CRISP: Critical Step Perception for Training Efficient Deep Search Agents

Haosi Mo; Hexuan Deng; Min Zhang; Ruiqing Zhang; Xuebo Liu; Zhongli Li; Zihao Yan

Deep search agents can cut redundant tool calls and hold accuracy by learning to recognize which steps actually carry evidence for the final answer.

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 ↗

CRISP reduces deep-search agent tool turns by teaching the agent to keep evidence-critical steps and prune redundant ones, cutting turns 15-33% with no accuracy loss.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Solid new variant in reward shaping for deep search agents, grounded on BCP but with an unproven transfer to HLE-Verified. the 5 major comments →

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

CRISP: Critical Step Perception for Training Efficient Deep Search Agents

Haosi Mo , Zihao Yan , Ruiqing Zhang , Zhongli Li , Hexuan Deng , Xuebo Liu , Min Zhang This is my paper

classification cs.CL

keywords deep search agentscritical step perceptionBackward Evidence Inductionefficiency-aware rewardreinforcement learningtool-use efficiencyevidence-critical stepsinteraction turns

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.

The reading

The paper claims that the high interaction cost of deep search agents comes not from trajectory length itself but from a mix of evidence-gathering steps and redundant ones, and that these can be told apart. It introduces CRISP, which uses Backward Evidence Induction—starting from the final answer, a strong model judges each tool-interaction step for whether it provides or preserves evidence—then distills those judgments into a small recognizer that labels a whole trajectory in one pass. This recognizer feeds an efficiency-aware reward, applied only to successful rollouts, that favors trajectories with a higher proportion of evidence-critical steps. On two deep-search benchmarks, the method reduces average interaction turns by 15.1% and 33.2% while keeping final-answer accuracy essentially unchanged. A sympathetic reader would care because it points to a way to make agents cheaper and faster without sacrificing the evidence needed for correct answers.

Core claim

The central discovery is that uniform penalties on tool use—the standard efficiency lever—suppress steps that gather necessary evidence, so a better training signal is one that distinguishes evidence-critical from redundant interactions. CRISP operationalizes this by having a strong teacher trace each completed trajectory backward from the final answer, marking a step critical only if it directly contributes evidence or enables a later confirmed step to access evidence, while enforcing non-redundancy against already-confirmed future steps. A distilled student recognizer then reproduces this backward judgment for a full trajectory in a single pass, and an efficiency-aware reward Rcrit = Ki/(K

What carries the argument

Backward Evidence Induction: the teacher model evaluates each tool-interaction step from the last to the first, conditioned on the already-confirmed valid future steps, labeling a step critical if it provides or preserves evidence for the final answer or directly enables a later confirmed step; the per-step judgments are aggregated into a formatted target and distilled into a smaller generative critical-step recognizer. The efficiency-aware reward Rcrit = Ki/(Ki+αTi), where Ki is the number of predicted critical steps, Ti the redundant ones, and α the penalty strength, is applied only to successful rollouts and combined with the answer reward as Ri = Rans_i + λ·Rans_i·Rcrit_i. The recognizer

Load-bearing premise

The teacher's backward judgments and the distilled recognizer's single-pass predictions are causally correct and transfer across benchmarks—the evidence for this is direct only on the web-evidence benchmark with gold document IDs, while the text-only benchmark's gains rely on the same notion of 'evidence-critical' applying to reasoning and code-execution steps.

What would settle it

Run the paper's trajectory-intervention test on rollouts sampled from the trained CRISP policy itself, not just Vanilla RL: if deleting the predicted critical steps preserves the final answer as often as the full context while deleting predicted redundant steps drops accuracy sharply, the critical-step labels are causally meaningful. If that asymmetry disappears on CRISP rollouts, the recognizer has co-adapted to the reward rather than identifying evidence.

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

If this is right

If CRISP is correct, efficiency gains in deep search do not have to come at the cost of accuracy; distinguishing evidence-critical steps from redundant ones yields a better accuracy–efficiency trade-off than uniform call penalties.
The training pipeline—backward annotation, distillation into a single-pass recognizer, and reward shaping—can transfer across benchmarks: the paper reports gains on both a web-evidence benchmark and a text-only reasoning benchmark with code tool calls.
Trained policies exhibit higher evidence density (critical-step ratio up from 34% to 41% on BrowseComp and from 42% to 53% on HLE-Verified, on correct trajectories), including a shift from broad repeated search to focused browsing on web-evidence tasks.
Lower failure rates (around 22% relative reduction) follow from pruning redundant interactions, because fewer turns reduce context overflow and tool-formatting error opportunities.
Removing the efficiency reward after training leaves the shorter-trajectory behavior largely intact, suggesting the policy change is structural, not a transient response to the auxiliary signal.

