PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Aggregation for Parallel Scaling of Long-Horizon Agentic Tasks

Yoonsang Lee, Howard Yen, Xi Ye, Danqi Chen

arXiv:2604.11753v3cs.CL

TL;DR

Long-horizon agentic tasks generate multi-turn, tool-augmented trajectories whose evidence is distributed, making final-answer-only aggregation inadequate and full concatenation impractical. AggAgent treats trajectories as an environment and selectively navigates them with lightweight tools, outperforming existing aggregation methods across six benchmarks and three model families with bounded aggregation cost.

  • Problem

    Long-horizon agentic trajectories are multi-turn and tool-augmented, while useful evidence is distributed across rollouts and cannot be preserved by final-answer-only aggregation.

  • Method

    AggAgent treats parallel trajectories as an interactive environment and uses lightweight tools to inspect solutions and retrieve trajectory segments on demand.

  • Results

    AggAgent consistently outperforms baselines across six benchmarks and three models, improving over the strongest baseline by 2.4–5.3 points on average.

  • Takeaways & Limitations

    Agentic aggregation provides a scalable and cost-efficient approach to parallel test-time scaling for long-horizon tasks.

Abstract

from arXiv · show

We study parallel test-time scaling for long-horizon agentic tasks such as agentic search and deep research, where multiple rollouts are generated in parallel and aggregated into a final response. While such scaling has proven effective for chain-of-thought reasoning, agentic tasks pose unique challenges: trajectories are long, multi-turn, and tool-augmented, and outputs are often open-ended. Aggregating only final answers discards rich information from trajectories, while concatenating all trajectories exceeds the model's context window. To address this, we propose AggAgent, an aggregation agent that treats parallel trajectories as an environment. We equip it with lightweight tools to inspect candidate solutions and search across trajectories, enabling it to navigate and synthesize information on demand. Across six benchmarks and three model families (GLM-4.7, Qwen3.5, MiniMax-M2.5), AggAgent outperforms all existing aggregation methods-by up to 5.3% absolute on average and 10.3% on two deep research tasks-while adding minimal overhead, as the aggregation cost remains bounded by a single agentic rollout. Our findings establish agentic aggregation as an effective and cost-efficient approach to parallel test-time scaling.

1 Introduction

Parallel scaling is effective for long-horizon agentic tasks, but aggregating their lengthy, distributed trajectories requires more than voting or final-answer synthesis. AggAgent addresses this by navigating trajectories on demand and consistently improving over existing aggregation methods.

  • Long-horizon agentic trajectories span hundreds of multi-turn steps with interleaved tool calls and observations.
  • Pass@8 improves GLM-4.7-Flash from 27% to 59% on BrowseComp and from 25% to 51% on HLE, showing that parallel rollouts often contain correct solutions.
  • Evidence is sparse and distributed across trajectories, while individual rollouts may each contain only partial progress, requiring cross-trajectory synthesis.
  • AggAgent treats trajectories as an interactive environment and uses lightweight tools to retrieve solutions, search trajectories, and read selected step ranges.
  • 5.3 points on average and 10.3 points on deep research tasks are achieved over existing baselines across six benchmarks and three model families.

2 Problem Formulation

The paper formalizes parallel scaling as independently generating K trajectories and applying an aggregation function to synthesize a final solution. Existing methods range from shallow heuristics to LLM-based synthesis, with different losses in trajectory information and applicability.

  • An agent trajectory records the problem, internal reasoning, tool calls, observations, and final output across multiple steps.
  • Parallel scaling runs the same agent independently K times, producing multiple candidate solutions for aggregation.
  • Aggregation is formalized as f : (q, T) 7→ ŷ, producing a final solution evaluated against answers or task-specific rubrics.
  • Heuristic baselines use shallow signals such as answer frequency, confidence, or trajectory length without inspecting trajectory content.
  • Solution Aggregation discards intermediate evidence, whereas Summary Aggregation compresses trajectories with irreversible information loss and K additional LLM calls.

3 Our Approach: AggAgent

AggAgent keeps trajectories outside the context window and lets the aggregator inspect them selectively through lightweight tools. Its coarse-to-fine workflow supports full-fidelity cross-trajectory reasoning while keeping aggregation cost independent of the number of trajectories.

  • AggAgent stores trajectories in an environment and retrieves them on demand, keeping aggregation cost bounded by a single context window independent of K.
  • AggAgent preserves full-fidelity trajectory reasoning without preloading all trajectories into context.
  • The tools provide final solutions, keyword-based trajectory search, and raw contiguous step ranges for detailed inspection.
  • The workflow first surveys metadata and solutions, then selectively inspects trajectories to verify claims against reasoning and tool observations.
  • AggAgent uses no external-service API calls for aggregation because its tools operate over the in-memory trajectory array.

