning framework for CSKB population that uses a teacher model pre-trained on CSKBs to provide pseudo labels on the unlabeled candidate dataset for a student model to learn from. The teacher can be a generative model rather than restricted to discriminative models as previous works. In addition, we design a new filtering procedure for pseudo labels based on influence function and the student model's prediction to further improve the performance. The framework can improve the backbone model KG-BERT (RoBERTa-large) by 3.3 points on the overall performance and especially, 5.3 points on the out-of-domain performance, and achieves the state-of-the-art. Codes and data are available at https://github.com/HKUST-KnowComp/PseudoReasoner.

Chunkit Chan, Jiayang Cheng, Weiqi Wang et al. · tencent-ai

This paper aims to quantitatively evaluate the performance of ChatGPT, an interactive large language model, on inter-sentential relations such as temporal relations, causal relations, and discourse relations. Given ChatGPT's promising performance across various tasks, we proceed to carry out thorough evaluations on the whole test sets of 11 datasets, including temporal and causal relations, PDTB2.0-based, and dialogue-based discourse relations. To ensure the reliability of our findings, we employ three tailored prompt templates for each task, including the zero-shot prompt template, zero-shot prompt engineering (PE) template, and in-context learning (ICL) prompt template, to establish the initial baseline scores for all popular sentence-pair relation classification tasks for the first time. Through our study, we discover that ChatGPT exhibits exceptional proficiency in detecting and reasoning about causal relations, albeit it may not possess the same level of expertise in identifying the temporal order between two events. While it is capable of identifying the majority of discourse relations with existing explicit discourse connectives, the implicit discourse relation remains a formidable challenge. Concurrently, ChatGPT demonstrates subpar performance in the dialogue discourse parsing task that requires structural understanding in a dialogue before being aware of the discourse relation.

Haochen Shi, Weiqi Wang, Tianqing Fang et al. · tencent-ai

Zero-shot commonsense Question-Answering (QA) requires models to reason about general situations beyond specific benchmarks. State-of-the-art approaches fine-tune language models on QA pairs constructed from CommonSense Knowledge Bases (CSKBs) to equip the models with more commonsense knowledge in a QA context. However, current QA synthesis protocols may introduce noise from the CSKBs and generate ungrammatical questions and false negative options, which impede the model's ability to generalize. To address these issues, we propose QADYNAMICS, a training dynamics-driven framework for QA diagnostics and refinement. Our approach analyzes the training dynamics of each QA pair at both the question level and option level, discarding machine-detectable artifacts by removing uninformative QA pairs and mislabeled or false-negative options. Extensive experiments demonstrate the effectiveness of our approach, which outperforms all baselines while using only 33% of the synthetic data, even including LLMs such as ChatGPT. Moreover, expert evaluations confirm that our framework significantly improves the quality of QA synthesis. Our codes and model checkpoints are available at https://github.com/HKUST-KnowComp/QaDynamics.

Ying Su, Zihao Wang, Tianqing Fang et al. · tencent-ai

Commonsense reasoning tasks such as commonsense knowledge graph completion and commonsense question answering require powerful representation learning. In this paper, we propose to learn commonsense knowledge representation by MICO, a Multi-alternative contrastve learning framework on COmmonsense knowledge graphs (MICO). MICO generates the commonsense knowledge representation by contextual interaction between entity nodes and relations with multi-alternative contrastive learning. In MICO, the head and tail entities in an $(h,r,t)$ knowledge triple are converted to two relation-aware sequence pairs (a premise and an alternative) in the form of natural language. Semantic representations generated by MICO can benefit the following two tasks by simply comparing the distance score between the representations: 1) zero-shot commonsense question answering task; 2) inductive commonsense knowledge graph completion task. Extensive experiments show the effectiveness of our method.

Mutian He, Tianqing Fang, Weiqi Wang et al. · tencent-ai

Conceptualization, or viewing entities and situations as instances of abstract concepts in mind and making inferences based on that, is a vital component in human intelligence for commonsense reasoning. Despite recent progress in artificial intelligence to acquire and model commonsense attributed to neural language models and commonsense knowledge graphs (CKGs), conceptualization is yet to be introduced thoroughly, making current approaches ineffective to cover knowledge about countless diverse entities and situations in the real world. To address the problem, we thoroughly study the role of conceptualization in commonsense reasoning, and formulate a framework to replicate human conceptual induction by acquiring abstract knowledge about events regarding abstract concepts, as well as higher-level triples or inferences upon them. We then apply the framework to ATOMIC, a large-scale human-annotated CKG, aided by the taxonomy Probase. We annotate a dataset on the validity of contextualized conceptualizations from ATOMIC on both event and triple levels, develop a series of heuristic rules based on linguistic features, and train a set of neural models to generate and verify abstract knowledge. Based on these components, a pipeline to acquire abstract knowledge is built. A large abstract CKG upon ATOMIC is then induced, ready to be instantiated to infer about unseen entities or situations. Finally, we empirically show the benefits of augmenting CKGs with abstract knowledge in downstream tasks like commonsense inference and zero-shot commonsense QA.

Ziqian Zeng, Weimin Ni, Tianqing Fang et al. · tencent-ai

Solving text classification in a weakly supervised manner is important for real-world applications where human annotations are scarce. In this paper, we propose to query a masked language model with cloze style prompts to obtain supervision signals. We design a prompt which combines the document itself and "this article is talking about [MASK]." A masked language model can generate words for the [MASK] token. The generated words which summarize the content of a document can be utilized as supervision signals. We propose a latent variable model to learn a word distribution learner which associates generated words to pre-defined categories and a document classifier simultaneously without using any annotated data. Evaluation on three datasets, AGNews, 20Newsgroups, and UCINews, shows that our method can outperform baselines by 2%, 4%, and 3%.

Zhaowei Wang, Hongming Zhang, Tianqing Fang et al. · tencent-ai

In this paper, we propose a new task of sub-event generation for an unseen process to evaluate the understanding of the coherence of sub-event actions and objects. To solve the problem, we design SubeventWriter, a sub-event sequence generation framework with a coherence controller. Given an unseen process, the framework can iteratively construct the sub-event sequence by generating one sub-event at each iteration. We also design a very effective coherence controller to decode more coherent sub-events. As our extensive experiments and analysis indicate, SubeventWriter can generate more reliable and meaningful sub-event sequences for unseen processes.

Tianqing Fang, Wenxuan Zhou, Fangyu Liu et al. · deepmind, tencent-ai

Data Augmentation (DA) is frequently used to provide additional training data without extra human annotation automatically. However, data augmentation may introduce noisy data that impairs training. To guarantee the quality of augmented data, existing methods either assume no noise exists in the augmented data and adopt consistency training or use simple heuristics such as training loss and diversity constraints to filter out "noisy" data. However, those filtered examples may still contain useful information, and dropping them completely causes a loss of supervision signals. In this paper, based on the assumption that the original dataset is cleaner than the augmented data, we propose an on-the-fly denoising technique for data augmentation that learns from soft augmented labels provided by an organic teacher model trained on the cleaner original data. To further prevent overfitting on noisy labels, a simple self-regularization module is applied to force the model prediction to be consistent across two distinct dropouts. Our method can be applied to general augmentation techniques and consistently improve the performance on both text classification and question-answering tasks.

