Pith. sign in

REVIEW 4 major objections 6 minor 300 references

HELP claims that two specialized operators supervising twelve robots, plus a learned rollout-segmentation critic, raise real-world manipulation success to 80–95% and throughput by 1.7–4.2× over the base model.

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 · deepseek-v4-flash

2026-08-02 08:08 UTC pith:KAAJOYE7

load-bearing objection HELP is a well-engineered human-efficiency pipeline with a genuinely useful benchmark, but the headline VLAC-CUT amplification claim is confounded by a missing unsegmented-rollout control arm. the 4 major comments →

arxiv 2607.09776 v2 pith:KAAJOYE7 submitted 2026-07-08 cs.RO cs.AI

HELP: Human-Efficient Large-Scale Robot Post-Training with Rollout Segmentation

Shaopeng Zhai , Qi Zhang , Tianyi Zhang , Haoran Zhang , Fuxian Huang , Zhanhui Lin , Zijun Xu , Weinan Zhang This is my paper
classification cs.RO cs.AI
keywords robot post-trainingvision-language-action modelshuman-in-the-looprollout segmentationprogress criticmulti-robot supervisionflow matchinghuman efficiency
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.

This paper tries to establish that the real bottleneck in adapting vision-language-action (VLA) robot policies to new tasks is human labor, not robot interaction count, and that existing post-training loops waste that labor. Its claim is that a two-operator, twelve-robot setup — a trained teleoperator delivering remote interventions and recovery demonstrations, plus a floor operator monitoring the fleet, triggering takeovers, and performing resets — raises task throughput by 1.7×–4.2× over the base model across four manipulation tasks. The paper further claims that a learned critic, VLAC-CUT, which segments autonomous rollouts into progress-making, idle, failure-inducing, and recovery segments, converts otherwise unusable robot-collected data into training signal, amplifying the gain from a fixed human recovery budget by 1.20×–3.43× in throughput and 1.50×–3.00× in success rate. If true, this outlines a practical route to scale real-world robot learning without scaling human operators proportionally.

Core claim

The central claim is that the human bottleneck in VLA post-training splits into supervision and data curation. HELP's role-specialized two-operator, twelve-robot setup handles supervision; VLAC-CUT, a learned critic, cuts each autonomous rollout into progress-making, idle, failure-inducing, and recovery segments, keeping only useful ones. On four manipulation tasks this brings success to 80–95% and throughput 1.7×–4.2× over the base model; under matched human-recovery budgets VLAC-CUT adds 1.20×–3.43× throughput and 1.50×–3.00× success-rate gains over HITL-only updates.

What carries the argument

Two pieces carry the argument. The first is the HELP supervision setup: one trained Teleoperator and one Floor Operator supervise twelve robots concurrently, the former giving VR interventions and recovery demonstrations, the latter monitoring the fleet, triggering takeovers, and resetting scenes; the role split is what lets two humans oversee twelve machines. The second is VLAC-CUT, a Vision-Language-Action Critic whose 'cut' operation segments task-conditioned rollouts into progress-making, idle, failure-inducing, and recovery segments using learned signed task-progress estimates; it converts autonomous robot data into curated training signal, and the paper's matched-budget comparisons iso

Load-bearing premise

The pipeline's extra gains rest on the paper's stated but unexperimented premise that training a flow-matching VLA on unsegmented autonomous rollouts would teach it to reproduce repetitive trial-and-error actions and lower throughput, rather than simply benefit from more data; if that premise fails, the segmentation critic loses its causal role.

What would settle it

Run a matched-budget experiment in which the same policy, same checkpoint, and same human recovery trajectories are trained on the same number of rollouts, but the autonomous portions are kept whole or cut at random instead of by VLAC-CUT; if the 1.20×–3.43× throughput and 1.50×–3.00× success-rate amplifications vanish, the critic's causal contribution is falsified. Alternatively, demonstrate a negative-learning-signal objective for flow-matching action decoders and show it matches or beats segmentation-based curation.

