acceptodds
Under review as a conference paper at ICLR 2027

PACT-GUI: Efficient GUI Agents with Action-Anchored Global–Local History

Abstract

GUI agents for multi-step tasks repeatedly encode high-resolution visual histories, incurring substantial inference costs. While token and KV-cache compression can reduce downstream computation, they leave the cost of visual encoding largely unchanged. However, downsampling historical screenshots can obscure local state changes, while visual redundancy-based pruning may discard seemingly similar regions that remain critical for action prediction. Our analysis shows that GUI histories are globally redundant but locally heterogeneous, and that visually similar content can still affect action prediction. These findings motivate PACT-GUI, a training-free framework that uses executed actions as spatial anchors for history compression. Specifically, before visual encoding, we represent each historical screenshot as a low-resolution global thumbnail paired with a high-resolution crop around the corresponding interaction. To achieve end-to-end acceleration, we pair this representation with template-based structured decoding, which separates the fixed output template from its variable fields, allowing the model to process known syntax in blocks and autoregressively predict only the action type and its parameters. Across four GUI benchmarks, PACT-GUI achieves a $6.62\times$ reduction in mean end-to-end computation, demonstrating substantial efficiency gains.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
GUI agents, Multi-Step GUI Interaction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fhjxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to reduce repeated processing of historical screenshots by GUI agents. Before visual encoding, PACT-GUI replaces each historical screenshot with a global thumbnail and a local crop anchored to the executed action, while retaining the current screen at its original resolution. It also uses template-based structured decoding. The method requires no training. Across four benchmarks of offline next-action prediction, mean computation falls to approximately 1/6.62 of the baseline. On AndroidControl-Curated, the full method improves step success from 55.20% to 63.90% and reduces per-instance inference time from 5.540 to 3.092 seconds.

Critical strengths
  1. Compressing history before visual encoding while retaining global layout and action-local detail is simple and training-free. It directly reduces visual processing that later token or KV compression cannot eliminate, giving the method practical value.
  2. Table 2 measures History Packets and structured decoding separately and supports their acceleration of different stages. History Packets alone improve AndroidControl step success from 55.20% to 59.80%, and the complete system achieves measured inference speedups. This provides evidence on both computation and measured runtime.
Critical weaknesses
  1. Action anchoring is the central design choice, but the ablations compare thumbnails, local crops, and their combination without directly comparing other region-selection strategies at the same budget. The results therefore do not establish that action information itself improves on a generic global–local representation, limiting the evidence for the core methodological contribution.
  2. Main-table quality comparisons combine differences in compression and decoding. Structured decoding alone outperforms the full system on AndroidControl and AITW, while compression baselines have not been compared under a shared decoding protocol. The full system's quality advantage therefore does not directly establish that its history representation is better than existing compression methods.
Questions for authors
  1. With image sizes, token budgets, and decoding fixed, how does action anchoring compare with center crops or other region-selection strategies?
  2. Does the relative quality–cost ranking persist when all compression baselines use the same structured decoding?
3 comments

We sincerely thank you for the constructive comments and valuable suggestions. We address each concern in detail below.

W1Q1: Controlled Comparison of Crop-Selection Strategies

Thank you for highlighting this question. To isolate the effect of region selection, we compared action-anchored crops with center, random, and interface-change-based crops on AndroidControl. The interface-change policy selects a window with high pixel differences between the historical screenshot and its successor. For each historical screenshot, all four strategies use the same crop dimensions and output resolution, and thus the same estimated local visual-token budget; the global thumbnail, model, and decoding settings are also fixed. The strategies differ only in the location of the local crop.

We first used the compact configuration in the paper, which was designed to prioritize acceleration. For example, given a 1080 × 2400 screenshot, a crop based on an action point and not clipped by the image boundary covers approximately 238 × 238 pixels, or 2.19% of the original image area. The crop is then resized to 224 × 224 pixels. Under this setting, action anchoring achieves a success rate of 59.7%, matching random cropping and outperforming center cropping by 0.2 percentage points, but falling 0.2 points below interface-change-based cropping. Thus, the original configuration does not establish that action anchoring outperforms all alternative strategies.

