acceptodds
Under review as a conference paper at ICLR 2027

WideSeek-R1: Exploring Width Scaling for Broad Information Seeking via Multi-Agent Reinforcement Learning

Abstract

Recent advances in large language models (LLMs) have largely focused on depth scaling, where a single agent solves long-horizon problems through multi-turn reasoning and tool use. However, as tasks grow broader, the key bottleneck shifts from individual competence to organizational capability. In this work, we explore width scaling through multi-agent systems as a complementary approach to broad information seeking. Existing multi-agent systems often rely on handcrafted workflows and turn-taking interactions that fail to parallelize work effectively. To bridge this gap, we propose WideSeek-R1, a lead-agent–subagent framework trained via multi-agent reinforcement learning (MARL) to jointly learn scalable orchestration and parallel execution. Using a shared LLM with isolated contexts and specialized tools, WideSeek-R1 jointly optimizes the lead agent and parallel subagents on a curated dataset of 20,000 broad information-seeking tasks. Extensive experiments show that WideSeek-R1-4B achieves an item F1 score of 40.0% on the WideSearch benchmark, comparable to that of single-agent DeepSeek-R1-671B. Furthermore, WideSeek-R1-4B exhibits consistent performance gains as the number of parallel subagents increases, highlighting the effectiveness of width scaling.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language model, width scaling, multi-agent system, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vaJeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

WideSeek-R1 uses a lead agent to divide broad search tasks among parallel subagents and combine their results. The agents share model parameters and are trained together with a GRPO-based objective. The authors also construct 20,000 synthetic training tasks. On WideSearch, Item F1 improves from 29.6 to 40.0 compared with the base model in the same multi-agent framework.

Critical strengths
  1. The gain over the base model is substantial: Item F1 increases from 29.6 to 40.0 in the same framework. This is useful evidence that RL training helps the system.

  2. Using the trained model only as the lead agent raises Item F1 to 34.7, while using it only as the subagents raises it to 34.2. The role-assignment experiment therefore shows gains from updating either role.

Critical weaknesses
  1. The lead agent handles task decomposition, further delegation, and final answer generation. Replacing it with the trained model changes all of these at once. Table 5 shows more subagents and tool calls after training, but does not measure the quality of the plans or searches. The paper would be more convincing with a test that separates these effects, such as comparing retrieval on the same subtasks or answer generation from the same retrieved results.

  2. Figure 4 compares turn limits and subagent counts without matching actual resource use. The depth curve also uses the base model rather than the trained SingleSeek-R1. In Table 5, WideSeek-R1 uses 51,104 response tokens, compared with 19,837 for the base multi-agent model. SFT uses even more tokens and scores lower, so token count alone does not explain the results. Still, a comparison at similar token and tool-call budgets is needed to assess the advantage of width scaling. Reporting elapsed time would also clarify the benefit of parallel execution.

Questions for authors
  1. At similar token and tool-call budgets, does WideSeek-R1 still outperform the base multi-agent model and SingleSeek-R1?

  2. What evidence shows that task decomposition improves? Can you separate this from changes in further delegation and final answer generation?

1 comment

Thank you for your constructive comments and questions! We appreciate your thorough review and hope the following responses address your concerns.


W1 / Q2: Evidence for improved decomposition, delegation, and final aggregation.

We manually compared the base model and WideSeek-R1 on the same questions, examining both task decomposition and final aggregation. The trained model more consistently produced self-contained subtasks that could actually run in parallel. The base model frequently placed dependencies within a batch, for example, "subtask 2: use the result of subtask 1 to ...". Because agents in the same batch run independently and return together, the second agent cannot obtain the first agent's result during execution.

Task decomposition and further delegation are closely related parts of the orchestration policy. Within a batch, assignments must be independent; across turns, the lead agent can use completed findings to formulate the next batch. WideSeek-R1 better respects this distinction, allowing parallel execution to gather useful information instead of assigning subtasks with unresolved dependencies.

For final aggregation, our manual analysis found that both the base and trained lead agents generally assembled accurate, sufficient subagent findings correctly. Cases where the evidence was correct but the lead agent failed to incorporate it were rare. The more prominent differences in the inspected trajectories were therefore the quality of decomposition and information gathering.