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

If this is right

  • If the results hold, a two-person team can sustain a twelve-robot post-training loop, removing the one-human-per-robot scaling constraint from real-world VLA adaptation.
  • Process-level curation means autonomous rollouts that fail are not wasted: recovery and progress segments become reusable training data instead of being discarded wholesale.
  • The matched-budget amplification factors imply the segmentation critic, not extra human data, drives the extra gain, so the same human recovery effort yields larger policy improvement.
  • Iterative post-training compounds: the second round in these experiments produced larger throughput jumps than the first because stronger policies generate higher-quality rollouts for the critic to curate.
  • The four-way segmentation schema and the VPB benchmark give the community a way to evaluate progress critics independently of the full pipeline.

Where Pith is reading between the lines

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

  • Editorial inference: a volume-matched ablation — training on the same number of unsegmented or randomly cut rollouts under the same HITL budget — would directly test whether VLAC-CUT's contribution is segmentation or merely more data.
  • Editorial inference: the paper's negative premise about flow-matching objectives is revisitable; if a reliable way to treat failure-inducing segments as negative supervision emerges, the need for a separate critic and its expensive training would shrink.
  • Editorial inference: the human-efficiency ratio defined here (policy gain per HITL recovery trajectory) could become a standard yardstick for comparing data-curation methods, and the VPB-style progress benchmark could generalize to other long-horizon embodied settings.
  • Editorial inference: external validity is bounded by the four tasks and the single VLA backbone used; testing the scheme on other action decoders would reveal whether the segmentation benefit is specific to flow-matching policies.

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

4 major / 6 minor

Summary. The paper proposes HELP, a pipeline for iterative real-world post-training of vision-language-action (VLA) policies, with the goal of maximizing policy improvement and task throughput per unit of human labor and time. HELP uses two specialized operators supervising twelve robots: a Teleoperator provides remote recovery demonstrations and a Floor Operator monitors the fleet, triggers takeovers, and performs resets. To use the resulting mixed-quality autonomous rollouts, the paper introduces VLAC-CUT, a process-level segmentation critic trained on signed progress annotations, which separates rollouts into progress-making, idle, failure-inducing, and recovery segments and retains only the useful ones. The paper reports a new Video Progress Benchmark (VPB) and evaluates VLAC-CUT on it, then reports real-world post-training experiments on four tasks (Refrigerator, Microplate, Test Tube, Stirrer), claiming 80%–95% success rates, 1.7×–4.2× throughput gains over the base model, and—under matched HITL recovery budgets—1.20×–3.43× amplification of throughput gains and 1.50×–3.00× amplification of success-rate gains relative to HITL-only updates.

Significance. If the central claims hold, the paper makes a practically important contribution: it reframes robot post-training around human efficiency and demonstrates a concrete two-operator/twelve-robot supervision protocol. The detailed annotation schema, the VPB benchmark with seen/unseen and expert/non-expert splits, and the effort to measure VLAC-CUT's contribution under matched HITL budgets are useful assets for the community. However, the headline quantitative attribution to VLAC-CUT is currently supported only by a comparison that also changes data volume, and the reported effect sizes lack statistical grounding. The paper's own premise for segmentation—that unsegmented rollouts are harmful—is asserted rather than tested, which is load-bearing for the main claim.

