PaperlayerView exact revision on arXiv

Source-linked AI summary

Judged Useless, Queried Anyway: Tool-Using Agents Rarely Turn Their Own Evidence Judgments into Stopping Decisions

Chubin Zhang, Zhenglin Wan, Xingrui Yu, Jingxuan Wu, Yaxin Zhou, Ivor Tsang, Bo An

arXiv:2610.06191v1cs.AIcs.CL

TL;DR

Agents can recognize useless retrieval results yet continue searching, so the paper tests whether stopping follows those judgments or merely time and deadlines. In controlled retrieval failures, it separates judgments from actions and measures stopping at matched steps. Agents usually judge failing-source results correctly but rarely stop on that evidence; an enforced integration rule makes stopping follow it, with the result confirmed on fresh questions.

  • Problem

    When retrieval sources fail silently, it remains unclear whether agents misjudge results, expect continued search to pay, or fail to turn judgments into stopping actions.

    Source: §1 · p. 1

  • Method

    The study controls retrieval-source failures, records agents’ judgments, beliefs, and actions separately, and compares stopping at the same step across histories with different useless-result runs.

    Source: §2 · p. 4

  • Results

    Recorded agents judged failing-source results useless 97–100% of the time, yet most rarely stopped on that evidence; enforcing integration made stopping follow judgments.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    The authors conclude that agent designs should make the integration of evidence into an action explicit rather than expect prompting alone to produce it.

    Source: §9 · p. 10

    Failures were simulated by replacing observations in fixed knowledge bases, so whether the pattern holds for natural partial, stale, or erroneous tool outputs remains untested.

    Source: §9 · p. 10

Abstract

from arXiv · show

An agent whose tool keeps returning nothing useful should stop relying on it. In a retrieval environment with controlled source failures, we separate how agents judge results from what they do. We compare stopping at the same step after longer and shorter runs of results the agent judged useless; this contrast is zero for clock- or deadline-driven stopping. Where we record their judgments, the seven agents we test call a failing source's results useless 97-100% of the time, yet most of them rarely stop on that judgment. Prompt cues change when they stop but not what they stop on. Permission to answer from memory and a reasoning mode can bring early stops regardless of evidence, a stated budget moves the 7-8B models' stops to the deadline, and a stopping rule or call cost in the prompt is followed at most partly. Stopping follows the evidence only when the harness enforces an integration step that makes the agent answer after five consecutive results it judged useless. This step raises failing-source success for every model, keeps the stopping point fixed when the budget doubles, and needs no extra judgment call when the agent states its judgments. A pre-registered replication on 300 fresh questions confirms the dissociation and the rule's effect.

1. Introduction

Agents often recognize that a retrieval source is returning useless results but keep searching rather than acting on that evidence. The paper separates judgments from stopping behavior and tests whether additional prompt information or an enforced integration step changes that behavior.

1. Introduction

  • A useful stopping policy should respond to the current run of useless results, not simply stop after a fixed time or never leave the source.A single useless result is weak evidence because a reformulated query may succeed or the source may recover.Source: §1 · p. 1
  • The study controls source failures and compares stopping at the same step after longer versus shorter runs of results judged useless.The time-matched contrast is zero for clock- or deadline-driven stopping and positive when stopping follows the useless run.Source: §1 · p. 2
  • Every model with recorded judgments called a failing source’s results useless 97–100% of the time, yet most agents rarely stopped on that evidence.After five consecutive useless judgments, five of seven agents answered on at most 7% of questions.Source: §1 · pp. 2–3
  • Prompt cues can change when agents stop without making their stopping track the evidence; the enforced integration rule does make it track the evidence.The rule makes the harness answer after five consecutive useless judgments; the result replicates on 300 fresh questions.Source: §1 · p. 3

2. Setting

The study uses retrieval tasks with controlled source failures so that result usefulness and recovery are known, while recording judgments, beliefs, and actions separately. It measures whether stopping responds to accumulated useless evidence rather than elapsed steps or the deadline.

