DLLM-Searcher: Adapting Diffusion Language Model for Efficient Search Agents
sigir26-119 · Full research paper · Jiahao Zhao, Shaoxuan Xu, Zhongxiang Sun, Fengqi Zhu, Jingyang Ou, Yuling Shi, Chongxuan Li, Xiao Zhang, Jun Xu
Efficient and Long-Context RAG · 25% Agentic and Deep Search RAG · 25% Reward Modeling and Process Supervis · 25% Reasoning Evaluation and Benchmarks · 25%
Abstract
Recently, Diffusion Large Language Models (dLLMs) have demonstrated unique efficiency advantages, enabled by their inherently parallel decoding mechanism and flexible generation paradigm. Meanwhile, despite the rapid advancement of Search Agents, their practical deployment is constrained by a fundamental limitation, termed as 1) Latency Challenge : the serial execution of multi-round reasoning, tool calling, and tool response waiting under the ReAct agent paradigm induces severe end-to-end latency. Intuitively, dLLMs can leverage their distinctive strengths to optimize the operational efficiency of agents under the ReAct agent paradigm. Practically, existing dLLM backbones face the 2) Agent Ability Challenge. That is, existing dLLMs exhibit remarkably weak reasoning and tool-calling capabilities, preventing these advantages from being effectively realized in practice. In this paper, we propose DLLM-Searcher, an optimization framework for dLLM-based Search Agents. To solve the Agent Ability Challenge, we design a two-stage post-training pipeline encompassing Agentic Supervised Fine-Tuning (Agentic SFT) and Agentic Variance-Reduced Preference Optimization (Agentic VRPO), which enhances the backbone dLLM's information seeking and reasoning capabilities. To mitigate the Latency Challenge, we leverage the flexible generation mechanism of dLLMs and propose a novel agent paradigm termed Parallel-Reasoning and Acting (P-ReAct). P-ReAct guides the model to prioritize decoding tool_call instructions, thereby allowing the model to keep thinking while waiting for the tool's return. Experimental results demonstrate that DLLM-Searcher achieves performance comparable to mainstream LLM-based search agents and P-ReAct delivers approximately 15% inference acceleration. Our code is available at https://github.com/bubble65/DLLM-Searcher
Assigned reviewers
| # | Reviewer | Match | Why (closest evidence publications in the taxonomy) | |
|---|---|---|---|---|
| 1 | Haritz Puerto | 56 | 51 Fine-Tuning on Diverse Reasoning Chains Drives Within-Inference CoT Refinement in LLMs (2025) Reward Modeling and Process Supervision 51 Code Prompting Elicits Conditional Reasoning Abilities in Text+Code LLMs (2024) Reward Modeling and Process Supervision 47 C-SEO Bench: Does Conversational SEO Work? (2025) Reasoning Evaluation and Benchmarks |
|
| 2 | Archiki Prasad | 52 | 54 MAgICoRe: Multi-Agent, Iterative, Coarse-to-Fine Refinement for Reasoning (2025) Reward Modeling and Process Supervision 52 ReCEval: Evaluating Reasoning Chains via Correctness and Informativeness (2023) Reasoning Evaluation and Benchmarks 51 LASeR: Learning to Adaptively Select Reward Models with Multi-Arm Bandits (2025) Reward Modeling and Process Supervision |
|
| 3 | Yong Li 0004 | 42 | 56 PG-Agent: An Agent Powered by Page Graph (2025) Multi-Agent and Tool-Use Orchestration 52 ProBench: Benchmarking GUI Agents with Accurate Process Information (2026) Reasoning Evaluation and Benchmarks 51 RFS-Guard: Detecting Reasoning Hallucinations via Cross-Phase Routing Focus in Large Reasoning Models (2026) Reward Modeling and Process Supervision |
Area chair
| Area chair | Match | |
|---|---|---|
| Hongjin Qian | 52 |
Area-chair candidates (4)
| Hongjin Qian assigned | 52 | 12 papers | |
| Xiaoxi Li 0005 | 51 | 15 papers | |
| Dacheng Tao | 44 | 5 papers | |
| Pengyue Jia | 35 | 10 papers |
Candidates from the panel ranked by match score
| # | Reviewer | Match | Load | Why | |
|---|---|---|---|---|---|