major comments (4)
  1. [§3.3.1 and §5.2.2 (Tables 5–6)] The central causal claim is not identified by the matched-budget comparison. In Table 5, HELP adds 100–334 curated rollout segments per task-round while HITL-only adds none; the two arms differ in data volume in addition to curation. The only justification that unsegmented rollouts are harmful is the assertion in §3.3.1 that 'we did not find a sufficiently reliable way to directly use failure-inducing action segments as negative learning signals' and that training on unsegmented rollouts would reinforce repetitive trial-and-error. No experiment, ablation, or citation tests this premise. If a flow-matching VLA improves from additional in-distribution rollouts regardless of segmentation, the amplification factors in Table 6 reduce to a volume-of-data effect. Please add control arms matched by segment count, e.g., HELP-raw using the same number of unsegmented autonomous rollouts, or randoml
  2. [§5.2.2 (Tables 4 and 6)] All real-world results are single-run and report no trial counts or confidence intervals. Many amplification factors are computed from small absolute differences—e.g., +4 vs +9 throughput and +15pp vs +25pp success rate in Refrigerator round 1, or +3 vs +5 throughput in Microplate round 1. Without episode counts or variance estimates, it is impossible to assess whether the 1.20×–3.43× range is statistically distinguishable from 1.0. Please report the number of evaluation episodes per cell and provide standard errors, confidence intervals, or repeated-seed/bootstrap estimates at least for the matched comparisons in Table 6.
  3. [§5.2.2 and §3.3.2] The matched comparison does not state whether HITL-only training uses the same optimizer and data-mixing recipe as HELP. Section 3.3.2 introduces ConSFT (Eq. 3) and a 1:1:1 base:HITL:curated-rollout mixture, but §5.2.2 only says 'standard flow-matching objective is adopted.' If HITL-only is trained with plain SFT while HELP uses ConSFT, the amplification factor conflates the effect of VLAC-CUT with the effect of the conservative objective and the mixture protocol. Specify the optimizer and mixing used in each arm and hold them fixed across the comparison.
  4. [§4.2, §4.4, and Table 7] VPB is constructed from the same annotation inventory used to train VLAC-CUT, with held-out records and task units. While the split discipline is careful, the evaluation shares the annotation schema, annotators, and data sources with training, and the comparison baselines are not trained or tuned on this schema. This weakens the claim that VLAC-CUT's large VPB advantage reflects general progress-estimation capability rather than familiarity with the annotation style. Please add an independently annotated or cross-institution held-out evaluation, or at least discuss this limitation explicitly in the benchmark interpretation.
minor comments (6)
  1. [§5.2.2 / Table 6] The arithmetic mean row reports 2.15× and 2.21×, but given the small deltas and likely correlated errors, a geometric mean or per-task confidence intervals would be more informative.
  2. [§4.2 vs Appendix A.1] The main text states the curated dataset contains 28,167 records, while Appendix A.1 reports a full inventory of 35,230 records and the curated split in Table 7 sums to 28,167. Please make the relationship between these numbers explicit in the main text to avoid apparent inconsistency.
  3. [§3.3.1] The phrase 'we did not find a sufficiently reliable way...' is a negative result reported without supporting experiments. If it remains, it should be presented as an explicit limitation; better, it should be accompanied by the ablation suggested in the major comments.
  4. [Author contributions] There is a typo: 'eaperiments' should be 'experiments.'
  5. [§5.2.1] The Stirrer task is described as 'relatively simple' and receives only one iteration, but the base model already has 55% success rate and 66 throughput, which is much higher than the other tasks. A sentence explaining why this task was easier or why the base model was already strong would help.
  6. [§5.1.2 / Table 2] Chrono-GVL-Gemini-3.5-Flash outperforms VLAC-CUT on terminal-state accuracy and successful-terminal F1. The narrative emphasizes macro F1, which is fair, but the abstract and introduction state the segmentation critic is superior without this nuance. Please make the trade-off explicit in the main text.

Circularity Check

0 steps flagged

No significant circularity: the central human-efficiency and amplification claims are empirical measurements, not derived from fitted quantities; self-citations are present but not load-bearing.

full rationale

The paper's main claims are empirical: HELP's absolute gains (Table 4) and the VLAC-CUT amplification factors (Table 6) are measured on real-world tasks rather than derived from a fitted parameter or an equation that encodes the result. Equation (4) defines the amplification factor as the ratio of measured improvements; it does not force any particular outcome. VLAC-CUT is trained on a progress-annotation dataset and evaluated on the held-out VPB split; Section A.3 explicitly excludes held-out annotations from training, augmentation, and fine-tuning, so the benchmark result is not a self-fulfilling fit. The paper does cite prior work by overlapping authors: ConSFT in Eq. (3) and VLAC in related work. However, these citations are not the load-bearing justification for the central human-efficiency derivation: the ConSFT objective is stated in the paper itself, and the main HELP results do not reduce to the content of those citations. The matched HITL-budget comparison in Section 5.2.2 does not equalize total data volume, since HELP additionally uses VLAC-CUT-curated rollouts (Table 5); this is a potential confound for attributing the improvement to segmentation rather than data volume, but it is an experimental-design limitation, not a circular reduction. No step in the paper's derivation chain reduces by construction to its inputs, so no circularity is flagged.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The HELP pipeline rests on hand-chosen system and evaluation parameters (τ, 90% terminal threshold, 1:1:1 data mix, per-task time limits), on the asserted premise that failure-inducing segments are unusable as negative signals, on annotation-defined progress ground truth, and on self-cited components (VLAC, ConSFT) without independent replication.

