Pith. sign in

REVIEW 1 major objections 5 minor 25 references

Delegation Drives Multi-Agent Search, Not Execution

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-09 07:10 UTC pith:VDMWPXFA

load-bearing objection Controlled capacity sweep shows delegation matters more than execution in hierarchical search agents; the asymmetry ratio is real but the sweep design inflates it somewhat. the 1 major comments →

arxiv 2607.07548 v1 pith:VDMWPXFA submitted 2026-07-08 cs.CL

Think Big, Search Small: Where Capacity Matters in Hierarchical Search Agents?

Qinnan Cai , Yibo Zhao , Xiang Li This is my paper
classification cs.CL
keywords multi-agent searchcapacity allocationhierarchical agentstask decompositiontrajectory distillationmulti-hop question answeringPareto frontierrole factorization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper studies how to allocate model capacity across roles in a hierarchical search agent—a system where a main agent decomposes a complex question into sub-queries and dispatches them to parallel sub-agents for retrieval. The authors factorize the task into three roles: delegation (task decomposition), execution (retrieval and evidence extraction), and answer generation (held fixed as a control). Through controlled capacity sweeps on five multi-hop QA benchmarks, they find a sharp asymmetry: scaling the delegation backbone improves exact match by about 11 points, while scaling the execution sub-agent moves it by only about 2.6 points. This identifies decomposition—not retrieval quality—as the capability bottleneck. Building on this, the authors train a 1.7B-parameter executor via quality-filtered trajectory distillation that matches a frontier-scale sub-agent in accuracy while consuming 37% fewer tokens, demonstrating that execution can be radically downsized without sacrificing performance. The paper's central claim is that the capability boundary of a hierarchical search agent is governed not by uniform component strength but by where capacity is placed in the hierarchy.

Core claim

The paper's central discovery is the capacity sensitivity asymmetry between delegation and execution in hierarchical search agents. Scaling the backbone (the model that decomposes questions and plans sub-queries) yields roughly four times the accuracy gain of scaling the executor (the model that retrieves and reads documents), across the same range of model scales. This asymmetry means the Pareto frontier of accuracy-versus-cost favors concentrating capacity in the backbone while keeping the executor compact. A 1.7B-parameter executor trained with targeted trajectory distillation—preserving single-search competence while injecting multi-search correction behavior—matches a frontier sub-agent

What carries the argument

The experimental design holds the answer-generation model fixed (Qwen3-32B, chosen for low parametric contamination) and varies only one of the two studied roles at a time. The delegation-capacity sweep fixes the executor at 1.7B and scales the backbone; the execution-capacity sweep fixes the backbone at a frontier model and scales the executor. The contrast between these two sweeps localizes the bottleneck. The compact executor training pipeline uses quality-filtered trajectory distillation from a stronger teacher, with a counterfactual filter that retains multi-search demonstrations only when a single-search variant of the same query fails—ensuring the student learns genuine iterative-retr

Load-bearing premise

The claim that decomposition is the bottleneck depends on the experimental design holding the answer-generation model fixed and restricting its input to only the question and sub-query reports. If the answerer were weaker, or if it had access to raw retrieved passages or reasoning traces, execution quality might become a larger bottleneck because the answerer would need to do more synthesis work from noisier evidence.

What would settle it

If a weaker answerer or one with access to raw passages were substituted, execution capacity might become the dominant factor in end-task accuracy.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners building multi-agent search systems should allocate their largest, most capable model to the delegation role and use a compact specialized executor, rather than instantiating all roles from one shared model.
  • A 1.7B executor trained with in-framework trajectory distillation can match a frontier sub-agent at a fraction of the token cost, making hierarchical search substantially cheaper to deploy.
  • The finding that off-the-shelf end-to-end search agents (like SearchR1) underperform when placed in the executor role suggests that execution-role training data must match the input distribution of atomic sub-queries, not full multi-hop questions.
  • The asymmetry implies that improving a search agent's capability is more efficiently achieved by improving decomposition quality—better relational reasoning, better sub-query scoping—than by improving retrieval or reading comprehension.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The bottleneck attribution depends on the answerer being strong enough to synthesize correct answers from sub-query reports alone. If the answerer were weaker or had access to noisier evidence, execution quality might become a larger bottleneck, since the system would rely more on the executor to pre-filter and organize evidence.
  • The asymmetry might shift on open-web search tasks (as opposed to the fixed Wikipedia corpus used here): if retrieval is noisier and documents less structured, execution quality—particularly multi-search correction and evidence extraction—could become more capacity-sensitive.
  • The same factorization logic may apply to other hierarchical agent systems beyond search (e.g., coding agents, tool-use agents): if decomposition is the bottleneck there too, compact specialized executors trained via in-framework distillation could replace frontier models in those executors.
  • The counterfactual filtering technique—retaining multi-step demonstrations only when the single-step variant fails—could be a general principle for distilling iterative-correction behavior without teaching spurious retries.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 5 minor