Where Pith is reading between the lines

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

The distilled recognizer is a reusable artifact: it could be applied offline to analyze any agent's rollouts, potentially serving as a cheap process-supervision signal for other long-horizon tasks beyond the two benchmarks tested.
If critical steps are truly causal, then trajectory intervention on the trained CRISP policy itself (not just Vanilla RL rollouts) should show the same asymmetry; a failure there would indicate recognizer-policy co-adaptation. This is a testable extension of the paper's own validation.
The backward-looking, marginal-evidence criterion may also be useful for interpreting agent behavior in human-auditable terms, since it yields a minimal evidence chain that can be shown to users alongside the final answer.
One open direction the paper does not pursue: using the recognizer's step-level labels as dense reward for intermediate steps rather than only as a trajectory-level efficiency bonus, which could sharpen credit assignment in longer horizons.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Solid new variant in reward shaping for deep search agents, grounded on BCP but with an unproven transfer to HLE-Verified.

read the letter

Core idea is well positioned and the experiments mostly support it, but the headline efficiency gain on HLE-Verified is softer than it looks because the recognizer is never independently validated there.

What's actually new: the evidential notion of criticality. Prior efficiency methods penalize tool use or define criticality by structural/planning role; CRISP labels steps by whether they provide or preserve evidence for the final answer, and it turns that into a reward without uniformly suppressing tool calls. The backward evidence induction is a clever decomposition that makes annotation manageable on long noisy trajectories, and distilling it into a single-pass student recognizer is a practical move. The recognizer evaluation in Appendix A.2 is the strongest part of the paper: on BrowseComp-Plus, teacher backward beats teacher direct on evidence recall and coverage, and the student matches the teacher on a held-out split, all against gold evidence document IDs. That gives the web-search version of the mechanism real grounding. The trajectory intervention on 100 Vanilla-RL trajectories also points in the right direction: critical-only preserves accuracy much better than redundant-only, which is not what you'd expect from a pure length penalty.

Soft spots, in rough order of importance. First, no error bars or significance tests anywhere on the headline turn reductions, although they say results are averaged over three runs. That's a small fix but matters. Second, alpha=0.7 is selected on HLE-Verified via the Appendix A.3 sensitivity run, and HLE-Verified is also one of the two benchmarks used for the main claims. That's at least tuned-on-test selection; a validation split would make the claim cleaner. Third, the bigger substantive concern: criticality on HLE-Verified is never checked. The recognizer is validated against gold document IDs on BCP, but HLE-Verified includes text-only reasoning and code-execution steps, and the 33.2% turn reduction rests entirely on the assumption that 'evidence-critical' transfers. The evidence density analysis in Section 4 uses the same recognizer that shapes the reward, so it is partly circular. The stress-test worry that Rcrit degenerates to a smooth length penalty on successful rollouts is not fully refuted. Having said that, the intervention study and the non-monotonic alpha sensitivity make me think the mechanism is real but the transfer is unproven, not disproven.

Who it's for: anyone training search agents or doing reward design for long-horizon tool use. It's a solid new variant rather than a fundamental breakthrough. It deserves a serious referee; I'd send it with a request for variance reporting, a validation split for alpha, and at least a small independent check of criticality labels on HLE-Verified trajectories.

Referee Report

5 major / 6 minor

Summary. The paper proposes CRISP, a framework for making deep search agents more interaction-efficient by learning to identify evidence-critical steps. It uses Backward Evidence Induction, where a strong teacher labels each tool-interaction step in a completed trajectory by working backward from the final answer; these labels are distilled into a smaller single-pass recognizer; and the recognizer's predictions are used in an auxiliary reward (Eq. 6–7) during GRPO training. Experiments on BrowseComp and HLE-Verified report average turn reductions of 15.1% and 33.2% over Vanilla RL with comparable or slightly better accuracy. Additional analyses include a held-out evaluation of the recognizer on BrowseComp-Plus, a trajectory intervention study, a sensitivity analysis of the redundancy penalty α, and a reward-removal check.