We hypothesized that the small local region retained by the original configuration might limit the contextual information available to each crop and make the benefit of action anchoring harder to observe. To investigate this possibility, we progressively increased the crop coverage using two additional configurations, Large and Huge. For the example screenshot and action-point crop described above, Large expands the region selected from the original image to approximately 432 × 432 pixels, or 7.20% of its area; Huge expands it to approximately 724 × 724 pixels, or 20.22% of its area. The resulting crops are then resized to 448 × 448 and 704 × 704 pixels, respectively. Actual crop coverage varies with the action type, action bounding box, and crop location; these values illustrate the scale of the three configurations. The size configuration is:

ConfigurationOriginal screenshot sizeCrop size in original screenshotPercentage of original image areaResized crop size
PACT-GUI1080 × 2400238 × 238≈ 2.19%224 × 224
Large1080 × 2400432 × 432≈ 7.20%448 × 448
Huge1080 × 2400724 × 724≈ 20.22%704 × 704

The SR results of different crop strategy on AndroidControl are shown below.

Crop-selection strategyPACT-GUILargeHuge
Action-anchored59.760.159.6
Center59.559.559.1
Random59.759.959.3
Interface change59.959.959.2

With greater crop coverage, action anchoring achieves the highest observed success rate in both the Large and Huge configurations, exceeding the strongest alternative in each configuration by 0.2 and 0.3 percentage points, respectively. This provides direct evidence for the effect of action anchoring: under these two configurations, the choice of local region affects the performance of strategies that otherwise share the same multiscale representation. The margins are small, however, and repeated runs or uncertainty estimates would be needed to assess their stability.

When moving from Large to Huge, the success rates of all four strategies decrease. This indicates that retaining more local content does not necessarily improve performance and is consistent with the possibility that excessive context introduces redundancy. The paper uses the more compact crop configuration to prioritize inference acceleration; the additional experiments illustrate the trade-off among crop selection, local context, and efficiency.

W2Q2: Decoding fairness and comparison under a shared SFD protocol

We thank you for highlighting the importance of fair comparisons in decoding. In the original Table 1 in paper, all baseline methods were evaluated within our unified evaluation framework using the same Qwen3-VL backbone and standard autoregressive decoding, without SFD. This design reflects the fact that PACT-GUI is an end-to-end acceleration framework comprising two complementary components: Action-Conditioned History Packets (HP) and Structured Fast Decoding (SFD). Accordingly, Table 1 is intended to compare PACT-GUI as a complete end-to-end system against the corresponding baseline systems, while Table 2 separately evaluates the contributions of HP, SFD, and their combination. Nevertheless, we acknowledge that the end-to-end comparisons in Table 1 alone do not establish the superiority of HP over alternative compression methods under identical decoding settings. In particular, the fact that SFD-only performs better than HP+SFD on AndroidControl and AITW in Table 2 makes such a controlled comparison necessary.

We therefore add a shared-decoder experiment to explicitly isolate the effect of the decoding protocol. Specifically, we apply the same SFD decoding protocol to all baselines and PACT-GUI while keeping all other evaluation settings unchanged. Given the rebuttal timeline, we conduct this additional evaluation on AndroidControl-Curated High, AITW, and Mind2Web. The results are shown below.

MethodAndroidControl SR ↑AndroidControl TFLOPs ↓AITW Avg. Acc. ↑AITW TFLOPs ↓Mind2Web SR ↑Mind2Web TFLOPs ↓Avg. Acc. ↑Avg. TFLOPs ↓
GUI-KV59.30184.8848.9437.8147.5252.1251.9291.60
ST-Lite59.00184.6148.4037.5346.1351.8651.1891.33
HistPrune58.9097.1847.8321.0841.8527.0049.5348.42
GUIPruner57.5042.4547.2510.6840.2613.2448.3422.12
FastV59.10131.8247.6025.5345.1236.5350.6164.63
SparseVLM57.80108.7148.0125.7543.8732.9649.8955.81
LLaVA-PruMerge32.5092.2944.4622.1614.2928.0930.4247.51
DivPrune43.6091.7144.8322.8716.6928.1935.0447.59
PACT-GUI63.9055.7250.9514.9444.4417.2353.1029.30

The results show that, under the shared SFD protocol, PACT-GUI still achieves the highest average accuracy (53.10), outperforming the strongest baseline, GUI-KV (51.92), and other history-compression methods, e.g., HistPrune (49.53).