W2 / Q1: Token/tool budgets, trained depth scaling, and elapsed time.

Our study examines how learning enables agents to translate additional test-time computation into better answers. Tables 1, 3, and 5 report the corresponding quality and resource usage:

ModelItem F1 Avg@4 (%)Response tokensSearch / access calls
SingleSeek-R1-4B28.15,3413.6 / 4.9
Qwen3-4B, multi-agent29.619,83723.3 / 11.9
WideSeek-SFT-4B31.557,68283.7 / 29.8
WideSeek-R1-4B40.051,10483.9 / 74.1

These are WideSearch measurements using online Serper/Jina tools; tokens and calls are per-query averages across the system. WideSeek-R1-4B improves Item F1 by 10.4 points over the base multi-agent model. Compared with WideSeek-SFT-4B, it gains 8.5 points while generating fewer response tokens and making more access calls. The SFT comparison highlights that generating more text alone does not explain the gain.

We further investigated single-agent scaling in three settings:

  • Longer depth: A turn-limit sweep of trained SingleSeek-R1 shows saturation similar to the base model.
  • Parallel tools: We performed full single-agent RL training for SingleSeek-R1-para5, allowing up to five parallel tool calls per turn to match each subagent's allowance. Its Item F1 Avg@4 is 30.7, versus 28.1 for SingleSeek-R1-4B and 40.0 for WideSeek-R1-4B.
  • Context compression: Our MiroThinker-8B evaluation uses context compression and no turn cap. It achieves 19.1 Item F1 with 9,058 average response tokens, versus 20.1 Item F1 and 4,721 tokens for single-agent Qwen3-4B.

In these tests, extra turns and context compression do not sustain gains; even fully RL-trained SingleSeek-R1-para5 trails WideSeek-R1 by 9.3 points. This supports scaling through isolated contexts and learned delegation.

Figure 4 first compares single- and multi-agent execution with the same base Qwen3-4B: width scaling improves performance beyond the plateau of depth scaling. MARL then further unlocks multi-agent width scaling, sustaining gains as the number of parallel subagents increases.

Elapsed time also depends on GPU parallelism, serving configuration, and variable online API delays. We therefore use measured tokens and tool calls to characterize resource use; the central finding is improved answer quality through learned parallel orchestration.


We sincerely appreciate your thoughtful and detailed review. We have carefully addressed each of your concerns and believe these clarifications strengthen our contribution. We hope you will consider adjusting your rating accordingly.

Reviewer_vhmVOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies broad information-seeking tasks that require collecting many facts rather than following one long reasoning chain. WideSeek-R1 trains a shared Qwen3-4B model for lead-agent and worker-agent roles using outcome-based reinforcement learning. It also introduces a pipeline for synthesizing table-valued search tasks, mixes these with deep-search data, and evaluates parallel search on WideSearch and standard QA benchmarks. On WideSearch, item-level Avg@4 F1 increases from 28.1 for the matched-data SingleSeek-R1 baseline to 40.0. The parallel system improves search quality. What remains less clear is how width compares with depth after training and at a similar inference cost.

Critical strengths
  1. The task and training data are well aligned. Broad table completion is a reasonable setting for parallel decomposition, and the paper includes a single-agent model trained on the same hybrid data. The gain from 28.1 to 40.0 item-level F1 is substantial (Table 1). Appendix E also includes an SFT baseline, so the comparison is not limited to an untrained multi-agent workflow.

  2. Figure 5 shows that using trained parameters in both lead and worker roles matters. The SFT comparison is also useful: WideSeek-R1 has higher item-level F1, while Table 5 reports fewer output tokens than the SFT variant. These comparisons help assess what RL adds to the parallel architecture, although the evaluation setting for the usage statistics needs clarification.