free parameters (4)
  • ConSFT temperature τ = not specified
    Eq. 3; controls conservative scaling sensitivity in the online incremental objective; chosen by hand with no ablation reported.
  • Terminal-state near-completion threshold = 90%
    Eq. 9; hand-chosen tolerance for VPB terminal-state accuracy and F1 metrics; no ablation.
  • Data mixture ratio base:HITL:curated-rollout = ~1:1:1
    Section 5.2.2 states the datasets are 'approximately balancing them at a 1:1:1 ratio to achieve strong model performance'; not ablated.
  • Per-task time limits = 100s / 100s / 200s / 60s
    Section 5.2.1 defines success within a time limit for each task; these cutoffs set the throughput ceiling and directly affect reported throughput numbers.
axioms (5)
  • domain assumption Failure-inducing action segments cannot be reliably used as negative learning signals under the flow-matching objective
    Section 3.3.1: 'we did not find a sufficiently reliable way to directly use failure-inducing action segments as negative learning signals'; asserted without experiment or citation, and it motivates discarding all failure-inducing data.
  • domain assumption Human-annotated signed progress labels are a consistent and valid ground truth for task progress
    Appendix A.2; the entire VLAC-CUT training and VPB evaluation rest on annotation quality; no inter-annotator agreement statistics are reported.
  • standard math Linear interpolation between sparse progress keyframes yields a valid dense progress trajectory
    A.5, Eq. 5; standard for dense proxy labels but masks non-linear sub-keyframe progress; used for both training interpolation and VPB evaluation.
  • ad hoc to paper ConSFT objective preserves capabilities without reference networks or historical data buffers
    Eq. 3 and Zhang et al., 2026b (same team); load-bearing for the online incremental optimization path; no independent verification or ablation appears in this paper.
  • domain assumption The π0.5 VLA family and flow-matching setup are representative of VLA post-training
    Section 5.2: all real-world experiments use π0.5; generalization to token-output or other action-interface VLAs is not demonstrated.
invented entities (3)
  • VLAC-CUT independent evidence
    purpose: Automatic rollout segmentation critic that classifies trajectory segments as progress-making, idle, failure-inducing, or recovery, and decides what enters the training set
    A new model artifact; evaluated on the held-out VPB benchmark (Section 5.1) and on real-world amplification (Section 5.2), so it is not an unexplained postulate.
  • Predictive takeover model φ and task termination model ψ no independent evidence
    purpose: Predict at each control step whether a takeover is required and whether the task is complete (Eq. 1)
    Components of the architecture mentioned in Section 3.1 but no training data, model details, or standalone evaluation are provided.
  • ARX-data independent evidence
    purpose: In-house real-world dataset of physically executed non-monotonic robot trajectories (grasp failures, drops, wrong-object interactions, recoveries) to train the critic
    Described with statistics and visual examples in Appendix A.1 and Figure 6, but the dataset is not released.

pith-pipeline@v1.3.0-alltime-deepseek · 30306 in / 15051 out tokens · 138506 ms · 2026-08-02T08:08:22.281045+00:00 · methodology

0 comments
read the original abstract

When adapting Vision Language Action (VLA) models to downstream tasks, multiple rounds of post-training are often required to progressively address policy weaknesses. In this report, we focus on maximizing human efficiency during this iterative process, measured by policy improvement and task throughput per unit of human labor and time. We propose HELP, a Human-Efficient Large-scale robot Post-training pipeline in which two specialized operators supervise twelve robots concurrently. A trained Teleoperator provides high-value remote interventions and recovery demonstrations, while a Floor Operator monitors the robot fleet, triggers takeovers, and performs physical resets. This role specialization improves human efficiency by reducing task switching, lowering operator training costs, and expanding robot interaction coverage. Beyond increasing rollout volume, concurrent supervision also broadens the range of policy behaviors observed by the human team, making recurring failure modes easier to identify and enabling more targeted takeovers, resets, and recovery demonstrations. To efficiently utilize the large and mixed-quality rollout data, HELP incorporates \vlac, an automatic rollout segmentation critic specifically designed for this setting. It separates autonomous trajectories into progress-making, idle, failure-inducing, and recovery segments. Useful rollout segments are retained and combined with Human-in-the-Loop data for the next post-training round. Across four real-world manipulation tasks, HELP achieves 80\%--95\% success rates and improves task throughput by 1.7$\times$--4.2$\times$ over the base model. Under matched HITL recovery budgets, VLAC-CUT further amplifies throughput gains by 1.20$\times$--3.43$\times$ and success-rate gains by 1.50$\times$--3.00$\times$ over HITL-only updates.