In the revised manuscript, we will explicitly state the decoding protocol used by each baseline in Table 1 and include this shared-SFD comparison in the main paper or appendix. These results demonstrate that PACT-GUI’s gains are not solely attributable to SFD: under the same structured decoder, HP still provides a consistent overall benefit.

Thank you for your detailed responses. After considering your rebuttal, I have decided to maintain my original score.

Reviewer_pzfbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the inference cost of processing high-resolution visual histories in multi-step GUI agents. The authors propose PACT-GUI, a training-free framework that replaces each historical screenshot with a global thumbnail and a local crop anchored to the corresponding historical action before visual encoding, while preserving the current screenshot at full resolution. The framework also incorporates an existing template-based structured decoding method that processes fixed syntax in blocks and autoregressively generates only variable fields. Experiments using Qwen3-VL-4B-Instruct on four GUI benchmarks report a 6.62× reduction in mean end-to-end FLOPs relative to the uncompressed baseline, with competitive quality–computation trade-offs.

Critical strengths

Compressing historical screenshots before visual encoding reduces both the cost of encoding historical images and the sequence length processed by subsequent multimodal layers. The training-free design and preservation of the full-resolution current screenshot also make the approach relatively straightforward to integrate into existing agents.

Critical weaknesses
  1. The benefit of action anchoring lacks direct validation. Section 3.2 and Appendix A motivate preserving local regions but do not establish that action-centered crops outperform other selection strategies. Likewise, Figure 4 evaluates representation components rather than crop-selection policies. Comparing action-anchored, central, random, and interface-change-based crops under otherwise identical settings would help distinguish action anchoring from multiscale compression more generally.

  2. The main comparison does not fully isolate the historical representation’s contribution. Table 2 shows that SFD alone achieves higher accuracy than the full method on AndroidControl and AITW, although the full method further reduces latency. Meanwhile, Table 1 compares methods with different decoding protocols and screenshot-compression policies: competing methods may compress the current screen, whereas PACT-GUI preserves it at full resolution. Matched comparisons are needed to disentangle the effects of historical compression, structured decoding, and current-screen preservation.

Questions for authors

please refer to the paper weakness.

2 comments

We sincerely thank you for the constructive comments and valuable suggestions. We address each concern in detail below.

W1: Direct Validation of Action-Anchored Crop Selection

Thank you for this suggestion. We agree that the representation-component ablation in Figure 4 does not directly establish whether action-anchored crops outperform other crop-selection strategies. We therefore conducted an additional comparison on AndroidControl, AITW, Mind2web using four strategies: action-anchored, center, random, and interface-change-based cropping. The interface-change policy selects a window with high pixel differences between the historical screenshot and its successor. For each historical screenshot, the four strategies use crops of the same dimensions and keep the remaining settings unchanged; only the crop location varies.

We first used the compact configuration in the paper, which was designed to prioritize acceleration. For example, given a 1080 × 2400 screenshot, a crop based on an action point and not clipped by the image boundary covers approximately 238 × 238 pixels, or 2.19% of the original image area. The crop is then resized to 224 × 224 pixels. In the full-run measurements, this history compression retained approximately 10.29% and 6.95% of the historical visual tokens on AITW and Mind2Web, respectively, relative to using full historical screenshots. Under this setting, action anchoring achieves the best success rate on AITW and Mind2Web, but on AndroidControl, the success rate is 59.7%, falling 0.2 points below interface-change-based cropping. Detailed results are shown below:

Crop-selection strategyAndroidControlAITWMind2Web
Action-anchored59.746.8444.06
Center59.546.5140.89
Random59.746.2940.02
Interface change59.946.7540.69

We hypothesized that the super small local region retained by the original configuration for AndroidControl might limit the contextual information available to each crop and make the benefit of action anchoring harder to observe. To investigate this possibility, we progressively increased the crop coverage using two additional configurations, Large and Huge. For the example Large expands the region selected from the original image to approximately 432 × 432 pixels, or 7.20% of its area; Huge expands it to approximately 724 × 724 pixels, or 20.22% of its area. The resulting crops are then resized to 448 × 448 and 704 × 704 pixels, respectively. Actual crop coverage varies with the action type, action bounding box, and crop location; these values illustrate the scale of the three configurations. The size configuration is:

ConfigurationOriginal screenshot sizeCrop size in original screenshotPercentage of original image areaResized crop size
PACT-GUI1080 × 2400238 × 238≈ 2.19%224 × 224
Large1080 × 2400432 × 432≈ 7.20%448 × 448
Huge1080 × 2400724 × 724≈ 20.22%704 × 704

The SR results of different crop strategy on AndroidControl are shown below.

