Medical reasoning models remain constrained by parametric knowledge and are thus susceptible to forgetting and hallucinations. DeepResearch (DR) models ground outputs in verifiable evidence from tools and perform strongly in general domains, but their direct transfer to medical field yields relatively limited gains. We attribute this to two gaps: task characteristic and tool-use scaling. Medical questions require evidence interpretation in a knowledge-intensive clinical context; while general DR models can retrieve information, they often lack clinical-context reasoning and thus "find it but fail to use it," leaving performance limited by medical abilities. Moreover, in medical scenarios, blindly scaling tool-call can inject noisy context, derailing sensitive medical reasoning and prompting repetitive evidence-seeking along incorrect paths. Therefore, we propose DeepMed. For data, we deploy a multi-hop med-search QA synthesis method supporting the model to apply the DR paradigm in medical contexts. For training, we introduce a difficulty-aware turn-penalty to suppress excessive tool-call growth. For inference, we bring a monitor to help validate hypotheses within a controlled number of steps and avoid context rot. Overall, on seven medical benchmarks, DeepMed improves its base model by 9.79% on average and outperforms larger medical reasoning and DR models.

Recent reinforcement learning (RL) methods have substantially enhanced the planning capabilities of Large Language Models (LLMs), yet the theoretical basis for their effectiveness remains elusive. In this work, we investigate RL's benefits and limitations through a tractable graph-based abstraction, focusing on policy gradient (PG) and Q-learning methods. Our theoretical analyses reveal that supervised fine-tuning (SFT) may introduce co-occurrence-based spurious solutions, whereas RL achieves correct planning primarily through exploration, underscoring exploration’s role in enabling better generalization. However, we also show that PG suffers from diversity collapse, where output diversity decreases during training and persists even after perfect accuracy is attained. By contrast, Q-learning provides two key advantages: off-policy learning and diversity preservation at convergence. We further demonstrate that careful reward design is necessary to prevent Q-value bias in Q-learning. Finally, applying our framework to the real-world planning benchmark Blocksworld, we confirm that these behaviors manifest in practice.

As LLM agents proliferate, built by different parties and with different capabilities and costs, orchestrating them is more like assembling labor across the economy than a computer calling a subroutine. Existing orchestration is typically centralized, with a single planner assigning every task, but this creates a bottleneck as agent pools grow, requires private information (e.g., agents' execution costs), and can easily be manipulated, such that a single inserted preference nearly doubles a favored agent's task share under a centralized LLM allocator. We introduce AgentLance, a repeated labor market in which agents bid on tasks using their private costs and self-maintained strategy notes, an allocator selects winners from bids and public reputation records, and a VCG-style payment rule rewards cost-aware bidding. Complex tasks are handled by hierarchical delegation: winning agents can decompose work and subcontract it through the same mechanism. Across mathematical reasoning, code generation, knowledge-intensive QA, and agentic tasks, AgentLance matches agents to their specializations, shifts work toward cheaper agents as cost sensitivity rises, and consistently outperforms single-model, centralized-orchestration, and market baselines. Diagnosing market failures, including inaccurate cost self-estimation and sub-optimal bidding, then correcting them in controlled experiments yields further gains, charting a path toward more efficient agent economies.

Augmenting large language models (LLMs) with browsing tools substantially improves their potential as deep search agents to solve complex, real-world tasks. Yet, open LLMs still perform poorly in such settings due to limited long-horizon reasoning capacity with browsing tools and the lack of sufficiently difficult supervised data. To address these challenges, we present DeepDive to advance deep search agents. First, we propose a strategy to automatically synthesize complex, difficult, and hard-to-find questions from open knowledge graphs. Second, we apply end-to-end multi-turn reinforcement learning (RL) to enhance LLMs' long-horizon reasoning with deep search. Experiments show that DeepDive-32B achieves a new open-source competitive result on BrowseComp, outperforming WebSailor, DeepSeek-R1-Browse, and Search-o1. We demonstrate that multi-turn RL training improves deep search ability and significantly contributes to the performance improvements across multiple benchmarks. We observe that DeepDive enables test-time scaling of tool calls and parallel sampling. All datasets, models, and code are publicly available at https://github.com/THUDM/DeepDive.

In combinatorial causal bandits (CCB), the learning agent chooses a subset of variables in each round to intervene and collects feedback from the observed variables to minimize expected regret or sample complexity. Previous works study this problem in both general causal models and binary generalized linear models (BGLMs). However, all of them require prior knowledge of causal graph structure. This paper studies the CCB problem without the graph structure on binary general causal models and BGLMs. We first provide an exponential lower bound of cumulative regrets for the CCB problem on general causal models. To overcome the exponentially large space of parameters, we then consider the CCB problem on BGLMs. We design a regret minimization algorithm for BGLMs even without the graph skeleton and show that it still achieves $O(\sqrt{T}\ln T)$ expected regret. This asymptotic regret is the same as the state-of-art algorithms relying on the graph structure. Moreover, we sacrifice the regret to $O(T^{\frac{2}{3}}\ln T)$ to remove the weight gap covered by the asymptotic notation. At last, we give some discussions and algorithms for pure exploration of the CCB problem without the graph structure.

