acceptodds
Under review as a conference paper at ICLR 2027

How Harnesses Reshape Agentic Search: Training Stateless Large Reasoning Models for Stateful Search

Abstract

Agentic search requires reasoning over evidence accumulated across interactions, yet the underlying language model is stateless across invocations. Existing approaches typically bridge this gap by carrying interaction histories across turns and applying post-training techniques. However, growing contexts can obscure earlier evidence and force the model to reconstruct search progress at each decision. Moreover, when supervision relies solely on final outcomes, training provides limited corrective feedback, leaving intermediate errors unaddressed and potentially reinforcing flawed reasoning or redundant actions. We introduce H-Search, a framework that addresses these challenges through the joint design of a stateful search harness and decision-level policy training. H-Search organizes multi-turn search into self-contained decisions: each model invocation conditions on the current structured state, with cross-turn dependencies preserved through explicit state transitions. This separation makes the same state-conditioned decision the basic unit of both search execution and policy supervision. To train within this structure, we develop harness-grounded on-policy self-distillation, where a frozen copy of the initial model uses privileged information about evidence coverage and termination readiness to provide dense token-level guidance along student-generated trajectories. The resulting policy operates within the same harness at inference without privileged information or teacher access. Across five open-domain benchmarks over Wikipedia, H-Search with a 27B backbone achieves 79.8% average recall, outperforming the evaluated same-backbone baselines by at least 11.2 percentage points. These results demonstrate the potential of coupling explicit search-state management with supervision at the level of individual decisions.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agentic Search, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_uEwnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a new harness for agentic search scenarios that abstracts away long evidence contexts into concise observations, and proposes a paired OPSD training method to encourage the base LLMs to familiarize themselves with such a harness. Experiments using Qwen3.6-27B as the base policy across multi-hop and open-domain QA benchmarks show the effectiveness of the harness compared to the Search-R1 harness as well as the OPSD training.

Critical strengths
  • The problem formulation is clean in Section 3.1.
  • Figures are easy to follow across the paper.
  • Experiments consider multiple benchmarks and different base LLMs to show general applicability and effectiveness.
Critical weaknesses

W1 The challenges are not clearly stated. Challenges i) and ii), memory and context degradation and policy errors and failures, are common bottlenecks for long-context scenarios, not specifically tied to agentic search; and iii) the credit-assignment issue is a general and open problem for policy training, as different tasks/policies may require careful, distinct designs to balance correct rewards while avoiding possible hacking issues.

W2 For methodology, my understanding is that the authors propose a new harness that lets an external module manage the context while the policy only cares about the next action conditioned on the abstracted context, together with OPSD training to become familiar with such a harness. That is fair. Yet how the context is compacted and abstracted is unclear or hard to follow, and how it compares with existing agentic context management works like summary, context folding, etc., is unclear.

W3 The biggest drawback of this paper is the empirical validation of the method. I think to show H-Search is effective, there are two aspects: i) the effectiveness of the H-Search harness alone, and ii) the effectiveness of the proposed OPSD in mastering the H-Search harness compared to other learning paradigms. For both conditions, the authors should use the same base policy and identical training/eval configurations for a fair comparison, and report end-to-end QA performance (e.g., answer accuracy) for policy capability, while intermediate evidence recall can serve as an extra metric (currently the paper only reports retrieval recall and leaves downstream answer quality to future work). However, the current experiments in the main paper are very mixed, with mixed-up agent harnesses and different base policies, and only upon reading the Appendix can I find the base policy direct inference scores. At least in the main paper, the authors should clearly provide:

  • Base model direct inference, e.g., the standard ReAct workflow;
  • Base model direct inference w/ H-Search and w/ Search-R1 harness across the same set of policies to show the general effectiveness of the harness;
  • Then for training, under the same base policy and the H-Search harness, compare OPSD, OPD (where the teacher can be a more capable Qwen model), standard GRPO, and offline imitation methods like SFT/BC. Comparing with these in the main text can better show the actual utility of OPSD. I saw the BC-only and GRPO results in Appendix Table 3, GRPO should be more expensive, and that efficiency advantage and training details should be highlighted in the main paper.

Beyond performance alone, different harnesses/training paradigms may shape different agentic behaviors, like the avg. searches, and avg. thinking and output lengths. All of these can help us understand the policy behaviors, what it learns, and how to better improve it, yet all of them are lacking in the paper. I still do not know the actual context length for training/eval in the paper.

