W&D: Scaling Parallel Tool Calling for Efficient Deep Research Agents (Feb 2026, arXiv)

Forum topic · 小凯 · 2026-07-05

Summary

W&D is a research paper by Xiaoqiang Lin, Jun Hao Liew, Silvio Savarese, and Junnany Li (listed as Junnan Li) focused on scaling parallel tool calling to improve the efficiency of deep research agents. The work targets agentic search systems that must gather information from external tools, where sequential tool calls create latency and cost bottlenecks. By issuing tool calls in parallel, W&D aims to reduce wall-clock time and token overhead while preserving answer quality for multi-step research tasks. The paper is indexed on arXiv at https://arxiv.org/abs/2602.07359 and belongs to the Deep Research category alongside surveys of LLM-based search and scientific agents. This forum entry summarizes the paper's context within the evolution from pipeline retrieval to agentic paradigms, where the number and strategy of retrieval steps become learnable decisions. Key discussion themes include the decomposition of agent systems into retrievers, rerankers, planners, and generators; trade-offs among latency, cost, and answer trustworthiness; and open challenges such as evaluation reliability, hallucination, safety on the open web, and multilingual extension. Readers are advised to consult the original PDF for quantitative results, as this summary is based on the abstract and public metadata.

W&D: Scaling Parallel Tool Calling for Efficient Deep Research Agents

Paper: W&D: Scaling Parallel Tool Calling for Efficient Deep Research Agents (arXiv, Feb 2026) Authors: Xiaoqiang Lin, Jun Hao Liew, Silvio Savarese, Junnan Li Category: Deep Research

Key points

Core idea: W&D addresses the efficiency bottleneck of deep research agents by scaling parallel tool calling, allowing multiple external tool invocations to proceed concurrently instead of strictly sequentially.
Problem context: Agentic search systems face challenges in efficiency, scalability, and intent understanding. Traditional pipelines separate retrieval, ranking, and generation, which poorly fits LLM-era requirements for natural-language interaction, multi-hop reasoning, and real-time knowledge.
Positioning: The work sits at the intersection of agentic search and large-scale search/recommendation systems, asking how responsibilities among retrieval, ranking, generation, and tool calling should be re-allocated in the LLM era. New variables are the inference budget and action space (whether to retrieve, how often, and which tools to call).

System perspective

Typical agentic methods follow a pipeline of problem formalization, model/system design, training or construction, and inference. Components may include:

1. Input & representation: encoding queries, documents, and user context into dense/sparse representations or structured prompts. 2. Core modules: retriever, reranker, planner, memory, and tool interfaces, arranged sequentially or in parallel. 3. Learning strategies: supervised fine-tuning, contrastive learning, distillation, reinforcement learning (including process rewards), and synthetic data bootstrapping. 4. Inference strategies: single-pass retrieval, iterative retrieval, parallel sub-queries, early stopping, and budget control.

W&D's parallel tool calling targets the last category, reducing end-to-end latency for multi-step research workflows.

Evaluation considerations

For papers in this area, typical setups include:

Datasets: MS MARCO, BEIR, Natural Questions, domain corpora, recommendation benchmarks.
Metrics: nDCG@10, MRR, Recall@k, task success rate, latency, and token cost.
Baselines: BM25, dense retrieval, cross-encoder reranking, retrieval-free LLMs, commercial search APIs.

> Note: Quantitative results should be verified against the original PDF; this summary is based on the abstract and public metadata.

Insights for Search / Rec / Personalization

1. Architecture: Cascaded retrieval + rerank + generation remains mainstream, but the agentic paradigm makes the retrieval count and policy themselves learnable. 2. Data: High-quality instruction data and session logs matter; synthetic data requires care against leakage and distribution shift. 3. Evaluation: The gap between offline metrics and online satisfaction is widening; LLM-as-judge should be cross-validated with human evaluation. 4. Deployment: Latency, cost, interpretability, and safety are hard constraints for production systems.

Limitations and open questions

Possible limitations include experiment scale constrained by compute budgets, benchmark–real-user distribution mismatch, English-centric data limiting cross-lingual generalization, and safety risks of agents operating on the open web. Future directions include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for recommendation.

Related entries

A Comprehensive Survey of Deep Research: Systems, Methodologies, and Applications (arXiv:2506.12594)
A Survey of LLM-based Deep Search Agents (arXiv:2508.05668)
A Survey of Scientific Large Language Models (arXiv:2508.21148)
Towards Scientific Intelligence: A Survey of LLM-based Scientific Agents (arXiv:2503.24047)
AgentIR: Reasoning-Aware Retrieval for Deep Research Agents (arXiv:2603.04384)
Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning (arXiv:2502.04644)

Glossary

| Term | Meaning | |------|---------| | IR | Information Retrieval | | RAG | Retrieval-Augmented Generation | | nDCG | Normalized Discounted Cumulative Gain | | Agentic Search | Modeling search as sequential decision-making and tool calling | | Gen-IR | Generative Information Retrieval |