Reasoning-intensive ranking models built on Large Language Models (LLMs) have made notable progress. However, existing approaches often rely on large-scale LLMs and explicit Chain-of-Thought (CoT) reasoning, resulting in high computational cost and latency that limit real-world use. To address this, we propose TFRank, an efficient pointwise reasoning ranker based on small-scale LLMs. To improve ranking performance, TFRank effectively integrates CoT data, fine-grained score supervision, and multi-task training. Furthermore, it achieves an efficient "Think-Free" reasoning capability by employing a "think-mode switch" and pointwise format constraints. Specifically, this allows the model to leverage explicit reasoning during training while delivering precise relevance scores for complex queries at inference without generating any reasoning chains. Experiments show that TFRank achieves performance comparable to models with four times more parameters on the BRIGHT benchmark, and demonstrates strong competitiveness on the BEIR benchmark. Further analysis shows that TFRank achieves an effective balance between performance and efficiency, providing a practical solution for integrating advanced reasoning into real-world systems.

As a core module of web search, query understanding aims to bridge the semantic gap between user queries and web page documents, thereby enhancing the ability to deliver more relevant results.Recently, Large Language Models (LLMs) have achieved significant breakthroughs that have fundamentally altered the workflow of existing search ranking tasks.However, few researchers have explored the integration of LLMs into the field of query understanding.In this paper, we investigate the potential of LLMs in query understanding by conducting a comprehensive evaluation across three dimensions: term, structure, and topic.This evaluation includes several representative tasks such as segmentation, term weighting, error correction, query expansion, and intent recognition.The experimental results reveal that LLMs are particularly effective in query expansion and intent recognition but show limited improvement in other areas.This limitation may be attributed to LLMs' primary focus on modeling the semantic knowledge of entire queries, while lacking the capability to capture token-level information with finer granularity.Additionally, we explore potential practical applications of LLMs in query understanding, such as integrating the evaluation and training capabilities of smaller models with LLMs and constructing unsupervised samples.Based on comprehensive empirical results, collaborative training emerges as a promising approach to leverage LLMs for query understanding.We hope this research will advance the practical application of LLMs in query understanding and contribute to the development of this field.

Relevance plays a crucial role in commercial search engines by identifying documents related to user queries and fulfilling their search needs.Traditional approaches employ encoder-only models like BERT, which process concatenated query-document pairs to predict relevance scores.While autoregressive large language models (LLMs) have revolutionized numerous NLP domains, their direct application to web-scale search systems presents significant challenges.On one hand, the relevance modeling capabilities of LLMs have not been fully explored.On the other, the high computational costs and inference times make deploying LLMs in online search systems, which demand extremely low latency, nearly impossible.In this work, we address these challenges through two key contributions.First, we develop a comprehensive evaluation framework to systematically assess the effectiveness of LLMs in query-document relevance ranking.By conducting assessment experiments to LLMs in four perspectives: ranking objectives, model size, domain-specific continuous pre-training, and the integration of prior knowledge, we identify the best resource allocation strategy given a restricted budget and develop practical LLMs in a more efficient way.Second, we propose a novel framework to transfer the capabilities of LLMs in the ranking aspect to existing BERT models to avoid directly deploying LLMs.Finally, to fully leverage the improvements in relevance ranking brought by LLMs, we successfully nearline deploy LLMs in Tencent QQ Browser search engine using query-based ondemand computing and quantization.Experiments on real-world datasets and online A/B tests demonstrate that our approach significantly enhances search engine performance while maintaining practical operational efficiency.Our findings provide actionable insights for integrating LLMs into production search engines.

Visual question answering is an important task in both natural language and vision understanding. However, in most of the public visual question answering datasets such as VQA, CLEVR, the questions are human generated that specific to the given image, such as 'What color are her eyes?'. The human generated crowdsourcing questions are relatively simple and sometimes have the bias toward certain entities or attributes [1, 55].

Fine-grained entity typing (FET) aims to deduce specific semantic types of the entity mentions in the text. Modern methods for FET mainly focus on learning what a certain type looks like. And few works directly model the type differences, that is, let models know the extent that which one type is different from others. To alleviate this problem, we propose a type-enriched hierarchical contrastive strategy for FET. Our method can directly model the differences between hierarchical types and improve the ability to distinguish multi-grained similar types. On the one hand, we embed type into entity contexts to make type information directly perceptible. On the other hand, we design a constrained contrastive strategy on the hierarchical structure to directly model the type differences, which can simultaneously perceive the distinguishability between types at different granularity. Experimental results on three benchmarks, BBN, OntoNotes, and FIGER show that our method achieves significant performance on FET by effectively modeling type differences.

The task of travel time estimation (TTE), which estimates the travel time for a given route and departure time, plays an important role in intelligent transportation systems such as navigation, route planning, and ride-hailing services. This task is challenging because of many essential aspects, such as traffic prediction and contextual information. First, the accuracy of traffic prediction is strongly correlated with the traffic speed of the road segments in a route. Existing work mainly adopts spatial-temporal graph neural networks to improve the accuracy of traffic prediction, where spatial and temporal information is used separately. However, one drawback is that the spatial and temporal correlations are not fully exploited to obtain better accuracy. Second, contextual information of a route, i.e., the connections of adjacent road segments in the route, is an essential factor that impacts the driving speed. Previous work mainly uses sequential encoding models to address this issue. However, it is difficult to scale up sequential models to large-scale real-world services. In this paper, we propose an end-to-end neural framework named ConSTGAT, which integrates traffic prediction and contextual information to address these two problems. Specifically, we first propose a spatial-temporal graph neural network that adopts a novel graph attention mechanism, which is designed to fully exploit the joint relations of spatial and temporal information. Then, in order to efficiently take advantage of the contextual information, we design a computationally efficient model that applies convolutions over local windows to capture a route's contextual information and further employs multi-task learning to improve the performance. In this way, the travel time of each road segment can be computed in parallel and in advance. Extensive experiments conducted on large-scale real-world datasets demonstrate the superiority of ConSTGAT. In addition, ConSTGAT has already been deployed in production at Baidu Maps, and it successfully keeps serving tens of billions of requests every day. This confirms that ConSTGAT is a practical and robust solution for large-scale real-world TTE services.