W4 Finally, a minor point: in my experience, all these benchmarks can mostly fit within a 32K context, which is quite small and does not align with the long-context bottlenecks repeatedly mentioned in the paper. More challenging agentic benchmarks, like BrowseComp, should definitely be considered to strengthen the evaluation. Also, the harness is limited to corpus search only; I am not sure whether it sacrifices generalization capability across broader agentic tasks.

Questions for authors

Please see my weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer uEwn

We thank the reviewers for their thoughtful feedback and respond point by point below.


Weaknesses & Detailed Review

W1: How the three challenges arise in agentic search

We agree that these challenges also arise in other sequential tasks. In agentic search, they have concrete consequences. (i) As new results arrive, the agent must retain evidence from earlier turns because it may be needed to formulate the next query. (ii) The appropriate action depends on the current evidence state: the agent must decide whether to inspect a candidate, search again, or stop. (iii) The same low final recall can result from failing to retrieve, inspect, or retain a relevant document. OPSD provides action-level guidance at the search states visited by the student.


W2: Context compaction and its relation to context folding

Appendix C.1 describes the state representation, and Appendix D provides the prompts.

Document summaries and reading. The harness deterministically extracts query-conditioned, 200-character excerpts after retrieval. The policy invokes read_document based on the question and available evidence when snippets seem insufficient. The tool returns full content when available and the available snippet in snippet-only mode or after full-text retrieval fails. The context-budget threshold is approximately 8,192 tokens.

How H-Search compacts context. Each policy input renders the current query plan, selected document IDs, a bounded set of top candidates with read status and brief summaries, remaining information gaps, and recent searches and outcomes, as illustrated in Appendix D’s Working-Memory Representation. Tool actions update this record: Read marks candidates as inspected, Select adds evidence, and Judge Sufficiency determines whether to stop or continue. The policy acts on the updated state, and OPSD uses the same representation during training and inference. Context Folding summarizes completed subtask trajectories. H-Search renders the search state needed for the next decision.


W3: Controlled comparisons for the harness and OPSD

We will adjust the descriptions of the key controlled comparative experiments—presented in Appendix Tables 5 and 3—within the main text.

Harness effectiveness. Table 5 compares H-Search, FrugalRAG, and CoRAG using the same Qwen3.6-27B checkpoint without task-specific training, achieving macro-average document recall of 76.3%, 64.7%, and 62.3%, respectively. Table 2 compares H-Search with Search-R1 on shared backbones, obtaining 83.1% vs. 72.0% with GLM-5.2 and 71.6% vs. 70.5% with GPT-OSS-120B. These comparisons control for backbone choice.

Training effectiveness. Table 3 fixes the Qwen3.6-27B backbone and H-Search harness. Macro-average recall is 76.3% without training, 77.3% with BC alone, 77.6% with GRPO, and 79.8% with OPSD, showing gains over BC and GRPO under the evaluated setup. Distillation from a stronger teacher would test a separate question.

Inference budgets. We additionally report TriviaQA recall, maximum rounds, and mean token consumption under our experimental configurations.

MethodRecallMax. roundsMean tokens/query
FrugalRAG72.5%53,337
Search-R159.7%104,504
CoRAG45.4%65,106
IRCoT65.2%107,721
H-Search-Lite73.9%85,907
H-Search-OPSD79.2%2013,697

With thinking disabled and an eight-round limit, H-Search-Lite achieves 73.9% recall, exceeding the strongest baseline, FrugalRAG, by 1.4 percentage points. H-Search-OPSD reaches 79.2% with a larger budget.

To assess the effect of round limits, we reran all four baselines with a maximum of 20 rounds.

MethodRecallMax. rounds
FrugalRAG73.8%20
Search-R160.3%20
CoRAG48.9%20
IRCoT67.1%20
H-Search-OPSD79.2%20

All baselines improve, while H-Search-OPSD retains a 5.4-percentage-point lead over FrugalRAG, the strongest baseline. The advantage thus persists on TriviaQA under a common 20-round limit. We will add these results and budget settings to the revision.


W4

Context length. Even within 32K tokens, earlier evidence can become harder to use as unrelated tool outputs accumulate (Liu et al. [1]). Removing structured state reduces average document recall from 76.3% to 46.7% (Table 4), measuring its value on the evaluated tasks.

Generalization. H-Search’s state fields and action constraints target corpus retrieval. Explicit state tracking could transfer to other agentic tasks with task-specific state representations and tools.