Summary. This paper studies how model capacity should be distributed across roles in hierarchical (multi-agent) search agents. The authors factorize the search task into three roles—delegation (task decomposition by a main agent), execution (retrieval and evidence extraction by sub-agents), and answer generation (held fixed as a confound control). Through controlled capacity sweeps on five multi-hop QA benchmarks, they find that (1) role factorization consistently outperforms a single-agent baseline, (2) scaling the delegation backbone improves EM by ~11 points whereas scaling the execution sub-agent moves EM by only ~2.6 points, identifying decomposition as the capability bottleneck, and (3) a 1.7B-parameter executor trained via quality-filtered trajectory distillation matches a frontier sub-agent in accuracy while consuming 37% fewer tokens. The experimental design is well-controlled: the answerer is held fixed, the retrieval corpus and retriever are constant, and the exclusion of instances answerable without search is a reasonable confound control.

Significance. The paper addresses a practically important question for multi-agent search system design—where to concentrate model capacity in a delegation–execution hierarchy. The role-factorized experimental instrument is clean, and the finding that execution can be aggressively downsized while delegation remains the bottleneck is actionable. The SFT pipeline for the compact executor is well-motivated by the behavioral gap analysis (Table 4: 1.7B does single-search 99.83% of the time vs. 40.31% multi-search for the teacher), and the quality-filtering design (preserving single-search competence, injecting multi-search correction with a counterfactual filter) is a concrete, falsifiable contribution. The cross-backbone generalization check (Table 5, GLM-5.1) and the Pareto frontier analysis (Figure 3) add credibility. Code is publicly available.

major comments (1)
  1. §3.3, Interventions 2 and 3; §4.5, Fig. 2: The central claim—that decomposition is the capability bottleneck—rests on comparing Δ_D ≈ 11 EM (Table 2, backbone varied, executor fixed at Qwen3-1.7B) against Δ_E ≈ 2.6 EM (Table 3, executor varied, backbone fixed at DeepSeek-V4-Flash). These two sweeps are measured at different operating points of the held-fixed variable: the delegation sweep holds the executor at its weakest (1.7B), while the execution sweep holds the backbone at its strongest (DeepSeek-V4-Flash). This design choice maximizes Δ_D while minimizing Δ_E, which is a classic interaction-effect problem. The two slopes are measured on different cross-sections of the (c_D, c_E) surface, and the asymmetry could be an artifact of where each cross-section is taken rather than a property of the surface itself. The paper does not acknowledge this confound or test its robustness. At a 2×
minor comments (5)
  1. §4.1: The LLM-as-judge uses DeepSeek-V4-Flash, which is also used as the frontier backbone and sub-agent in several experiments. This creates a potential self-preference bias in the judge scores. Consider noting this as a limitation.
  2. Table 1: The EM gains from single-agent to main–sub are described as '4.5 to 8.6 points' in the abstract, but the table shows gains ranging from +4.52 to +8.63. Minor rounding inconsistency.
  3. Figure 9 caption: The caption reads 'Delegation prompt (part 3): worked example of a multi-hop question with sequential sub-queries,' but the figure is labeled as the execution prompt (search sub-agent) in §C.2. Caption appears mismatched.
  4. §4.4: The text states 'Scaling from Qwen3-1.7B to Qwen3-32B moves Avg. EM by only 1.27 points (from 39.18 to 40.45),' but Table 3 shows Qwen3-32B at 40.45 EM and Qwen3-1.7B at 39.18 EM, so the difference is 1.27. However, the abstract and §4.5 cite Δ_E ≈ 2.6, which appears to include the DeepSeek-V4-Flash sub-agent (41.77 EM). Clarify which comparison defines Δ_E.
  5. App. D: The SFT corpus contains 2,168 records from 1,591 unique source questions. Consider reporting the single-search vs. multi-search split in the filtered corpus to verify that the multi-search correction behavior was adequately represented.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for a careful and constructive reading of our manuscript. The referee correctly identifies a genuine confound in our experimental design: the delegation and execution sweeps are measured at different operating points of the held-fixed variable, which could inflate the apparent asymmetry. We agree this is a real concern and will address it through additional experiments and revised framing in the revision.