Figures

Figures reproduced from arXiv: 2607.09776 by Fuxian Huang, Haoran Zhang, Qi Zhang, Shaopeng Zhai, Tianyi Zhang, Weinan Zhang, Zhanhui Lin, Zijun Xu.

Figure 1
Figure 1. Figure 1: System architecture of the proposed distributed hierarchical HITL data collection frame￾work. • Centralized GPU workers: Deployed on a centralized server, these GPU workers manage par￾allel inference and continuous policy optimization using the Ray distributed framework. At each control step, an asynchronous coordination module aggregates multi modal observations from the distributed robotic fleet and rout… view at source ↗
Figure 2
Figure 2. Figure 2: Role division between the Teleoperator and the Floor Operator in the proposed data col [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Role specialization between the Teleoperator and the Floor Operator in HELP. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the VLAC-CUT rollout segmentation pipeline. Given a task-conditioned autonomous rollout trajectory, VLAC-CUT estimates process-level progress, diagnoses local state changes, and segments the trajectory into training-useful and training-harmful portions. 4 VLAC-CUT 4.1 VLAC-CUT AS A ROLLOUT SEGMENTATION CRITIC The framework in Section 3 relies on large-scale autonomous rollout reuse to amplify l… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of signed progress prediction on a representative VPB trajectory. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative examples of terminal-state recognition on VPB trajectories. Each panel shows [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Task execution progress. Each row shows execution states for tasks, and each column corresponds to a normalized task-completion progress from 0% to 100%. input to the VLA models for this task: Open the refrigerator door, grasp the beaker with the gripper, place it inside the refrigerator, and then close the door. Microplate: This task requires the robot to open the lid of the microplate reader, pick up a m… view at source ↗
Figure 6
Figure 6. Figure 6: Task execution progress. Each row shows execution states for tasks, and each column corresponds to a normalized task-completion progress from 0% to 100%. model performance. Finally, we train the current inference model on this combined dataset to obtain an updated policy, and the best-performing checkpoint is used as the inference model for the next round of data collection and training. For the Refrigerat… view at source ↗
Figure 6
Figure 6. Figure 6: Representative failure (Err.) and recovery modes (Rec.) in ARX-data. These examples [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Representative failure (Err.) and recovery modes (Rec.) in ARX-data. These examples [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: Full annotation inventory statistics. (a) Source composition and execution-quality distri [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Full annotation inventory statistics. (a) Source composition and execution-quality distri [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: Semantic and action diversity of the full annotation inventory. The corpus covers a broad [PITH_FULL_IMAGE:figures/full_fig_p025_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Semantic and action diversity of the full annotation inventory. The corpus covers a broad [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example of sparse progress annotation for a robot manipulation trajectory. Only selected [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example of sparse progress annotation for a robot manipulation trajectory. Only selected [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: Overview of the VLA-based teleoperation assistance module. Observed VR commands [PITH_FULL_IMAGE:figures/full_fig_p031_10.png] view at source ↗

Review history (2 revisions) →

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

300 extracted references · 21 linked inside Pith

  1. [1]

    AI Magazine , volume=

    Truth is a lie: Crowd truth and the seven myths of human annotation , author=. AI Magazine , volume=

  2. [2]

    Human-centered tools for coping with imperfect algorithms during medical decision-making , author=

  3. [3]

    Proceedings of the 1964 International Congress for Logic, Methodology, and Philosophy of Science , pages=

    Theories of meaning and learnable languages , author=. Proceedings of the 1964 International Congress for Logic, Methodology, and Philosophy of Science , pages=. 1965 , publisher=

