acceptodds
Under review as a conference paper at ICLR 2027

RIVO: Lightweight Adaptive Rerouting for Deep Research Agents

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 still 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 solely 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 first estimates its research state from lightweight state 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 the expected downstream research utility. A curated dataset, BrowseCore, with a paired local corpus is further constructed 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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Deep Research, Adaptive Rerouting, Long-Horizon Reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dYAvOpenReview ↗
Rating
3(+1)
Confidence
4(+2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RIVO, a plug-in controller designed for long-horizon deep-research agents. The core idea is to keep the main Host model frozen while using a lightweight 0.6B controller. This controller leverages confidence, evidence attention, and attention convergence to dynamically choose between STOP, VERIFY, CONTINUE, and REROUTE actions. When a reroute is triggered, RIVO applies a sparse Top-K logit correction to steer the Host toward an alternative search direction. Additionally, the authors propose BrowseCore, a compact dataset paired with a retrieval corpus. Experiments conducted across two Host models and three benchmarks demonstrate consistent improvements over existing trajectory-control baselines.

Critical strengths
  • The collaboration between a small controller and a larger, frozen generator is a very appealing direction for deep-research systems. Decoupling trajectory control from language generation is practical and could allow various Host models to benefit without requiring full, expensive retraining.
  • RIVO shows consistent improvements across two different Host models and three benchmarks. I also appreciate the ablation studies, which clearly show that both deciding when to reroute and learning how to reroute contribute meaningfully to the final performance.
Critical weaknesses
  • Potential test-set leakage in BrowseCore evaluation is my primary concern. The same 359 questions are referred to as a "test set" in the main text but as a "development set" in the appendix. More problematically, these exact 359 questions are used to select hyperparameters (β, Top-K, λ, and τ), and the resulting score (59.30) is reported as a main result. Without a strictly held-out test set, the current BrowseCore results are highly inflated and unconvincing.
  • RIVO's pipeline is actually quite heavy: it requires generating provisional answers, extracting attention from the final four Host layers, running an auxiliary 0.6B model, and prompting the Host to rewrite rerouting hints. Yet, there is no analysis of latency, throughput, memory footprint, or total training cost. Furthermore, while the aggregate ablations are nice, they don't explain when or why the system succeeds or fails at different stages. There is a noticeable lack of trajectory-level case studies and intervention statistics.
Questions for authors
  • Can the authors clarify whether the 359-question split is a dev set or a test set? Ideally, could you select all hyperparameters on a separate development split and evaluate once on a strictly untouched test set to validate the 59.30 score?
  • What is RIVO’s end-to-end inference and training overhead relative to the original Host and the strongest baselines? Additionally, could you provide stage-wise action statistics and a few representative successful/failed trajectories to help clarify how the controller actually drives the performance gains?
6 comments

dataset splitation

We thank the reviewer for raising this concern. We would like to clarify that BrowseCore contains 359 development questions and 359 separate test questions. The inconsistency in the appendix was an editorial oversight. Importantly, all hyperparameters (β, Top-K, λ, and τ) were tuned exclusively on the 359-question development set, while the reported main result (59.30) was obtained on the separate, held-out 359-question test set. Thus, no test questions were used for hyperparameter selection. To ensure transparency and reproducibility, we have publicly released the complete dataset, including both the development and test splits, which has already received 176 downloads.The complete BrowseCore dataset, including the training, development, and test splits (train.json, dev.json, and test.json), is publicly available at the anonymized link under BrowseCore/ourdata/data/.

RIVO's pipeline is actually quite heavy

We thank the reviewer for this valuable suggestion. We provide additional efficiency measurements and trajectory-level analyses on the 359 BrowseCore test questions. Inference efficiency. RIVO achieves an amortized inference time of approximately 40 seconds per question on an H200 GPU under concurrent execution. The provisional-answer and attention probes each account for approximately 13% of wall-clock time, while reroute retrieval adds less than 1%. The additional sidecar memory footprint is at most approximately 24 GB. Training cost. RIVO keeps the 4B Host frozen and trains only the lightweight 0.6B plugin, avoiding full-parameter fine-tuning of the Host. Stage-level analysis. RIVO triggers rerouting for 266 out of 359 questions, performing 483 reroutes, of which 471 successfully apply logit correction. Case studies. QID 2023 demonstrates successful evidence retrieval after rerouting, while QID 8624 illustrates premature stopping with high confidence in an incorrect answer. QID 371 shows that correct evidence can be retrieved but not committed due to insufficient stopping confidence. We will incorporate these efficiency measurements, intervention statistics, and representative case studies into the revised manuscript to better explain when and why RIVO succeeds or fails.

Thank you for your response. I will keep my score at 2.

As presented, RIVO requires access to the host model’s attention weights and the ability to modify token logits during decoding. These requirements limit its applicability to APIs that expose only generated text or read-only token log probabilities. The evaluation covers two host models, and it remains unclear whether the reported gains would generalize to a broader range of models.

