IterResearch-30B: Iterative Deep-Research Agent

• IterResearch-30B is a 30B parameter iterative deep-research agent that leverages a Markovian workspace reconstruction paradigm for autonomous long-horizon knowledge construction.
• It employs Efficiency-Aware Policy Optimization—including geometric reward shaping and adaptive downsampling—to ensure stable and scalable reinforcement learning.
• Built on the Qwen3-30B-A3B backbone, the model achieves state-of-the-art performance with significant benchmark improvements and efficient interaction scaling.

IterResearch-30B is a 30-billion parameter iterative deep-research agent that introduces a Markovian workspace reconstruction paradigm for long-horizon autonomous knowledge construction using external tools. Developed on the Qwen3-30B-A3B transformer backbone, IterResearch-30B addresses the limitations of mono-contextual research agents by reformulating the research process as a Markov Decision Process (MDP) and employing Efficiency-Aware Policy Optimization (EAPO) for stable, scalable reinforcement learning. The agent demonstrates state-of-the-art performance among open-source systems, unprecedented interaction scaling, and significant effectiveness as both a standalone agent and as a prompting strategy for frontier models (Chen et al., 10 Nov 2025).

1. Markovian Paradigm and State Reconstruction

IterResearch-30B conceptualizes long-horizon research as an MDP characterized by strategic workspace reconstruction. The agent’s state at each time step is defined as $s_t = (q, \mathcal{M}_t, \{a_{t-1}, TR_{t-1}\})$, where $q$ is the immutable research question, $\mathcal{M}_t$ is the evolving report (compressed memory), and $\{a_{t-1}, TR_{t-1}\}$ represents the last tool action and response. The agent’s action $d_t = (\text{Think}_t, \mathcal{M}_{t+1}, a_t)$ encapsulates its reasoning trace, the next synthesized report, and a tool invocation or final answer.

State transitions only preserve the evolving report $\mathcal{M}$, explicitly discarding full history to enforce $O(1)$ growth in context, circumventing the context suffocation and noise accumulation that degrade mono-contextual agents. This Markovian reconstitution ensures that reasoning capacity is consistently preserved across arbitrary exploration depths, supporting effective multi-step interaction at scale.

2. Reinforcement Learning via Efficiency-Aware Policy Optimization (EAPO)

IterResearch-30B employs EAPO, a reinforcement learning framework designed for iterative deep-research trajectories, introducing several key innovations:

• Geometric Reward Shaping: Rewards are defined as $r_t = \gamma^{T-t}R_T$ with $\gamma \approx 0.995$, incentivizing concise completions by favoring early-correct answers via cumulative gradient amplification. The terminal reward $R_T$ is binary (1 if correct).
• Adaptive Downsampling: Training batches are subsampled to an integer multiple of the data-parallel (DP) batch size, $q$0, ensuring stable distributed optimization while retaining $q$1 of samples.
• Policy Gradient (GSPO with Clipping): The gradient objective is

$q$2

yielding data-efficient, variance-controlled training directly aligned with concise, correct research plans.

The two-stage training protocol comprises supervised fine-tuning (SFT) on 110K report-synthesis trajectories and RL on 4,096 zone-of-proximal questions, initialized for stability and performance.

3. Model Architecture, Data Pipeline, and Compute Budget

The backbone model is Qwen3-30B-A3B, consisting of approximately 30B parameters, ~48 transformer layers, 12K hidden size, and 96 attention heads, supporting 40K-token contexts through augmented positional embeddings. The scaling rules preserve the depth/width ratios of Qwen3 while adapting to extended context and workspace synthesis needs.

The compute budget encompasses approximately $q$3 GPU-hours for SFT (on Alibaba clusters) and $q$4 GPU-hours for RL distributed over 1,024 GPUs. The primary dataset comprises 30K high-quality multi-step research QA pairs, 110K synthesized SFT trajectories (produced by Qwen3-235B-A22B), and RL questions selected by difficulty. Hyperparameters are precisely specified (LR, batch size, epochs, discount factor, rollout numbers), enabling reproducibility and robust scaling.

4. Experimental Performance and Ablations

IterResearch-30B demonstrates substantial empirical advantages:

• Six-Benchmark Suite (Accuracy %)
  • HLE: 28.8
  • BC: 37.3
  • BC-zh: 45.2
  • GAIA: 72.8
  • Xbench-DS: 71.0
  • SEAL-0: 39.6
  • Average: 49.1 (a 14.5 percentage point improvement over the strongest open-source baseline)
• Proprietary Comparison: Surpasses OpenAI DeepResearch on HLE and BC-zh, matches Gemini DeepResearch and Perplexity on BC and GAIA.
• Ablations:
  • EAPO outperforms GSPO, reducing average turn count by 5.7% without accuracy loss.
  • The iterative paradigm gains +12.6pp over mono-contextual approaches—even when the latter uses 64K-token windows.
  • Interaction scaling experiments show accuracy rising from 3.5% to 42.5% as the interaction budget increases from 2 to 2048 steps; average turns per question with a 2048 budget is 80.1.

Training Method	Avg. Acc.	Avg. Turns
SFT	45.5	16.45
GSPO	48.3	19.13
EAPO	49.1	18.04

5. Implementation Practices and Tool Integration

Key development insights and practices include:

• Two-Stage Training: Initial report-synthesis SFT stabilizes RL by constraining the agent to meaningful exploration early on.
• Workspace Management: Enforcing explicit prompt engineering and report compression to maintain $q$5 under 30K tokens is critical for preserving constant context size.
• Adaptive Downsampling: Balances long and short episodes for stable data-parallel optimization.
• Tool Usage: Batch processing of web and search tool queries minimizes API overhead. Explicit, goal-oriented tool calls (“Visit”) enhance focus in summarization and result curation.

6. Limitations and Prospective Developments

Identified constraints include:

• Potential gradual report size drift if synthesis is degraded, indicating a need for periodic report pruning.
• RL is currently focused on moderate-difficulty queries, with harder research challenges under-explored.
• The framework currently employs single-agent RL; multicriteria objectives (e.g., cost, latency vs. accuracy) remain to be jointly optimized.

Proposed directions encompass hierarchical RL for adaptive budget allocation, dynamic discounting in response to task complexity, hybrid on-/off-policy learning to maximize data reuse, incorporation of further tools (e.g., knowledge-graph queries), and auxiliary report compression objectives.

7. Significance in Long-Horizon Research Automation

IterResearch-30B establishes a new methodological foundation for deep-research agents operating over extended reasoning horizons. By combining principled Markovian workspace reconstruction with efficiency-aware policy optimization, the agent delivers both technical progress (interaction scaling up to 2048 steps) and methodological clarity. Its universal design supports modular extensibility and positions it as an effective solution for both autonomous research and advanced prompting strategies, outperforming incumbent open-source paradigms across multiple long-horizon benchmarks (Chen et al., 10 Nov 2025).