Critical weaknesses
  1. The width-versus-depth comparison still mixes training and inference-budget differences. Figure 4 sweeps depth for the base Qwen3-4B model, while the width analysis includes the RL-trained system. Table 5 reports 51,104 tokens for WideSeek-R1 versus 5,341 for SingleSeek-R1, with much larger search and access counts. These costs are disclosed, but the table needs to identify its measurement set before readers can relate them to Table 1. A depth sweep for trained SingleSeek-R1, together with quality curves at overlapping token and tool budgets, would show whether width offers an advantage at comparable cost.

  2. The dual-level reweighting objective has no direct ablation. The role-swap experiment changes the parameters used at evaluation time, while the SFT comparison tests the benefit of RL as a whole. Neither isolates the proposed normalization terms. Since the objective is one of the paper's methodological contributions, comparing it with simpler token- and agent-weighting choices would strengthen the case.

Questions for authors
  1. How does trained SingleSeek-R1 perform as its interaction budget increases? Can you compare it with WideSeek-R1 at overlapping token and search/access budgets, and report elapsed time as well? Please also identify the data and tool setting used for Table 5. Parallelism may still offer a useful latency advantage even where total compute is higher.

  2. What happens if you average turns equally instead of averaging all tokens within each agent, or sum agent contributions instead of averaging them within a rollout? A small controlled comparison would help separate objective design from the architecture and training data.

  3. Did you audit a sample of synthetic reference tables against the source pages? The filter compares two answers from the same teacher, so shared factual errors may pass. Please report factual errors and missing entries in an independently checked sample, if available.

  4. How should readers interpret the low complete-task success rates? In Table 1, WideSeek-R1's Avg@4 success rate is 0.4%, versus 0.3% for SingleSeek-R1, while item-level F1 improves considerably. It would help to explain which remaining errors prevent the cell-level gains from translating into complete answers.

1 comment

Thank you for your constructive comments and questions! We appreciate your thorough review and hope the following responses address your concerns.


W1 / Q1: Trained depth scaling, resource use, and Table 5.

Table 5 measures WideSearch with online Serper search and Jina access, using Table 1's evaluation setup. Below we summarize Tables 1, 3, and 5; tokens and calls are per-query averages across agents:

ModelItem F1 Avg@4 (%)Response tokensSearch / access calls
SingleSeek-R1-4B28.15,3413.6 / 4.9
Qwen3-4B, multi-agent29.619,83723.3 / 11.9
WideSeek-SFT-4B31.557,68283.7 / 29.8
WideSeek-R1-4B40.051,10483.9 / 74.1

WideSeek-R1-4B gains 8.5 Item F1 points over WideSeek-SFT-4B with fewer response tokens and more access calls, also achieving higher Row F1. This shows that using additional computation effectively matters beyond token quantity.

We also tested stronger single-agent scaling settings. A turn-limit sweep of trained SingleSeek-R1 showed saturation similar to base Qwen3-4B. We performed full single-agent RL training for SingleSeek-R1-para5, allowing up to five parallel tool calls per turn to match each subagent's allowance. Its Item F1 Avg@4 is 30.7, versus 28.1 for SingleSeek-R1-4B and 40.0 for WideSeek-R1-4B. Our MiroThinker-8B evaluation uses context compression and no turn cap. It achieves 19.1 Item F1 with 9,058 average response tokens, versus 20.1 Item F1 and 4,721 tokens for single-agent Qwen3-4B.

In these tests, extra turns and context compression do not sustain gains; even fully RL-trained SingleSeek-R1-para5 trails WideSeek-R1 by 9.3 points. This supports scaling through isolated contexts and learned delegation.

Figure 4 first compares single- and multi-agent execution with the same base Qwen3-4B: width scaling improves performance beyond the plateau of depth scaling. MARL then further unlocks multi-agent width scaling, sustaining gains as the number of parallel subagents increases.

Elapsed time depends on GPU parallelism, serving configuration, and online API delays; tokens and tool calls provide more reproducible resource accounting here.

W2 / Q2: Equal-turn averaging and summing agent contributions.

Both alternatives were tested in short controlled runs, where the adopted combination gave the strongest reward improvement. Averaging all tokens within each agent preserves the contribution of longer turns; averaging agents within each rollout prevents larger groups from dominating the loss. With a full MARL run costing about 3,000 H100 GPU hours, these comparisons guided the configuration used for our width-scaling experiments.