read point-by-point responses

  1. Referee: §3.3, Interventions 2 and 3; §4.5, Fig. 2: The central claim—that decomposition is the capability bottleneck—rests on comparing Δ_D ≈ 11 EM (Table 2, backbone varied, executor fixed at Qwen3-1.7B) against Δ_E ≈ 2.6 EM (Table 3, executor varied, backbone fixed at DeepSeek-V4-Flash). These two sweeps are measured at different operating points of the held-fixed variable: the delegation sweep holds the executor at its weakest (1.7B), while the execution sweep holds the backbone at its strongest (DeepSeek-V4-Flash). This design choice maximizes Δ_D while minimizing Δ_E, which is a classic interaction-effect problem. The two slopes are measured on different cross-sections of the (c_D, c_E) surface, and the asymmetry could be an artifact of where each cross-section is taken rather than a property of the surface itself. The paper does not acknowledge this confound or test its robustness.

    Authors: The referee raises a valid and important concern. We acknowledge that the two capacity sweeps are measured at different operating points of the held-fixed variable, and that this design could in principle amplify the apparent asymmetry between delegation and execution sensitivity. This is a genuine confound in our current experimental design, and we appreciate the referee identifying it so precisely. In the revision, we will address it in two ways. First, we will add additional cross-sections of the (c_D, c_E) surface: specifically, we will run the delegation sweep with the executor fixed at a stronger operating point (e.g., Qwen3-32B or DeepSeek-V4-Flash) and the execution sweep with the backbone fixed at a weaker operating point (e.g., Qwen3-8B or Qwen3-14B). This will test whether the asymmetry is robust to the choice of cross-section or whether it diminishes when the held-fixed variable is set at a more comparable level. Second, we will revise the manuscript to explicitly acknowledge the interaction-effect concern in §3.3 and §4.5, and we will qualify the central claim accordingly: rather than stating that decomposition is the bottleneck as an absolute property, we will frame it as holding across the operating points we test, pending the additional cross-sections. We expect the asymmetry to persist—because the case study in §4.7 provides a mechanistic explanation for why decomposition quality governs the performance ceiling regardless of executor strength—but we agree this must be demonstrated empirically rather than assumed. If the additional experiments show that the asymmetry narrows substantially at different operating points, we will revise the strength of the claim accordingly. revision: yes

Circularity Check

0 steps flagged

No circularity found; all central claims are empirical measurements on external benchmarks with disjoint train/test splits

full rationale

The paper's three principal findings are all empirical measurements on external, independently developed benchmarks (2WikiMultihopQA, HotpotQA, MuSiQue, PopQA, Bamboogle) using independently developed models (Qwen3 series, DeepSeek-V4-Flash/Pro, GLM-5.1). (1) The role-factorization gain (Table 1) is a direct A/B comparison with the same model in both conditions. (2) The capacity-sensitivity asymmetry (Δ_D ≈ 11, Δ_E ≈ 2.6) is measured from Tables 2–3; these are observed EM differences, not quantities derived from definitions or fitted parameters. The skeptic's concern about asymmetric operating points is a valid methodology critique (correctness risk), but the measurements themselves are not circular. (3) The SFT executor result (Table 5) uses quality-filtered training trajectories (EM=1 + LLM-judge acceptance) on 3,000 training queries, evaluated on a disjoint 3,869-instance test set (Appendices D, E). Filtering training data by correctness and evaluating on held-out data is standard practice, not circularity. The one self-citation (Zhao et al., 2026) provides evaluation infrastructure (benchmarks, corpus), not a load-bearing theoretical premise. No step in the derivation chain reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper introduces no new entities, particles, forces, or dimensions. The three roles (delegation, execution, answer generation) are functional decompositions of an existing search pipeline, not new postulated objects. The free parameters (T, K) are standard search budgets. The SFT corpus size is determined by the filtering criteria, not hand-tuned to fit a target result.

