cs.AIOct 18, 2026

Explore or Confirm? LLM Agents Misallocate Measurements Under Noise

Authors: Abdul Karim, Matt Keon

Organizations: Xstendyl

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

As model reasoning gets cheaper, the physical experiment becomes the slow and costly step in discovery. Each experiment returns a noisy reading: a high value can mean a good candidate or a lucky draw. With a fixed budget, an AI agent must keep choosing between exploring a new candidate and confirming a promising one by re-testing it. We measure that choice in Wellocate, a simulated assay campaign with known true values. It is built on a generated landscape and on two real protein datasets, GB1 and PhoQ, scrambled so no model can answer from memory. We test five language models as minimal agents, with no tools and no memory beyond the history in each prompt. They misallocate in both directions. Three over-confirm, spending most of the budget re-testing. On the generated landscape one under-confirms, trusting single readings, and there every model ends with a worse final pick than a basic scripted strategy. Each failure has its own cause. The over-confirmers run short of new candidates worth trying: a list of untried candidates cuts their re-testing and, on real protein data, makes two of them pick much better. A running summary of their readings or a description of the landscape changes it by at most five percentage points. For the under-confirmer, re-testing its top candidates in a final round, replayed on its own campaigns, recovers only about a quarter of what it loses by picking the wrong candidate. This suggests confirmation has to run through the whole campaign. We release Wellocate as a benchmark and reinforcement-learning environment with frozen splits and scripted baselines. With it, anyone can measure how an agent splits its budget before trusting it with real experiments, and train agents to split it better. Code, data and the environment are available at https://github.com/Xstendyl/wellocate.

Figures & tables

Figure 1: Five models on the grid. Left: the re-testing share, the fraction of used slots spent on a candidate the agent had already ordered, with 95% intervals. Right: the true value of each model’s final pick. The lines are the three scripted rules (Section 2 ), drawn the same way in both panels. The guessing rule is dashed, the model-guided rule dotted, and the informed rule, which is also told which candidates are viable, solid. The grey band is the scripted band, from the guessing rule’s share to the model-guided rule’s. The informed rule re-tests about as often as the guessing rule, so their lines overlap on the left. On the grid, Opus 4.7, Gemini 2.5 Flash and DeepSeek V3.1 re-test far above the band, GPT-6.1 Sol just above it and Gemini 3.1 Pro below it. Every model’s pick falls short of all three scripted rules. The figure and its band are for the grid only; on GB1 and PhoQ the best scripted rule re-tests at most 1% of its slots (Table 1 ).
Figure 2: One Wellocate campaign, with the lab counterpart of each part. Each round the agent orders four candidates, one per assay well. The simulated assay returns a noisy reading for each, and the agent sees every reading so far before ordering again. Orders for non-viable candidates are rejected and replaced, and about 15% of orders for viable candidates fail. After 15 rounds the agent nominates the candidate to take forward, scored on its true value.