Significance. If the empirical claims hold, CRISP offers a genuinely different efficiency signal: instead of uniformly penalizing tool calls, it attempts to preserve steps that carry or enable evidence. The paper has real strengths: the recognizer is evaluated against gold evidence document IDs on BCP (Appendix A.2), the trajectory intervention shows an asymmetric effect between critical and redundant steps, the α-sensitivity analysis is behaviorally informative, and the reward-removal experiment addresses policy-collapse concerns. These design choices are more rigorous than typical efficiency-optimization work. However, the central mechanism is only directly validated on BCP; the headline HLE-Verified efficiency gain relies on the untested assumption that the document-ID-based notion of criticality transfers to text-only reasoning and code-execution steps. The evidence-density analysis is partly circular, and the absence of uncertainty quantification makes it difficult to assess the reliability of the main numbers.

major comments (5)

[§4 Evaluation and Appendix A.4] The paper states 'All results are averaged over three independent runs' but Appendix A.4 says 'we fix the random seed to 42 in all experiments involving randomness.' This is contradictory: with a fixed seed, three runs are not independent and any variance would be zero. No error bars, confidence intervals, or significance tests are reported for the headline turn reductions (15.1%, 33.2%) or for the accuracy differences (35.69 vs 36.22; 33.41 vs 33.02). These numbers are the main empirical contribution; without uncertainty, the claimed 'comparable accuracy' and the size of the turn reduction are not statistically grounded. Please clarify the seed/running protocol and report variance or CIs.
[§4 Experimental Setup and Appendix A.3] The hyperparameters α and λ appear to be tuned on the same evaluation benchmarks used for the final comparison. In particular, the sensitivity analysis in Appendix A.3 evaluates α ∈ {0.5, 0.7, 1.0} on HLE-Verified and then adopts α = 0.7 as the main configuration; the setup section describes an 80/20 train/test split but no validation split. If the test set is used for hyperparameter selection, the reported efficiency and accuracy numbers are optimistically biased relative to a truly held-out evaluation. A validation split or a nested selection procedure is needed before the headline claims can be assessed.
[Appendix A.2 and §4 Main Results] The critical-step recognizer is validated only on BrowseComp-Plus, where gold evidence document IDs provide ground truth. On HLE-Verified, no such independent validation exists, yet the 33.2% turn reduction is the paper's strongest result. The teacher and student prompts assume 'evidence' is document-ID-bearing content (URLs, document fragments), which may not correspond to evidence in text-only reasoning or code-execution steps. If the recognizer systematically labels code-execution or reasoning steps as non-critical, then Eq. (6) degenerates into a smooth length penalty on successful rollouts and CRISP becomes a weaker version of OTC-GRPO rather than an evidence-preserving method. Please provide at least a sample-based human/LLM validation of HLE-Verified critical-step labels, or a counterfactual analysis showing that preserved steps are causally necessary.
[§4 Evidence Density] The 'evidence density' analysis computes the proportion of steps predicted critical by the same recognizer that shapes the training reward. Since Rcrit (Eq. 6) explicitly rewards a higher K/(K+αT), an increase from 34.06% to 41.09% on BrowseComp and from 42.06% to 53.09% on HLE-Verified is partly by construction. This analysis therefore does not independently support the mechanism of evidence preservation. To make the density claim informative, the critical-step ratios should be measured with an external label source (e.g., BCP gold IDs on a held-out set or human annotations on a sample).
[§4 Functional Validation, Table 3] The trajectory intervention uses only 100 examples from HLE-Verified and no error bars. The key comparison is Critical Only (33.00%) vs Redundant Only (27.00%), a 6-point gap; with n=100 and no variance reporting, this could be within noise. More importantly, the critical/redundant partition is produced by the same recognizer whose behavior is being validated, so the experiment is not fully independent evidence for the mechanism. The full-context control is a good idea, but the analysis should be repeated with more samples, reported with CIs, and ideally with an externally validated partition.

minor comments (6)