Tianqing Fang, Quyet V. Do, Zihao Zheng et al. · tencent-ai

Commonsense Knowledge Bases (CSKB) Population, which aims at automatically expanding knowledge in CSKBs with external resources, is an important yet hard task in NLP. Fang et al. (2021a) proposed a CSKB Population (CKBP) framework with an evaluation set CKBP v1. However, CKBP v1 relies on crowdsourced annotations that suffer from a considerable number of mislabeled answers, and the evaluationset lacks alignment with the external knowledge source due to random sampling. In this paper, we introduce CKBP v2, a new high-quality CSKB Population evaluation set that addresses the two aforementioned issues by employing domain experts as annotators and incorporating diversified adversarial samples to make the evaluation data more representative. We show that CKBP v2 serves as a challenging and representative evaluation dataset for the CSKB Population task, while its development set aids in selecting a population model that leads to improved knowledge acquisition for downstream commonsense reasoning. A better population model can also help acquire more informative commonsense knowledge as additional supervision signals for both generative commonsense inference and zero-shot commonsense question answering. Specifically, the question-answering model based on DeBERTa-v3-large (He et al., 2023b) even outperforms powerful large language models in a zero-shot setting, including ChatGPT and GPT-3.5.

Cheng Jiayang, Lin Qiu, Tsz Ho Chan et al. · tencent-ai

Analogy-making between narratives is crucial for human reasoning. In this paper, we evaluate the ability to identify and generate analogies by constructing a first-of-its-kind large-scale story-level analogy corpus, \textsc{StoryAnalogy}, which contains 24K story pairs from diverse domains with human annotations on two similarities from the extended Structure-Mapping Theory. We design a set of tests on \textsc{StoryAnalogy}, presenting the first evaluation of story-level analogy identification and generation. Interestingly, we find that the analogy identification tasks are incredibly difficult not only for sentence embedding models but also for the recent large language models (LLMs) such as ChatGPT and LLaMa. ChatGPT, for example, only achieved around 30% accuracy in multiple-choice questions (compared to over 85% accuracy for humans). Furthermore, we observe that the data in \textsc{StoryAnalogy} can improve the quality of analogy generation in LLMs, where a fine-tuned FlanT5-xxl model achieves comparable performance to zero-shot ChatGPT.

Sehyun Choi, Tianqing Fang, Zhaowei Wang et al. · tencent-ai

Large Language Models (LLMs) have demonstrated remarkable human-level natural language generation capabilities. However, their potential to generate misinformation, often called the hallucination problem, poses a significant risk to their deployment. A common approach to address this issue is to retrieve relevant knowledge and fine-tune the LLM with the knowledge in its input. Unfortunately, this method incurs high training costs and may cause catastrophic forgetting for multi-tasking models. To overcome these limitations, we propose a knowledge-constrained decoding method called KCTS (Knowledge-Constrained Tree Search), which guides a frozen LM to generate text aligned with the reference knowledge at each decoding step using a knowledge classifier score and MCTS (Monte-Carlo Tree Search). To adapt the sequence-level knowledge classifier to token-level guidance, we also propose a novel token-level hallucination detection method called RIPA (Reward Inflection Point Approximation). Our empirical results on knowledge-grounded dialogue and abstractive summarization demonstrate the strength of KCTS as a plug-and-play, model-agnostic decoding method that can effectively reduce hallucinations in natural language generation.

Tianshi Zheng, Rui Wang, Xiyun Li et al. · tencent-ai

Frontier scientific reasoning is rapidly emerging as a key foundation for advancing AI agents in automated scientific discovery. Deep research agents offer a promising approach to this challenge. These models develop robust problem-solving capabilities through post-training on information-seeking tasks, which are typically curated via knowledge graph construction or iterative web browsing. However, these strategies face inherent limitations in frontier science, where domain-specific knowledge is scattered across sparse and heterogeneous academic sources, and problem solving requires sophisticated computation and reasoning far beyond factual recall. To bridge this gap, we introduce SciResearcher, a fully automated agentic framework for frontier-science data construction. SciResearcher synthesizes diverse conceptual and computational tasks grounded in academic evidence, while eliciting information acquisition, tool-integrated reasoning, and long-horizon capabilities. Leveraging the curated data for supervised fine-tuning and agentic reinforcement learning, we develop SciResearcher-8B, an agent foundation model that achieves 19.46% on the HLE-Bio/Chem-Gold benchmark, establishing a new state of the art at its parameter scale and surpassing several larger proprietary agents. It further achieves 13-15% absolute gains on SuperGPQA-Hard-Biology and TRQA-Literature benchmarks. Overall, SciResearcher introduces a new paradigm for automated data construction for frontier scientific reasoning and offers a scalable path toward future scientific agents.

Yuxuan Wan, Tianqing Fang, Zaitang Li et al.

Recent advances in Deep Research Agents (DRAs) are transforming automated knowledge discovery and problem-solving. While the majority of existing efforts focus on enhancing policy capabilities via post-training, we propose an alternative paradigm: self-evolving the agent's ability by iteratively verifying the policy model's outputs, guided by meticulously crafted rubrics. This approach gives rise to the inference-time scaling of verification, wherein an agent self-improves by evaluating its generated answers to produce iterative feedback and refinements. We derive the rubrics based on an automatically constructed DRA Failure Taxonomy, which systematically classifies agent failures into five major categories and thirteen sub-categories. We present DeepVerifier, a rubrics-based outcome reward verifier that leverages the asymmetry of verification and outperforms vanilla agent-as-judge and LLM judge baselines by 12%-48% in meta-evaluation F1 score. To enable practical self-evolution, DeepVerifier integrates as a plug-and-play module during test-time inference. The verifier produces detailed rubric-based feedback, which is fed back to the agent for iterative bootstrapping, refining responses without additional training. This test-time scaling delivers 8%-11% accuracy gains on challenging subsets of GAIA and XBench-DeepResearch when powered by capable closed-source LLMs. Finally, to support open-source advancement, we release DeepVerifier-4K, a curated supervised fine-tuning dataset of 4,646 high-quality agent steps focused on DRA verification. These examples emphasize reflection and self-critique, enabling open models to develop robust verification capabilities.

