Pith. sign in

REVIEW 3 major objections 5 minor 13 references

CRISP: Critical Step Perception for Training Efficient Deep Search Agents

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper argues that the right measure of search-agent efficiency is not fewer tool calls but a higher proportion of evidence-critical tool calls, and it trains a recognizer to reward that proportion, cutting average interaction turns…

desk verdict A well-motivated new step-credit mechanism for efficient search agents; the core results are plausible but need error bars and on-policy recognizer validation before I'd call them established. read the letter →

arxiv 2608.01867 v2 pith:4WDN3CQY submitted 2026-08-03 cs.CL

Haosi Mo , Zihao Yan , Ruiqing Zhang , Zhongli Li , Hexuan Deng , Xuebo Liu , Min Zhang This is my paper · ORCID
classification cs.CL
keywords criticalstepperceptiondeepsearchagentsefficiency-awarereinforcementlearningbackwardevidenceinductiontool-useefficiencycritical-steprecognizerrewardshapingBrowseComp
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

CRISP is a training framework for deep search agents—LLM agents that answer hard questions through repeated web search and browsing. The paper's central claim is that efficiency should be pursued by preserving tool interactions that supply evidence for the final answer and pruning only the redundant ones, rather than by uniformly penalizing tool use. To do this, CRISP labels each step of a completed search trajectory as evidence-critical or not by tracing backward from the final answer, distills those labels into a smaller recognizer that can judge a full trajectory in one pass, and uses the recognizer to add an efficiency-aware reward during policy optimization. On BrowseComp and HLE-Verified, the method keeps final-answer accuracy competitive while cutting average interaction turns by 15.1% and 33.2% relative to a vanilla reinforcement-learning baseline. A fair reader would care because the result points to a way of reducing search cost and failure rates without forcing agents to act less.

What carries the argument

The central mechanism is the critical-step recognizer plus the efficiency-aware reward it enables. Backward Evidence Induction gives stable per-step labels by decomposing the annotation into local decisions: each step is judged against the set of future steps already confirmed as evidence-critical, with the teacher requiring both evidence value (direct facts, data, or a pointer that enables later access) and marginal non-redundancy. Distillation compresses this $O(n)$ backward procedure into a single-pass student. In policy optimization, the recognizer predicts $\hat{c}_{i,j} \in \{0,1\}$ for each tool-interaction step of a successful rollout, and the reward $$R_i = R^\mathrm{ans}_i + \$\lambda$ R^\mathrm{ans}_i \frac{K_i}{K_i + \$\alpha$ T_i}$$—where $K_i$ is the number of predicted critical steps and $T_i$ the redundant ones—shapes behavior toward higher evidence density rather than shorter trajectories. The whole framework rests on this reward being computed from evidence-level judgments rather than from turn counts.

What would settle it

Run training with periodic recognition evaluation: at checkpoints, sample rollouts from the current policy, have the recognizer label them, and compare the labels against gold evidence document IDs, as the paper does held-out in Appendix A.2. If label quality degrades as the policy shifts, then the criticality signal used in Eq. (6) is stale, and any efficiency gain would no longer be explained by preserving evidence.

Watch

Extended reading notes

Core claim

The discovery is that critical-step perception—the ability to tell which tool calls form a minimal evidence chain supporting the final answer—can be learned and used as a training signal. The paper establishes this by constructing critical-step labels with Backward Evidence Induction, where a strong teacher model evaluates each trajectory step from last to first, conditioned on already-confirmed valid future steps, and marks a step critical only if it provides or preserves evidence not already carried by later steps. It then distills these per-step judgments into a smaller student recognizer that reads a full trajectory once and outputs the critical-step list. During GRPO training, the recognizer is applied only to successful rollouts, and the reward augments answer correctness with an efficiency term $R^\mathrm{crit}_i = K_i/(K_i+\alpha T_i)$ that rewards a high proportion of critical tool-interaction steps. The evidence that the partition is meaningful comes from a trajectory intervention: removing predicted critical steps costs little accuracy, while removing predicted redundant steps costs 8 accuracy points and forces extra turns to recover.

Load-bearing premise

The load-bearing premise is that the critical-step recognizer, trained on offline teacher labels, continues to identify evidence-carrying steps accurately on the new trajectories sampled by the policy as it changes during GRPO training; if the policy distribution shifts, the reward's criticality judgments can become stale and the efficiency gains would not reflect genuine evidence preservation.

