Pith. sign in

REVIEW 3 major objections 6 minor 24 references

GUI agents often fail not from wrong reasoning but from decoding too late; pre-compiling a small policy tree off the critical path recovers contested windows.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 08:38 UTC pith:FGTFXX7K

load-bearing objection Solid causal isolation of decode-on-critical-path for contested GUI windows; novelty is the controlled manipulation and pre-registered evidence, not the tree machinery itself. the 3 major comments →

arxiv 2607.28399 v1 pith:FGTFXX7K submitted 2026-07-30 cs.LG

Why Are GUI Agents Correct but Late? Decode on the Decision-Time Critical Path, Tested with Pre-Compiled Policy Trees

Zihan Dong , Rui Qian , Qishi Zhan , Dongshen Peng , Kaixin Li , Yu Li This is my paper
classification cs.LG
keywords GUI agentscomputer-use agentsdecision-time critical pathpolicy treesanticipatory planningdecode latencytransient eventscontested windows
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Computer-use agents frequently compute the right key or click only after a short-lived prompt, toast, or dialog has already vanished. This paper argues the dominant cause is not a lack of foresight but expensive autoregressive decoding sitting on the decision-time critical path. Adaptive Anticipatory Policy Trees (AAPT) keep the same frozen multimodal model and instead move generation into quiet screen time: the model compiles a bounded tree of observable guards, pre-authorized actions, and per-branch deadlines sized to cover its own latency. At event time a lightweight observer routes a change-gated frame to one prepared branch and fires the action with no new text generation. In paired, pre-registered trials the method lifts success from 0.50 to 0.79 inside a contested roughly 650 ms window with zero incorrect actions, while open-loop and predict-and-replan baselines that still decode at decision time score zero. The gain appears only when three gates hold—fast observer decode, valid flat-tree planning, and accurate branch routing—and only when candidate actions can be enumerated before the event. A sympathetic reader cares because many real interfaces run on the environment’s clock, not the model’s; fixing scheduling can turn correct-but-late into correct-and-on-time without retraining.

Core claim

Holding model, task, and information fixed, removing full autoregressive generation from the event-time critical path recovers otherwise-missed transient GUI actions. On a contested window AAPT raises paired success from 0.50 to 0.79 (exact McNemar p=1.8e-3) with no incorrect actions; baselines that still decode at decision time remain at zero. The effect follows a latency-coverage sizing rule, replicates on an untuned same-class generalist, and appears only where fast observer decoding, valid tree planning, and accurate branch routing all hold.

What carries the argument

Adaptive Anticipatory Policy Trees (AAPT): during idle periods the frozen model compiles a flat conditional tree of observable guards, pre-authorized allowlisted actions, and branch deadlines sized so coverage time meets planner p95 latency plus a safety margin; at event time a change gate plus low-token observer routes one frame to a branch and executes without generation.

Load-bearing premise

During a quiet prep interval long enough to cover the planner’s own latency, the true outcomes and actions can be enumerated in advance as schema-valid guards with pre-authorized actions; if the right response is revealed only at fire time, the tree cannot bind it.

What would settle it

A paired contested-window trial in which candidate actions are enumerable and prep time exceeds measured planner latency plus margin, yet AAPT still fails to beat a live reactive baseline once observer routing is accurate and trees are valid—or conversely succeeds when outcomes cannot be pre-enumerated.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Missed transient GUI events can be treated as a scheduling failure rather than a pure reasoning failure.
  • Anticipation becomes a pipeline property: valid tree, true outcome covered, correct route, route before deadline.
  • A hybrid controller should first decide whether the next transition is pre-enumerable, then choose AAPT or reactive execution.
  • Branch budget must match the enumerable outcome-set size; mis-setting it converts misses into wrong actions.
  • Reproducing thresholded agent results requires freezing both functional identity (kernels, schemas) and regime identity (latency quantiles, contested windows).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Distilled or template-based guard matchers could close much of the remaining oracle-to-live routing gap without raising planning cost.
  • The same critical-path logic likely applies to any embodied or tool-using agent facing irreversible sub-second windows, not only GUIs.
  • Long-lived skill memories could amortize AAPT preparation by promoting repeatedly validated sub-trees, complementary to one-shot contingency trees.
  • Serving non-stationarity of only ~100 ms can flip a window from contested to ceiling, so latency monitoring belongs in the evaluation contract itself.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper argues that many GUI-agent failures on transient events are scheduling failures: correct actions arrive after the window closes because full autoregressive decoding sits on the decision-time critical path. Holding model, task, and information fixed, it introduces Adaptive Anticipatory Policy Trees (AAPT): during idle periods the same frozen multimodal model compiles a bounded flat policy tree (observable guards, pre-authorized actions, branch deadlines) sized by T_cover ≥ L_p95 + M; at event time a low-token observer routes a change-gated frame and fires a prepared action with no free-form generation. On a custom contested-window benchmark, paired exact McNemar tests raise success from 0.50 to 0.79 at 650 ms (p=1.8e-3, zero incorrect actions at matched budget), while open-loop and predict-and-replan baselines that still decode at decision time score 0/42. The effect tracks a prep-budget crossover, a three-gate capability matrix (fast observer, valid flat trees, routing ≥0.85), and replicates on an untuned same-class generalist (126 pairs, p=4.9e-13). On DynaCU-Bench the aggregate is a tie with complementary category strengths when actions are pre-enumerable.