Zhaowei Wang, Haochen Shi, Weiqi Wang et al. · tencent-ai

Cognitive research indicates that abstraction ability is essential in human intelligence, which remains under-explored in language models. In this paper, we present AbsPyramid, a unified entailment graph of 221K textual descriptions of abstraction knowledge. While existing resources only touch nouns or verbs within simplified events or specific domains, AbsPyramid collects abstract knowledge for three components of diverse events to comprehensively evaluate the abstraction ability of language models in the open domain. Experimental results demonstrate that current LLMs face challenges comprehending abstraction knowledge in zero-shot and few-shot settings. By training on our rich abstraction knowledge, we find LLMs can acquire basic abstraction abilities and generalize to unseen events. In the meantime, we empirically show that our benchmark is comprehensive to enhance LLMs across two previous abstraction tasks.

Wenkai Wang, Xiyun Li, Hongcan Guo et al. · tencent-ai

Graphical User Interface (GUI) grounding requires mapping natural language instructions to precise pixel coordinates. However, due to visually homogeneous elements and dense layouts, models typically grasp semantic intent yet struggle with achieving precise localization. While scaling sampling attempts (Pass@k) reveals potential gains, static self-consistency strategies derived from geometric clustering often yield limited improvements, as the model's predictions tend to be spatially dispersed. In this paper, we propose replacing static consistency strategies with a learnable selection mechanism that selects the optimal target by critiquing its own proposals rendered on the screenshot. Given the significant disparity between the model's grounding and critiquing capabilities, we propose a co-evolving Propose-then-Critic framework. To jointly optimize these, we introduce a maturity-aware adaptive co-evolutionary reinforcement learning paradigm. This approach dynamically balances the training objectives of proposer and critic, where the diversity of the proposer's outputs enhances critic robustness, while the critic's maturing discrimination capability conversely unlocks the proposer's potential for extensive spatial exploration, fostering the mutual reinforcement and co-evolution of both capabilities, thereby ensuring generalizability to adapt to diverse and complex interface layouts. Extensive experiments over 6 benchmarks show that our method significantly enhances both grounding accuracy and critic reliability.

Qifan Zhang, Dongyang Ma, Tianqing Fang et al. · tencent-ai

Most agents today ``self-evolve'' by following rewards and rules defined by humans. However, this process remains fundamentally dependent on external supervision; without human guidance, the evolution stops. In this work, we train agents to possess an intrinsic meta-evolution capability to spontaneously learn about unseen environments prior to task execution. To instill this ability, we design an outcome-based reward mechanism that measures how much an agent's self-generated world knowledge improves its success rate on downstream tasks. This reward signal is used exclusively during the training phase to teach the model how to explore and summarize effectively. At inference time, the agent requires no external rewards or human instructions. It spontaneously performs native self-evolution to adapt to unknown environments using its internal parameters. When applied to Qwen3-30B and Seed-OSS-36B, this shift to native evolution yields a 20% performance increase on WebVoyager and WebWalker. Most strikingly, the generated world knowledge even enables a compact 14B Qwen3 model to outperform the unassisted Gemini-2.5-Flash, establishing a new paradigm for truly evolving agents.

Xingbo Wang, Renfei Huang, Zhihua Jin et al. · tencent-ai

Recently, large pretrained language models have achieved compelling performance on commonsense benchmarks. Nevertheless, it is unclear what commonsense knowledge the models learn and whether they solely exploit spurious patterns. Feature attributions are popular explainability techniques that identify important input concepts for model outputs. However, commonsense knowledge tends to be implicit and rarely explicitly presented in inputs. These methods cannot infer models' implicit reasoning over mentioned concepts. We present CommonsenseVIS, a visual explanatory system that utilizes external commonsense knowledge bases to contextualize model behavior for commonsense question-answering. Specifically, we extract relevant commonsense knowledge in inputs as references to align model behavior with human knowledge. Our system features multi-level visualization and interactive model probing and editing for different concepts and their underlying relations. Through a user study, we show that CommonsenseVIS helps NLP experts conduct a systematic and scalable visual analysis of models' relational reasoning over concepts in different situations.

Tianshi Zheng, Jiaxin Bai, Yicheng Wang et al. · tencent-ai

While large language models (LLMs) have demonstrated impressive capabilities across various natural language processing tasks by acquiring rich factual knowledge from their broad training data, their ability to synthesize and logically reason with this knowledge in complex ways remains underexplored. In this work, we present a systematic evaluation of state-of-the-art LLMs' complex logical reasoning abilities through a novel benchmark of automatically generated complex reasoning questions over general domain and biomedical knowledge graphs. Our extensive experiments, employing diverse in-context learning techniques, reveal that LLMs excel at reasoning over general world knowledge but face significant challenges with specialized domain-specific knowledge. We find that prompting with explicit Chain-of-Thought demonstrations can substantially improve LLM performance on complex logical reasoning tasks with diverse logical operations. Interestingly, our controlled evaluations uncover an asymmetry where LLMs display proficiency at set union operations, but struggle considerably with set intersections - a key building block of logical reasoning. To foster further work, we will publicly release our evaluation benchmark and code.

Mukai Li, Qingcheng Zeng, Tianqing Fang et al.

As large language model agents tackle increasingly complex long-horizon tasks, effective post-training becomes critical. Prior work faces fundamental challenges: outcome-only rewards fail to precisely attribute credit to intermediate steps, estimated step-level rewards introduce systematic noise, and Monte Carlo sampling approaches for step reward estimation incur prohibitive computational cost. Inspired by findings that only a small fraction of high-entropy tokens drive effective RL for reasoning, we propose Critical Step Optimization (CSO), which focuses preference learning on verified critical steps, decision points where alternate actions demonstrably flip task outcomes from failure to success. Crucially, our method starts from failed policy trajectories rather than expert demonstrations, directly targeting the policy model's weaknesses. We use a process reward model (PRM) to identify candidate critical steps, leverage expert models to propose high-quality alternatives, then continue execution from these alternatives using the policy model itself until task completion. Only alternatives that the policy successfully executes to correct outcomes are verified and used as DPO training data, ensuring both quality and policy reachability. This yields fine-grained, verifiable supervision at critical decisions while avoiding trajectory-level coarseness and step-level noise. Experiments on GAIA-Text-103 and XBench-DeepSearch show that CSO achieves 37% and 26% relative improvement over the SFT baseline and substantially outperforms other post-training methods, while requiring supervision at only 16% of trajectory steps. This demonstrates the effectiveness of selective verification-based learning for agent post-training.

Weiqi Wang, Tianqing Fang, Chunyang Li et al. · tencent-ai

The sequential process of conceptualization and instantiation is essential to generalizable commonsense reasoning as it allows the application of existing knowledge to unfamiliar scenarios. However, existing works tend to undervalue the step of instantiation and heavily rely on pre-built concept taxonomies and human annotations to collect both types of knowledge, resulting in a lack of instantiated knowledge to complete reasoning, high cost, and limited scalability. To tackle these challenges, we introduce CANDLE, a distillation framework that iteratively performs contextualized conceptualization and instantiation over commonsense knowledge bases by instructing large language models to generate both types of knowledge with critic filtering. By applying CANDLE to ATOMIC, we construct a comprehensive knowledge base comprising six million conceptualizations and instantiated commonsense knowledge triples. Both types of knowledge are firmly rooted in the original ATOMIC dataset, and intrinsic evaluations demonstrate their exceptional quality and diversity. Empirical results indicate that distilling CANDLE on student models provides benefits across four downstream tasks. Our code, data, and models are publicly available at https://github.com/HKUST-KnowComp/CANDLE.