[1] Liu, N. F., Lin, K., Hewitt, J., et al. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 2024, 12, 157–173.


Reviewer_CH4MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the challenges of state management and policy training in multi-turn agentic search. The authors propose H-Search, which combines a stateful search harness with decision-level policy training. The harness maintains persistent search states, organizes retrieved evidence, and constrains available actions at each step. It also introduces harness-grounded on-policy self-distillation, where a frozen teacher uses privileged evidence-state information to guide student decisions. Experiments across five open-domain retrieval benchmarks demonstrate improvements in document retrieval recall over existing search agents.

Critical strengths
  1. The paper addresses an important problem in multi-turn agentic search. Instead of focusing only on model capabilities, it considers the role of the search harness in shaping agent behavior. The joint design of state management and policy training is an interesting and well-motivated perspective.

  2. The proposed framework maintains explicit search states and provides state-conditioned contexts for each model decision. It also uses privileged evidence information to supervise intermediate decisions through OPSD. This combination feels more like a systematic design than a simple add-on module to me. The experimental results are strong across five benchmarks over the strongest same-backbone baseline.

Critical weaknesses
  1. Limited experimental validation. The evaluation mainly focuses on document-level retrieval recall, without showing whether the improvements translate into better downstream QA performance. The paper also lacks analysis of inference costs, such as latency, token consumption, and the number of tool calls. Moreover, while the harness is evaluated with larger models, OPSD training is only tested on the 27B backbone. It remains unclear whether the full approach can scale to larger models.

  2. Limitations in benchmark selection and baseline comparisons fairness. The paper emphasizes state management and context degradation in multi-turn search, but the experiments mainly use QA benchmarks with relatively short reasoning chains (questions that only need few hops). It is unclear whether these benchmarks sufficiently evaluate the proposed method in more challenging long-horizon search scenarios. In addition, Table 1 compares methods using different backbones, including substantially smaller or older models for some baselines.

Questions for authors
  1. Fairness of baseline comparisons. Are the baselines evaluated under comparable search budgets, including the number of tool calls and available retrieval operations?

  2. Details of the harness ablation. In Table 4, what exactly is removed in the "w/o structured state" setting? Does it only change how the search history is presented, or does it also change the information available to the model?

2 comments
AuthorsOpenReview ↗

Response to Reviewer CH4M

We sincerely thank all reviewers for their thoughtful and constructive feedback. Below we provide point-by-point responses to all comments.


Weaknesses & Detailed Review

W1 & Q1:

Thank you for highlighting the need to report inference budgets. We provide additional results on TriviaQA below. The first table reports recall, the maximum round limit, and mean token consumption under the baseline configurations used in our experiments.

MethodRecallMax. roundsMean tokens/query
FrugalRAG72.5%53,337
Search-R159.7%104,504
CoRAG45.4%65,106
IRCoT65.2%107,721
H-Search-Lite73.9%85,907
H-Search-OPSD79.2%2013,697

H-Search-Lite achieves 73.9% recall with thinking disabled and an eight-round limit. It exceeds the strongest baseline in this table, FrugalRAG, by 1.4 percentage points. The full H-Search-OPSD configuration reaches 79.2% recall with a larger inference budget.

To examine whether the different round limits account for the performance gap, we reran all four baselines with their maximum round limit increased to 20.

MethodRecallMax. rounds
FrugalRAG73.8%20
Search-R160.3%20
CoRAG48.9%20
IRCoT67.1%20
H-Search-OPSD79.2%20

All four baselines improve with the increased round limit. H-Search-OPSD retains a 5.4-percentage-point advantage over the strongest baseline, FrugalRAG. On TriviaQA, this result shows that the advantage persists when all methods are allowed up to 20 rounds. We will include these results and the explicit budget settings in the revised paper.

Regarding the question of "whether the full approach can scale to larger models," our goal in the experiments was to test whether harness-grounded post-training improves retrieval at a fixed model capacity. Qwen3.6-27B is comparable in scale to backbones used in prior search-agent post-training studies, such as Search-R1(Jin et al.[2])(3B and 7B) and ReSearch (Chen et al.[3])(7B and 32B). With the backbone and harness held fixed, OPSD raises macro-average document recall from 76.3% to 79.8% (Tables 3 and 6). The resulting 27B system also outperforms all Search-R1 configurations evaluated in Table 2, including GLM-5.2 at 72.0%, and comes within 3.3 percentage points of GLM-5.2 using H-Search (83.1%). Table 2 separately demonstrates that the harness works with larger backbones at inference.

