back

prompt placement dominates retention

a harness audit of gui-agent screenshot history. eccv, paper id #10. dm me for the pdf.

short version: i set out to compare four rules for deciding which past screenshots a gui agent should keep, and found they were all the same. then i found out why the comparison could not have told me anything either way. where the prompt was rendered mattered about twenty times more than any of the rules i was measuring.

the setup

a gui agent takes screenshots as it works, and that history gets expensive fast. the literature proposes rules for what to throw away: keep what looks like the instruction, re-weight by space and time, keep the most recent, or keep at random. i ran all four at a matched 20% budget with zero-shot ui-tars-1.5-7b on 23 gui-odyssey episodes, plus a full-history arm and a zero-history arm as bookends.

Six arms of the audit, from full history to zero history, all landing between 24.7 and 25.6 percent step accuracy: a 0.9 point spread. step accuracy, six arms, 344 shared steps random (20%) 25.6 full history (3 frames) 25.3 instruction similarity 25.0 spatio-temporal 25.0 fixed recency 25.0 zero history (0 frames) 24.7 0% bars start at 0, spread is 0.9 points
a six-way tie. the arm that keeps every frame and the arm that keeps none of them land within a point of each other. no difference against random was significant.

why the tie was not the finding

a tie is only interesting if the thing you removed was doing work. mine reached 24.0% step accuracy on a benchmark where a third-party reproduction reported 32.9%, and a gap that size means the harness, not the retention rule, is setting the ceiling. so i stopped comparing rules and started auditing my own setup, one variable at a time, with the template and parser held fixed.

the variable that moved was prompt placement. rendering the same prompt in the system role instead of in the user turn lifted official-evaluator accuracy from 17.4% to 39.1% on 46 identical steps. a control that changed only the ordering, keeping the user role, reached 23.9%, so ordering alone recovers about a third of the gain and the role does the rest.

On 46 identical steps, the prompt in the user turn scores 17.4 percent, reordering within the user turn scores 23.9 percent, and moving the prompt to the system role scores 39.1 percent. official step accuracy, 46 identical steps, one variable changed prompt in user turn 17.4% reordered, still user turn 23.9% prompt in system role 39.1% 0% 40% 21.7 points from moving one prompt. 0.9 points separated every retention rule.
same model, same template, same parser, same 46 steps. the only edit is which role the prompt is rendered in. 7:0 discordant steps, sign test p = 0.016.

what the repaired harness says about history

with the placement fixed, the configuration reaches 45.5% at scale over 378 steps. that is the version worth asking the original question of, so i re-ran full history against zero history on it. removing every historical screenshot, verified token-exact, shifted accuracy by 1.3 points down, with a 95% confidence interval running from 12.4 points down to 9.7 points up and mcnemar p = 0.72. keeping those frames cost 2.6x peak memory and 3.0x latency.

Full history versus zero history on the repaired harness: peak GPU memory 48.2 gigabytes against 18.6, and median step latency 17.7 seconds against 5.9, for an accuracy difference of 1.3 points. what those screenshots cost, repaired harness, 378 steps peak gpu memory 48.2 gb 18.6 gb median step latency 17.7 s 5.9 s black: full history (3 frames). grey: zero history. accuracy difference: 1.3 points, p = 0.72.
the screenshots are not free and, on this harness, not clearly worth it either.

what i am not claiming

this is a negative result and it is bounded like one. the study is single-seed, greedy decoding, one model, one benchmark. with 23 episode clusters the minimum detectable effect is 16.4 points, so a real 6.1-point history effect reported elsewhere would not show up here even if it exists. i did not detect that effect. that is different from ruling it out, and about 170 episodes would be needed to say anything stronger.

the point

a tie between methods is not evidence that the methods are equivalent. it can just as easily mean the harness is broken in a way that swamps them, and the way to tell the difference is to show that the input you are compressing changes the answer at all. a nuisance variable nobody reports, where the prompt goes, was worth an order of magnitude more than every effect under study.