Editorial extensions

If this is right

  • On BrowseComp, CRISP reduces repeated search calls while increasing browsing calls, so the efficiency gain comes from reallocating interaction toward evidence-bearing documents rather than from blanket tool suppression.
  • On HLE-Verified, search, browsing, and code calls all drop, showing the same reward can prune redundancy across heterogeneous tool use when trajectories are shorter.
  • Because the efficiency reward is applied only to correct rollouts, the policy is not rewarded for short but wrong answers, and the failure rate falls by roughly 21% on both benchmarks.
  • The evidence-density analysis shows CRISP raises the fraction of critical steps among correct trajectories from 34.06% to 41.09% on BrowseComp and from 42.06% to 53.09% on HLE-Verified, which is the direct mechanism behind the turn reductions.
  • After the efficiency reward is removed, the CRISP-trained policy keeps higher policy entropy, comparable accuracy, and lower average turns than vanilla RL, indicating the learned selectivity persists rather than being a temporary response to the auxiliary reward.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to move the recognizer from an offline training reward into the decoding loop, stopping or redirecting a rollout when predicted critical-step density is low; the paper's recognizer already runs in a single pass, so this is a latency-feasible next step the authors do not explore.
  • The criticality definition is anchored to a known final answer, so CRISP as presented applies where success is verifiable; extending it to open-ended research, where the final synthesis is not known in advance, would require an iterative or self-consistent evidence-chain construction that the framework itself does not provide.
  • Because the reward is gated on correctness and the recognizer is trained on gold-document evidence, the same pipeline should transfer to other tool-augmented settings with verifiable outcomes, such as code execution or structured database queries, whenever a success signal and a way to define evidence-bearing steps exists.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes CRISP, a three-stage framework for training deep search agents: Backward Evidence Induction (BEI) annotates tool-interaction steps in completed trajectories as evidence-critical or redundant using a strong teacher model; Critical-Step Recognizer Distillation trains a smaller generative model to reproduce these backward judgments in a single pass; and Efficiency-Aware Policy Optimization uses the recognizer's predictions on successful GRPO rollouts to construct an auxiliary reward (Eqs. 6-7) that favors trajectories with a higher proportion of critical steps. Experiments on BrowseComp and HLE-Verified report that CRISP reduces average interaction turns by 15.1% and 33.2% relative to Vanilla RL while maintaining comparable or slightly better final-answer accuracy. Supporting analyses include a trajectory intervention (Table 3), evidence-density comparisons, tool-use behavior changes, failure diagnostics, and a sensitivity study of the redundancy penalty alpha.

Significance. If the reported results hold, CRISP is a practically meaningful advance for deep search agents, which often consume 30+ interaction turns per query. The paper has notable strengths: the critical-step recognizer is validated against gold evidence document IDs from BrowseComp-Plus (Appendix A.2), the trajectory intervention provides functional evidence that predicted critical steps carry answer-relevant information, the alpha sensitivity analysis is honest about the accuracy-efficiency trade-off, and training details are sufficiently concrete for replication. However, two gaps weaken the central mechanism claim: the recognizer is not validated on the on-policy rollouts it is used to reward, and the main results lack error bars or significance tests. These are fixable within a revision, and the efficiency gains are large enough that the contribution remains valuable even before the mechanism is fully pinned down.