Crop-selection strategyPACT-GUILargeHuge
Action-anchored59.760.159.6
Center59.559.559.1
Random59.759.959.3
Interface change59.959.959.2

With greater crop coverage, action anchoring achieves the highest observed success rate in both the Large and Huge configurations, exceeding the strongest alternative in each configuration by 0.2 and 0.3 percentage points, respectively. This provides direct evidence for the effect of action anchoring: under these two configurations, the choice of local region affects the performance of strategies that otherwise share the same multiscale representation.

When moving from Large to Huge, the success rates of all four strategies decrease. This indicates that retaining more local content does not necessarily improve performance and is consistent with the possibility that excessive context introduces redundancy. The paper uses the more compact crop configuration on AndroidControl to prioritize inference acceleration; the additional experiments illustrate the trade-off among crop selection, local context, and efficiency.

W2: Disentangling History Compression, Structured Decoding, and Current-Screen Preservation

(1) Screen-shot compression policy (historical compression and current-screen preservation):

We thank you for raising this important concern. We would like to clarify that the baselines in Table 1 are not exclusively designed for historical compression. Several methods perform general visual-token pruning or KV-cache compression across the multimodal input. Restricting all of them to historical tokens would alter their original operating mechanisms and would therefore not constitute a faithful comparison of the complete methods.

Accordingly, our main comparison is intended to evaluate end-to-end quality–efficiency trade-offs under each method's original compression strategy, rather than to isolate historical compression alone. In practice, we maintained approximately matched overall token retention ratios across all methods, as detailed in Section 5.2 of the original paper.

(2) decoding protocols:

We thank you for highlighting the importance of fair decoding comparisons. In the original Table 1, all baseline methods were evaluated using the same Qwen3-VL backbone and the same standard autoregressive decoding procedure, with SFD disabled. Thus, Table 1 compares visual-history compression methods under a common conventional decoding setting. Table 2 separately evaluates HP, SFD, and their combination.

We agree that the original end-to-end comparison does not, by itself, establish that HP is superior to alternative compression methods under an identical structured decoding setting. In particular, the fact that SFD-only performs better than HP+SFD on AndroidControl and AITW in Table 2 makes such a controlled comparison necessary.

We therefore add a shared-decoder experiment to explicitly isolate the effect of the decoding protocol. Specifically, we apply the same Structured Fast Decoding (SFD) protocol to all baselines and PACT-GUI, while keeping the backbone, prompts, evaluation protocol, and all other non-compression settings unchanged. We conduct this additional evaluation on AndroidControl-Curated High, AITW, and Mind2Web. TTFLOPs denotes total FLOPs per sample, and the average is computed over these three benchmarks.

MethodAndroidControl SR ↑AndroidControl TFLOPs ↓AITW Avg. Acc. ↑AITW TFLOPs ↓Mind2Web SR ↑Mind2Web TFLOPs ↓Avg. Acc. ↑Avg. TFLOPs ↓
GUI-KV59.30184.8848.9437.8147.5252.1251.9291.60
ST-Lite59.00184.6148.4037.5346.1351.8651.1891.33
HistPrune58.9097.1847.8321.0841.8527.0049.5348.42
GUIPruner57.5042.4547.2510.6840.2613.2448.3422.12
FastV59.10131.8247.6025.5345.1236.5350.6164.63
SparseVLM57.80108.7148.0125.7543.8732.9649.8955.81
LLaVA-PruMerge32.5092.2944.4622.1614.2928.0930.4247.51
DivPrune43.6091.7144.8322.8716.6928.1935.0447.59
PACT-GUI63.9055.7250.9514.9444.4417.2353.1029.30

Under the shared SFD protocol, PACT-GUI achieves the highest average accuracy (53.10) with the second-lowest average TFLOPs (29.30). It lies on the quality–cost Pareto frontier alongside GUIPruner, which achieves lower computational cost but substantially lower accuracy (48.34). These results demonstrate that PACT-GUI's gains are not solely attributable to SFD, as HP maintains a superior quality–cost trade-off under identical decoding settings.