2. Setting

  • HotpotQA questions use individual ten-paragraph knowledge bases with two supporting paragraphs, and agents search, look up, or finish under an eight-action budget.Source failures replace observations with paragraphs from another question; six regimes vary failure and recovery, and success is averaged equally across them.Source: §2 · p. 3
  • At each decision point, the study records the latest-result judgment, beliefs about continuing or answering, and the agent’s action.A one-word usefulness judgment is also elicited on an unseen scratch copy, allowing judgments to be measured without changing the trajectory.Source: §2 · pp. 3–4
  • The time-matched contrast Δ compares answering after an all-useless history with answering after a history whose latest result is useless but an earlier result was useful.Δ is positive for stopping on the useless run, zero for stopping by time or deadline, and negative when stopping follows earlier useful evidence.Source: §2 · p. 4, Figure 2 · p. 4
  • Conditions vary whether agents receive permission, a budget, call costs, a stated stopping rule, a running count, or an enforced rule after five useless judgments.Two controls replace judgments with random or lexical signals; Table 1 presents conditions and stopping measures.Source: §2 · p. 4, Table 1 · p. 4
  • The experiments include four locally run open models and additional Claude, reasoning-mode, and reinforcement-learning search-agent conditions.Thresholds were set on development questions; confirmatory tests use disjoint test300 questions and FEVER claims, with a pre-registered replication on fresh300.Source: §2 · pp. 4–5
  • Stated judgments are labeled by two LLM annotators with author adjudication, while other outcomes rely on design labels, side-channel judgments, or deterministic action parsing.The authors report no errors in a 100-point audit of those other measures, and say the main conclusions are unchanged by a fixed answer-template repair.Source: §2 · p. 5

3. Agents judge correctly but do not act on their judgments

Agents generally identify failing-source results as useless, but their stopping behavior does not follow those judgments. The same dissociation appears when judgments are stated in the agents’ reasoning.

3. Agents judge correctly but do not act on their judgments

  • Every model with recorded side-channel judgments called failing-source results useless 97–100% of the time on HotpotQA and FEVER.Gold paragraphs were also usually judged useful, except by Qwen2.5-7B, which did so 57% of the time when first retrieved.Source: §3 · p. 5
  • Beliefs about whether to continue or answer varied across models, but actions did not follow those beliefs.When reports favored answering, Llama-3.1-8B answered at 32 of 595 decision points and Qwen2.5-7B at 267 of 2,443.Source: §3 · p. 5
  • After five consecutive useless judgments, unaided Qwen3-8B answered only 1 of 292 questions, and other Figure 1 models answered on at most 7%.Answering at that point succeeded 15–17% of the time, while continuing succeeded at most 7% for the 7–8B models in exploratory analyses.Source: §3 · p. 5, Figure 3 · p. 6
  • Judgments stated in reasoning agreed with side-channel judgments on 87–96% of decision points, yet Δ remained below zero.After five stated useless judgments in a row, agents answered on 0–0.4% of questions, with Qwen3-32B at 6%.Source: §3 · p. 5

4. Telling agents more changes when they stop but not what they stop on

Prompt cues can change when agents stop without making stopping reliably track their evidence judgments. Permission and reasoning can prompt early answers, budget cues shift stops to deadlines, and stated rules or costs are followed only partly.

4. Telling agents more changes when they stop but not what they stop on

  • Qwen3-8B answers before any real result on 55% of questions where the source recovers after three failures, with Δ falling to −0.32.Permission to answer from memory leaves Llama-3.1-8B and Qwen2.5-7B almost unchanged.Source: §4 · p. 6
  • 51–79% of budget-cued 7–8B answers on failing sources come on the last action, and doubling the budget doubles calls without improving success.Budget-cued answers shift from action eight to sixteen; the enforced rule and combination answer at the same action under both budgets.Source: §4 · p. 6, Figure 4 · p. 7
  • Stating a stopping rule or adding a call price to the prompt supplies the integration step only partly.Only Llama-3.1-8B and Qwen3-32B partly follow the stated rule; the 7–8B models still answer at the deadline with a stated price.Source: §4 · pp. 6–7
  • Adding an intact backup search does not make unaided agents more likely to leave as useless results accumulate (Δ ≤ +0.02).By the fourth action, Llama-3.1-8B and Qwen2.5-7B switched on only 17% and 49% of questions; disabling the failed tool after five useless judgments makes leaving evidence-driven for two models.Source: §4 · p. 7
  • Giving agents a running count of their useless judgments still does not make the 7–8B models stop on that count (Δ = −0.14 to −0.24).Qwen2.5-7B improves mean6 by stopping early indiscriminately; Table 2 cautions that similar success can conceal different stopping behavior.Source: Table 2 · p. 6, §4 · p. 7

