Pith. sign in

REVIEW 2 major objections 5 minor 29 references

Whole-Body Semantic-to-Actuation Grounding of Elephant-Inspired Soft-Trunk Motion via Lightweight Flow Matching

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Soft-trunk robots can turn open language into whole-body motion by reducing intent and actuation, then sampling with rectified flow.

desk verdict Usable dual-reduction + rectified-flow stack for whole-body soft-trunk social motion; ablations and 100-person study are real, but held-out “correctness” is partly library-circular. read the letter →

arxiv 2607.11018 v1 pith:E6VEWWGD submitted 2026-07-13 cs.RO

Tingcong Liu , Tongshun Chen , Siyi Ma , Yuhao Wang , Aye Phyu Phyu Aung , Ibrahim Alsarraj , J. Senthilnath , Bo An
show 1 more author
Ke Wu
This is my paper · ORCID
classification cs.RO
keywords softroboticswhole-bodysemantic-to-actuationgroundingbio-inspirationMLLMHRIrectifiedflowCatmull-Romsplinecontinuummanipulators
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Close-contact human-robot interaction needs bodies that can express intent through shape, not only through speech or a screen. Soft elephant-inspired trunks can do that, but ordinary language-to-action pipelines fail them: tip commands do not specify whole-body posture, and raw high-dimensional tendon trajectories are hard to keep feasible. This paper claims a dual reduction plus lightweight flow matching solves the mismatch. Open multimodal responses are first mapped into a small set of morphology-aligned intent-intensity pairs; continuous tendon profiles are then encoded as compact Catmull-Rom spline controls; a rectified-flow model learns the conditional distribution over those controls and samples smooth, executable motions. Ablations show held-out grounding correctness rising from 25 percent for raw dense regression to 77.2 percent for the full pipeline, with faster inference than a diffusion baseline. In a 100-person study, adding the generated trunk channel lifted positive overall satisfaction from 46 percent to 82 percent versus audiovisual-only interaction.

What carries the argument

Dual reduction plus rectified-flow grounding: morphology-aligned intent-intensity tuples (sixteen intents by four intensities) condition a lightweight velocity field over Catmull-Rom spline control matrices, so continuous tendon trajectories are reconstructed from low-dimensional samples rather than predicted densely.

What would settle it

If independent judges or a larger held-out library show that many high-RMS samples are still perceived as valid trunk gestures, or that many low-RMS samples look wrong on the real platform, the reported correctness gains and the social-validity claim would fail.

Watch

Extended reading notes

Core claim

The authors establish that open-vocabulary multimodal responses can be grounded into feasible whole-body soft-trunk actuation by first canonicalizing language into bounded intent-intensity tuples, then representing trajectories with compact Catmull-Rom controls, and finally sampling from a tuple-conditioned rectified-flow model. That pipeline yields higher held-out correctness than raw-response dense regression or a diffusion generator, faster inference than diffusion, retained diversity, and a clear user-facing gain when the trunk channel is added to audiovisual interaction.

Load-bearing premise

The claim rests on treating a small offline library of 256 expert trajectories, filtered by the same reduction rules and a fixed RMS tolerance of 0.35, as the gold standard for what counts as correct and socially valid whole-body motion.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes a whole-body semantic-to-actuation grounding pipeline for elephant-inspired soft-trunk HRI. Open-vocabulary MLLM responses are reduced to morphology-aligned intent–intensity tuples (16 intents × 4 intensity levels), continuous tendon trajectories are parameterized by compact Catmull–Rom control matrices (Nu=4, P=4), and a lightweight rectified-flow model samples from the conditional distribution over those controls. Ablations on a 256-sample library (Table II) show progressive gains: Raw+Dense+CNN 25.0% held-out success → Tuple+Spline+RF 77.2%, with RF also faster than DDPM (4.87 ms vs 7.86 ms) while retaining diversity. A 100-participant within-subject physical study reports that adding the generated trunk channel raises positive overall-satisfaction ratings from 46% to 82% versus audiovisual-only interaction.