Significance. If the result holds in the stated regime, it cleanly separates anticipation-as-pipeline from anticipation-as-model-trait and gives a falsifiable systems criterion (latency coverage, schema-valid trees, live routing) rather than another agent architecture. Strengths that raise the bar for this literature include: pre-registered endpoints and a prediction ledger that retains failed predictions at full weight (notably the UI-Venus oracle probe); paired per-seed exact McNemar tests; mechanism baselines R1/P1 at zero that rule out “act early” alone; quantitative agreement of the prep sweep with Eq. (1); and an honest transfer boundary on DynaCU. The work is significant for computer-use agents, speculative/anticipatory planning, and latency-critical multimodal systems, even though scope is deliberately narrow (enumerable outcomes, quiet prep ≥ L_p95+M, primarily Qwen3.5-MoE class).

major comments (3)
  1. [§5.5, Table 13, Abstract] §5.5 and Table 13: the abstract/conclusion claim that AAPT and reactive execution have complementary strengths, and that AAPT is preferred when candidates are pre-enumerable, rests on an aggregate tie (7/39 vs 6/39, p=1.0) and only 4 vs 5 discordant tasks, with all AAPT-only wins concentrated in one category (dashboards). That is enough to license a qualitative boundary sketch, but not a stable deployment rule. Either enlarge the external task set (or pre-register a larger majority-of-k design) or soften the abstract/conclusion language to match the underpowered split and state explicitly that the complementary-strengths claim is hypothesis-generating.
  2. [Abstract, §5.1, Appendix 12] Abstract and §5.1 vs Appendix 12: the headline “producing no incorrect actions” is true only at matched branch budget and under the suppress-on-late/unmatched design. Appendix 12 shows that mis-set K_max (under- or over-budget) converts misses into wrong keypresses (e.g., 13/42, 12/42, 15/42 wrong actions). The body already notes this is a design property at the operating point, not a general safety guarantee; the abstract and any claim ceiling should carry that caveat in one clause so the zero-error reading cannot be detached from budget matching.
  3. [§4.1, §5.1, Appendix 17.3] §4.1 and §5.1 primary-window handling: two pre-registered primaries left the contested regime (700 ms near R0 ceiling, n.s.; ablation 650 ms recalibrated to 600 ms by reactive-only scan). The paper reports deviations and confirms at a declared 650 ms primary, which is good practice, but the main text should state more sharply how many window choices were pre-registered secondaries versus post-scan selections, and that absolute rates are not portable across seed draws. Without that, readers may over-read the 0.50→0.79 headline as a stable operating point rather than a paired within-regime contrast.
minor comments (6)
  1. [Figure 1] Figure 1 latencies are clear; add the observer token cap (64) and change-gate bound (≤33 ms) to the caption so the critical-path cartoon is self-contained.
  2. [§3.1 Eq. (1)] Eq. (1): define M once as a fixed experimental constant (500 ms) and note it is not fit to the prep sweep, to avoid the appearance of a tuned threshold.
  3. [Table 2, §5.4] Table 2 gate (c) threshold 0.85 is used as a hard pass/fail; a one-sentence justification (or sensitivity) would help, since UI-Venus at 0.39 is far below and the oracle probe does the causal work.
  4. [§2 Table 1] Related work Table 1 is useful; a short note on how AAPT differs from classical behavior trees / POMDP contingency plans beyond GUI latency would orient non-agent readers.
  5. [Abstract, §1] Typos/spacing: “windowhasalreadyclosed”, “Weidentifythemaincause”, and similar missing spaces appear in the abstract/intro PDF text; clean for camera-ready.
  6. [§5, Appendix 18.1] Appendix 18.1 prediction ledger is a model of transparency; consider promoting a condensed version into the main text or supplement pointer earlier than §5.

