cs.CLOct 11, 2026

RIVO: Lightweight Adaptive Rerouting for Deep Research Agents

Authors: Mingchen Li, Jiatan Huang, Hong Yu

Organizations: University of Massachusetts, Amherst · University of Connecticut

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Deep Research agents rely on long-horizon reasoning and iterative tool use to solve complex information-seeking tasks. However, adaptive decision-making over these trajectories remains challenging, as agents may stop prematurely when further evidence is needed, continue unnecessarily after sufficient evidence has been gathered, or fail to redirect unproductive trajectories. Existing solutions either require costly model retraining and dense turn-level supervision, or rely on heuristic control rules that are not directly optimized for downstream research utility. This motivates a lightweight, utility-aware alternative for Deep Research control. We formulate adaptive rerouting in long-horizon Deep Research as a state-conditioned trajectory correction problem. Based on this view, we introduce RIVO, a lightweight plug-in framework for adaptive trajectory correction. The agent estimates its research state from lightweight signals and determines whether trajectory redirection is needed. Once triggered, the plug-in modifies the Host token logits to steer generation toward a more promising research direction. Theoretical analysis shows that, under a value-alignment condition and bounded local curvature, a sufficiently small logit intervention can locally improve expected downstream research utility. We further construct BrowseCore, a curated dataset with a paired local corpus, for efficient and controlled Deep Research experimentation. Building on this resource, the plug-in is trained from downstream outcomes, enabling it to learn corrections that improve subsequent research and final task performance. Extensive experiments across multiple Deep Research benchmarks and control settings demonstrate the effectiveness of our method, showing that lightweight logit intervention can improve long-horizon research without retraining the underlying Host model.

Figures & tables

Figure 1: Overview of RIVO . RIVO consists of three stages: state-guided control determines whether the current Deep Research trajectory should continue, stop, verify or reroute; logit intervention uses a lightweight plug-in to modify selected Host logits when rerouting is triggered; and trajectory-level optimization trains the plug-in while keeping the Host model frozen. At inference time, the learned intervention is activated only when rerouting is required.
Deep Research ModelMethodBrowseComp-PlusBrowseCoreDRBench
  SR ↑\uparrowSR ↑\uparrowIR ↑\uparrowFact ↑\uparrowDA ↑\uparrowRQ ↑\uparrowHM ↑\uparrow
LiteResearcher-4BSource32.2953.485.2454.4799.9081.2317.27
ASAG33.6154.6011.7687.0099.5776.6033.43
RAAC33.7354.046.4151.5299.6381.8920.23
PPDPP32.6555.153.3473.1499.6376.7411.91
ReAct31.5754.328.7651.3499.6080.7925.64
Table 1: Main results on BrowseComp-Plus, BrowseCore, and DRBench. SR denotes Success Rate. For DRBench, IR, Fact, DA, RQ, and HM denote Insight Recall, Factuality, Distractor Avoidance, Report Quality, and Harmonic Mean, respectively. Higher is better for all metrics.
Figure 2: Sensitivity analysis of RIVO under different hyperparameter settings.
Figure 3: Sensitivity to the state-control thresholds λ\lambda and τH\tau_{H} . Higher SR is better.
Figure 4: Ablation of state signals.
VariantSR (%)
Plug-in Function
w/o Reroute Predictor (Pre)53.63
w/o Logit Intervention56.82
w/o Logit Intervention+Reroute Pre55.71
Full RIVO59.30
Rerouting Representation
Table 2: Ablation of the rerouting design.
Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Controller DesignSR (%)
STOP-only56.27
REROUTE-only55.15
Standalone 0.6B 3-way47.91
w/o Verify54.32
Joint controller (Ours)59.30
Appendix
Table 3: Ablation study of different controller designs.
Figure 5: Structural ablations of RIVO .
Figure 6: System prompt used by LiteResearcher during Deep Research. The model iteratively reasons, invokes browser tools, and produces a final answer after sufficient evidence has been collected.
Figure 7: Browser tool specifications provided to LiteResearcher-4B. The agent uses browser.search for retrieval, browser.open for webpage inspection, and browser.find for locating specific information within a page.
Figure 8: Prompt template used to annotate each deep-research turn with Continue , or Reroute .
Figure 9: Prompt template for automated evidence-document verification.
Figure 10: Prompt template for atomic clue decomposition. Given an original question and its reference answer, the model decomposes the information need into a minimal set of self-contained factual clues for subsequent evidence retrieval and verification.
Figure 11: Overview of our benchmark and corpus construction pipeline. We start from OpenResearcher question–answer pairs, collect candidate evidence through clue-level retrieval, and combine GPT-5 mini assessment with human verification to filter unverifiable questions. Verified supporting documents are labeled as positive and gold evidence, while high-ranking non-supporting documents are retained as hard negatives. The final corpus augments the BrowseComp-Plus retrieval corpus with these verified evidence documents and mined distractors, resulting in a compact and efficient benchmark for controlled Deep Research experiments.
More in this topic

Deep Research Agents

Deep research agents plan long-horizon searches, synthesize cited evidence, and produce reports for complex open-ended questions.