Hongliang He, Wenlin Yao, Kaixin Ma et al. · tencent-ai

The rapid development of large language and multimodal models has sparked significant interest in using proprietary models, such as GPT-4o, to develop autonomous agents capable of handling real-world scenarios like web navigation. Although recent open-source efforts have tried to equip agents with the ability to explore environments and continuously improve over time, they are building text-only agents in synthetic environments where the reward signals are clearly defined. Such agents struggle to generalize to realistic settings that require multimodal perception abilities and lack ground-truth signals. In this paper, we introduce an open-source framework designed to facilitate the development of multimodal web agent that can autonomously conduct real-world exploration and improve itself. We first train the base model with imitation learning to gain the basic abilities. We then let the agent explore the open web and collect feedback on its trajectories. After that, it further improves its policy by learning from well-performing trajectories judged by another general-purpose model. This exploration-feedback-optimization cycle can continue for several iterations. Experimental results show that our web agent successfully improves itself after each iteration, demonstrating strong performance across multiple test sets.

Wenhao Yu, Zhenwen Liang, Chengsong Huang et al.

AI self-evolution has long been envisioned as a path toward superintelligence, where models autonomously acquire, refine, and internalize knowledge from their own learning experiences. Yet in practice, unguided self-evolving systems often plateau quickly or even degrade as training progresses. These failures arise from issues such as concept drift, diversity collapse, and mis-evolution, as models reinforce their own biases and converge toward low-entropy behaviors. To enable models to self-evolve in a stable and controllable manner while minimizing reliance on human supervision, we introduce R-Few, a guided Self-Play Challenger-Solver framework that incorporates lightweight human oversight through in-context grounding and mixed training. At each iteration, the Challenger samples a small set of human-labeled examples to guide synthetic question generation, while the Solver jointly trains on human and synthetic examples under an online, difficulty-based curriculum. Across math and general reasoning benchmarks, R-Few achieves consistent and iterative improvements. For example, Qwen3-8B-Base improves by +3.0 points over R-Zero on math tasks and achieves performance on par with General-Reasoner, despite the latter being trained on 20 times more human data. Ablation studies confirm the complementary contributions of grounded challenger training and curriculum-based solver training, and further analysis shows that R-Few mitigates drift, yielding more stable and controllable co-evolutionary dynamics.

Tianqing Fang, Zeming Chen, Yangqiu Song et al. · allen-ai, tencent-ai

Event commonsense reasoning requires the ability to reason about the relationship between events, as well as infer implicit context underlying that relationship. However, data scarcity makes it challenging for language models to learn to generate commonsense inferences for contexts and questions involving interactions between complex events. To address this demand, we present COM2 (COMplex COMmonsense), a new dataset created by sampling multi-hop logical queries (e.g., the joint effect or cause of both event A and B, or the effect of the effect of event C) from an existing commonsense knowledge graph (CSKG), and verbalizing them using handcrafted rules and large language models into multiple-choice and text generation questions. Our experiments show that language models trained on COM2 exhibit significant improvements in complex reasoning ability, resulting in enhanced zero-shot performance in both in-domain and out-of-domain tasks for question answering and generative commonsense reasoning, without expensive human annotations. Code and data are available at https://github.com/tqfang/complex-commonsense-reasoning.

Tianqing Fang, Hongming Zhang, Zhisong Zhang et al. · tencent-ai

Agent self-improvement, where the backbone Large Language Model (LLM) of the agent are trained on trajectories sampled autonomously based on their own policies, has emerged as a promising approach for enhancing performance. Recent advancements, particularly in web environments, face a critical limitation: their performance will reach a stagnation point during autonomous learning cycles, hindering further improvement. We argue that this stems from limited exploration of the web environment and insufficient exploitation of pre-trained web knowledge in LLMs. To improve the performance of self-improvement, we propose a novel framework that introduces a co-evolving World Model LLM. This world model predicts the next observation based on the current observation and action within the web environment. Leveraging LLMs' pretrained knowledge of abundant web content, the World Model serves dual roles: (1) as a virtual web server generating self-instructed training data to continuously refine the agent's policy, and (2) as an imagination engine during inference, enabling look-ahead simulation to guide action selection for the agent LLM. Experiments in real-world web environments (Mind2Web-Live, WebVoyager, and GAIA-web) show a 10% performance gain over existing self-evolving agents, demonstrating the efficacy and generalizability of our approach, without using any distillation from more powerful close-sourced models. Our work establishes the necessity of integrating world models into autonomous agent frameworks to unlock sustained adaptability. Code is available at https://github.com/Tencent/SelfEvolvingAgent

Tianqing Fang, Zhisong Zhang, Xiaoyang Wang et al. · tencent-ai

General AI Agents are increasingly recognized as foundational frameworks for the next generation of artificial intelligence, enabling complex reasoning, web interaction, coding, and autonomous research capabilities. However, current agent systems are either closed-source or heavily reliant on a variety of paid APIs and proprietary tools, limiting accessibility and reproducibility for the research community. In this work, we present \textbf{Cognitive Kernel-Pro}, a fully open-source and (to the maximum extent) free multi-module agent framework designed to democratize the development and evaluation of advanced AI agents. Within Cognitive Kernel-Pro, we systematically investigate the curation of high-quality training data for Agent Foundation Models, focusing on the construction of queries, trajectories, and verifiable answers across four key domains: web, file, code, and general reasoning. Furthermore, we explore novel strategies for agent test-time reflection and voting to enhance agent robustness and performance. We evaluate Cognitive Kernel-Pro on GAIA, achieving state-of-the-art results among open-source and free agents. Notably, our 8B-parameter open-source model surpasses previous leading systems such as WebDancer and WebSailor, establishing a new performance standard for accessible, high-capability AI agents. Code is available at https://github.com/Tencent/CognitiveKernel-Pro

Xiongye Xiao, Gengshuo Liu, Gaurav Gupta et al.

Integrating and processing information from various sources or modalities are critical for obtaining a comprehensive and accurate perception of the real world. Drawing inspiration from neuroscience, we develop the Information-Theoretic Hierarchical Perception (ITHP) model, which utilizes the concept of information bottleneck. Distinct from most traditional fusion models that aim to incorporate all modalities as input, our model designates the prime modality as input, while the remaining modalities act as detectors in the information pathway. Our proposed perception model focuses on constructing an effective and compact information flow by achieving a balance between the minimization of mutual information between the latent state and the input modal state, and the maximization of mutual information between the latent states and the remaining modal states. This approach leads to compact latent state representations that retain relevant information while minimizing redundancy, thereby substantially enhancing the performance of downstream tasks. Experimental evaluations on both the MUStARD and CMU-MOSI datasets demonstrate that our model consistently distills crucial information in multimodal learning scenarios, outperforming state-of-the-art benchmarks.