Planning is a crucial element of both human intelligence and contemporary large language models (LLMs). In this paper, we initiate a theoretical investigation into the emergence of planning capabilities in Transformer-based LLMs via their next-word prediction mechanisms. We model planning as a network path-finding task, where the objective is to generate a valid path from a specified source node to a designated target node. Our mathematical characterization shows that Transformer architectures can execute path-finding by embedding the adjacency and reachability matrices within their weights. Furthermore, our theoretical analysis of gradient-based learning dynamics reveals that LLMs can learn both the adjacency and a limited form of the reachability matrices. These theoretical insights are then validated through experiments, which demonstrate that Transformer architectures indeed learn the adjacency and an incomplete reachability matrices, consistent with our theoretical predictions. When applying our methodology to the real-world planning benchmark Blocksworld, our observations remain consistent. Additionally, our analyses uncover a fundamental limitation of current Transformer architectures in path-finding: these architectures cannot identify reachability relationships through transitivity, which leads to failures in generating paths when concatenation is required. These findings provide new insights into how the internal mechanisms of autoregressive learning facilitate intelligent planning and deepen our understanding of how future LLMs might achieve more advanced and general planning-and-reasoning capabilities across diverse applications.

In combinatorial causal bandits (CCB), the learning agent chooses at most K variables in each round to intervene, collects feedback from the observed variables, with the goal of minimizing expected regret on the target variable Y. We study under the context of binary generalized linear models (BGLMs) with a succinct parametric representation of the causal models. We present the algorithm BGLM-OFU for Markovian BGLMs (i.e., no hidden variables) based on the maximum likelihood estimation method and give regret analysis for it. For the special case of linear models with hidden variables, we apply causal inference techniques such as the do calculus to convert the original model into a Markovian model, and then show that our BGLM-OFU algorithm and another algorithm based on the linear regression both solve such linear models with hidden variables. Our novelty includes (a) considering the combinatorial intervention action space and the general causal graph structures including ones with hidden variables, (b) integrating and adapting techniques from diverse studies such as generalized linear bandits and online influence maximization, and (c) avoiding unrealistic assumptions (such as knowing the joint distribution of the parents of Y under all interventions) and regret factors exponential to causal graph size in prior studies.

Peer prediction refers to a collection of mechanisms for eliciting information from human agents when direct verification of the obtained information is unavailable. They are designed to have a game-theoretic equilibrium where everyone reveals their private information truthfully. This result holds under the assumption that agents are Bayesian and they each adopt a fixed strategy across all tasks. Human agents however are observed in many domains to exhibit learning behavior in sequential settings. In this paper, we explore the dynamics of sequential peer prediction mechanisms when participants are learning agents. We first show that the notion of no regret alone for the agents' learning algorithms cannot guarantee convergence to the truthful strategy. We then focus on a family of learning algorithms where strategy updates only depend on agents' cumulative rewards and prove that agents' strategies in the popular Correlated Agreement (CA) mechanism converge to truthful reporting when they use algorithms from this family. This family of algorithms is not necessarily no-regret, but includes several familiar no-regret learning algorithms (e.g multiplicative weight update and Follow the Perturbed Leader) as special cases. Simulation of several algorithms in this family as well as the $\epsilon$-greedy algorithm, which is outside of this family, shows convergence to the truthful strategy in the CA mechanism.

Independent cascade (IC) model is a widely used influence propagation model for social networks. In this paper, we incorporate the concept and techniques from causal inference to study the identifiability of parameters from observational data in extended IC model with unobserved confounding factors, which models more realistic propagation scenarios but is rarely studied in influence propagation modeling before. We provide the conditions for the identifiability or unidentifiability of parameters for several special structures including the Markovian IC model, semi-Markovian IC model, and IC model with a global unobserved variable. Parameter identifiability is important for other tasks such as influence maximization under the diffusion networks with unobserved confounding factors.