[1] Jin, B., Zeng, H., Yue, Z., et al. Search-R1: Training LLMs to reason and leverage search engines with reinforcement learning [C]. Conference on Language Modeling (COLM), 2025.

[2] Chen, M., Sun, L., Li, T., et al. ReSearch: Learning to reason with search for LLMs via reinforcement learning [C]. Advances in Neural Information Processing Systems (NeurIPS), 2025, 38.


W2 & R2:

No-structured-state ablation. Appendix C.1 describes the state representation, and Appendix D provides the corresponding prompts. We clarify the implementation below. While maintaining a constant total token budget for the input, the structured working memory representation is replaced by a chronological sequence of tool calls and observation records. History is retained until that budget is reached, after which the oldest complete interactions are removed while preserving the question and task instructions.

The model checkpoint, thinking setting, retriever, tools, and round limit will be held fixed. Backend state will remain available for tool execution, action constraints, and evaluation. The policy will receive its evidence history through the interaction transcript, including during sufficiency assessment.


Thank authors for the detailed response and additional experiments. I appreciate the clarifications and inference budget analysis. However, several of my original key concerns remain unresolved. First, improvements in retrieval recall are not validated on downstream QA performance. Second, the fairness issue of comparing different backbone models remains unaddressed. Third, the relatively simple, short-horizon benchmarks do not sufficiently demonstrate the benefits of H-Search in challenging long-horizon scenarios. For these reasons, I decided to maintain my original score.

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

The premise is that the model driving agentic search is stateless, so all cross‑turn information must be rebuilt by an external harness before each invocation. The authors argue that the harness, meaning how it stores search state, renders a context, restricts the action set, and handles termination, is a variable that shapes learning rather than a fixed implementation detail. From this they derive three failure modes: evidence buried as history grows, policies that must reconstruct progress from a messy history, and a supervision gap where one terminal reward cannot localize which decision went wrong. They propose H‑Search, which keeps persistent state in the harness and exposes a self‑contained context per decision, making the state‑conditioned decision the unit of both execution and supervision. Training applies on‑policy self‑distillation, where a frozen teacher of identical weights differs from the student only in receiving privileged state information such as gold evidence coverage and whether it is safe to finish, and guides the student along the student's own trajectory at the token level; teacher and privileged input are removed at inference. Evaluated across five Wikipedia‑based QA benchmarks using Qwen3.6‑27B as the backbone, the approach yields strong retrieval performance and outperforms the best baseline built upon the same model. Ablations attribute most of the gain to the harness rather than to training.