    1964

  4. [4]

    The Logic of Decision and Action , year=

    The logical form of action sentences , author=. The Logic of Decision and Action , year=

  5. [5]

    Philosophy, Language, and Artificial Intelligence: Resources for Processing Natural Language , pages=

    Truth and meaning , author=. Philosophy, Language, and Artificial Intelligence: Resources for Processing Natural Language , pages=. 1967 , publisher=

    1967

  6. [6]

    arXiv preprint arXiv:2307.01952 , year=

    Sdxl: Improving latent diffusion models for high-resolution image synthesis , author=. arXiv preprint arXiv:2307.01952 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2309.02591 , volume=

    Scaling autoregressive multi-modal models: Pretraining and instruction tuning , author=. arXiv preprint arXiv:2309.02591 , volume=

    Pith/arXiv arXiv

  8. [8]

    Microsoft

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll. Microsoft

  9. [9]

    arXiv preprint arXiv:2310.18235 , year=

    Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-image generation , author=. arXiv preprint arXiv:2310.18235 , year=

    Pith/arXiv arXiv

  10. [10]

    Interacting with computers , volume=

    Steps to take before intelligent user interfaces become real , author=. Interacting with computers , volume=. 2000 , publisher=

    2000

  11. [11]

    Help Me Help the AI

    "Help Me Help the AI": Understanding How Explainability Can Support Human-AI Interaction , author=

  12. [12]

    Re-examining whether, why, and how human-AI interaction is uniquely difficult to design , author=

  13. [13]

    2009 , publisher=

    The Quest for Artificial Intelligence , author=. 2009 , publisher=

    2009

  14. [14]

    Artwhisperer: A dataset for characterizing human-ai interactions in artistic creations , author=

  15. [15]

    arXiv preprint arXiv:2310.07653 , year=

    Mini-dalle3: Interactive text to image by prompting large language models , author=. arXiv preprint arXiv:2310.07653 , year=

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2404.18919 , year=

    TheaterGen: Character Management with LLM for Consistent Multi-turn Image Generation , author=. arXiv preprint arXiv:2404.18919 , year=

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2406.01388 , year=

    AutoStudio: Crafting Consistent Subjects in Multi-turn Interactive Image Generation , author=. arXiv preprint arXiv:2406.01388 , year=

    Pith/arXiv arXiv

  18. [18]

    Instruct-Imagen: Image generation with multi-modal instruction , author=

  19. [19]

    Intelligent Systems with Applications , volume=

    DSG-GAN: Multi-turn text-to-image synthesis via dual semantic-stream guidance with global and local linguistics , author=. Intelligent Systems with Applications , volume=. 2023 , publisher=

    2023

  20. [20]

    arXiv preprint arXiv:2403.08857 , year=

    DialogGen: Multi-modal Interactive Dialogue System for Multi-turn Text-to-Image Generation , author=. arXiv preprint arXiv:2403.08857 , year=

    Pith/arXiv arXiv

  21. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    EgoExoLearn: A Dataset for Bridging Asynchronous Ego-and Exo-centric View of Procedural Activities in Real World , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  22. [22]

    Communications of the ACM , volume=

    How might people interact with agents , author=. Communications of the ACM , volume=. 1994 , publisher=

    1994

  23. [23]

    Guidelines for human-AI interaction , author=

  24. [24]

    arXiv preprint arXiv:2406.07882 , year=

    Designing a Dashboard for Transparency and Control of Conversational AI , author=. arXiv preprint arXiv:2406.07882 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2305.02469 , year=

    The system model and the user model: Exploring AI dashboard design , author=. arXiv preprint arXiv:2305.02469 , year=

    Pith/arXiv arXiv

  26. [26]

    Statistical science , pages=

    Bayesian experimental design: A review , author=. Statistical science , pages=. 1995 , publisher=

    1995

  27. [27]

    The SAGE Handbook of Quantitative Methods in Psychology , pages=

    Experimental design , author=. The SAGE Handbook of Quantitative Methods in Psychology , pages=. 2009 , publisher=

    2009

  28. [28]

    Preference learning with Gaussian processes , author=

  29. [29]

    Pairwise preference learning and ranking , author=

  30. [30]