Significance. If the claims hold, the work supplies a practical dual-reduction interface that bridges open-vocabulary MLLM responses and high-dimensional continuum actuation for close-contact social HRI—an under-served setting relative to tip-centric VLA. The clean ablation isolation of semantic reduction, spline reduction, and RF versus CNN/DDPM, the explicit one-to-many sampling, and the sizable physical user study are concrete strengths. The contribution is primarily systems-level rather than a new theoretical result, but it is timely for soft-robot HRI and demonstrates deployable inference latency.

major comments (2)
  1. [Sec. III-C.1, IV-A.3, Table II] Sec. III-C.1 and IV-A.3 / Table II: Held-out correctness (normalized RMS ≤ τ_rms=0.35) is measured against library trajectories that were themselves produced by the same MLLM reduction under C, Catmull–Rom fitting, fixed-prompt MLLM screening, and manual verification used at train time. The threshold was calibrated on the same hardware motions. This makes the 25.0%→77.2% claim largely a measure of library-manifold reproduction rather than independently judged social validity. The progressive ablations and DDPM comparison remain informative, but the central quantitative claim needs either an external human/LLM-as-judge evaluation of generated motions or an explicit statement that success is library-relative.
  2. [Sec. IV-B] Sec. IV-B: The physical HRI study compares AV-only versus AV+generated trunk. It therefore shows that adding any trunk motion improves satisfaction (46%→82%), not that the proposed grounding is preferable to scripted or alternative generators. The manuscript itself notes that a direct perceptual comparison with manually scripted motions is left for future work; without that (or at least a no-grounding trunk baseline), the user-study claim cannot be read as validation of the semantic-to-actuation pipeline specifically.
minor comments (5)
  1. [Sec. III-A.2] Eq. (3) is presented as a conceptual argmin but is never optimized; the practical implementation is the few-shot MLLM call in Eq. (4). Clarify that the former is only notational.
  2. [Fig. 4] Fig. 4 qualitative labels (Wrong/Marginal/Correct) are useful but the corresponding RMS values are given only for selected cases; reporting RMS for every panel would strengthen the visual–quantitative link.
  3. [Sec. IV-A.3] Diversity (Eqs. 15–16) is mean pairwise RMS; a brief note on whether higher diversity is always desirable (vs. occasional outliers noted for DDPM) would help interpretation of the 0.148 vs 0.165 comparison.
  4. [Table I] Table I lists free parameters (P=4, S=50, τ_rms=0.35, network width) without sensitivity analysis; a short appendix or sentence on robustness to P and S would be useful.
  5. [Abstract / Introduction] Minor wording: Abstract and Introduction use both “intent-intensity” and “intent–intensity”; standardize the en-dash.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild evaluation self-reference only: held-out ‘correctness’ is RMS match to the same MLLM-screened library used for training; the method derivation itself is not forced by construction.

  1. other [Sec. III-C.1; Sec. IV-A.3; Eqs. (12)–(14); Table II]
    "For each condition (m_i, z_i), candidate actuation trajectories are designed and then fitted into Catmull–Rom matrices c*_i using (5). ... retained after actuation-bound, actuation-smoothness, and social-appropriateness checks for R(c*_i;τ) via fixed-prompt MLLM screening for semantic agreement with (m_i, z_i)[28], and manual verification... Succ_held-out = 1/N_held-out Σ I[RMS_j ≤ τ_rms]. ... The selected threshold τ_rms = 0.35 showed the closest agreement between RMS-based success and perceived motion validity."

    Held-out ‘grounding correctness’ is operationally defined as normalized RMS ≤ 0.35 to trajectories that were produced by the identical MLLM-under-C reduction, Catmull–Rom fitting, fixed-prompt MLLM screening, and manual verification used to build the training library. The tolerance itself was calibrated on those same hardware motions. Thus the 25.0%→77.2% claim largely measures how well generators reproduce the library manifold for a given (m,z), not independently adjudicated social validity. This is evaluation self-reference, not a derivation that equates method outputs to inputs by construction; ablations and the separate HRI study remain informative.

full rationale

This is an engineering pipeline paper (MLLM semantic reduction → Catmull–Rom actuation reduction → conditional rectified flow), not a first-principles derivation. The RF objective (Eq. 8–9), Euler sampling (Eq. 10), and spline reconstruction (Eq. 5, 11) are standard generative-model machinery and do not reduce to their inputs by definition. Self-citations [18],[19] are prior soft-robot flow-matching applications by overlapping authors; they are not load-bearing uniqueness theorems—the RF method is taken from external sources [16],[17]. The only mild circularity is evaluative: Table II’s held-out success rate treats library trajectories built by the same reduction/screening pipeline (Sec. III-C.1) as gold targets, with τ_rms calibrated on those motions (Sec. IV-A.3). That makes ‘grounding correctness’ partly a measure of library reproduction rather than an independent social-validity oracle. Progressive ablations (Raw→Tuple→Spline→DDPM→RF), the DDPM speed/accuracy contrast, and the 100-participant AV-only vs. trunk HRI study still supply independent content, so the central claim is not forced. Score 2 reflects that single non-load-bearing evaluation self-reference.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central empirical claims rest on a small hand-curated demonstration library, a discrete 16×4 semantic space chosen by the authors, a calibrated RMS threshold, and standard generative-model and spline assumptions. No new physical constants are introduced; the free parameters and invented discrete spaces are the main degrees of freedom that make the reported percentages possible.