Critical strengths
  1. The harness and training effects are separated cleanly and reported honestly. Table 5 keeps the model checkpoint unchanged and only replaces the workflow. Under this setup H‑Search reaches 76.3 percent, compared to 62.3 percent for CoRAG and 64.7 percent for FrugalRAG. Table 3 further shows training brings an additional 3.5 percentage point improvement. This reverses common attribution patterns within existing literature. Most performance gains stem from state representation and action constraints, rather than the optimization objective itself. Presenting such decomposition which tempers the paper’s core headline result improves result auditability. It also provides a clean baseline for future work focused on harness comparison.
  2. The parameter efficiency argument follows rigorous control practices. The authors retrain the strongest competing method using the exact same backbone model instead of directly quoting published results. They establish a 68.6 percent reference performance for this shared backbone, then their method outperforms this baseline by 11.2 percentage points. Combined with results from Table 2. Here the 27B model outperforms a 120B model configuration, and sits only 3.3 points behind a far larger system running solely on the harness. These observations support a practically useful conclusion. For this class of tasks, well‑designed harness can offset a large portion of the demand for larger model backbones.
Critical weaknesses
  1. The cross-model comparison rests on a harness effect that is inconsistent across backbones. Table 2 and Figure 4 carry RQ2 and RQ3, and both depend on swapping the framework while holding the backbone fixed. Measured against Search-R1 on the same backbone, that swap moves GLM5.2 from 72.0% to 83.1%, an 11.1 point gain, while it moves GPT-OSS-120B only from 70.5% to 71.6%, a gain of 1.1 points. On two of the five benchmarks the 120B configuration is not better at all under H-Search, scoring 0.809 versus 0.798 on 2Wiki and 0.811 versus 0.820 on NQ, so this is not a case of a harness that helps modestly but consistently. Three readings are available and the paper offers no diagnostic to separate them. The large-model configuration may have received less implementation tuning, which would cast doubt on Table 2 as a whole. The harness benefit may be specific to particular backbone families, in which case RQ2 cannot be stated as a general property. Or one entry may be an outlier being read as a property of the design. Every claim the paper makes about parameter efficiency inherits this ambiguity, including the headline 27.5 times figure, and the paper reports no per-backbone adaptation record that would resolve it.
  2. Recall alone cannot support the system level claims, and the component driving most performance gains lacks clear specification. The main paper only reports document recall for the final evidence set. It provides no answer accuracy results for any compared method. The failure cases that motivate this work involve termination timing, redundant retrieval, and unsupported conclusions. Recall metrics cannot capture any of these issues. The paper also fails to normalize experimental cost for its comparison with FrugalRAG. FrugalRAG is inherently designed to keep operation costs under control. Removing structured state brings the biggest performance change observed in the paper. Performance falls from 76.3% to 46.7%. This ablation only gets covered within a single short paragraph. The exact setup of this ablated variant is never defined. The renderer module responsible for this major performance shift comes without pseudocode. It also offers no concrete details about summary source, truncation length or window size.
Questions for authors
  1. What explains the uneven harness gain across backbones, where the untrained framework lifts GLM5.2 by roughly 11 points but GPT-OSS-120B by about 1 point, and leaves the 120B configuration at or below Search-R1 on 2Wiki and NQ? Were prompts, tool schemas, and turn budgets adapted separately for each backbone, and were the large-model Search-R1 numbers reproduced by you under identical retrievers and settings? Turn counts, retrieval-call volume, and token usage would show whether the 120B setup simply failed to use its assigned budget, and those statistics are also needed for a fair comparison against cost-aware FrugalRAG.

  2. Please describe the renderer concretely. How are document summaries generated? What rule decides when to read full content and when to use a short snippet? What are the exact truncation threshold and context window size, and where does the ranking score come from? Relatedly, please define the no-structured-state variant in Table 4, since it delivers the largest performance change in the paper and its magnitude depends on whether it behaves like a full-history agent hitting context limits or an agent with no history management at all.

  3. Do answer-level results exist for any portion of the test data? Retrieval recall is a reasonable proxy for evidence coverage, but the failures the paper discusses concern termination timing and unsupported inference, so it would be valuable to know whether the improvements hold once final-answer quality is measured. Relatedly, privileged training includes information about safe termination while inference relies only on the judge gate, so on the hardest MuSiQue items, what fraction of errors come from premature finish and what fraction from missing evidence?

I would raise my score if the authors address the points above

1 comment
AuthorsOpenReview ↗

Response to Reviewer KRwF

We sincerely thank all reviewers for their thoughtful and constructive feedback. Below we provide point-by-point responses to all comments.


Weaknesses & Detailed Review

W1 && Q1: Uneven harness gains across backbones and consistency of the experimental setup.

Thank you for raising this point. Across backbones, we used the same configurations and budget settings, including prompts and tool schemas, without backbone-specific tuning.

We carefully inspected the GPT-OSS-120B trajectories and identified instruction-following issues that affected its execution of the harness workflow. This evidence suggests that instruction following contributes to its smaller gain.

We would also like to clarify one numerical detail in Table 2. On 2Wiki, GPT-OSS-120B improves from 79.8% with Search-R1 to 80.9% with H-Search. On NQ, it decreases from 82.0% to 81.1%. H-Search therefore improves this backbone on four of the five benchmarks.

To address the concern about comparison with cost-aware FrugalRAG, we additionally evaluated all four baselines on TriviaQA with a common maximum of 20 rounds.

MethodTriviaQA recallMaximum rounds
FrugalRAG73.8%20
Search-R160.3%20
CoRAG48.9%20
IRCoT67.1%20
H-Search-OPSD79.2%20

Under this common round limit, H-Search-OPSD retains a 5.4-percentage-point advantage over the strongest baseline, FrugalRAG. This comparison controls the maximum number of rounds.


W2 && Q2: Renderer implementation and the structured-state ablation.

Appendix C.1 describes the state representation, and Appendix D provides the corresponding prompts. We clarify the implementation below.