5. Enforced integration makes stopping follow the evidence

An enforced integration step converts accumulated judgments into a stopping action. It raises success, remains stable when the budget doubles, and can be combined with a budget cue.

5. Enforced integration makes stopping follow the evidence

  • Mean6 rises for the four open models and Claude Haiku 4.5 (+0.026 to +0.097, Holm-significant), while stopping stays at the same action when the budget doubles.Where sources recover or never fail, the rule is non-inferior to unaided agents in 18 of 20 tests.Source: §5 · p. 7
  • The enforced rule beats a random useless-signal rule by 0.011 to 0.029 (Holm-significant), while a cheap lexical detector performs about as well as the agent.The comparison indicates that the timing of the accurate signal matters; the lexical-detector result is exploratory.Source: §5 · pp. 7–8
  • The budget-and-rule combination is non-inferior to both components in all 14 pre-registered tests and beats the enforced rule by +0.138 for Qwen2.5-7B.The stated budget prevents premature answers while the enforced rule supplies the evidence-driven stop.Source: §5 · p. 8
  • A keyword reader can avoid the enforced rule’s three to five judgment calls per question while keeping success within 0.01 for Llama-3.1-8B and Qwen3 models.Combined with the budget statement, the reader beats the budget-only condition at 0.05 per call for Llama-3.1-8B and Qwen3-8B.Source: §5 · p. 8

6. The pattern replicates on fresh questions, a second task and harder failures

The dissociation between judging useless results and stopping on those judgments replicates across fresh questions, FEVER fact verification, and harder failure settings. The enforced rule continues to raise success, with a boundary on answerless sources.

6. The pattern replicates on fresh questions, a second task and harder failures

  • All 19 pre-registered tests pass on 300 fresh questions, with Δ below zero for unaided and budget-cued agents and +0.30 to +0.35 under the enforced rule.The enforced rule beats unaided mean6 by +0.030 to +0.061; unaided 7–8B agents answer after five useless judgments on at most 3% of questions.Source: §6 · p. 8
  • On FEVER, the enforced rule raises mean6 by +0.129 to +0.201, while unaided agents answer after five useless judgments on at most 3% of claims.The budget condition again answers mostly on the last action (69–96%).Source: §6 · p. 8
  • With plausible or answerless failures, judgments are less accurate but the dissociation remains, and the enforced rule still raises success in both settings.After five useless judgments, unaided agents answer on 2–23% of plausible-regime questions and 5–21% of answerless-regime questions; on answerless sources, the budget condition does better.Source: §6 · p. 8

7. Larger, reasoning and search-trained agents rarely integrate either

Larger models, search training, and reasoning mode do not reliably make agents integrate their own evidence judgments. The enforced rule still raises success for the larger model tested.

7. Larger, reasoning and search-trained agents rarely integrate either

  • Qwen3-32B answered after five useless judgments on only 19 of 284 questions; the enforced rule raised Claude Sonnet 5 failing-source success from 0.40 to 0.51, though its six-regime gain was not significant.Unaided Claude Sonnet 5 exhausted its budget on 45% of questions where its enforced rule fired.Source: §7 · pp. 8–9
  • With its budget stated, Qwen3-32B answers around action six, but the time-matched contrast is Δ = 0.00 [−0.06, +0.05].Doubling the budget shifts answers a step later and increases tool calls by 46% without a success gain.Source: §7 · p. 9
  • Search-R1 answers from memory after one or two searches on 27% of failing-source questions and searches to the budget’s end on 45%, with no positive failure-length contrast.Its judgments call most real results useless (89%), and the enforced rule still raises mean6 by +0.024.Source: §7 · p. 9
  • In thinking mode, Qwen3-32B answers from memory on 97% of failing-source questions, versus 11% without thinking, while Δ remains near zero.When sources recover after two failures, it answers before any real result on 47% of questions.Source: §7 · p. 9

8. Related work

Prior work documents over-search and explores retrieval controls, but this paper asks whether agents act on their own judgments that a source keeps failing. Related work also identifies gaps between knowing and doing across tool-use and bandit decisions.