free parameters (6)
  • tau_rms success threshold = 0.35
    Normalized RMS cutoff of 0.35 chosen after calibration against human annotations and LLM-as-judge so that RMS success matches perceived validity; absolute success rates change with the threshold even though ordering is stable.
  • Catmull-Rom control count P = 4
    Number of control points per actuation channel fixed at 4; directly determines the reduced target dimensionality and reconstruction fidelity.
  • RF Euler steps S = 50
    Number of explicit Euler integration steps used for both RF and DDPM sampling; set to 50.
  • intent-intensity cardinality = 16 × 4
    Authors define 16 discrete social intents and 4 ordered intensity levels that form the entire conditioning space; size is a design choice, not derived.
  • demonstration library size = 256
    N_total = 256 paired samples (192 train / 64 held-out) constructed offline; all quantitative correctness numbers are conditioned on this finite set.
  • RF network width and depth = 4×256
    4-layer MLP of width 256 with SiLU and 128-d time embedding; architecture hyperparameters chosen by the authors.
assumptions (4)
  • domain assumption A Catmull-Rom spline with P control points per channel can reconstruct continuous multi-tendon trajectories with sufficient fidelity for social trunk motion while remaining low-dimensional.
    Invoked in Sec. III-B.2 and Eq. (5); standard spline property but the claim that P=4 is adequate for the platform is platform-specific.
  • domain assumption Rectified flow learns a straight-line transport map that yields a usable conditional distribution over spline controls with few Euler steps.
    Taken from cited flow-matching literature (Liu et al., Lipman et al.) and applied in Eqs. (8)–(10).
  • ad hoc to paper An MLLM under a fixed few-shot grounding specification C can map open-vocabulary responses onto the discrete executable tuple space without catastrophic loss of social intent.
    Core of Sec. III-A.2 and Eqs. (3)–(4); the paper treats the reduction as reliable once C is written.
  • domain assumption Elephant-inspired kinematic primitives (bend, curl, extend, retract, contact) plus the chosen 16 social classes span the socially appropriate whole-body behaviors needed for close-contact HRI.
    Stated in Introduction and Sec. III-A/B with citations to elephant biomechanics; not independently validated for the full open-vocabulary range.
invented entities (2)
  • morphology-aligned intent-intensity tuple space (M × Z)
    purpose: Canonical discrete semantic condition that replaces raw open-vocabulary embeddings for conditioning the flow model.
    Defined by the authors as 16 intents × 4 intensities; no external standard or independent measurement validates that this partition is complete or natural.
  • dual-reduction representation (semantic tuple + Catmull-Rom control matrix)
    purpose: Bridge the representation mismatch between open language and high-dimensional tendon trajectories.
    The joint construction is introduced in this paper; each half has precedents, but the paired interface is new and paper-specific.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Whole-Body Semantic-to-Actuation Grounding of Elephant-Inspired Soft-Trunk Motion via Lightweight Flow Matching." pith.science (2026). https://pith.science/paper/E6VEWWGD

@misc{pith2026260711018,
  author       = {Pith},
  title        = {Pith review of: Whole-Body Semantic-to-Actuation Grounding of Elephant-Inspired Soft-Trunk Motion via Lightweight Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E6VEWWGD}},
  note         = {Machine review of arXiv:2607.11018}
}
read the original abstract

For close-contact human-robot interaction (HRI), trunk-like continuum manipulators provide a physical channel for diverse whole-body expression, but grounding open-vocabulary responses into such robots is difficult: end-effector motion underspecifies body shape, whereas direct whole-body commands are high-dimensional and hard to keep feasible. We propose a whole-body semantic-to-actuation grounding framework for elephant-inspired soft-trunk HRI based on lightweight flow matching. The framework converts responses from a multimodal large language model into bounded, morphology-aligned intent-intensity tuples, parameterizes tendon-actuation trajectories with compact Catmull-Rom spline controls, and uses a rectified-flow generator to sample feasible whole-body trunk motions. Experiments show that the proposed framework improves held-out grounding correctness from 25.0% to 77.2% over a raw-response dense-regression baseline. Compared with a denoising-diffusion baseline, it improves correctness from 71.9% to 77.2% and reduces inference time from 7.86 ms to 4.87 ms while preserving motion diversity. A 100-participant physical HRI study further shows that adding the generated soft-trunk motion channel increases the positive overall-satisfaction rating from 46% to 82% over the audiovisual-only baseline.