Q3: Human checks of synthetic reference tables.

Question generation explicitly required pre-2018 information and answers retrievable from Wikipedia. We selected Gemini-3-Pro-Preview after comparing teachers and cross-verified two independent answers. Human rechecking confirmed retrievability from Wiki2018 and found over 95% factual agreement among inspected cells, with almost no missing rows. Most discrepancies involved source-dependent numerical values.

Q4: Interpreting complete-task success versus Item F1.

Item F1 measures progress across individual facts, while SR requires complete correctness. We optimize Item F1 as a dense answer-quality reward. WideSeek-R1 improves Item F1 from SingleSeek-R1's 28.1 to 40.0 and Row F1 from 6.5 to 15.3, showing gains in complete entries as well as individual cells.

WideSearch reports mean answer volumes of 2,001.2 facts for Chinese and 938.6 for English tasks, or about 1,470 facts per task across the balanced benchmark [1]. A missing entity or an unresolved attribute can prevent whole-table success. Rows vary greatly in retrieval difficulty, source availability, and required search depth; recovering more facts does not automatically resolve every difficult row. This explains why the SR difference (0.4% versus 0.3%) is much smaller than the Item/Row F1 gains. Even the best system in the original benchmark evaluation reached only 5.1% SR [1], illustrating the difficulty of complete retrieval.

[1] Wong et al. WideSearch: Benchmarking Agentic Broad Info-Seeking, Sections 3.4-3.5 and Table 2.


We sincerely appreciate your thoughtful and detailed review. We have carefully addressed each of your concerns and believe these clarifications strengthen our contribution. We hope you will consider adjusting your rating accordingly.

Reviewer_9zPdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper trains a shared-weight lead-agent/subagent search system with a GRPO-based objective and constructs 20,000 broad information-seeking tasks. It reports substantial WideSearch gains and improved performance as more subagents are allowed.

Critical strengths
  • The within-framework improvement is substantial: WideSearch Item F1 rises from 29.6 to 40.0, while the trained single-agent variant reaches 28.1.
  • The SFT comparison is useful. RL achieves higher Item F1 despite using fewer output tokens than SFT in Table 5, suggesting that effective training matters beyond simply generating longer trajectories.
  • The task setting is well motivated. The data pipeline addresses a useful training-data gap, and the implementation details are clearly described in the appendices.
Critical weaknesses
  1. The main width-versus-depth claim is not established under comparable budgets. Table 5 reports 51,104 output tokens for WideSeek versus 5,341 for SingleSeek, and roughly 18.6 times as many search/access calls. Appendix D also allows multiple parallel tool calls per subagent but only one tool call per turn for SingleSeek. Figure 4 uses untrained Qwen3-4B for the depth curve, rather than SingleSeek. These results show gains with additional resources, but do not isolate the benefit of allocating a fixed budget to width. Performance-versus-budget curves and measured latency would make the claim much stronger.
  2. The proposed training components are not isolated experimentally. Sharing the final reward is a straightforward GRPO extension, and the token averaging follows DAPO. There is no ablation of agent-level averaging or the two reweighting levels. Figure 5a swaps trained and base checkpoints at inference; it shows that both roles benefit from training, but does not establish that joint training is necessary or better than training the roles separately.
  3. The synthetic ground truth needs stronger validation. Two responses from the same Gemini model can agree on incorrect facts. The paper does not report a factual audit against independent sources or measure how much of the generated answer content is retrievable from Wiki2018, the training environment. Both matter because these labels directly determine the RL reward.
  4. The DeepSeek comparison should be narrower. Item F1 is close (40.0 versus 41.3), but Row F1 is noticeably lower (15.3 versus 20.7). The parameter-count comparison also does not establish inference efficiency.
Questions for authors
  1. Does WideSeek retain its advantage over trained SingleSeek at matched token or tool-call budgets, with comparable tool interfaces?
  2. What changes when agent-level averaging is removed, or the two roles are trained separately?
  3. What are the audited factual accuracy and Wiki2018 answerability of the synthetic training data?
1 comment

Thank you for your constructive comments and questions! We appreciate your thorough review and hope the following responses address your concerns.


