Organizations: School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai 519087, China · College of Computing and Data Science, Nanyang Technological University, Singapore
Web agents increasingly edit images by calling open editing models in a critique-and-retry loop, and these models differ in cost by more than an order of magnitude. After every rejected edit, a service must decide whether to retry the same model, escalate to another or give up. Existing routers pick one model per request, and cascades escalate in a fixed order. However, both treat a retry as an independent draw. We identify sticky failures, meaning that a retry after a failure succeeds only 17-53% as often as a first try across six open models, three benchmarks and two judges. Based on this finding, we propose RetryRoute, a retry-aware router that estimates each model's chance of success given the request's failure history. It plans retries, escalations and abandonment over failure states and adds live queue waits to the cost, so the same plan becomes load-aware dispatch. At the main budget, RetryRoute's success under the in-loop judge exceeds the strongest of eleven prior baselines by 2.8 points on average over eight groups. On a GPU serving testbed it raises the share of edits delivered within the service-level objective by 5.0-6.3 points over the best prior baseline at every load.
Figures & tables
Figure 1. Sticky failures. A model that fails an edit mostly fails it again, while a different model often succeeds. Top: grouped bars; for each of six editing models the success rate of a retry after a failure is 0.23 to 0.39 times the first-try rate per model, while switching to another model of the pool succeeds more often than the retry. Bottom: a sign edit where FLUX.2 klein fails twice and Qwen Lightning succeeds.
Figure 2. Four findings behind RetryRoute. Panels a and b use English GEdit-Bench with bands over four pool-1 groups, and panel c replays the logged edits under load on 3 GPUs. Four panels: the retry pass rate of the cheap model collapses after a failure while other models stay near 30 to 45 percent; its cost per success crosses the next model's after one or two failures; fixed plans collapse at different loads; late passes are confirmed less often by a single judge.
Figure 3. Overview of RetryRoute. It (a) estimates p(a∣s) offline, (b) plans over failure states and (c) re-plans with queue waits. In the bottom row a5 fails twice and RetryRoute escalates to a4. System diagram: offline estimation of p(a|s) from logged attempts, the RetryRoute planner over failure states with a Bellman equation and LinUCB variant, load-aware re-planning from queue waits, GPU workers with an ensemble judge, and the running example of three attempts at the bottom.
FLUX.1 Kontext-dev ( Black Forest Labs and others, 2025 )
28
60.2
a2
a1 + Turbo-Alpha ( Alimama Creative, 2024 )
8
25.7
b1
VIBE ( Alekseenko et al., 2026 )
20
0.9
Table 1. Editing models, called arms, and GPU seconds per edit, with pool 1 above the rule and pool 2 below.
Pool 1 (50 s)
Pool 2 (120 s)
Method
GEdit-CN
GEdit-EN
EN 5 arms
GEdit-v2
ImgEdit
GEdit-CN
GEdit-EN
ImgEdit
Oracle, per-item hindsight
92.466
91.553
93.151
88.261
87.778
87.671
88.584
92.778
Retry or route one model
Same-model retry
81.735
81.735
82.648
80.435
78.889
77.397
73.973
79.444
Rewrite-retry, blind
80.822
82.420
84.703
79.565
81.667
77.397
73.973
77.778
Rewrite-retry, critique
78.311
81.279
83.333
77.391
80.000
77.397
73.973
79.444
Table 2. Offline success rate (%) at the main budget of 50 s for pool 1 and 120 s for pool 2.
Figure 4. RetryRoute against the best of eleven prior baselines at each budget, with the main budget as a vertical line. Eight panels in four rows and two columns, one per group, five of pool 1 (P1) and three of pool 2 (P2), of success rate against GPU seconds per request; the RetryRoute curve lies on or above the envelope of the prior baselines at most budgets.
Figure 5. Lead of RetryRoute in APGR and AUC over four prior baselines in each group. Two bar panels over the eight groups, five groups of pool 1 and three of pool 2, with four bars per group for the lead of RetryRoute over RouteLLM-MF, CA-Routing, FrugalGPT and AutoMix. Every bar is above zero, from about 0.07 to 0.37 in APGR and from about 1.2 to 8.9 points in AUC.
Figure 6. How requests end at the main budget for pool 1 and pool 2, with success in percent on top of each bar. Stacked bars per method and pool: share of requests passing on the first try, after a same-model retry, or after escalation.
Figure 7. Lead of RetryRoute over the strongest prior baseline in each GEdit-Bench edit category, pooling Chinese and English, with 95% CIs. Grouped bars per edit category for the two pools; the lead is positive in most categories.