Figures

Figures reproduced from arXiv: 2607.11018 by the authors.

Figure 1
Figure 1. Overview of the whole-body semantic-to-actuation grounding framework. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Details of three proposed modules: morphology-aligned semantic reduction, compact actuation-space reduction, and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Tendon-driven soft-trunk platform. All methods are evaluated after their outputs are recon￾structed or represented in the common discrete actuation format specified in Table I, consisting of a T = 150-step trajectory over Nu = 4 tendon-driven actuation channels. 2) Compared Methods: Each method is named following the pattern: Semantic Condition + Actuation Representation + Generator. Semantic Condition: Raw means th… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Held-out semantic-to-actuation grounding examples. The proposed Tuple + Spline + RF model more closely follows [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Elephant-inspired Robotic Interaction Companion [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Representative interaction episodes between the user and ERIC. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Physical HRI study with 100 participants. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 5 linked inside Pith

  1. [1]

    Human-robot proxemics: Physical and psychological distancing in human-robot interaction,

    J. Mumm and B. Mutlu, “Human-robot proxemics: Physical and psychological distancing in human-robot interaction,” inProceedings of the 6th International Conference on Human-Robot Interaction. ACM, 2011, pp. 331–338

    2011

  2. [2]

    Survey of social touch interaction between humans and robots,

    M. Shiomi, H. Sumioka, and H. Ishiguro, “Survey of social touch interaction between humans and robots,”Journal of Robotics and Mechatronics, vol. 32, no. 1, pp. 128–135, 2020

    2020

  3. [3]

    Designing robots with movement in mind,

    G. Hoffman and W. Ju, “Designing robots with movement in mind,” Journal of Human-Robot Interaction, vol. 3, no. 1, pp. 91–122, 2014

    2014

  4. [4]

    A concise overview of safety aspects in human-robot interaction,

    M. Hamad, S. Nertinger, R. J. Kirschner, L. Figueredo, A. Naceri, and S. Haddadin, “A concise overview of safety aspects in human-robot interaction,” inInternational Workshop on Human-Friendly Robotics. Springer, 2023, pp. 1–18

    2023

  5. [5]

    A survey of nonver- bal signaling methods for non-humanoid robots,

    E. Cha, Y . Kim, T. Fong, and M. J. Matari ´c, “A survey of nonver- bal signaling methods for non-humanoid robots,”Foundations and Trends® in Robotics, vol. 6, no. 4, pp. 211–323, 2018

    2018

  6. [6]

    Design, fabrication and control of soft robots,

    D. Rus and M. T. Tolley, “Design, fabrication and control of soft robots,”Nature, vol. 521, no. 7553, pp. 467–475, 2015

    2015

  7. [7]

    Rt-1: Robotics transformer for real-world control at scale,

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsuet al., “Rt-1: Robotics transformer for real-world control at scale,”arXiv preprint arXiv:2212.06817, 2022

    arXiv 2022

  8. [8]

    Manisoft: Towards vision-language manipulation for soft continuum robotics,

    Z. Wei, L. Wang, C. Gao, L. Wen, and S. Liu, “Manisoft: Towards vision-language manipulation for soft continuum robotics,”arXiv preprint arXiv:2605.18617, 2026

    arXiv 2026

Show all 29 references
  1. [9]

    Design and kinematic modeling of constant curvature continuum robots: A review,

    R. J. Webster III and B. A. Jones, “Design and kinematic modeling of constant curvature continuum robots: A review,”The International Journal of Robotics Research, vol. 29, no. 13, pp. 1661–1683, 2010

    2010

  2. [10]

    Model-based control of soft robots: A survey of the state of the art and open challenges,

    C. Della Santina, C. Duriez, and D. Rus, “Model-based control of soft robots: A survey of the state of the art and open challenges,”IEEE Control Systems Magazine, vol. 43, no. 3, pp. 30–65, 2023

    2023

  3. [11]

    Do as i can, not as i say: Grounding language in robotic affordances,

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausmanet al., “Do as i can, not as i say: Grounding language in robotic affordances,”arXiv preprint arXiv:2204.01691, 2022

    2022 arXiv

  4. [12]

    Dynamical movement primitives: Learning attractor models for motor behaviors,

    A. J. Ijspeert, J. Nakanishi, H. Hoffmann, P. Pastor, and S. Schaal, “Dynamical movement primitives: Learning attractor models for motor behaviors,”Neural Computation, vol. 25, no. 2, pp. 328–373, 2013