W1 / Q1: Width versus depth and inference budgets.

MARL enables effective use of additional test-time computation. Tables 1, 3, and 5 summarize quality and resource use:

ModelItem F1 Avg@4 (%)Response tokensSearch / access calls
SingleSeek-R1-4B28.15,3413.6 / 4.9
Qwen3-4B, multi-agent29.619,83723.3 / 11.9
WideSeek-SFT-4B31.557,68283.7 / 29.8
WideSeek-R1-4B40.051,10483.9 / 74.1

These are WideSearch results with online Serper/Jina tools; tokens and calls are per-query averages across the system. WideSeek-R1-4B outperforms WideSeek-SFT-4B by 8.5 Item F1 points, with fewer response tokens and more access calls. Thus, using a larger budget effectively matters beyond generating more text.

We further investigated single-agent scaling:

  • Parallel tools: We performed full single-agent RL training for SingleSeek-R1-para5, allowing up to five parallel tool calls per turn to match each subagent's allowance. Its Item F1 Avg@4 is 30.7, versus 28.1 for SingleSeek-R1-4B and 40.0 for WideSeek-R1-4B.
  • Longer depth: Increasing trained SingleSeek-R1's turn limit yields a saturation trend similar to the base model.
  • Context compression: Our MiroThinker-8B evaluation uses context compression and no turn cap. It achieves 19.1 Item F1 with 9,058 average response tokens, versus 20.1 Item F1 and 4,721 tokens for single-agent Qwen3-4B.

In these tests, extra turns and context compression do not sustain gains; even fully RL-trained SingleSeek-R1-para5 trails WideSeek-R1 by 9.3 points. This supports scaling through isolated contexts and learned delegation.

Figure 4 first compares single- and multi-agent execution with the same base Qwen3-4B: width scaling improves performance beyond the plateau of depth scaling. MARL then further unlocks multi-agent width scaling, sustaining gains as the number of parallel subagents increases. Since wall-clock time depends on GPU parallelism, serving configuration, and online API delays, we use measured tokens and tool calls to characterize resource use.

W2 / Q2: Objective design and shared versus separate models.

We tested equal-turn averaging and unnormalized agent contributions in short preliminary runs; the adopted dual-level combination gave the strongest reward improvement. DAPO-style token averaging preserves the contribution of longer turns, while agent averaging prevents larger groups from dominating the loss. These comparisons guided our final configuration.

We view parameter sharing as a trade-off. Separate models allow role-specific specialization. A shared model learns both orchestration and information seeking, reduces training, serving, and deployment costs, and can be instantiated in either role as the system scales. We consider this a more general design than maintaining a separate model for each role. Figure 5a supports its effectiveness: either trained role improves performance, and using the trained model for both gives the best result. More broadly, learning parallel orchestration for general-purpose agents aligns with Kimi K2.5 Agent Swarm.

W3 / Q3: Synthetic-label accuracy and Wiki2018 answerability.

Question generation explicitly required pre-2018 information and answers retrievable from Wikipedia. Human rechecking also verified that the inspected answers could be retrieved from the Wiki2018 training corpus.

We compared available teachers and selected Gemini-3-Pro-Preview because it performed best in our checks. Over 95% of inspected cells agreed with human verification, with almost no missing rows. Most discrepancies involved source-dependent numerical values.

W4: Purpose of the DeepSeek-R1 comparison.

The comparison shows that our trained 4B model can approach DeepSeek-R1-671B on broad information seeking. WideSeek-R1-4B achieves 40.0 Item F1 versus 41.3 for DeepSeek-R1. We use Item F1 as the main answer-quality reward because it captures progress across individual facts. Row F1 is also reported (15.3 versus 20.7). The key result is the capability achieved with only 4B parameters.


We sincerely appreciate your thoughtful and detailed review. We have carefully addressed each of your concerns and believe these clarifications strengthen our contribution. We hope you will consider adjusting your rating accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.