Ce Zhang, Kaixin Ma, Tianqing Fang et al. · tencent-ai

Recent Large Vision-Language Models (LVLMs) have advanced multi-modal understanding by incorporating finer-grained visual perception and encoding. However, such methods incur significant computational costs due to longer visual token sequences, posing challenges for real-time deployment. To mitigate this, prior studies have explored pruning unimportant visual tokens either at the output layer of the visual encoder or at the early layers of the language model. In this work, we revisit these design choices and reassess their effectiveness through comprehensive empirical studies of how visual tokens are processed throughout the visual encoding and language decoding stages. Guided by these insights, we propose VScan, a two-stage visual token reduction framework that addresses token redundancy by: (1) integrating complementary global and local scans with token merging during visual encoding, and (2) introducing pruning at intermediate layers of the language model. Extensive experimental results across four LVLMs validate the effectiveness of VScan in accelerating inference and demonstrate its superior performance over current state-of-the-arts on sixteen benchmarks. Notably, when applied to LLaVA-NeXT-7B, VScan achieves a 2.91$\times$ speedup in prefilling and a 10$\times$ reduction in FLOPs, while retaining 95.4\% of the original performance. Code is available at https://github.com/Tencent/SelfEvolvingAgent/tree/main/VScan.

Xinyu Geng, Yanjing Xiao, Yuyang Zhang et al.

Deep research agents integrate fragmented evidence through multi-step tool use. BrowseComp offers a text-only testbed for such agents, but existing multimodal benchmarks rarely require both weak visual cues composition and BrowseComp-style multi-hop verification. Geolocation is a natural testbed because answers depend on combining multiple ambiguous visual cues and validating them with open-web evidence. Thus, we introduce GeoBrowse, a geolocation benchmark that combines visual reasoning with knowledge-intensive multi-hop queries. Level 1 tests extracting and composing fragmented visual cues, and Level 2 increases query difficulty by injecting long-tail knowledge and obfuscating key entities. To support evaluation, we provide an agentic workflow GATE with five think-with-image tools and four knowledge-intensive tools, and release expert-annotated stepwise traces grounded in verifiable evidence for trajectory-level analysis. Experiments show that GATE outperforms direct inference and open-source agents, indicating that no-tool, search-only or image-only setups are insufficient. Gains come from coherent, level-specific tool-use plans rather than more tool calls, as they more reliably reach annotated key evidence steps and make fewer errors when integrating into the final decision. The GeoBrowse bernchmark and codes are provided in https://github.com/ornamentt/GeoBrowse

Liyu Zhang, Weiqi Wang, Tianqing Fang et al. · tencent-ai

Knowledge Editing (KE) aims to adjust a Large Language Model's (LLM) internal representations and parameters to correct inaccuracies and improve output consistency without incurring the computational expense of re-training the entire model. However, editing commonsense knowledge still faces difficulties, including limited knowledge coverage in existing resources, the infeasibility of annotating labels for an overabundance of commonsense knowledge, and the strict knowledge formats of current editing methods. In this paper, we address these challenges by presenting ConceptEdit, a framework that integrates conceptualization and instantiation into the KE pipeline for LLMs to enhance their commonsense reasoning capabilities. ConceptEdit dynamically diagnoses implausible commonsense knowledge within an LLM using another verifier LLM and augments the source knowledge to be edited with conceptualization for stronger generalizability. Experimental results demonstrate that LLMs enhanced with ConceptEdit successfully generate commonsense knowledge with improved plausibility compared to other baselines and achieve stronger performance across multiple question answering benchmarks. Our data, code, and models are publicly available at https://github.com/HKUST-KnowComp/ConKE.

Chunyang Li, Yilun Zheng, Xinting Huang et al. · tencent-ai

The paradigm of LLM-as-a-judge is emerging as a scalable and efficient alternative to human evaluation, demonstrating strong performance on well-defined tasks. However, its reliability in open-ended tasks with dynamic environments and complex interactions remains unexplored. To bridge the gap, we introduce WebDevJudge, a systematic benchmark for assessing LLM-as-a-judge performance in web development, with support for both non-interactive evaluation based on static observations and continuous interactive evaluation with a dynamic web environment. WebDevJudge comprises human preference labels over paired web implementations, annotated with structured and query-grounded rubrics to ensure high-quality ground truth. Using this benchmark, we comprehensively evaluate various evaluators, including LLMs, MLLMs, and agentic workflows. We systematically investigate the impact of different paradigms and guidance mechanisms. Our experiments reveal a significant gap between LLM judges and human experts. In-depth analysis indicates this gap stems from fundamental model limitations, including failures in recognizing functional equivalence, verifying task feasibility, and mitigating bias. Overall, WebDevJudge presents a significant challenge to LLM-as-a-judge, offering insights to guide future research toward developing more reliable and capable automated evaluators for complicated scenarios. Code and data are available at https://github.com/lcy2723/WebDevJudge.

Rui Wang, Ce Zhang, Jun-Yu Ma et al. · tencent-ai

Deep research web agents not only retrieve information from diverse sources such as web environments, files, and multimodal inputs, but more importantly, they need to rigorously analyze and aggregate knowledge for insightful research. However, existing open-source deep research agents predominantly focus on enhancing information-seeking capabilities of web agents to locate specific information, while overlooking the essential need for information aggregation, which would limit their ability to support in-depth research. We propose an Explore to Evolve paradigm to scalably construct verifiable training data for web agents. Begins with proactive online exploration, an agent sources grounded information by exploring the real web. Using the collected evidence, the agent then self-evolves an aggregation program by selecting, composing, and refining operations from 12 high-level logical types to synthesize a verifiable QA pair. This evolution from high-level guidance to concrete operations allowed us to scalably produce WebAggregatorQA, a dataset of 10K samples across 50K websites and 11 domains. Based on an open-source agent framework, SmolAgents, we collect supervised fine-tuning trajectories to develop a series of foundation models, WebAggregator. WebAggregator-8B matches the performance of GPT-4.1, while the 32B variant surpasses GPT-4.1 by more than 10% on GAIA-text and closely approaches Claude-3.7-sonnet. Moreover, given the limited availability of benchmarks that evaluate web agents' information aggregation abilities, we construct a human-annotated evaluation split of WebAggregatorQA as a challenging test set. On this benchmark, Claude-3.7-sonnet only achieves 28%, and GPT-4.1 scores 25.8%. Even when agents manage to retrieve all references, they still struggle on WebAggregatorQA, highlighting the need to strengthen the information aggregation capabilities of web agent foundations.