| 1 | Haritz Puerto assigned | 56 | 3 | 51 Fine-Tuning on Diverse Reasoning Chains Drives Within-Inference CoT Refinement in LLMs (2025) Reward Modeling and Process Supervision 51 Code Prompting Elicits Conditional Reasoning Abilities in Text+Code LLMs (2024) Reward Modeling and Process Supervision |
|
| 2 | Archiki Prasad assigned | 52 | 3 | 54 MAgICoRe: Multi-Agent, Iterative, Coarse-to-Fine Refinement for Reasoning (2025) Reward Modeling and Process Supervision 52 ReCEval: Evaluating Reasoning Chains via Correctness and Informativeness (2023) Reasoning Evaluation and Benchmarks |
|
| 4 | Shu Zhou 0002 | 51 | 3 | 57 The Retrieval Bottleneck: Scaling Laws for Reinforcement Learning in RAG (2026) Efficient and Long-Context RAG 42 Activation Caching for Retrieval-Augmented Generation (2026) Efficient and Long-Context RAG |
|
| 5 | Wonduk Seo | 51 | 3 | 52 Automated Visualization Code Synthesis via Multi-path Reasoning and Feedback-Driven Optimizations (2026) Reasoning Evaluation and Benchmarks 42 SPIO: Ensemble and Selective Strategies via LLM-Based Multi-Agent Planning in Automated Data Science (2026) Multi-Agent and Tool-Use Orchestration |
|
| 7 | Farima Fatahi Bayat | 48 | 3 | 66 From Proof to Program: Characterizing Tool-Induced Reasoning Hallucinations in Large Language Models (2026) Reward Modeling and Process Supervision 31 FactBench: A Dynamic Benchmark for In-the-Wild Language Model Factuality Evaluation (2025) Automated Fact Checking and Claim Verification |
|
| 8 | Peitian Zhang | 48 | 3 | 53 Tackling the Length Barrier: Dynamic Context Browsing for Knowledge-Intensive Task (2025) Efficient and Long-Context RAG 52 Search-o1: Agentic Search-Enhanced Large Reasoning Models (2025) Agentic and Deep Search RAG |
|
| 9 | Jialong Wu 0007 | 48 | 3 | 68 WebWalker: Benchmarking LLMs in Web Traversal (2025) Agentic and Deep Search RAG 51 SEED: Accelerating Reasoning Tree Construction via Scheduled Speculative Decoding (2025) Reward Modeling and Process Supervision |
|
| 10 | Qingyi Si | 45 | 3 | 53 S-GRPO: Early Exit via Reinforcement Learning in Reasoning Models (2025) Reward Modeling and Process Supervision 51 Test-time Prompt Intervention (2026) Reward Modeling and Process Supervision |
|
| 11 | Kaixin Ma | 45 | 3 | 69 Retrieval-augmented GUI Agents with Generative Guidelines (2025) Agentic and Deep Search RAG 52 DSBench: How Far Are Data Science Agents from Becoming Data Science Experts? (2025) Reasoning Evaluation and Benchmarks |
|
| 13 | Yongqi Li 0001 | 43 | 3 | 51 Parallel Test-Time Scaling for Latent Reasoning Models (2026) Reward Modeling and Process Supervision 51 TokenSkip: Controllable Chain-of-Thought Compression in LLMs (2025) Reward Modeling and Process Supervision |
|
| 14 | Yong Li 0004 assigned | 42 | 3 | 56 PG-Agent: An Agent Powered by Page Graph (2025) Multi-Agent and Tool-Use Orchestration 52 ProBench: Benchmarking GUI Agents with Accurate Process Information (2026) Reasoning Evaluation and Benchmarks |
|
| 15 | Xi Ye 0003 | 42 | 3 | 55 Beyond Single-shot Writing: Deep Research Agents are Unreliable at Multi-turn Report Revision (2026) Deep Research and Report Generation 52 SatLM: Satisfiability-Aided Language Models Using Declarative Prompting (2023) Reasoning Evaluation and Benchmarks |
|
| 16 | Yu Gu 0016 | 42 | 3 | 55 Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (2025) Deep Research and Report Generation 52 AgentBench: Evaluating LLMs as Agents (2024) Reasoning Evaluation and Benchmarks |
|
| 17 | Daiting Shi | 42 | 3 | 51 ConsistRM: Improving Generative Reward Models via Consistency-Aware Self-Training (2026) Reward Modeling and Process Supervision 51 ReflectRM: Boosting Generative Reward Models via Self-Reflection within a Unified Judgment Framework (2026) Reward Modeling and Process Supervision |
|
| 18 | Keqin Bao | 41 | 3 | 54 Teaching Language Models to Reason with Tools (2025) Reasoning Evaluation and Benchmarks 52 MTR-Bench: A Comprehensive Benchmark for Multi-Turn Reasoning Evaluation (2026) Reasoning Evaluation and Benchmarks |
|