Understanding users' context is essential for successful recommendations, especially for Online-to-Offline (O2O) recommendation, such as Yelp, Groupon, and Koubei. Different from traditional recommendation where individual preference is mostly static, O2O recommendation should be dynamic to capture variation of users' purposes across time and location. However, precisely inferring users' real-time contexts information, especially those implicit ones, is extremely difficult, and it is a central challenge for O2O recommendation. In this paper, we propose a new approach, called Mixture Attentional Constrained Denoise AutoEncoder (MACDAE), to infer implicit contexts and consequently, to improve the quality of real-time O2O recommendation. In MACDAE, we first leverage the interaction among users, items, and explicit contexts to infer users' implicit contexts, then combine the learned implicit-context representation into an end-to-end model to make the recommendation. MACDAE works quite well in the real system. We conducted both offline and online evaluations of the proposed approach. Experiments on several real-world datasets (Yelp, Dianping, and Koubei) show our approach could achieve significant improvements over state-of-the-arts. Furthermore, online A/B test suggests a 2.9% increase for click-through rate and 5.6% improvement for conversion rate in real-world traffic. Our model has been deployed in the product of "Guess You Like" recommendation in Koubei.

Matrix factorization on user-item rating matrices has achieved significant success in collaborative filtering based recommendation tasks. However, it also encounters the problems of data sparsity and scalability when applied in real-world recommender systems. In this paper, we present the Localized Matrix Factorization (LMF) framework, which attempts to meet the challenges of sparsity and scalability by factorizing Block Diagonal Form (BDF) matrices. In the LMF framework, a large sparse matrix is first transformed into Recursive Bordered Block Diagonal Form (RBBDF), which is an intuitionally interpretable structure for user-item rating matrices. Smaller and denser submatrices are then extracted from this RBBDF matrix to construct a BDF matrix for more effective collaborative prediction. We show formally that the LMF framework is suitable for matrix factorization and that any decomposable matrix factorization algorithm can be integrated into this framework. It has the potential to improve prediction accuracy by factorizing smaller and denser submatrices independently, which is also suitable for parallelization and contributes to system scalability at the same time. Experimental results based on a number of real-world public-access benchmarks show the effectiveness and efficiency of the proposed LMF framework.

The statistics and analysis of historical air traffic flow are significant for Air Traffic Flow Management. In this paper, we propose a system design and implementation solution for analyzing air traffic flow based on historical 4-D radar tracking data. Using database and Object-Oriented programming technology, radar data is successively analyzed and extracted into flow data of smaller spatial dimension and time dimension with multiple granularities. Air traffic flow statistics are generated by automatic database procedures and can be inquired flexibly. Field test shows that hundreds of thousands of lines of raw data can be analyzed efficiently in an hour and user interface is friendly. © Springer-Verlag Berlin Heidelberg 2012.

Ontology matching plays a key role for semantic interoperability. Many methods have been proposed for automatically finding the alignment between heterogeneous ontologies. However, in many real-world applications, finding the alignment in a completely automatic way is highly infeasible. Ideally, an ontology matching system would have an interactive interface to allow users to provide feedbacks to guide the automatic algorithm. Fundamentally, we need answer the following questions: How can a system perform an efficiently interactive process with the user? How many interactions are sufficient for finding a more accurate matching? To address these questions, we propose an active learning framework for ontology matching, which tries to find the most informative candidate matches to query the user. The user’s feedbacks are used to: 1) correct the mistake matching and 2) propagate the supervise information to help the entire matching process. Three measures are proposed to estimate the confidence of each matching candidate. A correct propagation algorithm is further proposed to maximize the spread of the user’s “guidance”. Experimental results on several public data sets show that the proposed approach can significantly improve the matching accuracy (+8.0% better than the baseline methods).

In this paper, we introduce the method on how to integrate travel information data embedded in web pages using approaches of information extraction and instance matching. Furthermore we extend the concept of instance matching to find the connotative relationship between instances extracted from different sources in order to improve the result of integration. We extracted more than 145,000 pieces of travel data terms of sight, route, agent, hotel, restaurant and ticket from several different sources, and integrated them into a piece of travel data with comprehensive information. 1 Overview

In this report, we give a brief explanation of how RiMOM obtains the results at OAEI 2009 Campaign, especially in the new Instance Matching track. At first, we show the basic alignment process of RiMOM and different alignment strategies in RiMOM. Then we give new features in instance matching compared with traditional ontology matching (schema matching) and introduce the specific techniques we used for the 3 different subtracks of Instance Matching Track. At last we give some comments on our results and discuss some future work about RiMOM.

The semantic web has become the focus of attention for its strong ability of data sharing and reusing. But the acquisition of semantic data is a big problem. In this work, we introduce a series of methods on how to extract travel information data from web sites, how to transform the common travel information data to semantic data, how to integrate the semantic data through instance matching. The main contribution is the introduction to the whole process of the semantic data integration, including the technology of information extraction, instance matching and transformation from common data to the coherent semantic data. We extracted more than 20,000 pieces of common data of tourist routes, sights, and travel agents from different travel web sites, and integrated more than 4,000 pieces of tourist routes data with comprehensive information.