See all 115 papers

Explore similar work

CardsList
Sep 28, 2026

DeepRewind: Predicting and Repairing Premature Commitments in Deep Research Agents

Deep-research agents conduct long-horizon investigations through iterative search, evidence evaluation, belief revision, and synthesis. However, they may commit to claims before sufficient evidence is available, causing later reasoning to reinforce an incorrect interpretation. We introduce DeepRewind, an additive control layer for reversible deep research that represents the agent's evolving epistemic state as a typed graph of sources, evidence, claims, hypotheses, assumptions, commitments, plans, and drafts. Before accepting an intermediate conclusion, a prompt-based world model predicts its impact and estimates reversibility based on hypothesis narrowing, information loss, recovery cost, and contradiction-trigger coverage. A binary controller blocks risky commitments, while a consistency monitor performs dependency-aware rollback when later evidence invalidates them. Across DRBench and LiveDRBench, DeepRewind improves insight recall by 3.6 percentage points and reduces premature commitments by 59.1% relative to Open Deep Research.

Amirhossein Abaskohi, Amirhossein Dabiriaghdam, Lele Wang +2

University of British Columbia

Deep Research AgentsLLM Agent Reliability
AskPDF
Sep 1, 2026

Explore Before Committing: Hypothesis-Guided Search for Deep Research Agents

Deep-research agents answer complex questions by interacting with search and browsing tools, yet they often search along a single evolving trajectory. Our trajectory-level analysis reveals a common failure mode in which the agent may encounter an early search state with several plausible directions, but follow one direction before collecting enough comparative evidence. Once this happens, subsequent tool calls tend to reinforce the same path, increasing the chance of failure when the initial direction is misleading. We further find that successful trajectories reduce this risk through two behaviors: grounding vague exploration in concrete candidates and shifting directions when the current path is weak or incomplete. Based on these findings, we propose HypoSearch, which generates lightweight hypotheses as soft search hints, explores them through bounded independent branches, and compares branch-level evidence before commitment. Across four deep-research benchmarks and three backbone models, HypoSearch consistently outperforms single-trajectory search and standard parallel baselines, improving Qwen3.5-122B from 46.7 to 60.0 on BC-small while using fewer tool calls than five independent trajectories. A pilot supervised fine-tuning study further shows that these behavioral signals can curate compact training trajectories and reduce degradation from unfiltered data.

Ruochen Zhou, Zhengyu Chen, Luan Zhang +3

City University of Hong Kong · Meituan · Independent +1

Inference-Time SearchDeep Research AgentsAgentic Search
AskPDF
Feb 23, 2026

How to Train Your Deep Research Agent? Prompt, Reward, and Policy Optimization in Search-R1

Deep Research agents tackle knowledge-intensive tasks through multi-round retrieval and decision-oriented generation. While reinforcement learning (RL) has been shown to improve performance in this paradigm, its contributions remain underexplored. To fully understand the role of RL, we conduct a systematic study along three decoupled dimensions: prompt template, reward function, and policy optimization. Our study reveals that: 1) the Fast Thinking template yields greater stability and better performance than the Slow Thinking template used in prior work; 2) the F1-based reward underperforms the EM due to training collapse driven by answer avoidance; this can be mitigated by incorporating action-level penalties, ultimately surpassing EM; 3) REINFORCE outperforms PPO while requiring fewer search actions, whereas GRPO shows the poorest stability among policy optimization methods. Building on these insights, we then introduce Search-R1++, a strong baseline that improves the performance of Search-R1 from 0.403 to 0.442 (Qwen2.5-7B) and 0.289 to 0.331 (Qwen2.5-3B). We hope that our findings can pave the way for more principled and reliable RL training strategies in Deep Research systems.

Yinuo Xu, Shuo Lu, Jianjie Cheng +5

NLPR & MAIS, CASIA · School of AI, UCAS · Meituan Inc.

RL for LLM AgentsAutonomous Research AgentsReward Design for RLRL for IRPolicy Gradient Methods
AskPDF