free parameters (3)
  • T (delegation round budget) = 4
    Set by hand as the maximum number of delegation rounds; controls the depth of the search hierarchy.
  • K (execution loop budget) = 3
    Set by hand as the maximum number of assistant turns per sub-agent; controls the depth of individual sub-agent search.
  • SFT corpus size = 2,168 records from 1,591 unique questions
    Determined by the quality filtering criteria applied to 3,000 source questions; not a fitted parameter in the traditional sense but a design choice that affects the SFT executor's capability.
axioms (4)
  • domain assumption The answer-generation module π_A (Qwen3-32B) faithfully synthesizes answers from reports without relying on parametric memory.
    Sec. 4.1 and App. A: verified by a no-retrieval control showing 18.58 EM, but this is an assumption that the answerer does not inject knowledge when retrieval is present.
  • domain assumption Sub-queries are self-contained atomic questions interpretable without access to the delegation context H_D.
    Sec. 3.1: the decomposition rules (App. C) assume sub-queries can be executed in isolation, which is necessary for the context isolation design.
  • domain assumption The LLM-as-judge (DeepSeek-V4-Flash) provides reliable semantic correctness assessment.
    Sec. 4.1 and App. B: the judge score is a primary metric, and its reliability is assumed, not independently validated against human judgment.
  • domain assumption The Wiki-fixed corpus is a complete retrieval environment for the five benchmarks.
    Sec. 4.1: the corpus is described as a 'completed version' of Wikipedia-2018, but completeness is assumed rather than verified.

pith-pipeline@v1.1.0-glm · 19302 in / 3114 out tokens · 491965 ms · 2026-07-09T07:10:53.088083+00:00 · methodology

0 comments
read the original abstract

Large language model based search agents increasingly adopt multi-agent architectures in which a main agent decomposes a complex question into sub-queries and dispatches them to parallel sub-agents. However, existing systems instantiate all roles from a single model of identical scale, leaving open how model capacity should be distributed across roles. We factorize hierarchical search into three roles: a delegation role responsible for task decomposition, an execution role responsible for retrieval and evidence extraction, and an answer generation role held fixed as a confound control. We then conduct controlled capacity sweeps along the delegation and execution axes on five multi-hop QA benchmarks. The experiments yield three findings. First, role factorization consistently outperforms a single-agent baseline, improving exact match from 4.5 to 8.6 points across six model scales. Second, capacity sensitivity is asymmetric: scaling the delegation backbone improves EM by ~11 points, whereas scaling the execution sub-agent moves EM by only ~2.6 points, identifying decomposition as the capability bottleneck. Third, a 1.7B-parameter executor trained via quality-filtered trajectory distillation matches a frontier sub-agent in accuracy while consuming 37% fewer sub-agent tokens, advancing the Pareto frontier. These results suggest a concrete recipe for building hierarchical search agents: concentrate capacity at delegation and downsize execution without sacrificing accuracy. Our code is available at https://github.com/QinnanCai0115/role-factorized-search.

Figures

Figures reproduced from arXiv: 2607.07548 by Qinnan Cai, Xiang Li, Yibo Zhao.

Figure 1
Figure 1. Figure 1: Overview of our hierarchical search framework. The delegation policy [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Capacity sensitivity of each role. Scaling Execution (blue) fixes the backbone to DeepSeek-V4-Flash and varies the sub-agent; Scaling Decomposition (red) fixes the sub-agent to Qwen3-1.7B and varies the backbone. Arrows mark each sweep’s total gain (∆D vs. ∆E). V4-Flash sub-agent lifts EM by only 2.59 total. This pattern is consistent across all five benchmarks. Once the backbone produces well-scoped sub-q… view at source ↗
Figure 3
Figure 3. Figure 3: Effectiveness–efficiency trade-off across sub-agent configurations under a DeepSeek-V4- [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Backbone reasoning error on a multi-hop question. Left: DeepSeek-V4-Flash correctly [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Prompt of LLM-as-Judge C ROLE PROMPTS This appendix lists the system prompts of the delegation policy πD (backbone) and the execution policy πE (search sub-agent) described in Sec. 3.1. Both prompts are identical across all experi￾mental conditions; only the model instantiating each role varies. When the backbone judges the collected evidence sufficient, it emits a stop signal rather than an answer; the de… view at source ↗
Figure 6
Figure 6. Figure 6: Delegation prompt (part 1): role description, output format, and decomposition rules. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Delegation prompt (part 2): worked example of a comparison question with parallel sub [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Delegation prompt (part 3): worked example of a multi-hop question with sequential sub [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Delegation prompt (part 3): worked example of a multi-hop question with sequential sub [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 12 internal anchors