Have a wonderful day

Reviewer_2q7EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RIVO improves Deep Research agents through state-based control, a learned reroute predictor, and sparse logit corrections to a frozen Host model. The correction head is trained with trajectory-level rewards. The paper also introduces BrowseCore, an evidence-verified local retrieval dataset, and evaluates the framework on three benchmarks with two Host models.

Critical strengths
  1. The frozen-Host design is practical for adapting existing agents. Table 2 shows that both reroute prediction and logit correction contribute to performance.

  2. RIVO improves the overall metric in all six model–benchmark settings, including benchmarks beyond BrowseCore. This provides useful evidence that the benefit extends beyond its training environment.

Critical weaknesses
  1. The evaluation split is unclear. Section 5 calls the 359 BrowseCore questions a test set, while Appendix G.2.1 calls them development data, and Section 5.2 uses them for hyperparameter sweeps. I therefore cannot tell whether the main result is held out. Uncertainty estimates are also missing, which matters for small gains such as 52.00 to 52.40.

  2. The efficiency claim needs stronger support. Freezing the Host reduces trainable parameters, but RIVO still uses turn-level supervision, full-trajectory GRPO, repeated answer probes, and attention extraction. Training GPU-hours, end-to-end latency, and actual tool/token usage are not reported, leaving its practical cost advantage unclear.

Questions for authors
  1. Were the hyperparameters fixed before evaluating the 359 BrowseCore questions? Please clarify the development/test split and provide uncertainty estimates for the main comparisons.

  2. What are the training GPU-hours, inference latency, and tool/token usage compared with RAAC and PPDPP under the same resource limits?

3 comments

The evaluation split is unclear.Please clarify the development/test split and provide uncertainty estimates for the main comparisons.

We thank the reviewer for raising this concern. We would like to clarify that BrowseCore contains 359 development questions and 359 separate test questions. The inconsistency in the appendix was an editorial oversight. Importantly, all hyperparameters (β, Top-K, λ, and τ) were tuned exclusively on the 359-question development set, while the reported main result (59.30) was obtained on the separate, held-out 359-question test set. Thus, no test questions were used for hyperparameter selection. To ensure transparency and reproducibility, we have publicly released the complete dataset, including both the development and test splits, which has already received 176 downloads.The complete BrowseCore dataset, including the training, development, and test splits (train.json, dev.json, and test.json), is publicly available at the anonymized link under BrowseCore/ourdata/data/.

What are the training GPU-hours, inference latency, and tool/token usage compared with RAAC and PPDPP under the same resource limits?

We thank the reviewer for this suggestion. All methods use the same Host, retrieval server, and 80-round budget. RIVO averages 28.8 tool calls and 5.2K generated tokens per question, compared with 30.6 calls and 5.5K tokens for RAAC, and 32.9 calls and 5.6K tokens for PPDPP. RIVO achieves an amortized inference latency of approximately 40 seconds per question on an H200 GPU. For training, RIVO keeps the Host frozen and trains only the lightweight 0.6B plugin, avoiding full-parameter fine-tuning. We will include a more detailed comparison of inference latency and computational costs across methods in the final version of the manuscript.

Dear Authors,

Thank you very much for your detailed explanations, however, I think my concern on the process of the evaluation has not been addressed. Hence I remain the orginial rating.

thanks, If you believe that our response has not adequately addressed your concerns, could you please specify which aspects remain unresolved so that we can provide further clarification?

Reviewer_A7crOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RIVO is a plug-in controller for frozen Deep Research agents. A rule-based state controller (answer confidence, evidence attention, attention convergence) decides among STOP, VERIFY, REROUTE and CONTINUE; a 0.6B reroute predictor can promote CONTINUE to REROUTE; and when rerouting fires, the same 0.6B model adds a sparse low-rank correction to the Host's top-K logits while generating a structured rerouting hint, trained with GRPO from trajectory rewards. The paper also introduces BrowseCore, a compact benchmark with a verified local corpus, and gives a first-order result on when a small logit intervention improves expected utility. Experiments on two Hosts and three benchmarks report gains of up to 4.36 points over inference-time control baselines.

Critical strengths
  • A timely problem (when to stop, verify or reroute) with a clear framing as state-conditioned trajectory correction.
  • A sensible, well-engineered system that leaves Host weights untouched.
  • BrowseCore is a useful resource for cheap controlled experiments, and code and data are released.
  • The ablations cover the natural questions.
Critical weaknesses
  1. Closest related work is not discussed. Adding a small model's logit offset to a frozen large model is proxy tuning (Liu et al. 2024) and emulated fine-tuning (Mitchell et al. 2023), and reward-trained steering of a frozen decoder is controlled decoding (Mudgal et al. 2024). RIVO's additions (top-K mask, gating by a state controller, the Deep Research setting) are reasonable, but positioning against this line would clarify the contribution.