Document summaries and reading. Summaries are query-conditioned excerpts extracted deterministically from retrieved text. The harness stores a 200-character excerpt after retrieval. read_document is exposed to the policy as a tool. The policy decides whether and when to invoke it based on the current question and available evidence, requesting additional document content when it considers the retrieved snippets insufficient. Full content is obtained when available, while snippet-only mode or failed full-text retrieval uses the available snippet. The harness uses a context-budget threshold of approximately 8,192 tokens.

No-structured-state ablation. While maintaining a constant total token budget for the input, the structured working memory representation is replaced by a chronological sequence of tool calls and observation records. History is retained until that budget is reached, after which the oldest complete interactions are removed while preserving the question and task instructions.

The model checkpoint, thinking setting, retriever, tools, and round limit will be held fixed. Backend state will remain available for tool execution, action constraints, and evaluation. The policy will receive its evidence history through the interaction transcript, including during sufficiency assessment.


Q3:

Our primary objective is to evaluate the search agent’s ability to recover supporting evidence. Document-level recall directly measures the fraction of gold supporting documents present in the final evidence set and therefore closely matches this objective.

Final-answer correctness also depends on the downstream reader and may benefit from parametric knowledge or a correct guess despite incomplete retrieved evidence. Direct evidence evaluation helps identify such gaps, including missing supporting documents in multi-hop questions. We also report evidence precision and F1 in Appendix B.5, Table 7, to assess the quality of the selected evidence alongside its coverage.

The same evidence annotations support our training design. By tracking which gold documents have been covered and which remain missing at each state, we can provide state-specific supervision for intermediate decisions and termination. This offers more localized feedback than a single final-answer correctness signal.

We apologize that, due to time constraints, we are currently unable to provide answer-level results. However, we believe that recall is positively correlated with answer accuracy to some extent and provides a more demanding measure, as achieving full recall requires retrieving all gold supporting documents.


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

The paper proposes H-Search for agentic search. A stateful harness keeps all cross-turn state outside the model, re-renders a self-contained context at each step, and restricts the admissible tools by control phase. The policy is trained with on-policy self-distillation (OPSD): a frozen teacher that sees privileged information derived from gold documents supervises the student. With LoRA training on 5k examples from the five evaluation benchmarks, Qwen3.6-27B reaches 79.8% average document recall, 11.2 points above the strongest same-backbone baseline.

Critical strengths
  1. The paper makes a useful point: the harness is a confounding factor in agentic search and should be treated as part of the decision process.
  2. Appendices B.3–B.4 attempt to separate harness and training effects with a shared backbone.
  3. The method is cheap (LoRA, 5k examples), and the prompts and code are provided.
Critical weaknesses

W1. The main comparisons are not controlled.

  • Training distribution. H-Search is trained on all five evaluation benchmarks. According to their original papers, the trained baselines cover only part of them: Search-R1 is trained on NQ + HotpotQA, and CoRAG and FrugalRAG on the multi-hop datasets. The paper does not say whether the baselines were retrained or released checkpoints were used. Several benchmarks are therefore out-of-distribution for the baselines but in-distribution for H-Search.
  • Retrieval setup. The FrugalRAG (Qwen2.5-7B) results on HotpotQA / 2Wiki / MuSiQue (0.704 / 0.588 / 0.533) are identical to those reported in the FrugalRAG paper with an E5-base-v2 retriever. This conflicts with the claim that all methods share the same retriever.
  • Inference budget. H-Search uses thinking and up to 20 rounds, while the baselines' budgets are not reported. The same untrained model in the same harness drops from 76.3% to about 65.5% under the Lite setting (no thinking, 8 rounds). This is close to the FrugalRAG workflow on the same backbone (64.7%), so much of the reported gain may come from the larger budget.
  • Larger models. The frontier models in Table 2 appear to be prompted rather than trained. In Fig. 4, Qwen2.5-32B differs from Qwen3.6-27B by a model generation, so the claimed "gap reduction" mixes backbone, harness and training effects.

W2. Training contributes little, yet the paper is framed around training. On the same backbone, the harness alone raises average recall from 62.3–64.7% to 76.3% (Table 5). All training together adds only 3.5 points (Table 3), and OPSD beats GRPO by only about 2.2 points, with no variance reported. The reward used for GRPO is not specified. The title and contributions emphasize the training method, but the evidence mainly supports a harness-design contribution.