Circularity Check

0 steps flagged

No significant circularity: causal isolation rests on external server-side scoring, independent latency measurements, and pre-registered falsifiable predictions (including failed ones).

full rationale

This is an experimental systems paper that isolates decode-on-critical-path as a cause of missed GUI windows by holding model, task, and information fixed and moving generation off the event-time path. Success is scored by a deterministic server-side keypress-before-deadline criterion, not a model judge. Eq. (1) (T_cover ≥ L_p95 + M) is not a fit renamed as prediction: L_p95 is independently measured planner latency and the prep-budget sweep is a genuine out-of-sample check that the advantage appears only above that threshold. R1/P1 baselines share anticipation but keep decode on-path and score zero, so the central contrast is not definitional. Capability gates, the k×n knee at outcome-set size, the oracle-on-Venus probe, and the Qwen generalist confirmation were pre-registered; the oracle probe explicitly rejected the authors' committed hypothesis and is reported at full weight. No load-bearing uniqueness theorem, self-citation chain, or fitted parameter is smuggled in as a first-principles result. The enumerability/prep-coverage contract is an explicit scope boundary (falsified by click_target and the DynaCU split), not a circular premise. Score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The claim rests on a systems factorization (valid tree ∧ coverage ∧ correct route ∧ before deadline) plus a latency-coverage inequality built from measured planner latency and hand-chosen margins/thresholds. No new physical entities; AAPT is an engineered construct. Free parameters are operational thresholds (safety margin, confidence floors, change-gate pixels, token caps, branch budgets) that the paper partly sweeps but that still define the operating point where zero incorrect actions and the reported gains hold.

free parameters (6)
  • Safety margin M in T_cover ≥ L_p95 + M = 500 ms
    Fixed at 500 ms in all experiments; enters the quantitative onset prediction for the prep sweep.
  • Planner-confidence floor τ_m = 0.45
    Runtime suppresses branches below τ_m (0.45 in reported runs); uneven-prior arm is zeroed when priors are copied into confidence.
  • Observer token cap = 64 tokens
    Chosen by sweep; 64 tokens adopted as operating point affecting branch accuracy and latency.
  • Change-gate thresholds = 0.005 / 20 / 500 ms / 30 FPS
    min_changed_fraction=0.005, pixel_threshold=20, static_probe_ms=500, 30 FPS; bound observer input timing (≤33 ms).
  • Branch budget K_max and depth n_max = K_max=3 (3-outcome), n=1 recommended
    Must match enumerable outcome-set size; mis-set budgets produce wrong actions, not only misses.
  • cover_ms / T_cover validation budget = 3500 ms (validation reference)
    Trees whose declared coverage misses measured budget (3500 ms cited) are rejected at validation.
axioms (5)
  • domain assumption Success ⇔ correct action ∧ action before environment deadline
    Stated in §3.1; defines the contested-window criterion and server-side scoring.
  • ad hoc to paper Tree must bridge planner unavailability: T_cover ≥ L_p95 + M
    §3.1 Eq. (1); quantitative content of the hypothesis, tested by prep sweep.
  • ad hoc to paper Flat independent guards suffice; routing is single-shot classification over siblings, not deep traversal
    §3.2; motivated by schema-following failures on recursive trees under the same budget.
  • domain assumption Only allowlisted pre-authorized actions may execute; unmatched/expired/low-confidence routes abstain and replan
    §3.4 execution discipline; underpins the zero incorrect-action property at matched budgets.
  • standard math Paired per-seed McNemar comparisons are the valid unit when absolute rates are seed- and latency-sensitive
    §4.1 experimental design; standard exact test on discordant pairs.
invented entities (2)
  • Adaptive Anticipatory Policy Trees (AAPT) independent evidence
    purpose: Bounded conditional policy with guards, pre-authorized actions, and deadlines compiled off the critical path; used as the experimental manipulation.
    Engineered construct combining known pieces; independent handle is the falsifiable latency-coverage and gate predictions, not a new natural kind.
  • key_prompt contested-window benchmark no independent evidence
    purpose: Millisecond-controllable timed GUI task isolating deadline-sensitive key selection with hidden seed-derived schedules.
    Custom benchmark because OSWorld-class suites lack ms control; external DynaCU transfer partially de-risks benchmark-specificity.