Figure 3: The grid. (a) True value of all 160,000 candidates. A value averages four random position scores, so most candidates sit near the middle and very high or very low values are rare. The vertical axis is on a log scale so that the rare ones show. Only the top 7% (11,199) are viable, shown in blue; an order for any other candidate is rejected. Every campaign starts from GDIF , the median viable candidate (6.87), and the best possible is 8.762. (b, c) A candidate has four letters, so each grid shows one slice: letters 3 and 4 are fixed and letters 1 and 2 vary, 400 candidates per slice. Grey cells are not viable. Changing only letter 1 moves along the star’s row, and changing only letter 2 moves along its column. The outlined row and column thus hold the 38 candidates one letter away that a slice can show. (b) The slice through the best candidate, IOJO , with J and O fixed: 23% of the slice is viable, and 20 of the star’s 38 outlined neighbours. This is the richest region of the grid. (c) The slice through the start, GDIF , with I and F fixed: 16 of 400 cells are viable, and 4 of the star’s 38 outlined neighbours. 20 of the start’s 24 viable neighbours change letter 3 or 4 and so lie off this slice. The search problem is getting from (c) to (b).
 gridGB1PhoQ
 share (%)pickshare (%)pickshare (%)pick
Opus 4.774.97.6263.34.7862.81.89
Gemini 2.5 Flash94.07.7091.91.1263.72.13
DeepSeek V3.191.07.7696.00.8489.90.83
GPT-6.1 Sol44.17.6719.46.3623.92.37
Gemini 3.1 Pro14.07.6911.14.9014.71.72
Table 1: Re-testing share and true value of the final pick on each landscape, averaged over each model’s campaigns. The three scripted rules are shown at their best setting.
Figure 4: Where each model’s loss comes from, with 95% intervals. Exploration loss: how far the best candidate measured falls short of the landscape’s best. Selection loss: how far the pick falls short of the best candidate measured. The dashed line joins points with the guessing rule’s total loss. The informed and model-guided rules are shown for reference; the informed rule is told which candidates are viable. Opus 4.7, Gemini 2.5 Flash and DeepSeek V3.1 measure too few candidates and pick well among them; Gemini 3.1 Pro measures as widely as the guessing rule and picks poorly.
Figure 5: What we tested for each failure, each against the same model’s control on the same seeds, with 95% intervals. Over-confirmation. (a) The three aids on the grid, given to the three over-confirming models: change in re-testing share. (b, c) The Menu on scrambled GB1, given to all five models: change in re-testing share and in final pick. It cuts re-testing in the over-confirming models and lifts two of their picks by more than three units. It moves the re-testing of Gemini 3.1 Pro and Sol by at most four points. Under-confirmation. (d) The grid campaigns of Gemini 3.1 Pro and Sol, replayed with the last round spent re-testing the top mm candidates. With m=1m=1 the top one is re-tested four times, and with m=4m=4 each of the top four once. The y-axis is the change in the true value of the final pick against the agent’s own pick. Each dotted line is that model’s selection loss, the most better checking could win back. Gemini 3.1 Pro recovers about a quarter of it; Sol gains nothing.
Figure 6: The guessing rule’s final pick for each number of re-tests per round, with 95% intervals. Solid curves cover realistic noise; dashed curves go far beyond it, to show where re-testing starts to pay on the proteins. The star marks each curve’s best setting; where a curve is flat at the top, the star is a near tie.
Appendix figures & tables26 assets