Agents and unreliable tools

  • Models can carry wrong tool output into answers despite noticing conflicts [Yang et al., 2026], fail to diagnose recoverable faults, and be hurt by misleading feedback.Source: §8 · p. 9

When to stop searching

  • Search-augmented models over-search [Liu et al., 2026a, Xie et al., 2026], while proposed remedies include trained boundaries, sufficiency checks, controllers, adaptive retrieval, and learned stopping [Zhang et al., 2026a; Roh and Han, 2026; Soudani et al., 2026; Asai et al., 2024; Jiang et al., 2023; Yan et al., 2024; Jin et al., 2025].The paper distinguishes these system-level approaches by measuring whether agents act on their own judgments that a source keeps failing.Source: §8 · p. 9

Abstention and recovery

  • Benchmarks study abstention [Luo et al., 2026] and learned fallback strategies [Chen et al., 2026], while fixed search budgets set deadlines in this paper’s findings.The paper’s rule uses the agent’s own judgments rather than learned fallback strategies.Source: §8 · p. 9

Knowing and doing

  • Gaps between judgment and action have been reported in tool calling [Li et al., 2026] and bandit tasks.Here, an explicit judgment still does not lead to stopping even when the running count is shown.Source: §8 · p. 9

9. Discussion and conclusion

Agents often recognize useless results without acting on them; the paper argues that an explicit integration step is needed to connect those judgments to stopping. Its conclusions are bounded by the simulated failure setting, recovery distribution, and evaluation choices.

9. Discussion and conclusion

  • The enforced run-length rule maps consecutive useless judgments to an action, addressing the gap between recognizing failing evidence and stopping.The discussion recommends making this decision explicit and evaluating evidence-driven stopping separately from success.Source: §9 · p. 10
  • The rule’s threshold was chosen for sources recovering after at most three failures; later recovery can arrive just after the rule gives up.Natural partial, stale, or erroneous tool failures remain untested, although on-topic failures preserve the dissociation.Source: §9 · p. 10

A. Pre-registration, amendments and deviations

This section documents the study’s protocol and implementation, then examines agents’ usefulness judgments and how they relate to stopping behavior. It also reports a descriptive probe of usefulness judgments and continuation beliefs without changing recorded trajectories.

A. Pre-registration, amendments and deviations

  • The pre-registration set thresholds on 100 development questions and specified paired comparisons, bootstrap intervals, Holm correction across models, and a −0.05 non-inferiority margin.Qwen2.5-7B’s k=5 result was recorded before test runs but omitted from the protocol document, which the authors counted as a deviation; the stop-rule condition for withdrawing the intervention claim did not trigger. Amendments were frozen before their governed runs, and the append-only document was not deposited with a third-party registry.Source: A. Pre-registration, amendments and… · p. 13

B.1. Prompts and signals

  • The rule arm used five consecutive USELESS judgments to force a final answer; side-channel judgments were parsed and discarded without changing the acting history.Random and lexical controls substituted model-matched random judgments or a development-set lexical threshold, while the combination paired the rule with the budget prompt.Source: §B.1 · p. 18

B.2. Annotation of stated judgments

  • Qwen3-8B’s stated judgments labeled useless evidence useless 93.9% of the time and treated useful or partly useful evidence as such 98.1% of the time.These figures are based on 1,463 useless-evidence judgments and 212 useful-or-partly-useful observations, respectively.Source: §B.2 · p. 19