pith-pipeline@v1.2.0-daily-grok45 · 30936 in / 4076 out tokens · 96513 ms · 2026-07-31T08:38:26.506423+00:00 · methodology

0 comments
read the original abstract

Computer-use agents often fail on transient GUI events because they produce the correct action only after the relevant window has already closed. We identify the main cause as expensive autoregressive decoding on the decision-time critical path. We propose Adaptive Anticipatory Policy Trees (AAPT), which eliminates this delay without modifying the underlying model. During idle screen periods, the same frozen multimodal model constructs a bounded conditional policy tree with observable guards, pre-authorized actions, and branch-specific deadlines. The tree is sized to cover the model's own decoding latency. When an event occurs, a lightweight observer matches change-gated frames to a prepared branch and immediately executes the corresponding action without generating new text. In paired trials with pre-registered endpoints and exact McNemar tests, AAPT improves the success rate from 0.50 to 0.79 within a contested decision window ($p=1.8\times10^{-3}$), while producing no incorrect actions. Both open-loop and predict-and-replan baselines achieve zero success because they still decode during execution. A preparation-time sweep shows that the gain emerges where the latency-based tree-sizing rule predicts, and ablations reveal three key requirements: fast observer decoding, valid tree planning, and accurate branch routing. A pre-registered oracle probe rejects our initial hypothesis and instead points to branch routing as the causal bottleneck. We further reproduce the effect on an independent general-purpose multimodal model over 126 paired trials ($p=4.9\times10^{-13}$). On an external benchmark, AAPT matches the overall performance of a reactive baseline, although the two methods exhibit complementary strengths. Together, these results suggest that AAPT performs best when candidate actions can be enumerated in advance, whereas reactive execution remains stronger when they cannot.

Figures

Figures reproduced from arXiv: 2607.28399 by Dongshen Peng, Kaixin Li, Qishi Zhan, Rui Qian, Yu Li, Zihan Dong.

Figure 1
Figure 1. Figure 1: The critical path under a contested 600 ms win [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: AAPT dataflow. Off the critical path, the slow [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Documented real-world response windows (log [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mechanism isolation at the two contested windows (left: 650 ms; right: 600 ms; 42 pairs each; Wilson 95% intervals; [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Budget frontier at the 600 ms primary window (42 pairs per cell, shared reactive arm). (a) T2 success versus critical [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Pre-registered confirmation on Qwen3.6-35B-A3B, an untuned generalist of the same architecture class as Holo (42 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 17 linked inside Pith

  1. [2]

    NeurIPS2023DatasetsandBenchmarks Track, arXiv:2306.06070

    Mind2Web: Towards a Generalist AgentfortheWeb. NeurIPS2023DatasetsandBenchmarks Track, arXiv:2306.06070. Ding,H.;Liu,P.;Wang,J.;Ji,Z.;Cao,M.;Zhang,R.;Ai,L.; Yang, E.; Shi, T.; and Yu, L

    Pith/arXiv arXiv

  2. [5]

    arXiv:2606.08049

    SKILL.nb: Selective Formalization and Gated Execution for Durable Agent Workflows. arXiv:2606.08049. Gu, Y.; Zheng, B.; Gou, B.; Zhang, K.; Chang, C.; Srivas- tava, S.; Xie, Y.; Qi, P.; Sun, H.; and Su, Y

    Pith/arXiv arXiv

  3. [6]

    arXiv:2411.06559

    Is Your LLM Secretly a World Model of the Internet? Model-Based Planning for Web Agents. arXiv:2411.06559. Guan, Y.; Hua, W.; Lan, Q.; Fei, S.; Ding, D.; Acharya, D.; Wang, C.; and Wang, W. Y