major comments (3)
  1. [§4, Table 2] The paper states that all results are averaged over three independent runs but reports no standard deviations, confidence intervals, or significance tests. On BrowseComp, CRISP's accuracy (35.69) is numerically below Vanilla RL (36.22) by 0.53 points, while on HLE-Verified it is 0.39 points higher. These differences are small enough that the claim of "maintains competitive accuracy" is not statistically supported without variance information. Please report per-run values or standard deviations and, where possible, paired significance tests over the three runs.
  2. [Appendix A.2 and §3, Eqs. (4)-(7)] The critical-step recognizer is validated only on held-out BrowseComp-Plus trajectories generated by GLM-4.7 (Tables 5-6), a different model and benchmark from the rollouts it is used to reward. The efficiency-aware reward in Eq. (7) is applied to successful rollouts of the evolving CRISP policy on both BrowseComp and HLE-Verified, where no gold evidence labels exist. If recognizer predictions degrade under policy distribution shift, the reward would be based on stale or biased criticality judgments, and the claimed mechanism—preserving evidence-critical steps while pruning redundant ones—would not be established. Please add an on-policy validation: sample successful rollouts from the trained CRISP policy (ideally at several training checkpoints), have the teacher or gold evidence labels annotate them, and report recognizer accuracy/agreement on these trajectories. For HLE-Verified, teacher labels can serve as the reference since gold evidence document IDs are unavailable.
  3. [§4, Evidence Density and Table 3] The evidence-density analysis computes the critical-step ratio with the same recognizer that produced the reward, so the observed increase in this ratio is partly an artifact of directly optimizing that ratio. The trajectory intervention in Table 3 deliberately uses Vanilla RL trajectories, which is a good functional test of the recognizer's labels but does not test whether CRISP's own pruned trajectories preserve gold evidence. A direct test would measure gold evidence document coverage on BCP-style rollouts from the CRISP policy, or use teacher-based criticality labels on those rollouts, rather than relying on the recognizer's own predictions. As written, the central claim that CRISP improves efficiency by preserving evidence-critical steps is supported only indirectly.
minor comments (5)
  1. [§3, Eqs. (5)-(6)] The same index i is used for the rollout index and the step index in the definitions of K_i and T_i; please disambiguate the notation (e.g., use k for steps).
  2. [§4, Evaluation] The definition of "interaction turns" in Table 2 should be restated in the evaluation setup: a turn corresponds to a tool-call action (a_i), not a thought step. This is clear from Section 3 but should be explicit in the experiments for readers who skip the formalization.
  3. [Table 2] The HLE "Official Acc." column reports original HLE scores rather than HLE-Verified scores; the note explains this, but a footnote in the table itself would help avoid misreading.
  4. [Appendix A.1, Prompt 1] The "Strict Exclusions" mark all mine-sweeping or negative-result steps as non-critical. This is a reasonable design choice but is also a substantive assumption about what counts as evidence-critical behavior; please acknowledge in the main text that steps ruling out alternatives are intentionally excluded from the evidence-chain notion.
  5. [Appendix A.2] The paper does not state whether the trained recognizer and policies will be released; releasing them would substantially improve reproducibility, given that the backbone (GLM-4.5-Air-Midtrain) and benchmark subsets are not publicly available.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: headline efficiency gains are externally measured; one supporting analysis reuses the recognizer that defines the training reward, creating a mild internal loop.

  1. fitted input called prediction [Section 4, Evidence Density (after Table 3), compared with Eq. (6) in Section 3.3]
    "Since critical steps are defined with respect to the evidence supporting the final answer, we analyze correctly answered evaluation trajectories and compute the proportion of tool-interaction steps predicted as critical by the recognizer. Compared with Vanilla RL, CRISP increases the average critical-step ratio from 34.06% to 41.09% on BrowseComp and from 42.06% to 53.09% on HLE-Verified, corresponding to relative improvements of 20.6% and 26.2%, respectively. These consistent gains indicate that CRISP produces trajectories with higher evidence density."

    The evidence-density metric is the recognizer-predicted critical-step ratio, and the same recognizer's predictions define the efficiency reward in Eq. (6): R^crit_i = K_i / (K_i + alpha * T_i), where K_i and T_i are the recognizer's predicted counts from Eqs. (4)-(5). CRISP's policy is trained by GRPO to increase that reward on successful rollouts, so an increase in the recognizer-predicted critical ratio is largely the training objective being realized rather than an independent confirmation that the preserved steps carry evidence. This is a mild internal loop in the mechanism analysis, but it is not the source of the headline turn-reduction or accuracy numbers, which are measured externally against benchmark judges and are independent of the recognizer's predictions.

full rationale

The central claims of the paper—reduced average interaction turns (15.1% on BrowseComp and 33.2% on HLE-Verified) while maintaining competitive accuracy—are measured externally against benchmark judges, not against the critical-step recognizer's outputs, so they do not reduce to a fitted quantity. The recognizer itself is validated against gold evidence document IDs from BrowseComp-Plus in Appendix A.2, providing an external check on the label quality rather than relying on a self-citation chain or a uniqueness theorem. The only notable internal loop is the Evidence Density analysis, which uses the recognizer's own predictions as both the training reward target and as evidence of the mechanism; this supports but does not establish the mechanism independently. The trajectory-intervention experiment on Vanilla RL trajectories provides some independent functional validation, and no load-bearing self-citation was found. Overall, the derivation is substantially self-contained, and the circularity is minor rather than foundational.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method rests on a chain of modeling assumptions: final-answer-conditioned backward labeling is the right ground truth for evidence, a 9B student can absorb it, and the recognizer stays valid on-policy. These are domain assumptions rather than mathematical axioms. The only hand-fitted numbers that directly shape the reward are alpha and lambda, and alpha was tuned on a main evaluation benchmark.

