MiroThinker-1.7: Research Agent for Complex Tasks

• MiroThinker-1.7 is a research agent for complex long-horizon tasks that integrates planning, multi-turn reasoning, and tool interaction.
• It employs an extended ReAct loop and agentic mid-training to decompose planning, reasoning, and summarization for improved inference reliability.
• Its heavy-duty variant, MiroThinker-H1, adds local and global verification layers that enhance performance on benchmarks and reduce error propagation.

Searching arXiv for the specified paper and closely related background so the article can be grounded in current arXiv records. MiroThinker-1.7 is a research agent for complex long-horizon reasoning tasks, introduced together with the heavy-duty variant MiroThinker-H1 in “MiroThinker-1.7 & H1: Towards Heavy-Duty Research Agents via Verification” (Team et al., 16 Mar 2026). The system is built on the open-source Qwen3 MoE foundation model family, with up to 30B active parameters, and adds three agentic layers for planning, reasoning, and tool interaction. Its central design objective is to improve the reliability of each interaction step through agentic mid-training, and, in H1, to incorporate verification directly into inference at both local and global levels so that intermediate decisions can be refined and final answers can be audited against coherent chains of evidence (Team et al., 16 Mar 2026).

1. Model family and problem setting

MiroThinker-1.7 is positioned as a research agent rather than a conventional single-turn chat model. The paper defines it around complex multi-step interaction over open-web research, scientific reasoning, financial analysis, medical browsing, and long-report generation. In this framing, the operative problem is not only answer generation but sustained reasoning under partial observations, tool-mediated evidence acquisition, and repeated context updates across long trajectories (Team et al., 16 Mar 2026).

The base model family is the open-source Qwen3 MoE backbone, with the full system using up to 30B active parameters. On top of that backbone, MiroThinker-1.7 introduces three explicitly separated heads: a planning head, a reasoning head, and a tool-interface head. The planning head proposes a structured plan consisting of subgoals and tool calls given only the query; the reasoning head expands individual steps by evaluating partial observations, consolidating evidence, and deciding whether to call a tool or generate a thought; and the tool-interface head formats outputs into sanctioned API calls (Team et al., 16 Mar 2026).

The released model variants are MiroThinker-1.7 and MiroThinker-1.7-mini. The paper also presents MiroThinker-H1 as an extension of 1.7 rather than as merely another size point in the same family. A common misconception is to treat H1 as a renaming of the base system; the paper instead defines H1 by the addition of verification mechanisms that alter the inference process itself. Conversely, another misconception is to treat the mini model as equivalent to the full model. The reported results show that MiroThinker-1.7-mini is competitive and markedly more efficient, but it generally trails MiroThinker-1.7 and M-H1 on the principal benchmark suite (Team et al., 16 Mar 2026).

2. Architecture and inference loop

Inference is organized as an extended ReAct loop with two nested iterations: a step loop inside an episode and an episode loop across retries. At step $t$, the full thought–action–observation trace is

$$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$

The paper defines a sliding window over recent observations,

$$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$

and a context operator $\Phi_t$ that truncates older observations:

$$\Phi_t(O_i) = \begin{cases} \mathrm{Trunc}_L(O_i), & \text{if } i \in S_t(K), \ \emptyset, & \text{otherwise}. \end{cases}$$

The effective context is therefore

$C_t = \{(T_i, A_i, \Phi_t(O_i))\}_{i<t}.$

Given query $q$ and context $C_t$, the model computes a thought and then an action,

$$T_t = f_\theta(q, C_t), \qquad A_t = \pi_\theta(C_t, T_t),$$

after which the environment returns $O_t = \mathrm{Tool}(A_t)$ and appends $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$0 to the trajectory (Team et al., 16 Mar 2026).

The episode loop addresses failure modes in long-horizon search. If no valid final answer is produced within $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$1 steps, the context is reset, $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$2, and the agent retries up to $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$3 times. This design suggests that the model treats unsuccessful reasoning chains as disposable episodes rather than preserving all prior states indefinitely. A plausible implication is that the architecture is optimized not only for token-budget efficiency but also for recovery from early planning errors.

The structured planning stage is formalized by a single-turn planning algorithm. For up to $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$4 candidate samples, the model generates a plan $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$5, extracts its first tool call $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$6, and asks a lightweight judge model $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$7 to evaluate the plan. If the score passes a threshold $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$8, the system accepts that plan immediately; otherwise it falls back to the highest-scoring candidate. The paper states that the judge is taxonomy-aware, classifies each plan by domain, and verifies retrieval coverage (Team et al., 16 Mar 2026). This is significant because it locates quality control at the plan-selection stage rather than only after tool execution.