    Pith/arXiv arXiv

  4. [8]

    arXiv:2602.17365

    Computer-Using World Model. arXiv:2602.17365. HCompany.2026.Holo-3.1-35B-A3Bmodelcard.Hugging Face. He, R.; Brunskill, E.; and Roy, N

    arXiv 2026

  5. [9]

    arXiv:2410.00079

    Interactive Speculative Planning: Enhance Agent Efficiency through Co-design of System and User In- terface. arXiv:2410.00079. inclusionAI

    Pith/arXiv arXiv

  6. [11]

    arXiv:2606.29472

    Agent-Computer Observation In- terfacesEnableDynamicComputerUse. arXiv:2606.29472. Liang, Y.; Zhou, S.; Gu, Y.; Tan, H.; Wu, G.; Dernoncourt, F.; Kil, J.; Rossi, R. A.; and Zhang, R

    Pith/arXiv arXiv

  7. [12]

    arXiv:2603.16777

    Anticipatory Planning for Multimodal AI Agents. arXiv:2603.16777. Liu, E.; Pan, L.; Gao, Z.; Yang, Y.; Shi, C.; Liu, Y.; Wu, J.; and Li, Q

    arXiv

  8. [13]

    arXiv:2604.25380

    Benchmarking and Improving GUI Agents in High-Dynamic Environments. arXiv:2604.25380. Luo, D.; Tang, B.; Li, K.; Papoudakis, G.; Song, J.; Gong, S.; Hao, J.; Wang, J.; and Shao, K

    Pith/arXiv arXiv

  9. [14]

    arXiv:2504.13936

    ViMo: A Generative Visual GUI World Model for App Agents. arXiv:2504.13936. Meituan

    Pith/arXiv arXiv

  10. [16]

    arXiv:2405.14573

    AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents. arXiv:2405.14573. Tongyi-MiA

    Pith/arXiv arXiv

  11. [17]

    Hugging Face

    UI-Ins-32B model card. Hugging Face. Wang,Z.Z.;Mao,J.;Fried,D.;andNeubig,G.2024. Agent Workflow Memory. arXiv:2409.07429. Xiao, Z.; Tu, J.; Zou, C.; Zuo, Y.; Li, Z.; Wang, P.; Yu, B.; Huang, F.; Lin, J.; and Liu, Z

    Pith/arXiv arXiv 2024

  12. [18]

    arXiv:2602.14721

    Web- World:ALarge-ScaleWorldModelforWebAgentTraining. arXiv:2602.14721. Xie, T.; Zhang, D.; Chen, J.; Li, X.; Zhao, S.; Cao, R.; Hua, T. J.; Cheng, Z.; Shin, D.; Lei, F.; Liu, Y.; Xu, Y.; Zhou, S.; Savarese, S.; Xiong, C.; Zhong, V.; and Yu, T

    arXiv

  13. [19]

    NeurIPS 2024, arXiv:2404.07972

    OS- World: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. NeurIPS 2024, arXiv:2404.07972. Yang,H.;Tang,F.;Zhao,L.;Zhuang,X.;Lu,Y.;An,X.;Hu, M.; Zhang, X.; Swikir, A.; He, J.; Ge, Z.; Khan, M. H.; and Razzak,I.2025. StreamAgent:TowardsAnticipatoryAgents for Streaming Video Understanding. arXiv:2508.01875. Yao, J.; Huang,...

    Pith/arXiv arXiv 2024

  14. [20]

    arXiv:2606.04701

    Benchmarking Living-Screen-Native GUI Agents on Short-Video Platforms. arXiv:2606.04701. Ye, N.; Ahuja, A.; Liargkovas, G.; Lu, Y.; Kaffes, K.; and Peng, T

    Pith/arXiv arXiv

  15. [21]

    arXiv:2510.04371

    Speculative Actions: A Lossless Framework for Faster Agentic Systems. arXiv:2510.04371. Ye,N.;Somani,A.;Hsu,D.;andLee,W.S.2017. DESPOT: Online POMDP Planning with Regularization.Journal of Artificial Intelligence Research, 58: 231–266. Zheng,Y.;Zhong,L.;Wang,Y.;Dai,R.;Liu,K.;Chu,X.;Lv, L.;Torr,P.;andLin,K.Q.2026. Code2World:AGUIWorld ModelviaRenderableCod...