Supplementary material from the paper’s appendix.

Appendix

 informed (told which are viable)guessing (must guess)
best re-tests a round, k∗k^{\ast}1 of 4 slots1 of 4 slots
mean pick at k=0k{=}07.88627.7098
mean pick at k∗k^{\ast}7.99937.8659
distinct candidates at k∗k^{\ast}39.1428.94
rejections per campaign at k=1k{=}10.0083.02
lost slots per campaign at k=1k{=}10.0011.70
Appendix
Table 2: The guessing and informed rules on the grid. One knows more than the agents, the other knows the same.
 no re-testsone re-test a round
GB1, 8 times assay noise3.60 [3.53, 3.67]4.20 [4.11, 4.29]
grid, free rejections, assay noise7.79 [7.77, 7.82]7.95 [7.93, 7.97]
Appendix
Table 3: The guessing rule’s final pick with no re-tests and with one re-test a round, with 95% intervals.
Figure 7: Figure 6 with free rejections: a rejected order never costs a slot.
tracklandscapetrainvalidationtestheld outpaper
proteinscrambled GB1400 / 100,00010 / 1,50020 / 1,5001 / 1,5001 / 1,500
proteinscrambled PhoQ400 / 100,00010 / 1,50020 / 1,5001 / 1,5001 / 1,500
gridgenerated NK400 / 100,00010 / 1,50020 / 1,500K=2K=2 : 20 / 1,500 K=3K=3 : 20 / 1,5001 / 1,500
Appendix
Table 4: Wellocate’s frozen splits: landscapes / episodes per split. Train, validation and test use disjoint landscape seeds (scrambles for GB1 and PhoQ, NK seeds for the grid) and disjoint campaign seeds. Held out: the real amino-acid names for GB1 and PhoQ, and two rougher NK settings, K=2K=2 and K=3K=3 . Paper is the landscape and seeds used in this study.
modelreleasedreasoning as runmedian reasoning tokens per call
Opus 4.72026-02off0
Gemini 2.5 Flash2025-06off0
DeepSeek V3.12025-08off0
GPT-6.1 Sol2026-09minimal0
Gemini 3.1 Pro2026-02minimal240
Appendix
Table 5: The five models as run on the grid. GPT-6.1 Sol and Gemini 3.1 Pro ran at their lowest reasoning setting. Reasoning tokens are the median over every logged call of the runs with the Blind prompt. Sol reported zero reasoning tokens on at least half of its calls.
 share (%)share, slots not lost (%)final pickexploration lossselection loss
Opus 4.774.9 [73.5, 76.3]77.4 [76.1, 78.7]7.621 [7.543, 7.696]1.02 [0.96, 1.09]0.12 [0.08, 0.16]
Gemini 2.5 Flash94.0 [93.1, 94.8]94.6 [93.8, 95.3]7.701 [7.680, 7.720]1.04 [1.03, 1.06]0.02 [0.00, 0.03]
DeepSeek V3.191.0 [90.3, 91.6]91.2 [90.7, 91.8]7.760 [7.716, 7.799]0.94 [0.90, 0.98]0.06 [0.04, 0.08]
GPT-6.1 Sol44.1 [41.5, 46.7]45.7 [43.1, 48.4]7.672 [7.582, 7.764]0.93 [0.84, 1.01]0.16 [0.11, 0.22]
Gemini 3.1 Pro14.0 [12.6, 15.3]17.5 [15.9, 19.1]7.688 [7.580, 7.802]0.70 [0.60, 0.79]0.38 [0.30, 0.46]
Appendix
Table 6: The five models on the grid, with 95% intervals. The second column counts re-testing over slots not lost.
modelledgerlandscape noteMenu
Opus 4.7++ 4.4 [ ++ 2.6, ++ 6.2]−- 4.7 [ −- 7.8, −- 1.5]−- 18.2 [ −- 20.8, −- 15.5]
Gemini 2.5 Flash−- 0.9 [ −- 2.1, ++ 0.2]−- 0.5 [ −- 1.8, ++ 0.8]−- 26.2 [ −- 27.9, −- 24.5]
DeepSeek V3.1++ 0.8 [ ++ 0.1, ++ 1.6]++ 1.5 [ ++ 0.8, ++ 2.2]−- 31.4 [ −- 33.5, −- 29.4]
Appendix
Table 7: Change in re-testing share on the grid when each aid is added, in percentage points against the same model’s control on the same seeds, with 95% intervals. Negative values mean less re-testing. The Menu’s effect on scrambled GB1 is in Figure 5 .
 GB1PhoQ
registered: campaigns resampled by seed++ 0.28 [ ++ 0.06, ++ 0.53]++ 0.32 [ ++ 0.10, ++ 0.54]
whole nominees resampled++ 0.28 [ −- 0.01, ++ 0.74]++ 0.32 [ −- 0.09, ++ 0.55]
repeated nominee counted once++ 0.33 [ −- 0.01, ++ 0.71]++ 0.20 [ −- 0.14, ++ 0.57]