    Boutilier, Craig , booktitle=aaai, year=. A

  31. [31]

    2016 , publisher=

    Artificial intelligence: A modern approach , author=. 2016 , publisher=

    2016

  32. [32]

    Hu, Yushi and Liu, Benlin and Kasai, Jungo and Wang, Yizhong and Ostendorf, Mari and Krishna, Ranjay and Smith, Noah A , booktitle=iccv, year=

  33. [33]

    Hierarchical task and motion planning in the now , author=

  34. [34]

    Pre-trained Gaussian processes for Bayesian optimization , author=

  35. [35]

    A constraint-based method for solving sequential manipulation planning problems , author=

  36. [36]

    Combined task and motion planning under partial observability: An optimization-based approach , author=

  37. [37]

    arXiv preprint arXiv:2402.03719 , year=

    Empowering Language Models with Active Inquiry for Deeper Understanding , author=. arXiv preprint arXiv:2402.03719 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2403.10454 , year=

    Partially Observable Task and Motion Planning with Uncertainty and Risk Awareness , author=. arXiv preprint arXiv:2403.10454 , year=

    Pith/arXiv arXiv

  39. [39]

    Artificial Intelligence , volume=

    Planning and acting in partially observable stochastic domains , author=. Artificial Intelligence , volume=. 1998 , publisher=

    1998

  40. [40]

    arXiv preprint arXiv:1112.5745 , year=

    Bayesian active learning for classification and preference learning , author=. arXiv preprint arXiv:1112.5745 , year=

    Pith/arXiv arXiv

  41. [41]

    Machine learning , volume=

    Queries and concept learning , author=. Machine learning , volume=. 1988 , publisher=

    1988

  42. [42]

    Machine learning: An artificial intelligence approach , volume=

    Learning concepts by asking questions , author=. Machine learning: An artificial intelligence approach , volume=

  43. [43]

    2023 , publisher=

    Bayesian optimization , author=. 2023 , publisher=

    2023

  44. [44]

    ACM computing surveys (CSUR) , volume=

    A survey of deep active learning , author=. ACM computing surveys (CSUR) , volume=. 2021 , publisher=

    2021

  45. [45]

    2017 , organization=

    Deep bayesian active learning with image data , author=. 2017 , organization=

    2017

  46. [46]

    Active learning with statistical models , author=

  47. [47]

    Active learning for identifying function threshold boundaries , author=

  48. [48]

    Active Learning Literature Survey , Type =

    Burr Settles , Institution =. Active Learning Literature Survey , Type =

  49. [49]

    Chen, Sanxing and Wiseman, Sam and Dhingra, Bhuwan , journal=

  50. [50]

    Visual programming for step-by-step text-to-image generation and evaluation , author=

  51. [51]

    Artificial Intelligence , volume=

    Probabilistic logic , author=. Artificial Intelligence , volume=. 1986 , publisher=

    1986

  52. [52]

    John McCarthy and Patrick J. Hayes. Some Philosophical Problems from the Standpoint of Artificial Intelligence. Machine Intelligence 4. 1969

    1969

  53. [53]

    u tzner and Malte Helmert and Thomas Liebetraut and Robert Mattm \

    Yusra Alkhazraji and Matthias Frorath and Markus Gr \"u tzner and Malte Helmert and Thomas Liebetraut and Robert Mattm \"u ller and Manuela Ortlieb and Jendrik Seipp and Tobias Springenberg and Philip Stahl and Jan W \"u lfing. Pyperplan. 2020. doi:10.5281/zenodo.3700819

    doi:10.5281/zenodo.3700819 2020

  54. [54]

    1988 , publisher=

    The Society of Mind , author=. 1988 , publisher=

    1988

  55. [55]

    Formal theories of the commonsense world , pages=

    A formal theory of knowledge and action , author=. Formal theories of the commonsense world , pages=. 1985 , publisher=

    1985

  56. [56]

    1974 , institution=

    A framework for representing knowledge , author=. 1974 , institution=

    1974

  57. [57]

    Artificial Intelligence , volume=

    On the complexity of blocks-world planning , author=. Artificial Intelligence , volume=. 1992 , publisher=

    1992

  58. [58]