3. Agentic mid-training curriculum

A defining feature of MiroThinker-1.7 is an agentic mid-training stage designed to strengthen three atomic capabilities: cold-start structured planning, context-conditioned multi-turn reasoning, and intermediate summarization under partial observations (Team et al., 16 Mar 2026). The paper presents this as a large-scale, multi-phase curriculum rather than a single monolithic fine-tuning stage.

The training data are drawn from two explicitly named sources. The first source is single-turn planning examples consisting of a plan plus a first tool call. The second source is isolated reasoning or summarization turns taken from successful multi-step trajectories. This decomposition matters because it separates plan synthesis from later-step evidence processing, allowing the model to learn distinct agentic primitives under controlled supervision.

The core objective is per-turn next-token prediction over contexts that may represent either a bare query or a trajectory prefix. The paper denotes the mid-training loss as $$H_t = \{(T_1, A_1, O_1), \ldots, (T_{t-1}, A_{t-1}, O_{t-1})\}.$$9, with $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$0 as the context up to step $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$1, $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$2 as the target assistant output, and $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$3 as the model’s next-token distribution (Team et al., 16 Mar 2026). It also gives an illustrative structured planning loss with a diversity or retrieval-coverage regularizer $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$4 weighted by $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$5. In functional terms, this means that plan generation is not treated as unconstrained free-form text generation; it is explicitly tied to retrieval coverage and plan diversity.

The paper further states that mid-training is mixed with generic instruction-following and knowledge-intensive corpora to avoid forgetting. This is methodologically important because it frames the agentic curriculum as additive specialization rather than replacement of the model’s general capabilities. A plausible implication is that the system aims to preserve broad instruction fidelity while improving the reliability of tool-mediated reasoning steps.

4. Verification in MiroThinker-H1

MiroThinker-H1 extends MiroThinker-1.7 with two complementary verifiers: a local verifier at the step level and a global verifier at the trajectory level (Team et al., 16 Mar 2026). This verification layer is the central distinction between the base model and the heavy-duty variant.

The local verifier intervenes at each step $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$6. Instead of taking only the single highest-probability action, the model samples a small candidate set $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$7. Each candidate is executed in a sandbox to obtain $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$8, and a learned scoring head $$S_t(K) = \{ i \in [1,\ldots,t-1] \mid i \ge t-K \},$$9 estimates action quality. If the best score remains below $\Phi_t$0, the model resamples; otherwise it selects the best candidate. On a hard subset of BrowseComp containing 295 questions, the paper reports that Pass@1 jumps from 32.1% to 58.5%, while total steps drop from 1185.2 to 210.8 (Team et al., 16 Mar 2026). The combination of higher Pass@1 and sharply reduced step count indicates that local verification is used for both reliability and search-efficiency control.

The global verifier scores full candidate trajectories after they are generated. For up to $\Phi_t$1 full trajectories $\Phi_t$2, the system computes

$\Phi_t$3

Coverage measures the fraction of subgoals with confirmed evidence, while coherence measures chain logical consistency via an LLM judge. The system selects the trajectory with the largest $\Phi_t$4; if that score is below $\Phi_t$5, the agent is asked to extend or refine the chain (Team et al., 16 Mar 2026).

The reported effect of global verification is an additional +14.2 points on BrowseComp and +8.3 on SEAL-0, which the paper identifies as making MiroThinker-H1 the new SOTA (Team et al., 16 Mar 2026). Methodologically, this separates local action validity from trajectory-level evidential integrity. A plausible implication is that H1 treats long-horizon reasoning as a two-scale control problem: local decisions must be useful, but full solutions must also exhibit coverage and coherence when audited holistically.

5. Tool interface, context management, and long-horizon operation

MiroThinker-1.7 ships with a unified tool interface in the MiroThinker codebase. The tools are grouped into three categories. Information retrieval includes google_search(query), which returns ranked (title, URL, snippet), and scrape_and_extract_info(URL), which returns a raw page plus an LLM-distilled summary. Code execution is provided through a Linux sandbox with create_sandbox(), run_command(cmd), and run_python_code(script). File and data transfer includes upload_file_from_local_to_sandbox, download_file_from_sandbox_to_local, and download_file_from_internet_to_sandbox (Team et al., 16 Mar 2026).