B.3. Figure 1 selection rule

  • Figure 1’s example was selected as the first held-out question meeting all three criteria: budget exhaustion without intervention, five consecutive USELESS rule judgments, and a correct rule-arm answer.No other candidate was inspected.Source: §B.3 · p. 19

B.4. Claude Sonnet 5

  • Claude Sonnet 5’s rule improved persistent-regime performance by +0.11, while its overall rule effect was +0.022 and did not meet the significance threshold (p=0.051).The unaided agent answered on 69/100 persistent questions; among 65 questions where the rule fired, it exhausted its budget on 29 (45%). The rule’s S-H1 test failed, while S-H2 passed; non-inferiority could not be shown in three of four regimes.Source: §B.4 · p. 20

C.1. Time-matched contrast

  • After five consecutive results judged useless, unaided agents answered on 3/122, 1/282, 1/292, and 19/284 questions for Qwen2.5-7B, Llama-3.1-8B, Qwen3-8B, and Qwen3-32B, respectively.The time-matched contrast followed the pre-registered pattern for the rule and for unaided, permission, and budget arms across all four models, but not fully for stated-rule and stated-cost arms. Under every stratification, unaided and budget-cued contrasts had 95% upper bounds below +0.06; stated cost left the 7–8B models answering at the deadline. The combination exceeded stated cost in cost-adjusted success for 7–8B models and tied it for Qwen3-32B.Source: §C.1 · pp. 20–21, Table 6 · p. 22

C.2. Stopping hazards and the integration index

  • In Figure 5’s 8-action-budget setting, the enforced rule steps up at five consecutive failures; unaided agents generally do not rise, and prompt-cued rises occur only for Qwen3-32B.Claude Sonnet 5 shows a partial unaided rise; prompt cues stay flat or fall for 7–8B models. The hazard model adds question fixed effects in its second specification.Source: §C.2 · p. 21, Figure 5 · p. 23

C.3. Specificity of side-channel judgments

  • Real distractor paragraphs are usually judged useless, and repeated pages add nothing new, making low usefulness rates appropriate.Table 11 splits USEFUL judgments by gold versus distractor paragraphs and whether the agent had already seen the page.Source: §C.3 · p. 25

C.4. Beliefs probe

  • At each recorded decision point, the probe forks the history for separate usefulness and continuation assessments, discarding replies and recording unparseable responses as missing.The continuation prompt asks for p_continue and p_answer_now without stating the remaining budget or offering an action.Source: §C.4 · p. 25
  • The descriptive probe covers unaided Qwen2.5-7B, Llama-3.1-8B, and Qwen3-8B runs across four regimes; its ideal observer is a reference, not an optimum.It compares realized continuation success with answer-now success and an observer that updates regime probabilities using the model’s own usefulness judgments; Figure 6 shows these quantities.Source: §C.4 · p. 25

C.5. Answer repair

Placeholder repair affected answer scores but not stopping decisions; exploratory analyses found benefits to answering after five useless judgments and examined stated-judgment rules and other variants.

C.5. Answer repair

  • Placeholder repair changed answer scores but not stopping decisions; it could shift contrasts between answering arms, and placeholders’ concentration in those arms made H1 and H2 conservative.Raw results remained primary; repaired results were a sensitivity analysis.Source: §C.5 · pp. 25–26

D.1. Full result tables

  • Appendix tables report HotpotQA test300 at an eight-action budget and FEVER’s 292 claims; raw analyses count placeholder answers wrong and use 2,000 question-bootstrap resamples for intervals.The appendix defines success as the share answered correctly per regime and mean6 as the equal-weight average across six regimes.Source: §D.1 · p. 27

D.2. Fixed step caps, the best simple policy, and whether stopping pays

  • 0.171 vs. 0.065 for Qwen2.5-7B, 0.159 vs. 0.005 for Llama-3.1-8B, and 0.149 vs. 0.013 for Qwen3-8B: answering after five useless judgments outperformed continuing.This exploratory comparison concerns sources that do not recover after five failures; continuing used 2.5–2.6 further tool calls on average.Source: §D.2 · p. 29
  • Offline, the own-5 policy closely reproduced the rule arm’s success; with a doubled budget, budget-cue calls grew while rule and combination calls barely changed.The own-5 and rule-arm success values were 0.378 vs. 0.372, 0.271 vs. 0.271, and 0.216 vs. 0.224 for Qwen3-8B, Llama-3.1-8B, and Qwen2.5-7B, respectively.Source: §D.2 · p. 29