Wenxuan Ding, Weiqi Wang, Sze Heng Douglas Kwok et al.

Enhancing Language Models' (LMs) ability to understand purchase intentions in E-commerce scenarios is crucial for their effective assistance in various downstream tasks. However, previous approaches that distill intentions from LMs often fail to generate meaningful and human-centric intentions applicable in real-world E-commerce contexts. This raises concerns about the true comprehension and utilization of purchase intentions by LMs. In this paper, we present IntentionQA, a double-task multiple-choice question answering benchmark to evaluate LMs' comprehension of purchase intentions in E-commerce. Specifically, LMs are tasked to infer intentions based on purchased products and utilize them to predict additional purchases. IntentionQA consists of 4,360 carefully curated problems across three difficulty levels, constructed using an automated pipeline to ensure scalability on large E-commerce platforms. Human evaluations demonstrate the high quality and low false-negative rate of our benchmark. Extensive experiments across 19 language models show that they still struggle with certain scenarios, such as understanding products and intentions accurately, jointly reasoning with products and intentions, and more, in which they fall far behind human performances. Our code and data are publicly available at https://github.com/HKUST-KnowComp/IntentionQA.

Quyet V. Do, Tianqing Fang, Shizhe Diao et al.

Reasoning over Commonsense Knowledge Bases (CSKB), i.e. CSKB reasoning, has been explored as a way to acquire new commonsense knowledge based on reference knowledge in the original CSKBs and external prior knowledge. Despite the advancement of Large Language Models (LLM) and prompt engineering techniques in various reasoning tasks, they still struggle to deal with CSKB reasoning. One of the problems is that it is hard for them to acquire explicit relational constraints in CSKBs from only in-context exemplars, due to a lack of symbolic reasoning capabilities (Bengio et al., 2021). To this end, we proposed **ConstraintChecker**, a plugin over prompting techniques to provide and check explicit constraints. When considering a new knowledge instance, ConstraintChecker employs a rule-based module to produce a list of constraints, then it uses a zero-shot learning module to check whether this knowledge instance satisfies all constraints. The acquired constraint-checking result is then aggregated with the output of the main prompting technique to produce the final output. Experimental results on CSKB Reasoning benchmarks demonstrate the effectiveness of our method by bringing consistent improvements over all prompting methods. Codes and data are available at \url{https://github.com/HKUST-KnowComp/ConstraintChecker}.

Weiqi Wang, Tianqing Fang, Wenxuan Ding et al.

The task of zero-shot commonsense question answering evaluates models on their capacity to reason about general scenarios beyond those presented in specific datasets. Existing approaches for tackling this task leverage external knowledge from CommonSense Knowledge Bases (CSKBs) by pretraining the model on synthetic QA pairs constructed from CSKBs. In these approaches, negative examples (distractors) are formulated by randomly sampling from CSKBs using fairly primitive keyword constraints. However, two bottlenecks limit these approaches: the inherent incompleteness of CSKBs limits the semantic coverage of synthetic QA pairs, and the lack of human annotations makes the sampled negative examples potentially uninformative and contradictory. To tackle these limitations above, we propose Conceptualization-Augmented Reasoner (CAR), a zero-shot commonsense question-answering framework that fully leverages the power of conceptualization. Specifically, CAR abstracts a commonsense knowledge triple to many higher-level instances, which increases the coverage of CSKB and expands the ground-truth answer space, reducing the likelihood of selecting false-negative distractors. Extensive experiments demonstrate that CAR more robustly generalizes to answering questions about zero-shot commonsense scenarios than existing methods, including large language models, such as GPT3.5 and ChatGPT. Our codes, data, and model checkpoints are available at https://github.com/HKUST-KnowComp/CAR.

Weiqi Wang, Tianqing Fang, Baixuan Xu et al.

Commonsense reasoning, aiming at endowing machines with a human-like ability to make situational presumptions, is extremely challenging to generalize. For someone who barely knows about "meditation," while is knowledgeable about "singing," he can still infer that "meditation makes people relaxed" from the existing knowledge that "singing makes people relaxed" by first conceptualizing "singing" as a "relaxing event" and then instantiating that event to "meditation." This process, known as conceptual induction and deduction, is fundamental to commonsense reasoning while lacking both labeled data and methodologies to enhance commonsense modeling. To fill such a research gap, we propose CAT (Contextualized ConceptuAlization and InsTantiation), a semi-supervised learning framework that integrates event conceptualization and instantiation to conceptualize commonsense knowledge bases at scale. Extensive experiments show that our framework achieves state-of-the-art performances on two conceptualization tasks, and the acquired abstract commonsense knowledge can significantly improve commonsense inference modeling. Our code, data, and fine-tuned models are publicly available at https://github.com/HKUST-KnowComp/CAT.

Tianqing Fang, Weiqi Wang, Sehyun Choi et al.

Reasoning over commonsense knowledge bases (CSKB) whose elements are in the form of free-text is an important yet hard task in NLP. While CSKB completion only fills the missing links within the domain of the CSKB, CSKB population is alternatively proposed with the goal of reasoning unseen assertions from external resources. In this task, CSKBs are grounded to a large-scale eventuality (activity, state, and event) graph to discriminate whether novel triples from the eventuality graph are plausible or not. However, existing evaluations on the population task are either not accurate (automatic evaluation with randomly sampled negative examples) or of small scale (human annotation). In this paper, we benchmark the CSKB population task with a new large-scale dataset by first aligning four popular CSKBs, and then presenting a high-quality human-annotated evaluation set to probe neural models' commonsense reasoning ability. We also propose a novel inductive commonsense reasoning model that reasons over graphs. Experimental results show that generalizing commonsense reasoning on unseen assertions is inherently a hard task. Models achieving high accuracy during training perform poorly on the evaluation set, with a large gap between human performance. We will make the data publicly available for future contributions. Codes and data are available at https://github.com/HKUST-KnowComp/CSKB-Population.

Hongming Zhang, Xin Liu, Haojie Pan et al.

Commonsense knowledge acquisition and reasoning have long been a core artificial intelligence problem. However, in the past, there has been a lack of scalable methods to collect commonsense knowledge. In this paper, we propose to develop principles for collecting commonsense knowledge based on selectional preference. We generalize the definition of selectional preference from one-hop linguistic syntactic relations to higher-order relations over linguistic graphs. Unlike previous commonsense knowledge definition (e.g., ConceptNet), selectional preference (SP) knowledge only relies on statistical distribution over linguistic graphs, which can be efficiently and accurately acquired from the unlabeled corpus with modern tools. Following this principle, we develop a large-scale eventuality (a linguistic term covering activity, state, and event)-based knowledge graph ASER, where each eventuality is represented as a dependency graph, and the relation between them is a discourse relation defined in shallow discourse parsing. The higher-order selectional preference over collected linguistic graphs reflects various kinds of commonsense knowledge. Moreover, motivated by the observation that humans understand events by abstracting the observed events to a higher level and can thus transfer their knowledge to new events, we propose a conceptualization module to significantly boost the coverage of ASER. In total, ASER contains 648 million edges between 438 million eventualities. After conceptualization with Probase, a selectional preference based concept-instance relational knowledge base, our concept graph contains 15 million conceptualized eventualities and 224 million edges between them. Detailed analysis is provided to demonstrate its quality. All the collected data, APIs, and tools are available at https://github.com/HKUST-KnowComp/ASER.

