Hailong Yin

Tongyi DeepResearch Team, Baixuan Li, Bo Zhang et al.

We present Tongyi DeepResearch, an agentic large language model, which is specifically designed for long-horizon, deep information-seeking research tasks. To incentivize autonomous deep research agency, Tongyi DeepResearch is developed through an end-to-end training framework that combines agentic mid-training and agentic post-training, enabling scalable reasoning and information seeking across complex tasks. We design a highly scalable data synthesis pipeline that is fully automatic, without relying on costly human annotation, and empowers all training stages. By constructing customized environments for each stage, our system enables stable and consistent interactions throughout. Tongyi DeepResearch, featuring 30.5 billion total parameters, with only 3.3 billion activated per token, achieves state-of-the-art performance across a range of agentic deep research benchmarks, including Humanity's Last Exam, BrowseComp, BrowseComp-ZH, WebWalkerQA, xbench-DeepSearch, FRAMES and xbench-DeepSearch-2510. We open-source the model, framework, and complete solutions to empower the community.

Hailong Yin, Bin Zhu, Jingjing Chen et al.

Although Large Language Models (LLMs) demonstrate significant capabilities, their reliance on parametric knowledge often leads to inaccuracies. Retrieval Augmented Generation (RAG) mitigates this by incorporating external knowledge, but these methods may introduce irrelevant retrieved documents, leading to inaccurate responses. While the integration methods filter out incorrect answers from multiple responses, but lack external knowledge like RAG methods, and their high costs require balancing overhead with performance gains. To address these issues, we propose an Efficient Test-Time Retrieval-Augmented Generation Framework named ET2RAG to improve the performance of LLMs while maintaining efficiency. Specifically, ET2RAG is a training-free method, that first retrieves the most relevant documents and augments the LLMs to efficiently generate diverse candidate responses by managing response length. Then we compute the similarity of candidate responses and employ a majority voting mechanism to select the most suitable response as the final output. In particular, we discover that partial generation is sufficient to capture the key information necessary for consensus calculation, allowing us to effectively perform majority voting without the need for fully generated responses. Thus, we can reach a balance between computational cost and performance by managing the response length for the number of retrieved documents for majority voting. Experimental results demonstrate that ET2RAG significantly enhances performance across three tasks, including open-domain question answering, recipe generation and image captioning.

Guoshan Liu, Hailong Yin, Bin Zhu et al.

Given the potential applications of generating recipes from food images, this area has garnered significant attention from researchers in recent years. Existing works for recipe generation primarily utilize a two-stage training method, first generating ingredients and then obtaining instructions from both the image and ingredients. Large Multi-modal Models (LMMs), which have achieved notable success across a variety of vision and language tasks, shed light to generating both ingredients and instructions directly from images. Nevertheless, LMMs still face the common issue of hallucinations during recipe generation, leading to suboptimal performance. To tackle this, we propose a retrieval augmented large multimodal model for recipe generation. We first introduce Stochastic Diversified Retrieval Augmentation (SDRA) to retrieve recipes semantically related to the image from an existing datastore as a supplement, integrating them into the prompt to add diverse and rich context to the input image. Additionally, Self-Consistency Ensemble Voting mechanism is proposed to determine the most confident prediction recipes as the final output. It calculates the consistency among generated recipe candidates, which use different retrieval recipes as context for generation. Extensive experiments validate the effectiveness of our proposed method, which demonstrates state-of-the-art (SOTA) performance in recipe generation tasks on the Recipe1M dataset.

Bin Zhu, Hailong Yin, Jingjing Chen et al.

Recent advances in reasoning-centric models promise improved robustness through mechanisms such as chain-of-thought prompting and test-time scaling. However, their ability to withstand misleading user input remains underexplored. In this paper, we conduct a systematic evaluation of three state-of-the-art reasoning models, i.e., OpenAI's o4-mini, Claude-3.7-Sonnet and Gemini-2.5-Flash, across three multimodal benchmarks: MMMU, MathVista, and CharXiv. Our evaluation reveals significant accuracy drops (25-29% on average) following gaslighting negation prompts, indicating that even top-tier reasoning models struggle to preserve correct answers under manipulative user feedback. Built upon the insights of the evaluation and to further probe this vulnerability, we introduce GaslightingBench-R, a new diagnostic benchmark specifically designed to evaluate reasoning models' susceptibility to defend their belief under gaslighting negation prompt. Constructed by filtering and curating 1,025 challenging samples from the existing benchmarks, GaslightingBench-R induces even more dramatic failures, with accuracy drops exceeding 53% on average. Our findings reveal fundamental limitations in the robustness of reasoning models, highlighting the gap between step-by-step reasoning and belief persistence.