Reviewer_R1XhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PACT-GUI is a training-free framework that compresses historical screenshots into global thumbnails and local crops anchored to past actions, while preserving the current screen. It also adopts structured decoding to reduce output-generation cost. Across four GUI benchmarks, the paper reports a 6.62× reduction in mean end-to-end FLOPs.

Critical strengths

The proposed representation is simple, requires no additional training, and reduces historical visual input before encoding while preserving the full current screenshot for grounding. The paper provides direct evidence that the history representation offers a practical benefit beyond the adopted decoder. The paper evaluates four benchmarks and eight baselines. Component ablations help distinguish the effects of history compression and structured decoding.

Critical weaknesses

Although Table 2 separates HP and SFD, Table 1 does not establish whether competing methods use the same decoder. While SFD alone achieves higher accuracy than the full system on ablated datasets, it is necessary to have a shared-decoder comparison to establish HP’s advantage over competing compression methods.

The experiments permit baselines to compress current-screen tokens, whereas PACT-GUI preserves them and also uses more FLOPs than GUIPruner on every benchmark. GUIPruner’s separate current/history retention settings are unclear. A TAR-only comparison with full current-screen retention and a small budget sweep would clarify the claimed quality–efficiency advantage.

Questions for authors
  1. Which decoding protocol does each Table 1 baseline use? Does PACT-GUI’s advantage persist with identical SFD?
  2. What current/history retention settings were used for GUIPruner? How does HP compare with TAR-only under identical decoding, full current-screen retention, and comparable history budgets?
  3. Do applicable baselines reuse cached visual features across steps, including the caching described by Li et al.? How does this affect the efficiency comparison?
3 comments

We sincerely thank you for the constructive comments and valuable suggestions. We address each concern in detail below.

Q1: Decoding fairness and comparison under a shared SFD protocol

We thank you for highlighting the importance of fair comparisons in decoding. In the original Table 1 in paper, all baseline methods were evaluated within our unified evaluation framework using the same Qwen3-VL backbone and standard autoregressive decoding, without SFD. This design reflects the fact that PACT-GUI is an end-to-end acceleration framework comprising two complementary components: Action-Conditioned History Packets (HP) and Structured Fast Decoding (SFD). Accordingly, Table 1 is intended to compare PACT-GUI as a complete end-to-end system against the corresponding baseline systems, while Table 2 separately evaluates the contributions of HP, SFD, and their combination. Nevertheless, we acknowledge that the end-to-end comparisons in Table 1 alone do not establish the superiority of HP over alternative compression methods under identical decoding settings.

We therefore add a shared-decoder experiment to explicitly isolate the effect of the decoding protocol. Specifically, we apply the same SFD decoding protocol to all baselines and PACT-GUI while keeping all other evaluation settings unchanged. Given the rebuttal timeline, we conduct this additional evaluation on AndroidControl-Curated High, AITW, and Mind2Web. The results are shown below.

MethodAndroidControl SR ↑AndroidControl TFLOPs ↓AITW Avg. Acc. ↑AITW TFLOPs ↓Mind2Web SR ↑Mind2Web TFLOPs ↓Avg. Acc. ↑Avg. TFLOPs ↓
GUI-KV59.30184.8848.9437.8147.5252.1251.9291.60
ST-Lite59.00184.6148.4037.5346.1351.8651.1891.33
HistPrune58.9097.1847.8321.0841.8527.0049.5348.42
GUIPruner57.5042.4547.2510.6840.2613.2448.3422.12
FastV59.10131.8247.6025.5345.1236.5350.6164.63
SparseVLM57.80108.7148.0125.7543.8732.9649.8955.81
LLaVA-PruMerge32.5092.2944.4622.1614.2928.0930.4247.51
DivPrune43.6091.7144.8322.8716.6928.1935.0447.59
PACT-GUI63.9055.7250.9514.9444.4417.2353.1029.30

The results show that, under the shared SFD protocol, PACT-GUI still achieves the highest average accuracy (53.10), outperforming the strongest baseline, GUI-KV (51.92), and other history-compression methods, e.g., HistPrune (49.53).