both checks together–++ 0.20 [ −- 0.16, ++ 0.61]
Appendix
Table 8: Sol’s pick minus the guessing rule’s mean, with 95% intervals that include the guessing rule’s own uncertainty: as registered, and under two checks that count its repeated nominee as one. The checks were not registered.
 best value measuredselection loss
Gemini 2.5 Flash++ 0.17 [ ++ 0.11, ++ 0.23]++ 0.13 [ ++ 0.10, ++ 0.17]
DeepSeek V3.1++ 0.08 [ ++ 0.01, ++ 0.16]++ 0.10 [ ++ 0.05, ++ 0.14]
Opus 4.7–++ 0.09 [ ++ 0.03, ++ 0.16]
Appendix
Table 9: Change with the Menu on the grid, against each model’s control on the same seeds, with 95% intervals. For Opus 4.7 only the selection loss was computed.
   pick minus the guessing rule
 share (%)final pickregisteredpaired, PhoQpaired, GB1
Opus 4.762.8 [61.0, 64.6]1.893 [1.753, 2.025]−- 0.16 [ −- 0.30, −- 0.02]−- 0.26 [ −- 0.60, ++ 0.07]−- 1.41 [ −- 1.92, −- 0.89]
Gemini 2.5 Flash63.7 [60.6, 66.6]2.130 [2.050, 2.184]++ 0.08 [ −- 0.01, ++ 0.15]−- 0.02 [ −- 0.35, ++ 0.28]−- 5.07 [ −- 5.47, −- 4.68]
DeepSeek V3.189.9 [87.5, 91.9]0.829 [0.757, 0.942]−- 1.22 [ −- 1.31, −- 1.10]−- 1.32 [ −- 1.68, −- 0.98]−- 5.35 [ −- 5.75, −- 4.94]
GPT-6.1 Sol23.9 [20.8, 27.3]2.368 [2.155, 2.582]++ 0.32 [ ++ 0.10, ++ 0.54]++ 0.22 [ −- 0.21, ++ 0.62]−- 0.06 [ −- 0.65, ++ 0.52]
Gemini 3.1 Pro14.7 [13.1, 16.3]1.725 [1.601, 1.857]−- 0.33 [ −- 0.46, −- 0.19]−- 0.42 [ −- 0.78, −- 0.10]−- 1.52 [ −- 2.44, −- 0.55]
Appendix
Table 10: The five models on scrambled PhoQ, with 95% intervals. The last three columns are each model’s pick minus the guessing rule’s. The first sets it against that rule’s mean over 1,500 reference seeds (the registered comparison). The other two pair it with the guessing rule on the model’s own seeds, on PhoQ and on GB1.
modelreplaychange in final pick
Gemini 3.1 Prolast round, m=3m=3++ 0.088 [ ++ 0.011, ++ 0.171]
Gemini 3.1 Prolast round, m=4m=4++ 0.090 [ −- 0.004, ++ 0.190]
Gemini 3.1 Proregistered rule on unchanged readings++ 0.004 [ −- 0.028, ++ 0.036]
Gemini 2.5 Flashlast round, m=1m=1++ 0.014 [ ++ 0.002, ++ 0.032]
DeepSeek V3.1last round, m=1m=1++ 0.035 [ ++ 0.016, ++ 0.055]
Gemini 3.1 Prolast two rounds, m=3m=3++ 0.100 [ ++ 0.013, ++ 0.191]
Appendix
Table 11: Change in final pick against each model’s own pick in the confirmation simulation, with 95% intervals.
Figure 8: True value of the final pick against the re-testing share, one point per cell. The two lines are the informed rule (upper, told which candidates are viable) and the guessing rule (lower), spending k=0k=0 to 4 of the four slots each round re-testing their current best, with 95% intervals.
  share (%)  against the guessing rule
modelconditionallexcl. startdistinctpicktestedpickednet
Opus 4.7Blind75.462.811.37.65+0.33+0.33−0.12-0.12+0.21+0.21
 Quantitative78.465.510.47.73+0.25+0.25−0.11-0.11+0.14+0.14
 Mechanism81.668.19.07.69+0.31+0.31−0.14-0.14+0.17+0.17
 Aggregated85.870.47.07.66+0.37+0.37−0.17-0.17+0.20+0.20
 Decision-rule85.370.87.37.66+0.36+0.36−0.16-0.16+0.20+0.20
Appendix
Table 12: Results for each cell, one model under one prompt condition. Share is the re-testing share, the fraction of used slots spent re-testing a candidate already ordered. Excl. start leaves out re-orders of the start, whose reading the agent is given in advance. Distinct is the mean number of different candidates measured per campaign, and pick the mean true value of the final pick. The last three columns compare each cell with the guessing rule at its best setting, which measures 28.94 distinct candidates and reaches 7.87. Tested is how much further the best candidate the agents measured falls short of the global best. Picked is the change in how far their pick falls short of the best they measured, negative where they picked better. Net is the sum, the agents’ shortfall in the true value of the final pick; rounded parts may not add to it exactly.