D.3. The rule on stated judgments, and Claude Haiku 4.5 with the rule or cost stated

  • A keyword reader tracks five-useless runs from thought text; stated-judgment support requires firing on at least 20% of persistent failures, and all four models passed E2’s −0.05 non-inferiority margin.For Qwen3-8B development judgments, the reader’s useless classifications had 0.971 precision and 0.913 recall.Source: §D.3 · pp. 30, 32
  • Claude Haiku 4.5’s stated call cost raised utility by +0.154 over unaided and +0.213 over the enforced rule, but format fixes favored the new conditions, especially call-cost.The stated-rule verdict remained partial.Source: §D.3 · p. 33

D.4. Fresh questions, stated judgments, reasoning mode and Search-R1

  • On 300 disjoint fresh questions, stated judgments matched side-channel judgments at 0.87–0.96 across four models when judgments were explicit.The explicit-judgment counts ranged from 99 to 157 of 160 sampled decision points.Source: §D.4 · p. 33
  • Thinking shifted Qwen3-32B’s mean over four regimes by +0.076 without budget and −0.051 with budget; Search-R1’s own-judgment contrast was uninterpretable after it mislabeled 89% of real observations useless.Search-R1 answered on 53% of persistently failing questions unaided, with a median action of 3.Source: §D.4 · p. 34

D.5. Backup tool and harder failures

The backup-tool and harder-failure analyses show that switching and stopping responses vary by model; the replication supports some pre-registered findings while qualifying others.

D.5. Backup tool and harder failures

  • Qwen2.5-7B and Llama-3.1-8B did not abandon a failing primary source by the pre-registered criteria, whereas Qwen3-8B and Qwen3-32B switched.Switch-rule minus none mean-six-regime success differences were +0.011, +0.054, +0.016, and +0.007, respectively; the intervals include zero for Qwen2.5-7B and Qwen3-32B.Source: §D.5 · p. 35, Table 25 · p. 35
  • After five useless judgments on an answerless source, the four models answered on 41/202, 9/166, 26/170, and 26/122 questions, respectively.Such pages lack the question’s supporting-fact sentences; the predicted loss in recover_after_5 held for the Qwen3 models but not the two 7–8B models.Source: §D.5 · p. 35

D.6. Second and third runs

  • In the replication, the rule gained over none for all three 7–8B models and budget arms answered at the deadline; the rule-index threshold failed for Qwen2.5-7B.The combination’s non-inferiority tests passed five of six comparisons before repair, and all six after repair; two conclusions were not stable under the replication’s interpretation rule.Source: §D.6 · pp. 35–36
  • The third run repeated the amendment 10 tests and pooled the rule’s gain across all three runs using a bootstrap over runs and questions.The supplied passage describes the analysis but does not provide its numerical outcome.Source: §D.6 · p. 36

E. Licenses and compute

The appendix identifies the licenses covering the datasets and models and reports the local compute used for the experiments.

E. Licenses and compute

  • HotpotQA is distributed under CC BY-SA 4.0, and FEVER under CC BY-SA 3.0.The FEVER license covers Wikipedia text under CC BY-SA.Source: E. Licenses and compute · p. 36
  • Qwen2.5-7B, Qwen3-8B, Qwen3-32B, Search-R1, and its base model are released under Apache 2.0; Llama-3.1-8B uses the Llama 3.1 Community License.Claude models and annotators were accessed through official APIs under providers’ terms.Source: E. Licenses and compute · pp. 36–37
  • Local models ran on four NVIDIA RTX A5000 GPUs, using about 350 GPU-hours including development; no model was trained or fine-tuned.Qwen3-32B used tensor parallelism across all four GPUs.Source: E. Licenses and compute · p. 37

Related papers

Loading 2610.06191v1…