Tongyi DeepResearch: A New Era of Open-Source AI Researchers

TL;DR

Tongyi DeepResearch is Alibaba’s open-source 30B-A3B MoE web-research agent, posted as the first fully open Deep-Research-class system to roughly match OpenAI’s DeepResearch on a set of agentic benchmarks (32.9 on HLE, 43.4 on BrowseComp, 46.7 on BrowseComp-ZH, 75 on xbench-DeepSearch). The release’s main value to the team is the recipe: a complete Agentic-CPT → Agentic-SFT → on-policy Agentic-RL pipeline built almost entirely on fully synthetic data (no human-annotated trajectories), with explicit attention to the data-synthesis methodology — entity-anchored knowledge memory, multi-style QA construction, action synthesis, formalized graph-and-set-theoretic difficulty escalation, and PhD-level question synthesis via an iterative complexity-upgrade loop. Two inference modes ship: native ReAct (single 128K-context trajectory) and “Heavy Mode” (IterResearch — multiple parallel agents in streamlined per-round workspaces synthesized by a final aggregator).

Key claims

Tongyi DeepResearch-30B-A3B reports 32.9 on Humanity’s Last Exam, 43.4 on BrowseComp, 46.7 on BrowseComp-ZH, and 75 on xbench-DeepSearch — claimed on par with proprietary DeepResearch systems and ahead of prior open-source deep-research agents [§From Chatbot to Autonomous Agent].
Training is staged Agentic-CPT → Agentic-SFT → Agentic-RL, framed as a new paradigm: CPT installs tool-use priors via large-scale offline action synthesis (AgentFounder); SFT bootstraps with ReAct + IterResearch trajectories; on-policy GRPO does the final shaping [§End-to-End Agent Training Pipeline].
Continual-Pretraining data is built by reorganizing documents, crawled web data, knowledge graphs, and historical tool-invocation records into an entity-anchored open-world knowledge memory; multi-style (question, answer) pairs are generated from sampled entities, and first/higher-order action-synthesis data is produced offline to avoid live API costs [§Continual Pre-training Data].
Post-training QA is generated by a graph-based pipeline (WebSailor lineage): a highly interconnected knowledge graph is built from real websites via random walks + isomorphic-table fusion, subgraphs/subtables are sampled to seed QA, and difficulty is escalated by strategically obfuscating information; the difficulty operations are formalized as controllable “atomic operations” over entity relationships [§Post-training Data].
A formal set-theoretic model of the information-seeking problem (WebShaper, arXiv:2507.15061) is used to control difficulty/structure scaling and to enable efficient verification of synthetic QA correctness — the verification step is what closes the synthetic-data loop [§Post-training Data].
A separate automated engine generates PhD-level research questions via an iterative complexity-upgrade loop: a question-crafting agent equipped with web-search + academic retrieval + a Python sandbox expands each “seed” QA across rounds, deepening conceptual abstraction and adding computational tasks [§Post-training Data].
Two rollout paradigms ship: Native ReAct (strict Thought-Action-Observation cycles, 128K context, no prompt engineering — explicitly motivated by Sutton’s “Bitter Lesson”) and Heavy Mode (IterResearch: deconstructs a task into rounds, each rebuilding a streamlined workspace from the previous round’s distilled outputs into an evolving central report) [§Rollout Mode, §Heavy Mode].
IterResearch is framed as a fix for “cognitive suffocation” and “noise pollution” in single-context agents that accumulate all interaction history — every round is forced to keep only the essential outputs and write into the central report [§Heavy Mode].
Heavy-Mode “Research-Synthesis” parallelizes multiple Research Agents (each running IterResearch) and merges their refined reports with a final Synthesis Agent — pushing inference-time compute up at the cost of more rollouts per query [§Heavy Mode].
The RL stage uses a customized strictly on-policy GRPO with token-level policy-gradient loss, a leave-one-out advantage estimator for variance reduction, and a conservative negative-sample policy that excludes length-exceeded incomplete trajectories from the loss to prevent observed “format collapse” under unfiltered negatives [§On-Policy Agent Reinforcement Learning].
Dynamic-sampling is explicitly not used; the team instead leverages larger batch and group sizes to keep variance low and supervision adequate. Policy entropy stays high without explicit entropy regularization, which they attribute to the non-stationary web environment [§On-Policy Agent Reinforcement Learning].
Data and training-environment stability are claimed to matter more than the RL algorithm itself: training directly on the BrowseComp test distribution gives worse results than the synthetic distribution, attributed to the limited scale + noise of human-annotated data hindering generalization [§On-Policy Agent Reinforcement Learning].
Infrastructure: a simulated training environment on an offline Wikipedia database with a custom tool suite replaces live web APIs to make RL cheap, fast, and deterministic; a stable tool sandbox handles concurrency, retries failed calls, and falls back to redundant search-API providers to prevent tool errors from corrupting trajectories [§On-Policy Agent Reinforcement Learning — Infrastructure side].
A fully automated data-curation pipeline runs during training, dynamically adjusting the synthesis distribution and filtering set in response to training dynamics — the data generation and model training are explicitly closed-loop, not staged [§Automatic Data Curation].
A step-level asynchronous RL training loop is implemented on top of rLLM; multiple agent instances run in parallel against the (simulated or real) environment [§On-Policy Asynchronous Framework].
Limitations explicitly stated: 128K context is still insufficient for the hardest long-horizon tasks; the pipeline has not been validated above 30B-MoE; the RL framework would benefit from partial rollouts but that requires solving the off-policy distributional-shift problem [§Limitations].