Tianqing Fang, Hongming Zhang, Weiqi Wang et al.

Commonsense knowledge is crucial for artificial intelligence systems to understand natural language. Previous commonsense knowledge acquisition approaches typically rely on human annotations (for example, ATOMIC) or text generation models (for example, COMET.) Human annotation could provide high-quality commonsense knowledge, yet its high cost often results in relatively small scale and low coverage. On the other hand, generation models have the potential to automatically generate more knowledge. Nonetheless, machine learning models often fit the training data well and thus struggle to generate high-quality novel knowledge. To address the limitations of previous approaches, in this paper, we propose an alternative commonsense knowledge acquisition framework DISCOS (from DIScourse to COmmonSense), which automatically populates expensive complex commonsense knowledge to more affordable linguistic knowledge resources. Experiments demonstrate that we can successfully convert discourse knowledge about eventualities from ASER, a large-scale discourse knowledge graph, into if-then commonsense knowledge defined in ATOMIC without any additional annotation effort. Further study suggests that DISCOS significantly outperforms previous supervised approaches in terms of novelty and diversity with comparable quality. In total, we can acquire 3.4M ATOMIC-like inferential commonsense knowledge by populating ATOMIC on the core part of ASER. Codes and data are available at https://github.com/HKUST-KnowComp/DISCOS-commonsense.

Zhouzhou Shen, Xueyu Hu, Xiyun Li et al.

Web agents based on large language models have demonstrated promising capability in automating web tasks. However, current web agents struggle to reason out sensible actions due to the limitations of predicting environment changes, and might not possess comprehensive awareness of execution risks, prematurely performing risky actions that cause losses and lead to task failure. To address these challenges, we propose WAC, a web agent that integrates model collaboration, consequence simulation, and feedback-driven action refinement. To overcome the cognitive isolation of individual models, we introduce a multi-agent collaboration process that enables an action model to consult a world model as a web-environment expert for strategic guidance; the action model then grounds these suggestions into executable actions, leveraging prior knowledge of environmental state transition dynamics to enhance candidate action proposal. To achieve risk-aware resilient task execution, we introduce a two-stage deduction chain. A world model, specialized in environmental state transitions, simulates action outcomes, which a judge model then scrutinizes to trigger action corrective feedback when necessary. Experiments show that WAC achieves absolute gains of 1.8% on VisualWebArena and 1.3% on Online-Mind2Web.

Zhaowei Wang, Wei Fan, Qing Zong et al. · tencent-ai

Abstraction ability is crucial in human intelligence, which can also benefit various tasks in NLP study. Existing work shows that LLMs are deficient in abstract ability, and how to improve it remains unexplored. In this work, we design the framework AbsInstruct to enhance LLMs' abstraction ability through instruction tuning. The framework builds instructions with in-depth explanations to assist LLMs in capturing the underlying rationale of abstraction. Meanwhile, we introduce a plausibility estimator to select instructions that are more consistent with the abstraction knowledge of LLMs to be aligned. Then, our framework combines abstraction instructions with general-purpose ones to build a hybrid dataset. Extensive experiments and analyses demonstrate that our framework can considerably enhance LLMs' abstraction ability with strong generalization performance while maintaining their general instruction-following abilities.

Xiongye Xiao, Gengshuo Liu, Gaurav Gupta et al.

Integrating and processing information from various sources or modalities are critical for obtaining a comprehensive and accurate perception of the real world in autonomous systems and cyber-physical systems. Drawing inspiration from neuroscience, we develop the Information-Theoretic Hierarchical Perception (ITHP) model, which utilizes the concept of information bottleneck. Different from most traditional fusion models that incorporate all modalities identically in neural networks, our model designates a prime modality and regards the remaining modalities as detectors in the information pathway, serving to distill the flow of information. Our proposed perception model focuses on constructing an effective and compact information flow by achieving a balance between the minimization of mutual information between the latent state and the input modal state, and the maximization of mutual information between the latent states and the remaining modal states. This approach leads to compact latent state representations that retain relevant information while minimizing redundancy, thereby substantially enhancing the performance of multimodal representation learning. Experimental evaluations on the MUStARD, CMU-MOSI, and CMU-MOSEI datasets demonstrate that our model consistently distills crucial information in multimodal learning scenarios, outperforming state-of-the-art benchmarks. Remarkably, on the CMU-MOSI dataset, ITHP surpasses human-level performance in the multimodal sentiment binary classification task across all evaluation metrics (i.e., Binary Accuracy, F1 Score, Mean Absolute Error, and Pearson Correlation).

Zhisong Zhang, Yan Wang, Xinting Huang et al. · tencent-ai

Large language models have shown remarkable performance across a wide range of language tasks, owing to their exceptional capabilities in context modeling. The most commonly used method of context modeling is full self-attention, as seen in standard decoder-only Transformers. Although powerful, this method can be inefficient for long sequences and may overlook inherent input structures. To address these problems, an alternative approach is parallel context encoding, which splits the context into sub-pieces and encodes them parallelly. Because parallel patterns are not encountered during training, naively applying parallel encoding leads to performance degradation. However, the underlying reasons and potential mitigations are unclear. In this work, we provide a detailed analysis of this issue and identify that unusually high attention entropy can be a key factor. Furthermore, we adopt two straightforward methods to reduce attention entropy by incorporating attention sinks and selective mechanisms. Experiments on various tasks reveal that these methods effectively lower irregular attention entropy and narrow performance gaps. We hope this study can illuminate ways to enhance context modeling mechanisms.

Zhisong Zhang, Tianqing Fang, Kaixin Ma et al. · tencent-ai

With recent advancements in large language models, web agents have been greatly improved. However, dealing with complex and dynamic web environments requires more advanced planning and search abilities. Previous studies usually adopt a greedy one-way search strategy, which may struggle to recover from erroneous states. In this work, we enhance web agents with an explicit rollback mechanism, enabling the agent to revert back to a previous state in its navigation trajectory. This mechanism gives the model the flexibility to directly control the search process, leading to an effective and efficient web navigation method. We conduct experiments on two live web navigation benchmarks with zero-shot and fine-tuning settings. The results demonstrate the effectiveness of our proposed approach.