Questions for authors
  1. Were the baselines retrained or used as released checkpoints? What training data and retriever were used for each, including FrugalRAG-27B?
  2. Were the FrugalRAG-7B results re-run under your retrieval setup?
  3. What are the inference budgets (turns, tool calls, tokens) of all methods, and how do the methods compare at matched budgets?
  4. Can you report answer accuracy, and results for a baseline (e.g., Search-R1-style RL) trained on the same 5k data with Qwen3.6-27B?
1 comment
AuthorsOpenReview ↗

Response to Reviewer HxRM

We sincerely thank all reviewers for their thoughtful and constructive feedback. Below we provide point-by-point responses to all comments.


Weaknesses & Detailed Review

W1 && Q1,2: Baseline checkpoints, training data, and retrieval settings

Thank you for requesting clarification on baseline provenance, training data, and retrieval settings. For Search-R1 and CoRAG, some scores come from the FrugalRAG paper, while the remainder were obtained by evaluating author-released checkpoints under their original settings. Search-R1 was trained on NQ and HotpotQA, and CoRAG on HotpotQA, 2WikiMultiHopQA, and MuSiQue. H-Search used 5,000 examples, sampled equally from the five benchmark training splits, so training coverage differs across methods. For FrugalRAG, the three multi-hop scores come from Table 2 of the original paper using E5-base-v2. For NQ and TriviaQA, we reproduced training using the official repository’s sample IDs and configurations and evaluated under the original retrieval setup. We will revise §4.1, Appendix A, and Table 1 notes to specify each result’s source, training data, and retrieval settings.

To examine the effect of the search workflow, Appendix Table 5 runs H-Search, FrugalRAG, and CoRAG with the same Qwen3.6-27B checkpoint, without task-specific fine-tuning. This comparison holds the model weights and task-specific training status fixed across all three workflows.


W1 && Q3: Inference budgets and performance under a common round limit.

Thank you for highlighting the need to report inference budgets. We provide additional results on TriviaQA below. The first table reports recall, the maximum round limit, and mean token consumption under the baseline configurations used in our experiments.

MethodRecallMax. roundsMean tokens/query
FrugalRAG72.5%53,337
Search-R159.7%104,504
CoRAG45.4%65,106
IRCoT65.2%107,721
H-Search-Lite73.9%85,907
H-Search-OPSD79.2%2013,697

H-Search-Lite achieves 73.9% recall with thinking disabled and an eight-round limit. It exceeds the strongest baseline in this table, FrugalRAG, by 1.4 percentage points. The full H-Search-OPSD configuration reaches 79.2% recall with a larger inference budget.

To examine whether the different round limits account for the performance gap, we reran all four baselines with their maximum round limit increased to 20.

MethodRecallMax. rounds
FrugalRAG73.8%20
Search-R160.3%20
CoRAG48.9%20
IRCoT67.1%20
H-Search-OPSD79.2%20

All four baselines improve with the increased round limit. H-Search-OPSD retains a 5.4-percentage-point advantage over the strongest baseline, FrugalRAG. On TriviaQA, this result shows that the advantage persists when all methods are allowed up to 20 rounds. We will include these results and the explicit budget settings in the revised paper.


W1 && Q4: Answer accuracy and a same-data Search-R1-style RL baseline.

We trained a Search-R1-style RL baseline using Qwen3.6-27B and the same 5,000 training examples as H-Search. The baseline achieved 45.7% document recall on HotpotQA.

MethodBackboneTraining examplesHotpotQA document recall
Search-R1Qwen3.6-27B5,00045.7%
H-Search-OPSDQwen3.6-27B5,00083.9%

It should be noted that the official training set released for Search-R1 contains 169,615 samples. The smaller size of the training set used here may be one of the reasons for the baseline model's performance.


W2

Thank you for raising this concern. We would like to clarify that the title and contributions are intended to emphasize the joint design of the harness and its associated training method. Our central research question is how harness design and model adaptation together improve agentic search. Accordingly, the contribution should be assessed through the effectiveness of the harness, the additional gains from training, and the computational cost of achieving those gains.

The results highlight the respective contributions of these components. Table 5 shows that harness design provides a substantial improvement on the same backbone, increasing average recall from 62.3–64.7% to 76.3%. Table 3 further shows a 3.5 percentage-point improvement from training. We agree that the harness accounts for the larger performance gain.

The OPSD–GRPO comparison should also account for training cost. GRPO requires substantially more training computation, while OPSD achieves approximately 2.2 percentage points higher average recall at a lower training cost.


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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.