    2013

  5. [13]

    Conditional neural movement primitives

    M. Y . Seker, M. Imre, J. H. Piater, and E. Ugur, “Conditional neural movement primitives.” inRobotics: Science and Systems, vol. 10, 2019

    2019

  6. [14]

    Implicit behavioral cloning,

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, and J. Tompson, “Implicit behavioral cloning,” inConference on robot learning. PMLR, 2022, pp. 158– 168

    2022

  7. [15]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, 2020

    2020

  8. [16]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inInternational Conference on Learning Representations, 2023

    2023

  9. [17]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inInternational Conference on Learning Representations, 2023

    2023

  10. [18]

    A flow matching framework for soft-robot inverse dynamics,

    H. Yang, F. Yang, Y . Zhang, I. Alsarraj, Y . Wang, Z. Luo, Z. Chen, and K. Wu, “A flow matching framework for soft-robot inverse dynamics,” arXiv preprint arXiv:2604.03006, 2026

    2026 arXiv

  11. [19]

    Lightweight learning from actuation-space demonstrations via flow matching for whole-body soft robotic grasping,

    L. Yang, Y . Bai, Y . Wang, I. Alsarraj, G. Kutyniok, Z. Wang, and K. Wu, “Lightweight learning from actuation-space demonstrations via flow matching for whole-body soft robotic grasping,”IEEE Robotics and Automation Letters, vol. 11, no. 6, pp. 6720–6727, 2026

    2026

  12. [20]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “React: Synergizing reasoning and acting in language models,”arXiv preprint arXiv:2210.03629, 2022

    2022 arXiv

  13. [21]

    Elephants evolved strategies reducing the biomechanical complexity of their trunk,

    P. Dagenais, S. Hensman, V . Haechler, and M. C. Milinkovitch, “Elephants evolved strategies reducing the biomechanical complexity of their trunk,”Current Biology, vol. 31, no. 21, pp. 4727–4737, 2021

    2021

  14. [22]

    Biomimetic tapered soft manipulator with precision and load-bearing capacity,

    X. Li, S. Zhang, Q. Xiong, D. Sui, Q. Zhang, Z. Wang, L. Luan, T. Zheng, J. Fan, J. Zhaoet al., “Biomimetic tapered soft manipulator with precision and load-bearing capacity,”Cell Reports Physical Science, vol. 5, no. 10, 2024

    2024

  15. [23]

    A preprogrammable continuum robot inspired by elephant trunk for dexterous manipulation,

    J. Zhang, Y . Li, Z. Kan, Q. Yuan, H. Rajabi, Z. Wu, H. Peng, and J. Wu, “A preprogrammable continuum robot inspired by elephant trunk for dexterous manipulation,”Soft Robotics, vol. 10, no. 3, pp. 636–646, 2023

    2023

  16. [24]

    Spirobs: Logarithmic spiral- shaped robots for versatile grasping across scales,

    Z. Wang, N. M. Freris, and X. Wei, “Spirobs: Logarithmic spiral- shaped robots for versatile grasping across scales,”Device, vol. 3, no. 4, 2025

    2025

  17. [25]

    Static shape control of soft continuum robots using deep visual inverse kinematic models,

    E. Almanzor, F. Ye, J. Shi, T. G. Thuruthel, H. A. Wurdemann, and F. Iida, “Static shape control of soft continuum robots using deep visual inverse kinematic models,”IEEE Transactions on Robotics, vol. 39, no. 4, pp. 2973–2988, 2023

    2023

  18. [26]

    A synergistic framework for learning shape estimation and shape-aware whole-body control policy for continuum robots,

    M. Kasaei, F. Alambeigi, and M. Khadem, “A synergistic framework for learning shape estimation and shape-aware whole-body control policy for continuum robots,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 13 994–14 000

    2025

  19. [27]

    Geometric continuity, shape pa- rameters, and geometric constructions for catmull-rom splines,

    T. D. DeRose and B. A. Barsky, “Geometric continuity, shape pa- rameters, and geometric constructions for catmull-rom splines,”ACM Transactions on Graphics (TOG), vol. 7, no. 1, pp. 1–41, 1988

    1988

  20. [28]

    Judging llm-as-a-judge with mt-bench and chatbot arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica, “Judging llm-as-a-judge with mt-bench and chatbot arena,” inAd- vances in Neural Information Processing Systems, 2023

    2023

  21. [29]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” inProceedings of the 38th International Conference on Machine ...

    2021

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.