Kaiqiao Han, Tianqing Fang, Zhaowei Wang et al. · tencent-ai

While Large Language Models (LLMs) have showcased remarkable proficiency in reasoning, there is still a concern about hallucinations and unreliable reasoning issues due to semantic associations and superficial logical chains. To evaluate the extent to which LLMs perform robust reasoning instead of relying on superficial logical chains, we propose a new evaluation dataset, the Concept-Reversed Winograd Schema Challenge (CR-WSC), based on the famous Winograd Schema Challenge (WSC) dataset. By simply reversing the concepts to those that are more associated with the wrong answer, we find that the performance of LLMs drops significantly despite the rationale of reasoning remaining the same. Furthermore, we propose Abstraction-of-Thought (AoT), a novel prompt method for recovering adversarial cases to normal cases using conceptual abstraction to improve LLMs' robustness and consistency in reasoning, as demonstrated by experiments on CR-WSC.

Junyu Ma, Tianqing Fang, Zhisong Zhang et al. · tencent-ai

Mamba's theoretical infinite-context potential is limited in practice when sequences far exceed training lengths. This work explores unlocking Mamba's long-context memory ability by a simple-yet-effective method, Recall with Reasoning (RwR), by distilling chain-of-thought (CoT) summarization from a teacher model. Specifically, RwR prepends these summarization as CoT prompts during fine-tuning, teaching Mamba to actively recall and reason over long contexts. Experiments on LONGMEMEVAL and HELMET show RwR boosts Mamba's long-context performance against comparable Transformer/hybrid baselines under similar pretraining conditions, while preserving short-context capabilities, all without architectural changes.

Tianshi Zheng, Kelvin Kiu-Wai Tam, Newt Hue-Nam K. Nguyen et al. · tencent-ai

Large language models are emerging as powerful tools for scientific law discovery, a foundational challenge in AI-driven science. However, existing benchmarks for this task suffer from a fundamental methodological trilemma, forcing a trade-off between scientific relevance, scalability, and resistance to memorization. Furthermore, they oversimplify discovery as static function fitting, failing to capture the authentic scientific process of uncovering embedded laws through the interactive exploration of complex model systems. To address these critical gaps, we introduce NewtonBench, a benchmark comprising 324 scientific law discovery tasks across 12 physics domains. Our design mitigates the evaluation trilemma by using metaphysical shifts - systematic alterations of canonical laws - to generate a vast suite of problems that are scalable, scientifically relevant, and memorization-resistant. Moreover, we elevate the evaluation from static function fitting to interactive model discovery, requiring agents to experimentally probe simulated complex systems to uncover hidden principles. Our extensive experiment reveals a clear but fragile capability for discovery in frontier LLMs: this ability degrades precipitously with increasing system complexity and exhibits extreme sensitivity to observational noise. Notably, we uncover a paradoxical effect of tool assistance: providing a code interpreter can hinder more capable models by inducing a premature shift from exploration to exploitation, causing them to satisfice on suboptimal solutions. These results demonstrate that robust, generalizable discovery in complex, interactive environments remains the core challenge. By providing a scalable, robust, and scientifically authentic testbed, NewtonBench offers a crucial tool for measuring true progress and guiding the development of next-generation AI agents capable of genuine scientific discovery.

Baixuan Xu, Tianshi Zheng, Zhaowei Wang et al. · tencent-ai

Enabling LLMs to effectively operate long-horizon task which requires long-term planning and multiple interactions is essential for open-world autonomy. Conventional methods adopt planning with actions where a executable action list would be provided as reference. However, this action representation choice would be impractical when the environment action space is combinatorial exploded (e.g., open-ended real world). This naturally leads to a question: As environmental action space scales, what is the optimal action representation for long-horizon agents? In this paper, we systematically study the effectiveness of two different action representations. The first one is conventional planning with actions (PwA) which is predominantly adopted for its effectiveness on existing benchmarks. The other one is planning with schemas (PwS) which instantiate an action schema into action lists (e.g., "move [OBJ] to [OBJ]" -> "move apple to desk") to ensure concise action space and reliable scalability. This alternative is motivated by its alignment with human cognition and its compliance with environment-imposed action format restriction. We propose cognitive bandwidth perspective as a conceptual framework to qualitatively understand the differences between these two action representations and empirically observe a representation-choice inflection point between ALFWorld (~35 actions) and SciWorld (~500 actions), which serve as evidence of the need for scalable representations. We further conduct controlled experiments to study how the location of this inflection point interacts with different model capacities: stronger planning proficiency shifts the inflection rightward, whereas better schema instantiation shifts it leftward. Finally, noting the suboptimal performance of PwS agents, we provide an actionable guide for building more capable PwS agents for better scalable autonomy.

Ying Su, Tianqing Fang, Huiru Xiao et al. · tencent-ai

Commonsense knowledge graph completion is a new challenge for commonsense knowledge graph construction and application. In contrast to factual knowledge graphs such as Freebase and YAGO, commonsense knowledge graphs (CSKGs; e.g., ConceptNet) utilize free-form text to represent named entities, short phrases, and events as their nodes. Such a loose structure results in large and sparse CSKGs, which makes the semantic understanding of these nodes more critical for learning rich commonsense knowledge graph embedding. While current methods leverage semantic similarities to increase the graph density, the semantic plausibility of the nodes and their relations are under-explored. Previous works adopt conceptual abstraction to improve the consistency of modeling (event) plausibility, but they are not scalable enough and still suffer from data sparsity. In this paper, we propose to adopt textual entailment to find implicit entailment relations between CSKG nodes, to effectively densify the subgraph connecting nodes within the same conceptual class, which indicates a similar level of plausibility. Each node in CSKG finds its top entailed nodes using a finetuned transformer over natural language inference (NLI) tasks, which sufficiently capture textual entailment signals. The entailment relation between these nodes are further utilized to: 1) build new connections between source triplets and entailed nodes to densify the sparse CSKGs; 2) enrich the generalization ability of node representations by comparing the node embeddings with a contrastive loss. Experiments on two standard CSKGs demonstrate that our proposed framework EntailE can improve the performance of CSKG completion tasks under both transductive and inductive settings.

Chenlong Deng, Zhisong Zhang, Kelong Mao et al. · tencent-ai

Large language models are increasingly capable of handling long-context inputs, but the memory overhead of key-value (KV) cache remains a major bottleneck for general-purpose deployment. While various compression strategies have been explored, sequence-level compression, which drops the full KV caches for certain tokens, is particularly challenging as it can lead to the loss of important contextual information. To address this, we introduce UniGist, a sequence-level long-context compression framework that efficiently preserves context information by replacing raw tokens with special compression tokens (gists) in a fine-grained manner. We adopt a chunk-free training strategy and design an efficient kernel with a gist shift trick, enabling optimized GPU training. Our scheme also supports flexible inference by allowing the actual removal of compressed tokens, resulting in real-time memory savings. Experiments across multiple long-context tasks demonstrate that UniGist significantly improves compression quality, with especially strong performance in detail-recalling tasks and long-range dependency modeling.