Tongyi DeepResearch Technical Report

Summary

Technical report for Tongyi DeepResearch, an agentic large language model from Alibaba built for long-horizon, deep information-seeking research tasks. Instead of prompting a general model into an agent loop, the team trains agency into the model end-to-end through "agentic mid-training" followed by "agentic post-training". All training stages are fed by a fully automatic data-synthesis pipeline that needs no human annotation, and each stage runs in its own customized environment. The released model has 30.5 billion total parameters of which 3.3 billion are activated per token. The abstract claims state-of-the-art results on Humanity's Last Exam, BrowseComp, BrowseComp-ZH, WebWalkerQA, xbench-DeepSearch, FRAMES and xbench-DeepSearch-2510, and the model, framework and "complete solutions" are open-sourced. The arXiv record has three versions (v1 Oct 2025, v3 May 2026).

Contribution

Claimed: a complete training recipe (mid-training plus post-training), a scalable synthetic-data pipeline, per-stage environments, and state-of-the-art performance on seven agentic deep-research benchmarks, delivered as open weights. What the abstract actually supports: that the recipe exists and that benchmark results are reported. The abstract gives no scores, baselines or ablations, so the contribution of any single component (mid-training vs. post-training vs. data synthesis) cannot be isolated from it, and "state-of-the-art" is relative to an unspecified comparison set at the time of each version.

Method

Model training plus benchmark evaluation. The abstract states the two training phases, the automatic data-synthesis pipeline, the per-stage environments, the parameter counts and the benchmark names. It does not specify the RL or optimisation algorithm, the size or provenance of the synthetic data, the tools in the environment (search, browse, code), inference modes, compute, or any numbers. The catalog entry records ReAct and "Heavy" (IterResearch) inference modes and GRPO-based RL; those details come from the project page and technical blog, not from the abstract.

Relevance to RISE

Informs the rq-formulation, literature-discovery and literature-synthesis stages, matching the catalog entry tongyi-deepresearch, which files it in the literature/synthesis block with lifecycle coverage 1: it is a research-capable model, not a paper-producing pipeline. The closest catalog neighbour with an abstract in this KB is deepresearcher (zheng2025deepresearcher): both train deep-research agents end-to-end with reinforcement learning against real web environments; Tongyi DeepResearch adds an explicit agentic mid-training phase and a fully automatic data-synthesis pipeline and ships the weights. For the ISR question of how multi-agent structure shapes epistemic quality, the abstract describes a single trained agent and no multi-agent review, debate or aggregation mechanism, so any quality control is located in training rather than in structural aggregation; it is therefore a natural single-agent reference point against which structured designs can be contrasted.

Critique / open questions

Nothing in the abstract allows assessment of benchmark scores, contamination controls for the synthetic training data, or whether a "fully automatic" data pipeline reinforces its own errors. Faithfulness and citation accuracy of the produced research reports are not mentioned. Applicability to scholarly literature search (as opposed to open-web question answering) is not addressed. The abstract concedes no limitations. Corporate authorship (a team name plus 56 named authors) complicates attribution. The catalog entry notes that outputs depend on live web state and are not bitwise reproducible.

Key quotes

"We present Tongyi DeepResearch, an agentic large language model, which is specifically designed for long-horizon, deep information-seeking research tasks." (abstract)

"We design a highly scalable data synthesis pipeline that is fully automatic, without relying on costly human annotation, and empowers all training stages." (abstract)

"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." (abstract)