Figure 9: Re-testing share against distinct candidates measured, one point per cell, with the informed and guessing rules at their best setting and the three scripted lab protocols.
Figure 10: Re-testing after the first repeat when every reading is exact (the marker) and the level implied under noise (the bar), one row per cell. A repeat of an exactly known value returns no new information.
modelexact, statedexact, unstatednoisy, statednoise effect (paired)
Opus 4.724.2%24.2%61.7%+37.5+37.5 pp (SE 5.3)
Gemini 2.5 Flash45.0%50.8%95.0%+50.0+50.0 pp (SE 7.0)
DeepSeek V3.155.8%49.2%94.2%+38.3+38.3 pp (SE 8.3)
Appendix
Table 13: Share of the four chosen measurements that repeat an already-measured candidate.
modelcontrastseedsdiff95% CIHolm pp
Opus 4.7Aggregated695−0.0334-0.0334[−0.0606,−0.0061][-0.0606,-0.0061]0.0873
Opus 4.7Decision-rule695−0.0316-0.0316[−0.0583,−0.0049][-0.0583,-0.0049]0.0873
DeepSeek V3.1Aggregated181−0.0332-0.0332[−0.0757,+0.0092][-0.0757,+0.0092]0.3721
DeepSeek V3.1Decision-rule180+0.0699+0.0699[+0.0271,+0.1126][+0.0271,+0.1126]0.0075 †
Gemini 2.5 FlashAggregated139+0.0067+0.0067[−0.0130,+0.0265][-0.0130,+0.0265]1.0000
Gemini 2.5 FlashDecision-rule139−0.0044-0.0044[−0.0278,+0.0189][-0.0278,+0.0189]1.0000
Appendix
Table 14: The six confirmatory contrasts on the final pick’s true value, paired by seed, with Holm correction across the family. The primary pp comes from a bootstrap that resamples seeds (studentised wild cluster bootstrap, 20,000 draws). The intervals are unadjusted and the Holm pp adjusts for the family of six. Seed counts differ because each model’s cells were sized separately (Section H.1 ). † Did not hold up on 120 new seeds: +0.008+0.008 , 95% bootstrap interval [−0.036,+0.052][-0.036,+0.052] (Appendix J ).
modelconditiongridcheck Acheck Bsame-day rerun minus grid
Opus 4.7Blind76.076.674.8−1.1-1.1
 Mechanism82.181.180.9+0.2+0.2
Gemini 2.5 FlashBlind94.489.883.3−0.1-0.1
 Mechanism95.295.189.0−0.9-0.9
DeepSeek V3.1Blind90.890.490.0−0.9-0.9
 Mechanism93.693.693.5+0.1+0.1
Appendix
Table 15: Re-testing share in the prompt checks, as a percentage of used slots, on the same 60 seeds in every column. Grid is the original campaigns on those seeds. Check A states that the final pick is scored on its true value and corrects the replacement prompt’s order count. Check B also makes a rejected order free. Same-day rerun minus grid is the change, in points, for the original prompts run again on the same day as check A.
Figure 11: The shortfall to the global best in each cell, split into exploration loss and selection loss, which sum to the whole. The guessing rule’s own split, 0.67 to exploration and 0.23 to selection, is the reference point of Figure 4 .
Figure 12: The learnability simulation: mean final pick against kk , the slots per round spent re-testing the current best, with standard errors. Each panel is one setting of ruggedness KK , noise (as a multiple of assay noise) and rule, over 200 generated landscapes. The star marks the best kk . A panel marked tie is one where the paired gap to the runner-up is inside two of its own standard errors.