    Artificial Intelligence , volume=

    Reasoning about action I: A possible worlds approach , author=. Artificial Intelligence , volume=. 1988 , publisher=

    1988

  59. [59]

    Conference on Language Modeling , year=

    Andukuri, Chinmaya and Fr. Conference on Language Modeling , year=

  60. [60]

    Monographs of the Society for Research in Child Development , pages=

    Children's questions: A mechanism for cognitive development , author=. Monographs of the Society for Research in Child Development , pages=. 2007 , publisher=

    2007

  61. [61]

    Annual Symposium on Machine Programming (MAPS) , year=

    Large Language Models Should Ask Clarifying Questions to Increase Confidence in Generated Code , author=. Annual Symposium on Machine Programming (MAPS) , year=

  62. [62]

    arXiv:2311.09469 [cs.CL] , year=

    Clarify when necessary: Resolving ambiguity through interaction with lms , author=. arXiv:2311.09469 [cs.CL] , year=

    Pith/arXiv arXiv

  63. [63]

    Tong Zhang and Peixin Qin and Yang Deng and Chen Huang and Wenqiang Lei and Junhong Liu and Dingnan Jin and Hongru Liang and Tat-Seng Chua , year=

  64. [64]

    Min, Sewon and Michael, Julian and Hajishirzi, Hannaneh and Zettlemoyer, Luke , booktitle=

  65. [65]

    Kuhn, Lorenz and Gal, Yarin and Farquhar, Sebastian , journal=

  66. [66]

    Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation , author=

  67. [67]

    arXiv:2402.03271 [cs.CL] , year=

    Uncertainty of Thoughts: Uncertainty-Aware Planning Enhances Information Seeking in Large Language Models , author=. arXiv:2402.03271 [cs.CL] , year=

    Pith/arXiv arXiv

  68. [68]

    arXiv:2406.12639 [cs.CL] , year=

    Ask-before-Plan: Proactive Language Agents for Real-World Planning , author=. arXiv:2406.12639 [cs.CL] , year=

    Pith/arXiv arXiv

  69. [69]

    Practical transfer learning for

    Feurer, Matthias and Letham, Benjamin and Hutter, Frank and Bakshy, Eytan , journal=. Practical transfer learning for

  70. [70]

    Asymptotic analysis of maximum likelihood estimation of covariance parameters for

    Bachoc, Fran. Asymptotic analysis of maximum likelihood estimation of covariance parameters for. Advances in Contemporary Statistics and Econometrics: Festschrift in Honor of Christine Thomas-Agnan , pages=. 2021 , publisher=

    2021

  71. [71]

    Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization , author=

  72. [72]

    Adam: A method for stochastic optimization , author=

  73. [73]

    Mathematical Programming , volume=

    Benchmarking optimization software with performance profiles , author=. Mathematical Programming , volume=. 2002 , publisher=

    2002

  74. [74]

    CoRR , volume =

    Ciprian Chelba and Tomas Mikolov and Mike Schuster and Qi Ge and Thorsten Brants and Phillipp Koehn , title =. CoRR , volume =. 2013 , url =

    2013

  75. [75]

    LeCun, Yann and Cortes, Corinna and Burges, CJ , journal=

  76. [76]

    Ten lessons from three generations shaped

    Jouppi, Norman P and Yoon, Doe Hyun and Ashcraft, Matthew and Gottscho, Mark and Jablin, Thomas B and Kurian, George and Laudon, James and Li, Sheng and Ma, Peter and Ma, Xiaoyu and others , booktitle=. Ten lessons from three generations shaped

  77. [77]

    , booktitle=

    Nesterov, Yurii E. , booktitle=. A method of solving a convex programming problem with convergence rate. 1983 , organization=

    1983

  78. [78]

    2006 , publisher=

    Elements of Information Theory , author=. 2006 , publisher=

    2006

  79. [79]

    2010 , publisher=

    Graph kernels , author=. 2010 , publisher=

    2010

  80. [80]

    Yutian Chen and Xingyou Song and Chansoo Lee and Zi Wang and Qiuyi Zhang and David Dohan and Kazuya Kawakami and Greg Kochanski and Arnaud Doucet and Marc'Aurelio Ranzato and Sagi Perel and Nando de Freitas , title =

Showing first 80 references.