In the revised manuscript, we will explicitly state the decoding protocol used by each baseline in Table 1 and include this shared-SFD comparison in the main paper or appendix. These results demonstrate that PACT-GUI’s gains are not solely attributable to SFD: under the same structured decoder, HP still provides a consistent overall benefit.

Q3: Cross-Step Visual Feature Caching and Its Impact on Efficiency

Thank you for raising this important point. In our reported experiments, cross-step reuse of visual-encoder features was not enabled in the evaluated implementations. Thus, repeated historical screenshots were re-encoded when they appeared in subsequent steps. The KV cache used during autoregressive generation should not be confused with cross-step visual-feature caching.

We agree that this setting does not fully capture the efficiency of a deployment that reuses visual features, as described by Li et al. Such caching can eliminate repeated vision-encoder computation for unchanged historical screenshots, while the corresponding visual tokens may still incur language-model prefill and decoding costs. Therefore, the original efficiency results characterize a no-cross-step-cache setting; they should not be interpreted as the speedup over a cache-enabled baseline. We have clarified this limitation in the manuscript and restricted our efficiency claims to the no-cross-step-cache setting. Quantifying the comparison against cache-enabled baselines is left for future work.

Q2: Comparison with GUIPruner under Matched History Budgets

Thank you for the suggestion. We clarify the retention settings used for GUIPruner and add a TAR-only comparison under full current-screen retention and the same decoding configuration.

For the original GUIPruner comparison, our unofficial Qwen3-VL reproduction combines TAR and SSP. TAR resizes up to four history screenshots, while SSP prunes visual tokens from the current screenshot. We enabled a global prompt-retention target of η\eta=0.35, 0.45, and 0.40 for AndroidControl, AITW, and Mind2Web, respectively. Consequently, the effective current-screen retention was determined separately for each sample rather than fixed at the nominal μ\mu=0.75. The realized aggregate history visual-token retention rates were 2.00%, 1.89%, and 1.97%, respectively, measured using our image-grid accounting.

For TAR-only, we disabled SSP and retained 100% of the current-screen visual tokens. We kept up to four history screenshots and set the nominal TAR history budget to λ\lambda=0.0260, 0.111 and 0.076 for AndroidControl, AITW, and Mind2Web, respectively, with γ\gamma=0.20. The global prompt budget η\eta was disabled. Both GUIPruner variants used the same model and structured fast decoding (SFD).

The TAR budgets were selected to approximate HP’s history-retention rates of 2.36%, 10.29%, and 6.95%, respectively. On a fixed 20-step calibration subset, TAR’s realized image-grid history-retention rates were 2.45%, 10.25%, and 6.85%.

MethodAndroidControl SR ↑TFLOPs ↓AITW Avg Action Accuracy ↑TFLOPs ↓Mind2Web Step Success Rate ↑TFLOPs ↓AVG score ↑AVG TFLOPs ↓
GUIPruner (TAR + SSP)57.5042.4547.2510.6840.2613.2448.3422.12
GUIPruner (TAR-only)58.4048.5548.6114.2345.1215.9150.7126.23
PACT-GUI (HP)63.9055.7250.9514.9444.4417.2353.1029.30

TAR-only improves the average score from 48.34 to 50.71, with average FLOPs increasing from 22.12 to 26.23 TFLOPs. A plausible explanation is that TAR-only retains more history visual tokens under the selected budgets and preserves all current-screen visual tokens, whereas the original GUIPruner configuration additionally prunes the current screen with SSP. The history-budget difference is especially pronounced on AITW and Mind2Web.

This comparison changes both the history budget and current-screen pruning, so it does not isolate their individual contributions. The increased FLOPs are consistent with the additional retained input and computation, but do not by themselves establish the cause of the accuracy gains.

PACT-GUI achieves the highest average score, 53.10, outperforming TAR-only by 2.39 percentage points on average. Thus, PACT-GUI still achieves a better performance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.