free parameters (2)
  • redundancy penalty alpha = 0.7
    Penalty strength for redundant steps in Rcrit = K / (K + alpha * T); chosen via sensitivity analysis on HLE-Verified where it gave the best accuracy.
  • efficiency reward weight lambda = 0.1
    Weight of efficiency-aware reward in total reward Ri = Rans + lambda * Rans * Rcrit; set by hand to keep correctness primary, with no sensitivity analysis reported.
assumptions (4)
  • domain assumption A tool step is evidence-critical iff it provides or preserves evidence for the final answer, judged backward against already-confirmed future steps.
    Stage 1 (Backward Evidence Induction) defines criticality purely with respect to the final answer; if the final answer is wrong, the labels are not meaningful for evidence.
  • domain assumption A 9B student can reproduce the teacher's backward judgments in a single pass and generalize to trajectories beyond the offline set.
    Section 3.2 and Appendix A.2; the held-out evaluation only covers correct BrowseComp-Plus trajectories with gold document IDs.
  • domain assumption The recognizer remains accurate on on-policy rollouts sampled from the policy being optimized.
    Section 3.3 applies the recognizer to successful rollouts; no on-policy validation of label quality is reported.
  • domain assumption Efficiency reward should be applied only to successful rollouts.
    Section 3.3 gates Rcrit by Rans; assumes this does not suppress useful exploration on hard questions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CRISP: Critical Step Perception for Training Efficient Deep Search Agents." pith.science (2026). https://pith.science/paper/4WDN3CQY

@misc{pith2026260801867,
  author       = {Pith},
  title        = {Pith review of: CRISP: Critical Step Perception for Training Efficient Deep Search Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4WDN3CQY}},
  note         = {Machine review of arXiv:2608.01867}
}
read the original abstract

Large language models (LLMs) are increasingly extended into deep search agents that solve complex questions through multi-step interaction with external search and browsing tools. However, existing agents often incur substantial computational and interaction costs, generating lengthy trajectories that contain redundant queries, inefficient exploration, and irrelevant observations. Existing efficiency-oriented methods usually encourage agents to use tools less frequently, but treating all tool interactions uniformly may also suppress steps that gather necessary evidence. In this paper, we propose CRISP, a framework for training efficient deep search agents through critical step perception. Unlike prior efficiency methods that uniformly penalize tool use, CRISP distinguishes interactions that gather necessary evidence from redundant ones and shapes the training reward to preserve the former while pruning the latter, improving efficiency without sacrificing the evidence needed for correct answers. Specifically, CRISP first constructs critical-step labels with Backward Evidence Induction: starting from the final answer, a strong model traverses a completed search trajectory backward and judges whether each tool-interaction step provides or preserves evidence for the final answer. We then distill these step-wise judgments into a smaller critical-step recognizer, enabling full-trajectory analysis in a single pass. During policy optimization, an efficiency-aware reward is applied only to successful rollouts. Experiments on BrowseComp and HLE-Verified show that CRISP maintains competitive final-answer accuracy while reducing average interaction turns by 15.1% and 33.2%, respectively, demonstrating substantial improvements in interaction efficiency.

Figures

Figures reproduced from arXiv: 2608.01867 by the authors.

Figure 1
Figure 1. Overview of CRISP. CRISP introduces a critical-step recognizer identifying critical steps in policy rollouts. The recognized critical-step structure is combined with the outcome reward to perform efficiency-aware policy optimization, encouraging the policy to preserve necessary evidence while reducing redundant search behavior. Perception), a framework for training deep search agents to identify evidence-critical st… view at source ↗
Figure 2
Figure 2. Detailed procedure of CRISP. Left: Backward Evidence Induction obtains step-wise criticality annotations by traversing trajectories backward from the final response. Middle: Critical-Step Recognizer Distillation formats these annotations into full-trajectory supervision and trains a student recognizer for single-pass critical-step analysis. Right: Efficiency-Aware Policy Optimization uses the recognizer to estimate … view at source ↗
Figure 3
Figure 3. Composition of diagnostic failure modes on [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Teacher critic prompt (Stage 1). The teacher evaluates one candidate step at a time, conditioned on the confirmed valid [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Student recognizer prompt (Stage 2). The distilled recognizer reads the entire trajectory at once and reproduces the [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: The shared trajectory step schema used by both prompts. The teacher’s confirmed valid future steps and the student’s [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 9 canonical work pages