The tool stack is tightly integrated with the model’s context policy. All tool outputs are truncated to $\Phi_t$6 tokens via $\Phi_t$7, and the paper states that a [Result truncated] marker is inserted when truncation occurs. Malformed calls are caught and auto-corrected. Together with the full thought–action trace and observation truncation, this creates a distinction between preserved reasoning state and selectively retained environmental evidence.

Context management uses a sliding window of size $\Phi_t$8 for observations while preserving the full thought–action trace. The paper states that this supports token budgets of up to 256K and sustains deep reasoning over hundreds of turns (Team et al., 16 Mar 2026). This architectural choice is notable because it does not simply compress the entire interaction history into summaries; instead, it preserves a complete action-level trace while bounding direct observation exposure. That design suggests an attempt to maintain procedural continuity without allowing large retrieved artifacts to dominate the active context window.

A common misconception is that long-context capability alone is sufficient for deep research behavior. The system description indicates otherwise: the long context budget is paired with structured planning, multi-turn reasoning, tool execution, truncation control, and retry logic. In the paper’s formulation, these components operate jointly rather than as interchangeable substitutes.

6. Benchmarks, performance, and efficiency

The evaluation spans agentic open-web and domain-specific benchmarks: BrowseComp, BrowseComp-ZH, Humanity’s Last Exam, GAIA, DeepSearchQA, SEAL-0, xbench-DeepSearch-2510, FrontierSci-Olympiad, SUPERChem, FinSearchComp, MedBrowseComp, and DeepResearchEval (Team et al., 16 Mar 2026). The protocols distinguish between benchmark families: high-variance sets such as BrowseComp and HLE use avg@3, while others use avg@8. Long-report generation is evaluated with Report Quality and Factuality under LLM-as-judge rather than an accuracy-style metric.

The reported results place MiroThinker-H1 above MiroThinker-1.7 across the principal benchmark suite, with especially large gains on deep research tasks.

Benchmark	M-1.7	M-H1
BrowseComp	74.0	88.2
BrowseComp-ZH	75.3	84.4
HLE (avg@3)	42.9	47.7
GAIA	82.7	88.5
xbench-DeepSearch	62.0	72.0
SEAL-0	53.0	61.3
DeepSearchQA	72.1	80.6

On specialized domains, the same pattern continues. M-H1 scores 79.0 on FrontierSci-Olympiad, 51.3 on SUPERChem (text-only), 73.9 on FinSearchComp (T2/3), and 56.5 on MedBrowseComp, compared with 71.5, 42.1, 67.9, and 54.2 for M-1.7 respectively (Team et al., 16 Mar 2026). These results support the paper’s claim that H1 maintains strong results on specialized domains in addition to leading deep research tasks.

Relative to cited baselines in the tables, H1 exceeds Qwen3.5 on BrowseComp (88.2 vs. 78.6), BrowseComp-ZH (84.4 vs. 70.3), FrontierSci-Olympiad (79.0 vs. 60.6), and FinSearchComp (73.9 vs. 60.8). The comparison is not uniformly dominant across every listed baseline and task: for example, HLE is reported at 47.7 for M-H1 versus 50.4 for GLM-5.0, and SUPERChem is 51.3 for M-H1 versus 53.0 for Seed-2.0 (Team et al., 16 Mar 2026). This is important for interpreting the “state-of-the-art” claim precisely: the paper presents H1 as the new SOTA on deep research tasks, not as an unconditional leader on every benchmark and comparator.

Long-report generation on DeepResearchEval ($\Phi_t$9) is evaluated separately. ChatGPT-5.4 DeepRes. is reported at 76.4 Report Q, 85.5 Factuality Q, and 81.0 Overall Q; M-1.7-mini at 75.4, 78.4, and 76.9; M-1.7 at 76.5, 78.5, and 77.5; and M-H1 at 76.8, 79.1, and 78.0 (Team et al., 16 Mar 2026). A methodological distinction therefore runs through the evaluation suite: some tasks are benchmarked with accuracy percentages, whereas DeepResearchEval uses judge-based quality assessments.

Efficiency is a central part of the 1.7 family narrative. MiroThinker-1.7-mini uses only 3B active parameters versus 30B in full M-1.7. On five agentic benchmarks, it achieves +16.7% average accuracy with −43.0% average turns relative to its M-1.5 predecessor, and typical per-query latency is reduced by approximately 30–50% (Team et al., 16 Mar 2026). These figures clarify the role of the mini variant: it is not presented as a direct accuracy replacement for H1, but as an open-source model that offers competitive research-agent capability with significantly improved efficiency.