  2. Differences are within single-run noise, and hyperparameters seem tuned on the test set. BrowseCore has 359 questions (one standard deviation is about 2.6 points), BrowseComp-Plus 830, DRBench 100, and every comparison is a single run. Gains over the best baseline on Marco-DeepResearch-8B are 0.6 and 0.4 points on the SR benchmarks; the "up to 4.36" figure is a harmonic mean on 100 tasks. Section 5.2 sweeps lambda, tau_H, beta and K on the BrowseCore test set, and the chosen cell (59.30) is 4 to 7 points above every neighbour. A dev split and a few seeds would address most of this.

  3. The ablations point to the heuristic parts rather than the learned ones. The rule controller alone (55.71) already beats every baseline. Adding the logit intervention without the predictor lowers SR to 53.63, while adding both gives 59.30; this interaction is not discussed. Removing VERIFY (a prompt-based step mentioned in one line) costs more than removing the logit intervention. This sits a little awkwardly with an introduction that motivates the method by the limits of heuristic control.

  4. The sensitivity analysis is hard to read as support. Figure 2 reports success rates in steps of 4, which suggests roughly 25 replayed questions. The best beta is -10, i.e., the learned correction works best with its sign reversed, and the K curve is non-monotonic. If the GRPO-trained head were value-aligned, beta should be positive; an explanation of the sign convention would help.

  5. Minor: how PPDPP was adapted and how ReAct differs from Source are not explained; the REROUTE trigger rate is not reported; the number of runs averaged should be stated

Questions for authors
  1. Were lambda, tau_H, beta and K selected on the 359 test questions?

  2. Why is the best intervention strength negative, and what beta was used in training?

  3. Is the same plug-in used for both Hosts without retraining?

  4. What are the average turns and cost per question for Source versus RIVO?

1 comment

Closest related work is not discussed

We thank the reviewer for this suggestion. We will discuss these closely related works in the revised manuscript and clarify how RIVO differs through state-aware gating, top-K logit intervention, and its application to long-horizon Deep Research agents.

Differences are within single-run noise, and hyperparameters seem tuned on the test set.

We thank the reviewer for raising these concerns. On BrowseCore, RIVO achieves 59.30 ± 1.50 SR compared with 55.15 ± 1.20 for the strongest baseline (PPDPP), averaged over five independent runs. The improvement of 4.15 percentage points is statistically significant. We will moderate some of our claims and revise the manuscript to emphasize that RIVO achieves competitive performance on several benchmarks. We would like to clarify that BrowseCore contains 359 development questions and 359 separate test questions. The inconsistency in the appendix was an editorial oversight. Importantly, all hyperparameters (β, Top-K, λ, and τ) were tuned exclusively on the 359-question development set, while the reported main result (59.30) was obtained on the separate, held-out 359-question test set. Thus, no test questions were used for hyperparameter selection. To ensure transparency and reproducibility, we have publicly released the complete dataset, including both the development and test splits, which has already received 176 downloads.The complete BrowseCore dataset, including the training, development, and test splits (train.json, dev.json, and test.json), is publicly available at the anonymized link under BrowseCore/ourdata/data/.

The ablations point to the heuristic parts rather than the learned ones

We thank the reviewer for this observation. We would like to clarify that RIVO is designed as a unified framework in which the rule-based controller, reroute predictor, and logit intervention work jointly, rather than relying on any single component. As shown in Table 4, the full framework achieves the best performance (59.30), outperforming the rule-only controller (55.71) by 3.59 points. These results highlight the complementary roles of heuristic control and learned components.

The sensitivity analysis is hard to read as support

We thank the reviewer for this observation. In Eq. (5), is a signed scaling coefficient that controls both the direction and magnitude of the logit intervention. Our formulation allows both positive and negative values. Importantly, all GRPO training hyperparameters and inference-time settings were selected using the development set, and the selected inference configuration, including =-10, was kept fixed during evaluation on the held-out test set. Therefore, the negative optimal does not, by itself, indicate that the learned correction is misaligned. We will clarify the sign convention and hyperparameter selection procedure in the revised manuscript.

Minor: how PPDPP was adapted,the number of runs averaged should be stated

We thank the reviewer for these suggestions. We adapted PPDPP's policy guidance to the Deep Research setting, while ReAct explicitly prompts the Host to assess its progress and decide whether to reroute, unlike Source. On BrowseCore, REROUTE was triggered for 266 out of 359 questions, with 483 reroutes in total. We will clarify these implementation details and explicitly state the number of runs averaged in the revised manuscript.

Were lambda, tau_H, beta and K selected on the 359 test questions?

thanks for the questions, All hyperparameters were selected on the separate 359-question development set

Is the same plug-in used for both Hosts without retraining?

thanks for the questions, YES.

What are the average turns Source versus RIVO?

thanks for the questions. Source requires an average of 31.1 turns per question, compared with 28.8 turns for RIVO.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.