  1. [3]

    Wu, F.; Xuan, W.; Qi, H.; Lu, X.; Tu, A.; Li, L

    Browsecomp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv:2504.12516. Wu, F.; Xuan, W.; Qi, H.; Lu, X.; Tu, A.; Li, L. E.; and Choi, Y. 2025a. DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search.arXiv preprint arXiv:2509.25454. Wu,X.;Li,K.;Zhao,Y.;Zhang,L.;Ou,L.;Yi...

    arXiv 2025

  2. [4]

    Current Step

    Deepresearcher: Scaling deep research via reinforce- ment learning in real-world environments. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 414–431. ZhipuAI.2025. GLM-4.7:AdvancingtheCodingCapability. Zhu, K.; Li, H.; Wu, S.; Xing, T.; Ma, D.; Tang, X.; Liu, M.; Yang, J.; Liu, J.; Jiang, Y. E.; et al. 2025. Sca...

    arXiv 2025

  3. [5]

    [Condition A]: Evidence Value The step must satisfy at least ONE of the following:

    Confirmed Valid Future Steps (already extracted critical steps, back to front): {Valid_Future_Steps_String} # see Figure 3 for the step format [Evaluation Criteria] To be marked as critical (`true`), the Current Step MUST satisfy BOTH [Condition A] AND [Condition B]. [Condition A]: Evidence Value The step must satisfy at least ONE of the following:

    work page

  4. [7]

    Current Step to Evaluate (Step {Current_Step_Num}): Thought: {Current_Thought} Action: {Current_Action} Observation: {Current_Observation}

    work page

  5. [8]

    Thought From Next Step (Step {Next_Step_Num}): {Next_Thought}

    work page

  6. [11]

    brief_reasoning

    Evidence-Enabling Contribution: Its`Observation`provides a document identifier, URL, entity anchor, or other specific pointer that directly enables later steps to access, open, or identify necessary evidence. Mere help for search planning, query writing, or reasoning is NOT sufficient. When judging whether the Current Step provides useful information, you...

    work page

  7. [12]

    Target Question: {Question}

    work page

  8. [13]

    Final Answer (Step {Last_Step_Num}): {Final_Answer}

    work page

Show all 13 references
  1. [14]

    [Evaluation Criteria] A step is`critical`if it satisfies BOTH [Condition A] and [Condition B]

    Full Trajectory: {Full_Trajectory} # see Figure 3 for the step format [Goal] Identify the minimal set of trajectory steps that are necessary to preserve the evidence used to support the Final Answer. [Evaluation Criteria] A step is`critical`if it satisfies BOTH [Condition A] a...

  2. [15]

    Direct Evidence Contribution: Its`Observation`directly contains facts, data, document content, or document IDs that are necessary for supporting the Final Answer

  3. [16]

    smaller-is-always-shorter

    Evidence-Enabling Contribution: Its`Observation`provides a document identifier, URL, entity anchor, or other specific pointer that directly enables later steps to access, open, or identify necessary evidence. Mere help for search planning, query writing, or reasoning is NOT su...

  4. [2025]

    Chen, D.; Zong, Z.; Ma, Z.; Luo, L.; Li, Y.; Li, C.; Chen, P.; andJiang,J.2026a.EfficientAgenticReinforcementLearning with On-Policy Intrinsic Knowledge Boundary Enhancement

    gpt-oss-120b&gpt-oss-20bmodelcard.arXivpreprint arXiv:2508.10925. Chen, D.; Zong, Z.; Ma, Z.; Luo, L.; Li, Y.; Li, C.; Chen, P.; andJiang,J.2026a.EfficientAgenticReinforcementLearning with On-Policy Intrinsic Knowledge Boundary Enhancement. arXiv preprint arXiv:2605.26952. Che...

    2025 arXiv

  5. [2026]

    Huang, L.; Liu, Y.; Jiang, J.; Zhang, R.; Yan, J.; Li, J.; and Zhao, W

    DeepSearchQA:BridgingtheComprehensivenessGap for Deep Research Agents.arXiv preprint arXiv:2601.20975. Huang, L.; Liu, Y.; Jiang, J.; Zhang, R.; Yan, J.; Li, J.; and Zhao, W. X. 2025a. ManuSearch: Democratizing Deep Search in Large Language Models with a Transparent and Open M...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.