  1. [1]

    Advancing general-purpose reasoning models with modular gradient surgery

    Min Cai, Yu Liang, Longzheng Wang, Yan Wang, Yueyang Zhang, Long Xia, Zhiyuan Sun, Xi Ye, and Daiting Shi. Advancing general-purpose reasoning models with modular gradient surgery. arXiv preprint arXiv:2602.02301,

    work page arXiv

  2. [2]

    MAO-ARAG: Multi-Agent Orchestration for Adaptive Retrieval-Augmented Generation

    Yiqun Chen, Erhan Zhang, Lingyong Yan, Shuaiqiang Wang, Jizhou Huang, Dawei Yin, and Jiaxin Mao. Mao-arag: Multi-agent orchestration for adaptive retrieval-augmented generation.arXiv preprint arXiv:2508.01005,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    The Faiss library

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre- Emmanuel Mazar ´e, Maria Lomeli, Lucas Hosseini, and Herv ´e J ´egou. The faiss library.arXiv preprint arXiv:2401.08281,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Context length alone hurts llm perfor- mance despite perfect retrieval.arXiv preprint arXiv:2510.05381,

    Yufeng Du, Minyang Tian, Srikanth Ronanki, Subendhu Rongali, Sravan Bodapati, Aram Galstyan, Azton Wells, Roy Schwartz, Eliu A Huerta, and Hao Peng. Context length alone hurts llm perfor- mance despite perfect retrieval.arXiv preprint arXiv:2510.05381,

    work page arXiv

  5. [5]

    OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories

    Yuwen Du, Rui Ye, Shuo Tang, Keduan Huang, Xinyu Zhu, Yuzhu Cai, and Siheng Chen. Openseeker-v2: Pushing the limits of search agents with informative and high-difficulty trajecto- ries.arXiv preprint arXiv:2605.04036,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2508.07976

    Jiaxuan Gao, Wei Fu, Minyang Xie, Shusheng Xu, Chuyi He, Zhiyu Mei, Banghua Zhu, and Yi Wu. Beyond ten turns: Unlocking long-horizon agentic search with large-scale asynchronous rl.arXiv preprint arXiv:2508.07976,

    work page arXiv

  7. [7]

    Terminus-4B: Can a Smaller Model Replace Frontier LLMs at Agentic Execution Tasks?

    Spandan Garg, Vikram Nitin, and Yufan Huang. Terminus-4b: Can a smaller model replace frontier llms at agentic execution tasks?arXiv preprint arXiv:2605.03195,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Adaptation of agentic ai.arXiv preprint arXiv:2512.16301, 2025a

    Pengcheng Jiang, Jiacheng Lin, Zhiyi Shi, Zifeng Wang, Luxi He, Yichen Wu, Ming Zhong, Peiyang Song, Qizheng Zhang, Heng Wang, et al. Adaptation of agentic ai.arXiv preprint arXiv:2512.16301, 2025a. Pengcheng Jiang, Xueqiang Xu, Jiacheng Lin, Jinfeng Xiao, Zifeng Wang, Jimeng Sun, and Jiawei Han. s3: You don’t need that much data to train a search agent v...

    work page arXiv 2025

  9. [9]

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Websailor-v2: Bridging the chasm to proprietary agents via synthetic data and scalable reinforcement learning.arXiv preprint arXiv:2509.13305, 2025a

    Kuan Li, Zhongwang Zhang, Huifeng Yin, Rui Ye, Yida Zhao, Liwen Zhang, Litu Ou, Dingchu Zhang, Xixi Wu, Jialong Wu, et al. Websailor-v2: Bridging the chasm to proprietary agents via synthetic data and scalable reinforcement learning.arXiv preprint arXiv:2509.13305, 2025a. 13 Preprint. Under review. Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia ...

