MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling

TL;DR

MiroThinker is an open-source research agent (8B/30B/72B; the v1.5 release adds a 235B MoE variant on top of Qwen3) trained to do deep web-search–driven research with up to 600 tool calls per task inside a 256K context window. The paper’s framing thesis is that “interaction depth” — how often and how deep the agent loops with its environment — is a third scaling axis alongside model size and context length, distinct from LLM test-time reasoning scaling. Training is a three-stage pipeline: agentic SFT on synthesized multi-hop trajectories, DPO with an auxiliary SFT loss on preferred trajectories, then online GRPO with streaming rollouts. On BrowseComp, BrowseComp-ZH, HLE, and GAIA-Text-Only, the 72B v1.0 reaches 47.1 / 55.6 / 37.7 / 81.9, approaching closed systems like Claude Research and ChatGPT Agent.

Key claims

Interaction depth is proposed as a third scaling dimension for research agents — alongside model size and context length — with empirical scaling curves showing that accuracy improves predictably as the agent makes more tool calls per task [Abstract, §1].
A 256K context window combined with recency-based context retention enables up to 600 tool calls per task, compared to fewer than 100 in prior open-source research agents [§3.3, §1].
Recency-based retention keeps the full thought/action trace but masks all but the most recent K tool responses; the paper reports this simple strategy does not degrade performance and frees context for deeper trajectories [§3.3, Eq. 4–7].
A three-stage post-training pipeline — agentic SFT → DPO with auxiliary SFT loss → online GRPO with streaming rollouts — is used over a synthesized trajectory dataset built from a MultiDocQA pipeline that constructs multi-hop questions from interlinked Wikipedia/CommonCrawl subgraphs with constraint obfuscation [§4, §5].
Trajectory synthesis uses two paradigms (ReAct single-agent and MiroFlow multi-agent) and two tool-call mechanisms (function calling and MCP), with multiple driver LLMs (GPT-OSS, DeepSeek-V3.1, …) to diversify trajectory styles [§4.2].
A streaming rollout mechanism handles the long-tailed completion times of multi-round agentic RL by re-queuing unfinished tasks each iteration, rather than waiting for the slowest trajectory in a batch [§5.3].
On BrowseComp, MiroThinker-v1.0-72B reaches 47.1% accuracy, +2.0 over MiniMax-M2 (the strongest open-source comparison) [§1, Table on benchmarks].
On GAIA-Text-Only, MiroThinker-v1.0-72B reaches 81.9% vs. MiniMax-M2’s 75.7%, a +6.2 absolute gap [§1].
The v1.5 model card (HF) reports 39.2 / 69.8 / 71.5 / 80.8 on HLE-Text / BrowseComp / BrowseComp-ZH / GAIA-Val-165, and extends to 235B-MoE and a 400-tool-call budget per task [HF model card].

Method

The agentic loop is a standard ReAct trajectory of (thought, action, observation) triples; the model alternates internal thinking, MCP-style tool invocation, and observation until it emits no further action. The tool suite is a Linux sandbox (with create_sandbox, run_command, run_python_code), file management utilities, and information retrieval (google_search, plus a scrape_and_extract_info web-scraper that internally runs a small LLM, Qwen3-14B, to extract task-relevant content from pages before feeding it back to the agent). HuggingFace access is explicitly blocked in the tools to prevent benchmark-answer leakage.

Context management is two-fold: recency-based retention masks tool responses outside a sliding window K (keeping all thoughts/actions intact), and result truncation caps overly long tool outputs with a [Result truncated] tag. Training data is built by a MultiDocQA synthesis pipeline that samples document subgraphs from Wikipedia/CommonCrawl by following internal hyperlinks, extracts cross-document facts tied to a seed document’s theme, then obfuscates the facts (temporal/spatial generalization, referential indirection) before asking an LLM to compose multi-hop questions. Agentic trajectories are then synthesized using ReAct and MiroFlow multi-agent paradigms with multiple driver LLMs, filtered to remove repetition/duplication/malformed tool calls, and used for SFT. DPO is run with an auxiliary SFT loss on preferred trajectories (chosen on correctness without enforcing structural constraints like step counts). RL is online GRPO with one update per rollout, a streaming rollout queue to handle long-tailed agentic trajectories, a correctness-plus-format reward, and trajectory curation that filters both noisy-correct (repeated API failures, timeout loops) and trivially-incorrect (format failures, action loops) samples.

Results

On BrowseComp, MiroThinker-v1.0-72B reaches 47.1% (+2.0 over MiniMax-M2 at 44.0, +2.0 over GLM-4.6 at 45.1, ~2x over DeepSeek-V3.1 at 30.0). On BrowseComp-ZH it reaches 55.6%, +6.1 over GLM-4.6 (49.5). On HLE it reaches 37.7%, +4.8 over Tongyi-DeepResearch. On GAIA-Text-Only it reaches 81.9%, +6.2 over MiniMax-M2 (75.7). The v1.5 release pushes these further: 39.2 (HLE-Text) / 69.8 (BrowseComp) / 71.5 (BrowseComp-ZH) / 80.8 (GAIA-Val-165), with BrowseComp being claimed as the new world-leading open-source figure. The interaction-scaling figure in the paper shows accuracy on BrowseComp/HLE/GAIA continuing to climb up to several hundred tool calls per task — not saturating at the prior open-source ~100-call ceiling.

Why it’s interesting

The “interaction depth as a third scaling axis” framing is the part worth carrying forward, even if the curves are noisier than they look. Existing scaling stories are (i) parameter count, (ii) context length, (iii) test-time chain-of-thought tokens — all of which act on a single forward pass or a closed reasoning chain. MiroThinker’s claim is that the per-task budget of environment interactions is its own dimension, and that the bottleneck for prior open-source research agents (<100 tool calls) was not the model but the training+inference plumbing. For Luma, the relevant transferable pieces are the streaming GRPO rollout (handles the long-tail-trajectory problem that any video-agent or interactive-generation pipeline would also hit), and the recency-based context retention as a cheap baseline for long-horizon multi-step generative pipelines. The data construction pipeline — hyperlink subgraph → obfuscated multi-hop QA — is also a generally useful recipe for synthesizing reasoning tasks where ground truth is checkable. Worth watching: how much of the 30B → 72B → 235B scaling actually comes from base capacity vs. interaction-budget headroom that smaller models can’t usefully spend.