[Eq. (6)] The quantity K/(K+αT) is called a 'critical-step ratio' in the text, but for α ≠ 1 it is not the actual proportion of critical steps; it is a weighted precision. Please use a different term or define it as a weighted ratio.
[§4 Evaluation / A.4] The inconsistency between 'three independent runs' and 'fix the random seed to 42' should be resolved; if multiple seeds are used, report them; if not, remove the claim of averaging over independent runs.
[§4 Experimental Setup] The model 'GLM-4.5-Air-Midtrain' is not described; it is used as the backbone for all RL methods. Please provide at least a brief description of the mid-training procedure, data, and objective, or cite a public source if available.
[Appendix A.3] The sensitivity analysis reports accuracy, failure rate, and turns for α ∈ {0.5, 0.7, 1.0}, but no λ sensitivity is given. Since λ controls the strength of the auxiliary reward, a sentence or table justifying λ = 0.1 would improve reproducibility.
[§4 Tool-Use Behavior, Table 4] The table reports average calls by tool type, but the text interprets changes (e.g., +0.80 browsing calls on BrowseComp) as evidence of a shift toward 'more focused access.' Without variance or per-query distributions, some of these differences may be small relative to noise.
[Figure 3] The failure rate decomposition would be easier to interpret if the absolute counts or percentages were printed in the figure; currently only total failure rates are given, with relative reductions in text.

Circularity Check

1 steps flagged

Supporting evidence-density analysis is circular; headline efficiency results are external and not circular.

specific steps

self definitional [Section 4, Evidence Density paragraph]
"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."

The recognizer is the same fitted model used to compute the reward Rcrit = K/(K+αT) in Eq. (6), which the policy is trained to maximize via Eq. (7). The 'critical-step ratio' reported here is exactly the quantity the reward pushes upward; a policy trained with this reward will tend to have a higher predicted critical ratio by construction, whether or not the predicted steps are truly evidence-critical. The BCP gold-ID evaluation provides some independent grounding, but the comparative evidence-density gain itself is not independent evidence for the mechanism.

full rationale

The main claims—15.1% and 33.2% turn reductions with comparable accuracy—are measured externally (interaction turns and final-answer correctness) and are not derived from the recognizer's own outputs. The recognizer is independently validated on BCP gold evidence document IDs (Appendix A.2), limiting the burden of the critical-step definition. However, the 'Evidence Density' analysis in Section 4 measures the very proportion of predicted-critical steps that the reward in Eq. (6)-(7) is designed to increase; this supporting claim is tautological. Hyperparameter alpha is tuned on HLE-Verified (Appendix A.3), which makes the reported HLE-Verified numbers a selected configuration rather than a fresh prediction; this is a selection concern, not a definitional circularity. The transfer of the critical-step notion to HLE-Verified's text/code trajectories is not independently validated, but that is an empirical validity gap, not circularity. Overall, the central derivation is self-contained; only the supporting density analysis is circular, warranting a moderate score.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim rests on the reliability of teacher judgments, the fidelity of the distilled recognizer, and the correctness of LLM answer judges. Alpha and lambda are the key hand-set/fitted numbers in the reward. No new physical or structural entities are introduced; 'critical step' is a labeling construct, not an invented entity.

free parameters (2)

alpha (redundancy penalty) = 0.7
Controls penalty on redundant steps in R_crit = K/(K+alpha*T); chosen from sensitivity {0.5,0.7,1.0} on HLE-Verified.
lambda (auxiliary reward weight) = 0.1
Weights efficiency signal relative to answer correctness; set small by hand, no sensitivity reported.

axioms (4)

domain assumption A strong model can reliably judge a step's 'marginal evidence contribution' by backward traversal conditioned on the final answer and confirmed future steps.
Stage 1 assumes the teacher's local backward judgments are causally correct; no human agreement or counterfactual validation is presented for this judgment.
domain assumption The distilled 9B recognizer's single-pass predictions faithfully reproduce teacher backward judgments on unseen trajectories.
Only validated on BCP document-ID proxy for evidence, not on HLE-Verified trajectories.
domain assumption The LLM judges (GLM-4.7 for BrowseComp, GPT-OSS-120B for HLE-Verified) provide reliable answer-correctness rewards.
Binary correctness reward is the gate for the efficiency reward; judge errors would propagate.
domain assumption Document-ID overlap with BCP gold evidence is a valid operationalization of 'evidence'.
Used to evaluate the recognizer; not used in training reward, but supports the mechanism claim.

reviewed 2026-08-04 · 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}
}

Share X Bluesky LinkedIn Reddit HN

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.