| 19 | Farzad Shami | 41 | 3 | 54 GROKE: Vision-Free Navigation Instruction Evaluation via Graph Reasoning on OpenStreetMap (2026) Reasoning Evaluation and Benchmarks 33 Benchmarking Large Language Models for Knowledge Graph Validation (2026) KG-Augmented LLM Reasoning |
|
| 20 | Wenlin Zhang 0001 | 41 | 3 | 69 Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learning (2025) Agentic and Deep Search RAG 57 To Search or Not to Search: Aligning the Decision Boundary of Deep Search Agents via Causal Intervention (2026) Agentic and Deep Search RAG |
|
| 21 | Yingyi Zhang 0001 | 40 | 3 | 69 Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learning (2025) Agentic and Deep Search RAG 58 To Search or Not to Search: Aligning the Decision Boundary of Deep Search Agents via Causal Intervention (2026) Agentic and Deep Search RAG |
|
| 22 | Jing Liu 0022 | 37 | 3 | 38 BEE-RAG: Balanced Entropy Engineering for Retrieval-Augmented Generation (2026) Efficient and Long-Context RAG 37 Reinforced Informativeness Optimization for Long-Form Retrieval-Augmented Generation (2026) Evidence Selection and Context Refinement |
|
| 23 | Xinyu Ma 0001 | 37 | 3 | 58 Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning (2025) Agentic and Deep Search RAG 42 Probe-and-Fetch: Dynamic KV Cache Pruning for Accelerated Long-Context Inference in Web-Scale AI Search (2026) Efficient and Long-Context RAG |
|
| 24 | Nancy F. Chen | 37 | 3 | 53 Learning Planning-based Reasoning by Trajectories Collection and Process Reward Synthesizing (2024) Reward Modeling and Process Supervision 51 AdaMCoT: Rethinking Cross-Lingual Factual Reasoning Through Adaptive Multilingual Chain-of-Thought (2026) Reward Modeling and Process Supervision |
|
| 25 | Ori Yoran | 36 | 3 | 54 AssistantBench: Can Web Agents Solve Realistic and Time-Consuming Tasks? (2024) Reasoning Evaluation and Benchmarks 37 Answering Questions by Meta-Reasoning over Multiple Chains of Thought (2023) Reasoning-Intensive Retrieval and Ranking |
|
| 26 | Chen Zhao 0013 | 36 | 3 | 55 Beyond Single-shot Writing: Deep Research Agents are Unreliable at Multi-turn Report Revision (2026) Deep Research and Report Generation 48 Rethinking Reasoning-Intensive Retrieval: Evaluating and Advancing Retrievers in Agentic Search Systems (2026) Reasoning-Intensive Retrieval and Ranking |
|
| 27 | Weiwei Sun 0001 | 35 | 3 | 58 Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning (2025) Agentic and Deep Search RAG 39 ReasonRank: Empowering Passage Ranking with Strong Reasoning Ability (2026) LLM-Based Reranking |
|
| 28 | Qinggang Zhang | 35 | 3 | 48 LoSemB: Logic-Guided Semantic Bridging for Inductive Tool Retrieval (2026) Reasoning Evaluation and Benchmarks 36 FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation (2025) Knowledge Conflict and Faithfulness |
|
| 29 | Yujia Zhou 0002 | 35 | 3 | 38 Parametric Retrieval Augmented Generation (2025) Efficient and Long-Context RAG 36 Robust Fine-tuning for Retrieval Augmented Generation against Retrieval Defects (2025) Knowledge Conflict and Faithfulness |
|
| 30 | Yuanzhao Zhai | 35 | 3 | 54 Empowering Large Language Model Agent through Step-Level Self-Critique and Self-Training (2025) Reward Modeling and Process Supervision 54 Enhancing Decision-Making for LLM Agents via Step-Level Q-Value Models (2025) Reward Modeling and Process Supervision |
|
| 31 | Xingyu Tan 0001 | 35 | 3 | 36 HydraRAG: Structured Cross-Source Enhanced Large Language Model Reasoning (2025) KG-Augmented LLM Reasoning 36 Paths-over-Graph: Knowledge Graph Empowered Large Language Model Reasoning (2025) KG-Augmented LLM Reasoning |
|
| 32 | Di Wu 0054 | 35 | 3 | 50 Repoformer: Selective Retrieval for Repository-Level Code Completion (2024) Efficient and Long-Context RAG 36 Synchronous Faithfulness Monitoring for Trustworthy Retrieval-Augmented Generation (2024) Knowledge Conflict and Faithfulness |
|
| 33 | Weipeng Chen | 35 | 3 | 56 ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning (2025) RL Training of Search Agents 52 Facilitating Multi-turn Function Calling for LLMs via Compositional Instruction Tuning (2025) Reasoning Evaluation and Benchmarks |