    Pith/arXiv arXiv 2017

  16. [22]

    arXiv:2602.20502

    ActionEngine: From Reactive to Programmatic GUI Agents via State Ma- chine Memory. arXiv:2602.20502. Zhou, S.; Xu, F. F.; Zhu, H.; Zhou, X.; Lo, R.; Sridhar, A.; Cheng, X.; Ou, T.; Bisk, Y.; Fried, D.; Alon, U.; and Neubig, G

    arXiv

  17. [23]

    Press F12

    WebArena: A Realistic Web Environment for Building Autonomous Agents. arXiv:2307.13854. 8 Protocol details Stage ladder.The study ran as a pre-registered lad- der of stages, each with a written plan (goals, windows, seeds, endpoints, decision rules) committed before execu- tion: pilot and planner diagnostics; guided-decoding adop- tion and the below-floor...

    Pith/arXiv arXiv

  18. [24]

    an action or an action array

    is a meaningful fraction of the shortest ofthese.Atorabove10s(regulatedtakeover,WCAG,TOTP, push approval, SSH) the reactive loop suffices and AAPT’s premium is not justified. The survey bounds the claim hon- estly in both directions: real sub-second windows exist in deployed systems, and most windows above a few seconds do not need anticipation. 16 Method...

    2026

  19. [2003]

    Optimal Limited Contin- gencyPlanning.InProceedingsoftheNineteenthConference on Uncertainty in Artificial Intelligence. Qin,Y.;Ye,Y.;Fang,J.;Wang,H.;Liang,S.;Tian,S.;Zhang, J.;Li,J.;Li,Y.;Huang,S.;Zhong,W.;Li,K.;Yang,J.;Miao, Y.; Lin, W.; Liu, L.; Jiang, X.; Ma, Q.; Li, J.; Xiao, X.; Cai, K.; Li, C.; Zheng, Y.; Jin, C.; Li, C.; Zhou, X.; Wang, M.; Chen, H...

    Pith/arXiv arXiv 2025

  20. [2020]

    arXiv:2005.05842

    A Survey of Behavior Trees in Robotics and AI. arXiv:2005.05842. Kim,H.;Wu,Y.;andTambe,T.2026. AgenticCache:Cache- DrivenAsynchronousPlanningforEmbodiedAIAgents. In Proceedingsofthe9thConferenceonMachineLearningand Systems. Li,B.2026. PreAct:Computer-UsingAgentsthatGetFaster on Repeated Tasks. arXiv:2606.17929. Li, B.; and Shi, N

    Pith/arXiv arXiv 2005

  21. [2023]

    InProceedings of AREA 2023 (EPTCS 391), 89–101

    Rollout Heuristics for Online Stochastic Contingent Planning. InProceedings of AREA 2023 (EPTCS 391), 89–101. Cao, Y.; Zhong, Y.; Zeng, Z.; Zheng, L.; Huang, J.; Qiu, H.; Shi,P.;Mao,W.;andWan,G.2026. MobileDreamer:Gener- ativeSketchWorldModelforGUIAgent.arXiv:2601.04035. Chen,Q.;Bellucci,A.;Sun,Z.;andJacucci,G.2026. Skill- Droid: Compile Once, Reuse Forev...

    arXiv 2023

  22. [2024]

    arXiv:2411.15100

    XGrammar: Flexible and Efficient Structured Generation Engine for Large Language Models. arXiv:2411.15100. El Hattami, A.; Chapados, N.; and Pal, C

    Pith/arXiv arXiv

  23. [2025]

    arXiv:2509.01920

    Dynamic Speculative Agent Planning. arXiv:2509.01920. Guan, Y.; Yu, R.; Zhang, J.; Wang, L.; Li, L.; Qiao, B.; Qin, S.; Huang, H.; Yang, F.; Zhao, P.; Wutschitz, L.; Kessler, S.; Inan, H. A.; Sim, R.; Rajmohan, S.; Lin, Q.; and Zhang, D

    arXiv

  24. [2026]

    arXiv:2601.22149

    DynaWeb: Model-Based ReinforcementLearningofWebAgents. arXiv:2601.22149. Dong, Y.; Ruan, C. F.; Cai, Y.; Lai, R.; Xu, Z.; Zhao, Y.; and Chen, T

    Pith/arXiv arXiv