 Run A, gridRun A, GB1Run A, PhoQRun B, PhoQ
trained pick7.914.792.151.53
re-testing share (%)9.31.82.368.3
minus initialisation++ 0.86 [0.83, 0.89]++ 3.32 [3.24, 3.40]++ 1.49 [1.44, 1.54]++ 0.87 [0.84, 0.91]
minus random++ 0.40 [0.36, 0.44]++ 1.43 [1.34, 1.53]++ 0.65 [0.58, 0.73]++ 0.03 [ −- 0.03, 0.09]
minus guessing rule−- 0.03 [ −- 0.06, 0.01]−- 1.32 [ −- 1.45, −- 1.20]++ 0.07 [ −- 0.01, 0.15]−- 0.55 [ −- 0.62, −- 0.48]
minus informed rule−- 0.05 [ −- 0.10, 0.00]−- 1.16 [ −- 1.26, −- 1.07]−- 0.52 [ −- 0.63, −- 0.41]−- 1.14 [ −- 1.25, −- 1.03]
Appendix
Table 16: The registered test of the trained policy, 1,500 episodes per split, as a mean over three seeds. Run A trained on generated landscapes only; Run B on generated landscapes and scrambled GB1, tested on PhoQ. Differences are in final pick, with 95% intervals from a two-level bootstrap that resamples landscapes and then episodes (post-hoc; the registered intervals resample episodes only).
changeconditionpickshare (%)minus randomminus best scripted rule
noise, gridnoise halved (3.65%)8.119.2++ 0.45 [0.40, 0.49]−- 0.04 [ −- 0.08, 0.00]
 noise doubled (14.6%)7.669.0++ 0.35 [0.31, 0.40]−- 0.09 [ −- 0.14, −- 0.05]
stress, gridbatch 5%7.849.3++ 0.41 [0.36, 0.46]−- 0.13 [ −- 0.18, −- 0.07]
 batch 10%7.739.2++ 0.40 [0.35, 0.45]−- 0.13 [ −- 0.19, −- 0.08]
 floor7.899.3++ 0.40 [0.35, 0.44]−- 0.11 [ −- 0.16, −- 0.06]
 failed wells7.909.5++ 0.39 [0.35, 0.43]−- 0.10 [ −- 0.14, −- 0.06]
Appendix
Table 17: The trained policy (Run A, without retraining) under every change after the registered test, 1,500 episodes per row, at assay noise unless the row changes it. Rows that would repeat Table 16 are left out. The last two columns are Run A’s pick minus random ordering and minus the best scripted rule at its best number of re-tests, on the same episodes. Stress: each round’s readings share one multiplicative factor with this standard deviation (batch), or get an additive noise floor of 2% of the landscape maximum (floor). In the other two, 2% of readings are replaced by a near-zero value (failed wells) or by three times the reading (spikes). Real names: GB1 and PhoQ under their amino-acid names. At 4 rounds of 24 the scripted rules’ re-tests scale with the slots per round, and two of Run A’s input features exceed the range seen in training. Intervals as in Table 16 ; under real names, one landscape each, they resample episodes only.
runsplitshare (%)retries (%)same round (%)top three (%)parent (%)nominee (%)pick cost
Run Agrid9.3621556197++ 0.00 [0.00, 0.00]
Run AGB11.8582946351−- 0.71 [ −- 0.75, −- 0.67]
Run APhoQ2.3612326489−- 0.05 [ −- 0.06, −- 0.04]
Run Bgrid72.035584371−- 0.03 [ −- 0.04, −- 0.02]
Run BGB169.535423593−- 0.06 [ −- 0.07, −- 0.05]
Run BPhoQ68.333393193−- 0.02 [ −- 0.02, −- 0.01]
Appendix
Table 18: What the trained policies do on the test splits, as a mean over three seeds. Share: re-tests as a share of the budget. The next three columns split the re-tests: retries of an order that failed, repeats of a candidate first read earlier the same round, and re-tests of one of the three candidates with the highest mean reading. Parent: new candidates one letter from one of those three. Nominee: how often it is the registered selection rule’s nominee, and its pick minus that nominee’s.
 Run A, seeds 3 to 7Run B, registeredRun B, larger bonus
grid: minus random++ 0.42 [0.37, 0.46]–++ 0.32 [0.28, 0.37]
