Technical Summary: IterResearch

Problem Statement

Recent advances in deep-research agents have demonstrated the potential for autonomous knowledge construction through dynamic reasoning over external sources. However, existing approaches predominantly rely on a mono-contextual paradigm, where all retrieved information and intermediate reasoning steps are appended to a single, continuously expanding context window. The authors identify two critical failure modes in this paradigm for long-horizon tasks:

1. Context Suffocation: As the context window fills with prior interactions, the available space for active reasoning shrinks, forcing constrained responses that degrade into premature or superficial conclusions.
2. Noise Contamination: Irrelevant information from web searches and early exploration errors become permanently embedded in the context, creating cascading interference that dilutes signal quality throughout the reasoning process.

These limitations prevent existing agents from maintaining sustained reasoning capacity as exploration depth increases.

Methodology

The paper introduces IterResearch, a novel iterative deep-research paradigm that reimagines long-horizon research through the lens of Interaction Scaling. The methodology is built on two core components:

1. Iterative Deep-Research Paradigm (MDP-Inspired Architecture)

Instead of linear context accumulation, IterResearch adopts a Markov Decision Process (MDP) formulation defined by the tuple $\langle S, D, E, T, R \rangle$.

• State Space ($S$): Each state $s_t$ is a strategically reconstructed workspace containing only essential elements: the constant question $q$, an evolving report $M_t$ (serving as compressed memory of validated findings), and the immediate context from the previous step.
• Decision Space ($D$): The policy $\pi$ generates a composite decision $d_t$ consisting of internal thought ($Think_t$), an updated report ($M_{t+1}$), and an external action ($a_t$).
• Transition Function ($T$): Crucially, the transition function reconstructs the workspace rather than appending to history. The historical trajectory is deliberately discarded, with only the synthesized knowledge preserved in the updated report. This ensures the workspace size remains constant ($O(1)$) regardless of the number of interaction rounds, theoretically enabling unbounded exploration.

2. Efficiency-Aware Policy Optimization (EAPO)

To train this iterative paradigm effectively, the authors propose EAPO, a training strategy addressing two challenges:

• Geometric Reward Discounting: Recognizing that not all successful trajectories are equally valuable, the authors adapt standard geometric discounting ($r_t = \gamma^{T-t} \cdot R_T$). This creates an implicit efficiency pressure, incentivizing agents to reach correct conclusions through concise, focused exploration rather than exhaustive searching.
• Adaptive Downsampling: Since the iterative paradigm decomposes trajectories into multiple independent training samples (one per round), sample counts vary across questions. The authors employ adaptive downsampling to align training batches with the data-parallel size, ensuring stable distributed training while preserving over 99% of the data.

Key Contributions

1. Novel Paradigm: The proposal of IterResearch, which eliminates context suffocation and noise contamination by maintaining an evolving report memory and performing strategic workspace reconstruction.
2. Training Strategy: The development of EAPO, which combines geometric reward discounting for efficiency and adaptive downsampling for stable training on variable-length trajectories.
3. Empirical Validation: Extensive experiments demonstrating that IterResearch achieves substantial improvements over existing open-source agents and narrows the gap with frontier proprietary systems.

Experimental Results

The authors evaluated IterResearch (based on Qwen3-30B-A3B) across six challenging benchmarks: Humanity's Last Exam (HLE), BrowseComp, BrowseComp-zh, GAIA, Xbench-DeepSearch, and SEAL-0.

• Performance Gains: IterResearch achieved an average improvement of +14.5 percentage points (pp) over existing open-source agents across the six benchmarks. It outperformed all open-source baselines and demonstrated competitive or superior performance compared to proprietary systems (e.g., surpassing OpenAI's DeepResearch on HLE and BrowseComp-zh).
• Interaction Scaling: The paradigm exhibited unprecedented scaling capabilities, extending to 2048 interactions with only a 40K context length. Performance improved dramatically from 3.5% to 42.5% as the maximum interaction limit increased from 2 to 2048, a feat structurally infeasible for mono-contextual approaches due to context overflow.
• Cross-Paradigm Transfer: Trajectories generated by IterResearch were found to significantly enhance mono-contextual agents when used as training data, suggesting the paradigm induces superior exploration behaviors transferable across different architectures.
• Prompting Strategy: Without any training, applying the IterResearch prompting structure to frontier models (e.g., o3, DeepSeek-V3.1) yielded substantial improvements over standard ReAct prompting, particularly on long-horizon tasks (up to +19.2pp on BrowseComp).

Significance and Claims

The paper positions IterResearch as a versatile solution for long-horizon reasoning, effective both as a trained agent and as a model-agnostic prompting paradigm. The authors claim three fundamental insights:

1. Interaction Scaling: The iterative paradigm unlocks extreme interaction scaling, suggesting that the perceived difficulty of long-horizon tasks may stem from insufficient exploration capacity rather than inherent reasoning limits.
2. Superior Exploration Behaviors: The paradigm induces high-quality exploration behaviors that create transferable training signals, even for agents using different architectural paradigms.
3. Model-Agnostic Efficacy: The iterative structure with workspace reconstruction provides a fundamental cognitive advantage for long-horizon reasoning, improving performance across diverse model architectures without requiring specific fine-tuning.

The work concludes that iteration with strategic synthesis, rather than linear accumulation, is fundamental to conquering long-horizon reasoning challenges.