Figure 8. What each method delivers at the main budget on two pool-1 GEdit-Bench requests in English, framed blue for a pass and pink for a failure, with the arm and attempt on each badge. Rows of requests; for each, the input and the image delivered by six prior baselines and RetryRoute, framed blue for a pass and pink for a failure.
Figure 9. Sensitivity in points. Panels a and b give the change from the default RetryRoute, and panels c and d the lead over the best prior baseline. Four panels: shrinkage, attempt cap, judge threshold and training-data fraction.
In-domain
Unseen category
Cross-benchmark
Pool
Test set
Prior
Ours
Prior
Ours
Prior
Ours
1
GEdit-CN
84.25
88.81
84.25
87.67
85.16
87.56
1
GEdit-EN
83.79
88.81
83.79
88.81
84.47
88.01
1
EN 5 arms
87.44
90.64
87.44
90.87
–
–
1
GEdit-v2
83.04
83.91
82.61
84.35
83.26
83.70
1
ImgEdit
85.00
85.00
83.89
85.56
83.89
85.28
Table 3. Success (%) of Ours and the best prior baseline (Prior) in-domain, on unseen categories and across benchmarks.
Figure 10. Online learning from a cold start and after a shift. Learning curves of the offline and online planners and of LinUCB with and without retry history.
Steady 0.045
Steady 0.055
Bursty 0.05
Method
Good.
p95
Good.
p95
Good.
p95
Queue thr.
85.00
218.1
85.00
190.6
84.38
210.2
Cascade a5 → a4
81.25
167.4
80.00
248.4
78.75
277.5
Fixed ladder
80.62
447.3
76.88
777.4
81.25
361.7
HADIS
76.25
631.6
75.62
945.2
75.00
1064.3
Retry a5 × 8
73.12
869.7
70.62
1479.8
58.13
1938.6
Table 4. Testbed goodput (%) and p95 latency (s) at three loads, with p95 in italics when it exceeds the 300 s SLO.
Figure 11. Load replay against testbed over 30 runs on the same request streams, with latency in seconds on log axes. Scatter plots of replayed against measured success and latency, close to the equality line.
Figure 12. Goodput in load replay and the highest load sustained at each success floor within the service-level objective. Left: goodput curves over arrival rate. Right: bars of the highest sustainable arrival rate of RetryRoute and the best fixed ladder at four success floors.
Figure 13. Highest load sustained in load replay at each success floor within the service-level objective on three GPUs, under bursty arrivals and an arrival trace. Two bar panels over success floors of 78, 80, 82 and 84 percent. In each group the RetryRoute bar is the tallest or tied, above the best queue threshold, the best fixed ladder and the DP without load awareness, with ratios over the better heuristic from 1.0 to 1.9 written above the groups.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
EditScore
EditReward
Group
first
retry
ratio
first
retry
ratio
P1 GEdit-CN
65–73
16–30
24–41
66–75
15–22
22–29
P1 GEdit-EN
70–75
14–29
21–39
68–76
16–22
24–30
P1 EN 5 arms
48–75
14–29
21–42
41–79
11–23
24–37
P1 GEdit-v2
68–73
18–30
26–43
65–70
17–22
25–32
P1 ImgEdit
64–72
22–37
34–52
63–74
13–24
21–32
Appendix
Table 5. First-try and retry pass rates and their ratio in percent, as ranges over the arms of each group.
Figure 14. Cost–success frontiers of each method. Eight panels with one cost–success frontier per method.
Policy (load in req/s)
0.04
0.05
0.06
burst 0.055
worst
Queue thr. @0.045
80.50
77.60
49.30
61.70
49.30
Queue thr. @0.055
80.50
80.10
78.80
76.30
76.30
Queue thr. @burst 0.05
79.70
79.70
79.00
78.40
78.40
Ladder @0.045
80.70
79.60
74.80
73.50
73.50
Ladder @0.055
80.40
80.00
78.60
76.60
76.60
Ladder @burst 0.05
79.40
78.50
76.80
77.20
76.80
Appendix
Table 6. Goodput (%) in load replay when a policy tuned at load L, written @L, runs at other loads.
Judge
Agreement (%)
κ
AUC
FA (%)
FR (%)
In-loop ensemble
83.70 [79.95, 87.39]
0.621
0.935
2.61
13.69
EditScore
75.93 [72.48, 79.16]
0.450
0.853
5.74
18.33
EditReward
78.65 [75.14, 82.05]
0.512
0.890
4.37
16.97
VIEScore
79.53 [73.65, 85.25]
0.520
0.853
5.04
15.43
Appendix
Table 7. Agreement of each judge with the majority vote of two annotators, reweighted to all attempts, with 95% CIs.