grid: re-testing share (%)10.3–3.2
grid: seeds above random5 of 5–3 of 3
GB1: minus random++ 1.63 [1.53, 1.72]–++ 1.66 [1.56, 1.76]
GB1: re-testing share (%)1.7–2.7
GB1: seeds above random5 of 5–3 of 3
Appendix
Table 19: Five more seeds of Run A (three to seven), trained and tested as registered. Beside them, Run B retrained with an entropy bonus five times the registered one, tested with the registered code on all three test splits. Run B trains on the grid and GB1, so only PhoQ tests its transfer. Differences in final pick against random ordering, with intervals as in Table 16 . Entropies are at the end of training: means over the last updates, averaged over seeds.
landscapemodelcampaignsmodelRun ARun A minus model
gridOpus 4.7607.62 (75%)7.87 (10%)++ 0.25 [0.14, 0.35]
gridGemini 2.5 Flash1207.70 (94%)7.86 (9%)++ 0.16 [0.11, 0.21]
gridDeepSeek V3.11207.76 (91%)7.86 (9%)++ 0.10 [0.04, 0.15]
gridGPT-6.1 Sol607.67 (44%)7.87 (10%)++ 0.19 [0.07, 0.31]
gridGemini 3.1 Pro607.69 (14%)7.87 (10%)++ 0.18 [0.07, 0.29]
GB1, scrambledOpus 4.7604.78 (63%)4.79 (2%)++ 0.02 [ −- 0.37, 0.42]
Appendix
Table 20: The trained policy (Run A, without retraining) on each language model’s own campaigns, the same landscape, campaign seeds, noise and budget. Pick, with re-testing share; the last column is Run A’s pick minus the model’s, paired by campaign, with a 95% bootstrap interval over campaigns. Each of these landscapes is outside the training and validation splits.
landscapeconditionguessing, best kguessing, three minus bestinformed, best kinformed, three minus best
gridbatch 5%1 (24%)−- 0.19 [ −- 0.22, −- 0.15]1 (23%)−- 0.17 [ −- 0.21, −- 0.13]
gridbatch 10%1 (24%)−- 0.15 [ −- 0.19, −- 0.11]1 (23%)−- 0.14 [ −- 0.18, −- 0.11]
gridfloor1 (24%)−- 0.17 [ −- 0.21, −- 0.14]1 (23%)−- 0.17 [ −- 0.20, −- 0.14]
gridfailed wells1 (24%)−- 0.22 [ −- 0.26, −- 0.17]1 (23%)−- 0.21 [ −- 0.25, −- 0.17]
GB1batch 5%0 (1%)−- 2.21 [ −- 2.34, −- 2.08]0 (0%)−- 1.58 [ −- 1.70, −- 1.47]
GB1batch 10%0 (1%)−- 2.16 [ −- 2.30, −- 2.03]0 (0%)−- 1.48 [ −- 1.59, −- 1.36]
Appendix
Table 21: The guessing and informed rules under the stress conditions of Table 17 that keep their best number of re-tests, 1,500 episodes per landscape and condition, at assay noise. Without stress their best is 1 re-test a round on the grid and 0 on GB1. Best k is the best number of re-tests a round, with its re-testing share. Three minus best is the pick of three re-tests a round minus that of the best, with intervals as in Table 16 . Four re-tests a round are lower still in every row.
More in this topic

AI Agents for Scientific Discovery

AI agents for scientific discovery use language models, tools, and experiments to generate, test, and refine scientific knowledge.

See all 270 papers

Explore similar work

CardsList

  1. Exploration and Exploitation Errors Are Measurable for Language Model Agents

    Apr 14, 2026Jaden Park, Jungtaek Kim, Jongwon Jeong +3Exploration-Exploitation TradeoffLLM Agent Evaluation

    University of Wisconsin–Madison · KRAFTON · Ludo Robotics

  2. Sample, Mix, or Verify? Allocating Test-Time Compute under Imperfect Verifiers

    Oct 16, 2026Srinivasan IyengarInference-Time Compute AllocationLanguage Model Performance Prediction

    Microsoft Research India

  3. Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate's Log-Probabilities

    Oct 2, 2026Yikai Zhao, Saurabh Pandey, Pradeep Kumar MisraLLM Agent ReliabilityLLM Auditing

    Amazon