4 Experimental Setup

The evaluation covers six agentic search and deep research tasks using three model families, eight sampled rollouts, and LLM-based judging. Costs and latency include both rollout and aggregation components.

  • Tasks: The benchmark suite contains four agentic-search tasks and two deep-research tasks, including BrowseComp, HLE, Healthbench-Hard, and ResearchRubrics.
  • Tasks: Evaluation uses randomly sampled subsets, including 155 HLE questions, 101 ResearchRubrics problems, and 100 Healthbench-Hard instances.
  • Models and protocol: Rollouts use GLM-4.7-Flash, Qwen3.5-122B-A10B, and MiniMax-M2.5, with eight independent trajectories per problem.
  • Models and protocol: The setup permits at most 128K context length and 100 tool calls, while LLM-based aggregators use the same model as the rollout agent.
  • Cost and latency: Reported cost includes rollouts, tool calls, and aggregation, while latency includes rollout and aggregation.
  • Evaluation: LLM-as-a-judge evaluates outputs, with bootstrapped sampling used to calculate Metric@K for K ∈ {1, 2, 4, 8}.

5 Results

At K=8, AggAgent performs best across the evaluated long-horizon tasks and model families, while maintaining a strong performance–efficiency tradeoff. Its selective trajectory access keeps overhead low relative to methods that summarize every trajectory.

  • Heuristic baselines: Confidence-based heuristics improve over Majority Voting, but calibration weaknesses limit their gains on several difficult benchmarks.Voting methods also do not apply to multi-answer or long-form tasks.
  • LLM-based baselines: LLM-based methods outperform heuristic baselines, while AggAgent preserves trajectory detail without the compression costs of Summary Aggregation.Summary Aggregation is stronger on agentic search, but compression harms detail and coherence on deep research tasks.

6 Analysis and Discussion

The analyses show that stronger aggregation models can improve results, synthesis is especially valuable for deep research, and lightweight retrieval remains robust across backends. AggAgent’s tools are used selectively, while removing thinking traces causes only small performance losses.

  • Stronger Models for Aggregation: A stronger MiniMax-M2.5 aggregator generally improves performance while GLM-4.7-Flash remains the rollout agent.MiniMax-based AggAgent achieves the highest average score and surpasses Pass@8 on BrowseComp-Plus.
  • Synthesis vs. Selection: Synthesis outperforms selecting one trajectory overall, with the largest advantage on deep research benchmarks.Research quality is distributed across trajectories, so synthesis can combine local contributions instead of committing to one global winner.
  • Tool Design Ablations: No retrieval backend consistently dominates: ROUGE-L, BM25, and Qwen3-Embedding-8B produce only small performance differences.Trajectory search often requires exact strings, localized evidence, or tool-returned facts rather than broad semantic matches.
  • Tool Design Ablations: Removing thinking traces causes only a small average-performance drop, and the restricted variant still outperforms the baselines.The reported average scores change from 47.90 to 45.13 for GLM-4.7-Flash, 55.83 to 54.58 for Qwen3.5-122B, and 57.31 to 57.28 for MiniMax-M2.5.
  • Tool Usage: search_trajectory dominates tool usage, while get_solution and finish occur approximately once per query.get_segment is used selectively, consistent with a coarse-to-fine workflow that reads full content when keyword search is insufficient.

7 Related Work

Related work spans long-horizon interactive agents and test-time scaling for reasoning tasks. Prior parallel-scaling approaches address agentic tasks, while sequential scaling extends a single agent’s effective horizon through context and refinement techniques.

  • Long-horizon agents: Long-horizon agent work builds on interactive workflows that interleave reasoning, actions, and environment observations.ReAct is identified as a prominent workflow supporting applications including agentic RAG and software engineering.
  • Test-time scaling: Test-time scaling for CoT tasks has used search, extended thinking budgets, majority voting, and learning-based aggregation.These approaches were developed primarily for mathematical reasoning and coding.
  • Parallel scaling for agents: Long-horizon agent trajectories are lengthy and heterogeneous, motivating parallel generation followed by LLM-based reasoning across multiple trajectories.Existing approaches include concatenating final answers for synthesis, but the related-work passage frames the broader challenge as difficult comparison using surface-level signals.
  • Sequential scaling: Sequential scaling extends a single agent through context management, context folding, or iterative refinement, orthogonally to parallel scaling.The two directions could be combined by managing context within each rollout before aggregation.

8 Conclusion

AggAgent treats parallel trajectories as an interactive environment and navigates them with lightweight in-memory tools. Across six benchmarks and three models, it consistently outperforms baselines while achieving Pareto-optimal performance and efficiency.