Round limit R
Pool
Test set
Cost
1
2
3
4
1
GEdit-CN
60.12
+17.35
+10.05
+7.76
+6.85
1
GEdit-EN
65.49
+15.98
+9.82
+6.39
+5.02
1
EN 5 arms
97.44
+13.47
+5.48
+3.42
+2.97
1
GEdit-v2
47.18
+14.78
+11.74
+7.39
+5.65
1
ImgEdit
24.04
+5.56
+3.33
− 0.56
− 0.56
Appendix
Table 8. Lead of RetryRoute in points over OctoT2I at OctoT2I’s own cost for each round limit R , with the cost of OctoT2I at R=1 in GPU seconds per request.
EditScore in loop
EditReward in loop
Pool
Test set
Prior
Ours
Prior
Ours
1
GEdit-CN
83.56
84.93
82.65
84.70
1
GEdit-EN
85.62
87.44
82.65
82.88
1
EN 5 arms
85.39
87.90
84.70
87.21
1
GEdit-v2
85.65
90.43
78.70
81.74
1
ImgEdit
87.78
89.44
82.78
82.22
Appendix
Table 9. Success (%) at the main budget of the strongest prior baseline and of RetryRoute when a single judge replaces the ensemble in the loop.
Adaptive model routing selects models or specialized experts per input to balance prediction quality, computational cost, latency, and resource constraints.
Instruction-based image editing models exhibit pronounced sample-specific performance variation: different models can produce substantially different results for the same source image and editing instruction, and the best-performing model often varies across requests. These models also differ substantially in inference latency, making fixed-model deployment inefficient. We therefore study image editing model routing, which dynamically selects an editor from a candidate pool given an input image and editing instruction. We introduce EditJev, a System One routing model that directly predicts sample-level preference scores over candidate image editing models in a single forward pass. A lightweight preference head on a vision-language backbone produces a fixed-dimensional score vector without autoregressive model-name generation or response parsing. To train EditJev, we construct pairwise preferences from offline candidate outputs and optimize a pairwise ranking objective to learn relative model suitability. At inference time, EditJev combines the predicted preference scores with model-level speed scores to balance editing quality and generation latency. Experiments on ImgEdit and GEdit-En show that EditJev improves aggregate quality over the strongest fixed editor, from 4.540 to 4.592 and from 8.194 to 8.255, respectively, while reducing estimated end-to-end latency by 60.8% and 31.7%. These results demonstrate the effectiveness of direct, structured model routing for efficient image editing.
Zhile Guan, Xu Jiang, Yule Duan +4
School of Electronic and Computer Engineering, Peking University · Tsinghua University
Agentic image editors decompose complex requests into steps, but when a step fails, they mostly retry with the same narrow set of tools and a judge that scores each output by eye directly. We argue that such agents are bounded by two spaces: the action space of what the agent can do, and the observation space of what its judge can see. On our expert-annotated benchmark, a general-purpose editor yields no usable result in over 40% of the cases, and no judge is reliable across edit types. We present EditAgent, which expands both spaces. It routes each step to general editors and specialized methods, and attaches to each candidate a verification record. With these records, off-the-shelf VLM judges outperform finetuned reward models. Candidates are generated in parallel and judged in one listwise call: experts prefer this to sequential refinement in 78.5% of decided comparisons, at no more than half the judge calls. EditAgent improves two open-weight editors on three benchmarks and serves as a data engine for training stronger editors.
Small language models are increasingly deployed on edge devices and robotic systems to meet cost and latency constraints, yet their limited capabilities can make reliable deployment challenging. Stronger models can provide assistance when needed, but deciding when to route a weak-model response to a stronger model often relies on signals of response correctness. In open-ended and agentic tasks, however, obtaining such signals can require costly ground-truth supervision that is difficult to obtain at scale. We revisit the foundations of weak-to-strong routing and ask whether routing decisions can instead be guided by signals that do not require ground-truth supervision. We first characterize the optimal weak-to-strong routing policy, while allowing for multiple weak-model responses generated through local inference. Building on this characterization, we derive a principled approximation to the optimal routing based on model-to-model comparison. In particular, we introduce a notion of model disagreement that can be measured directly from model-generated responses, without ground-truth supervision. Importantly, the approximation becomes increasingly accurate as the strong model becomes a better proxy for the ground-truth distribution. Across agentic and single-turn tasks, our approach effectively navigates the trade-off between routing cost and response accuracy without ground-truth supervision. Moreover, even when ground-truth correctness signals are available for training, disagreement-based routing can outperform methods trained directly to predict correctness