Method

The training pipeline is a three-stage end-to-end loop. Agentic CPT builds AgentFounder by restructuring all available data (documents, crawled pages, knowledge graphs, historical trajectories with tool-invocation traces) into an entity-anchored “open-world knowledge memory,” from which the pipeline samples entities and generates multi-style QA pairs plus first/higher-order action-synthesis traces. Agentic SFT cold-starts from a rejection-sampled set of ReAct trajectories (instilling rigid format adherence and reasoning behaviors) plus IterResearch trajectories (instilling sustained planning and per-round workspace reconstruction). Agentic RL runs on-policy customized GRPO with the conservative-negatives + leave-one-out + token-level-PG modifications, against rewards delivered by the simulated Wikipedia tool environment.

The post-training data engine has two pillars: (a) the graph-based information-seeking pipeline (WebWalker → WebSailor → WebSailor-V2 → WebShaper lineage) generates QA with controllable difficulty via formal atomic operations on entity-relationship subgraphs and verifies correctness using set-theoretic structure, and (b) an iterative-complexity-upgrade engine where a question-crafting agent with search/retrieval/Python tools expands seed multidisciplinary QAs over rounds.

Inference offers Native ReAct (single 128K-context trajectory) and Heavy Mode. Heavy Mode runs the IterResearch paradigm: per round, the agent reconstructs a streamlined workspace from the previous round’s distilled outputs, analyzes the problem, integrates findings into an evolving central report, and decides next action. Research-Synthesis runs multiple parallel Research Agents and aggregates their reports via a final Synthesis Agent — explicit parallel inference-time scaling.

Companion artifacts: open-sourced 30B-A3B MoE checkpoint (HuggingFace + ModelScope), the DeepResearch GitHub repo, plus a series of technical reports — WebWalker (arXiv:2501.07572 in spirit; see Series Work), WebDancer (2505.22648), WebSailor (2507.02592), WebShaper (2507.15061), WebWatcher (2508.05748), WebResearch (2509.13309), ReSum (2509.13313), WebWeaver (2509.13312), Scaling Agents via Continual Pre-training (2509.13310), Environment Scaling (2509.13311).

Results

Benchmark headline numbers (proprietary DeepResearch comparison): HLE 32.9, BrowseComp 43.4, BrowseComp-ZH 46.7, xbench-DeepSearch 75 — claimed systematic outperformance over all existing open-source Deep Research agents [§From Chatbot to Autonomous Agent].
Reported as the first open-source Web Agent at OpenAI-DeepResearch-class on a comprehensive benchmark suite [§From Chatbot to Autonomous Agent].
Training-dynamics observations: reward trends upward consistently; policy entropy stays high without explicit regularization (attributed to non-stationary web environment) [§On-Policy Agent Reinforcement Learning].
Ablation observation: training on BrowseComp directly is worse than training on synthetic data — explicitly framed as evidence that “data and stability are likely the more critical components than the algorithm” [§On-Policy Agent Reinforcement Learning].

Why it’s interesting

This is the cleanest open-source articulation to date of the “Agentic CPT → SFT → RL with fully synthetic data + verification gate” recipe — the same lineage MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling sits in, but with explicit CPT stage and a formalized set-theoretic difficulty controller that MiroThinker does not have. Tongyi’s claim that data and environment stability dominate algorithm choice directly corroborates the Synthetic Training Data page’s “verification gate is the load-bearing piece” thesis, and the BrowseComp-test-set-is-worse-than-synthetic observation is a notable concrete data point against training on raw human-annotated agent traces at small scale. The IterResearch Heavy Mode + Research-Synthesis aggregator is a sequential+parallel hybrid that complements Kimi K2.5’s PARL parallel-agent recipe (see Tool-Use Agents) — both extend Reasoning RL beyond single-trajectory ReAct, but in different ways (workspace reconstruction per round vs orchestrator + sub-agent decomposition). The “simulated Wikipedia training environment + redundant-fallback tool sandbox” is also a useful contrast to OSGym: Scalable OS Infra for Computer Use Agents and other RL Environment Platforms work — Tongyi favors offline determinism over live realism, citing cost + signal-quality reasons.