    work page arXiv 2025

  11. [11]

    Ai-searchplanner: Mod- ular agentic search via pareto-optimal multi-objective reinforcement learning.arXiv preprint arXiv:2508.20368,

    Lang Mei, Zhihan Yang, Xiaohan Yu, Huanyao Zhang, and Chong Chen. Ai-searchplanner: Mod- ular agentic search via pareto-optimal multi-objective reinforcement learning.arXiv preprint arXiv:2508.20368,

    work page arXiv

  12. [12]

    SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research

    Pu Ning, Quan Chen, Kun Tao, Xinyu Tang, Tianshu Wang, Qianggang Cao, Xinyu Kong, Zujie Wen, Zhiqiang Zhang, and Jun Zhou. Searchswarm: Towards delegation intelligence in agentic llms for long-horizon deep research.arXiv preprint arXiv:2606.09730,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Measuring and narrowing the compositionality gap in language models

    Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models. InFindings of the Association for Computational Linguistics: EMNLP 2023, pp. 5687–5711,

    work page 2023

  14. [14]

    Small language models for agentic systems: A survey of archi- tectures, capabilities, and deployment trade offs.arXiv preprint arXiv:2510.03847,

    Raghav Sharma and Manan Mehta. Small language models for agentic systems: A survey of archi- tectures, capabilities, and deployment trade offs.arXiv preprint arXiv:2510.03847,

    work page arXiv

  15. [15]

    Deep research: A systematic survey.arXiv preprint arXiv:2512.02038,

    Zhengliang Shi, Yiqun Chen, Haitao Li, Weiwei Sun, Shiyu Ni, Yougang Lyu, Run-Ze Fan, Bowen Jin, Yixuan Weng, Minjun Zhu, et al. Deep research: A systematic survey.arXiv preprint arXiv:2512.02038,

    work page arXiv

  16. [16]

    Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG

    Aditi Singh, Abul Ehtesham, Saket Kumar, Tala Talaei Khoei, and Athanasios V Vasilakos. Agentic retrieval-augmented generation: A survey on agentic rag.2501.09136,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning

    Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning.arXiv preprint arXiv:2503.05592,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Simpledeepsearcher: Deep information seeking via web-powered reasoning trajectory synthesis.arXiv preprint arXiv:2505.16834,

    Shuang Sun, Huatong Song, Yuhao Wang, Ruiyang Ren, Jinhao Jiang, Junjie Zhang, Fei Bai, Jia Deng, Wayne Xin Zhao, Zheng Liu, et al. Simpledeepsearcher: Deep information seeking via web-powered reasoning trajectory synthesis.arXiv preprint arXiv:2505.16834,

    work page arXiv

  19. [19]

    Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Y Charles, HS Che, Cheng Chen, Guanduo Chen, et al. Kimi k2. 5: Visual agentic intelligence.arXiv preprint arXiv:2602.02276,

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Wideseek-r1: Exploring width scaling for broad information seeking via multi-agent reinforcement learning.arXiv preprint arXiv:2602.04634,

    Zelai Xu, Zhexuan Xu, Ruize Zhang, Chunyang Zhu, Shi Yu, Weilin Liu, Quanlu Zhang, Wenbo Ding, Chao Yu, and Yu Wang. Wideseek-r1: Exploring width scaling for broad information seeking via multi-agent reinforcement learning.arXiv preprint arXiv:2602.04634,

    work page arXiv

  21. [21]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 conference on empirical methods in natural language processing, pp. 2369–2380,

    work page 2018

  23. [23]

    ReAct: Synergizing Reasoning and Acting in Language Models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629,

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Retrieval, Reward, and Training Protocols: What Matters in Training Search Agents?

    Yibo Zhao, Zichen Ding, Jiayi Wu, Zun Wang, and Xiang Li. Retrieval, reward, and training proto- cols: What matters in training search agents?arXiv preprint arXiv:2605.27881,

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Under review

    15 Preprint. Under review. A CHOICE OFANSWER-GENERATIONMODEL As discussed in Sec. 4.1, we choose Qwen3-32B over DeepSeek as the fixed answer-generation modelπ A. To validate this choice, we run both models on the 3,869 evaluation instanceswithout any retrieval, so that correct answers can only come from parametric memory. Tab. 6 reports the results. Table...

    work page 2025