  • AggAgent enables cross-trajectory reasoning at full fidelity by interacting with parallel trajectories through lightweight in-memory tools.This avoids both information loss from compression and the cost of loading all trajectories into context.
  • AggAgent achieves Pareto-optimal performance and efficiency for parallel test-time scaling.The paper presents agentic aggregation as a cost-efficient paradigm for long-horizon agentic tasks.

A.1 Implementation

The implementation defines model-serving, dataset-tool, cost, latency, evaluation, and rollout procedures for comparing aggregation methods across agentic benchmarks.

  • Implementation: Models are served through vLLM with a 128K context window, 10K maximum output tokens, native function calling, and at most 100 tool calls per rollout.Experiments use the Tongyi DeepResearch scaffold and temperature 1.0 with top-p 0.95.
  • Dataset tools: BrowseComp-Plus uses a local corpus with top-5 retrieval and documents truncated to 4,096 tokens, while other datasets use Google Search and webpage visiting.The local corpus contains 100,195 documents; other tools return top-10 search results and relevant webpage snippets.
  • Cost calculation: Total cost per question is decomposed into rollout, tool-call, and aggregation costs.Tool-call cost is zero for BrowseComp-Plus and otherwise depends on search and visit calls; aggregation costs vary by method.
  • Latency: Latency is measured sequentially on sampled instances, with rollout latency covering trajectory execution and AggAgent aggregation latency measured equivalently.The reported latency statistic is the median over 30 instances.
  • Evaluation: Metric@K evaluates K ∈ {1, 2, 4, 8} trajectories using bootstrap samples from N = 8 independent trajectories.The scoring function compares aggregated predictions with ground truth.

C.1 Main Results

AggAgent remains competitive across aggregation sizes and consistently improves performance-cost trade-offs over baselines. Its advantage reflects adaptive inspection and synthesis of trajectory evidence rather than one-shot retrieval alone.

  • Main results: AggAgent achieves competitive performance at K = 2, K = 4, and K = 8.Full results are reported across the evaluated aggregation sizes.
  • Efficiency: AggAgent achieves optimal performance relative to cost and latency across per-benchmark settings.The per-benchmark results are consistent with the averaged figures in the main paper.
  • Efficiency: AggAgent incurs much lower aggregation cost than Summary Aggregation and comparable cost to Solution Aggregation.This comparison holds across the reported settings.
  • Model strength: AggAgent remains the most effective LLM-based aggregation method when a stronger aggregator model is used.Using MiniMax generally improves performance across LLM-based aggregation methods.
  • Retrieval ablation: AggAgent consistently outperforms one-shot trajectory-RAG, which retrieves evidence once before synthesis.Its advantage is attributed to adaptively inspecting candidate answers, following up on disagreements, and iteratively verifying evidence.
  • Sequential and parallel scaling: AggAgent provides additional gains over aggregation baselines after context management improves the base rollout’s Pass@1 from 40.25 to 46.50.The results support complementarity between sequential context management and agentic aggregation.

D.3 Failure Mode Analysis

The failure analysis examines cases where at least one trajectory is correct but AggAgent fails, identifying retrieval, interpretation, conflict-resolution, and synthesis failures.

  • Failure cases: AggAgent can fail even when Pass@K succeeds, meaning at least one trajectory contains a correct answer.The analysis identifies four failure modes across these cases.
  • Failure breakdown: Evidence interpretation is the dominant annotated failure source in MiniMax-M2.5 cases.The authors identify improving verification and reasoning over retrieved evidence as an important direction.

D.4 Qualitative Analysis

AggAgent succeeds by identifying minority correct answers, resolving disagreements, and synthesizing partial progress across trajectories. Its aggregation procedure prioritizes verifiable tool observations over confidence or majority agreement.

  • Qualitative behaviours: AggAgent exhibits minority answer identification, disagreement resolution, and cross-trajectory synthesis when producing correct answers.These behaviours are highlighted as qualitative examples of how the method combines information across trajectories.
  • Model comparison: Using a stronger LLM aggregator leads to better performance across varying numbers of parallel samples.The comparison uses GLM-4.7-Flash as the rollout model and compares GLM with MiniMax as the aggregator.
  • Aggregation procedure: The aggregator retrieves final solutions, searches trajectories, reads surrounding steps, and cross-checks observations, reasoning, arithmetic, and logic before finishing.This workflow uses selective inspection to validate candidate solutions rather than relying solely on final answers.
  • Evidence evaluation: Tool results are treated as stronger evidence than agent reasoning, confidence, or numerical majority agreement.The procedure instructs the aggregator to trust objective tool outputs and verify conclusions against them.
  • Failure modes: The system warns against hallucinated observations, silent failures, circular logic, arithmetic errors, and majority bias.These pitfalls motivate checking raw tool outputs and grounding conclusions in verified evidence.
Loading 2604.11753v3…