acceptodds
Under review as a conference paper at ICLR 2027

Beyond Verbs and Trajectories: Physical Action Primitives through Local Interaction Laws

Abstract

Similar actions need not yield reusable experience: interactions with similar action labels, trajectories, and nominal outcomes can respond differently when interventions change. We introduce Physical Action Primitives (PAPs), reusable action knowledge objects that preserve local interaction laws and how physical responses vary with intervention. A PAP retains this relationship as a persistent physical operator, supporting outcome prediction and intervention selection for new requests and contexts. Its explicit physical interface allows execution implementations to be replaced or calibrated while the response knowledge is retained. Independent physical audits in matched manipulation contexts show that PAP-based organization reduces audit discrepancy area by 26.7% relative to semantic organization and 19.1% relative to trajectory organization. Closely matched results from affine and nonlinear constructions support this organizing principle across model classes. Retained PAPs predict responses beyond their source probes, guide repeated invocation, and carry source information whose controlled exchange changes control accuracy. Real-robot experiments on a Franka Panda demonstrate prediction and invocation from real interaction evidence without real-world fine-tuning. These results establish a basis for robot experience to accumulate as libraries of persistent physical operators that planning, execution, and adaptation can access through a common interface. Code: https://anonymous.4open.science/r/PAP-B093/

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, action primitives, physical interaction, dynamics modeling, reusable skills, manipulation, system identification

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 2 reviews, above 27% of the others.
Reviewer_qtF3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

this paper presents PAP, which is basically an action primitive model for robot arm tasks combined with a learned local dynamics model. The dynamics model captures the system dynamics around a primitive action and predicts how the system responds when the interaction conditions change. With this capability, control can be computed based on model predictions, inverse dynamics, or rejection sampling. This idea is similar to model-based control methods like MPC, but the interesting part is that the paper proposes to learn a model to predict the system dynamics, which is different from previous work that trains a model to predict the next state. To achieve this, the paper first pretrains an encoder on a set of tens of millions of trajectories. This encoder can extract the necessary features from a few demonstrations, and the system dynamics prediction model is then trained on a couple of new demonstrations. The main paper does not provide comparisons with baseline methods (I am not sure if DMP and SPiRL are considered baseline methods, as the paper does not state so, nor does it discuss why they are chosen and what they are compared against). Based on my understanding, the system dynamics prediction model is quite overfitted to the specific tasks, and it can only handle small changes to the settings. The generalization and practicality of the proposed method are limited.

Critical strengths

the paper proposes to include a local system dynamics model as part of an action primitive, which is an interesting idea.

the paper proposes to train a model to predict the dynamics itself, which is a different approach from those that train a model to predict the next state.

Critical weaknesses

The main paper is not very well written. It appears to keep introducing new concepts and terminology (such as Interaction Laws, intervision, query, invoke, probe, etc.) whose meanings are not easy to guess, making it difficult to identify what is existing work can achieved and what the key technical contributions of the paper are. The key technical part, particularly how the system dynamics model is trained, is explained very vaguely and briefly. Even with the appendix, it is still hard to understand the details without guessing. Many metrics used in the paper, such as the audit discrepancy area in Fig. 4, do not appear to be commonly used metrics in the field, and the numbers are not obviously interpretable.

while I think the idea of learning a model to predict local system dynamics is interesting, I have strong concerns regarding the generalization that the proposed method can achieve. The three tasks used in the paper are simple and of similar types. The system dynamics are also very simple. The paper does not compare the method with more directly learned dynamics models, for example, as simple as a linear model trained using least squares with some intuitively selected features. On the other hand, if I understood correctly, the system dynamics model is kind of overfitted to the specific configuration of a task, which further limits its generalization ability. The paper may need to test on task with more complicated dynamics.

the paper lack of comparison with existing method, as I mentioned in summary

Questions for authors

please correct me if I am wrong with second paragraph in "Weakness"

8 comments
AuthorsOpenReview ↗

Short clarifications, and how the paper fits together

Thank you very much for your thoughtful review, and especially for highlighting the core idea of our paper: proposing to "include a local system dynamics model as part of an action primitive", and training a model "to predict the dynamics itself" rather than merely predicting the next state.

Thank you also for inviting us to clarify the points in the second paragraph of your Weaknesses. We provide brief, direct clarifications below, followed by full experimental details in the following comments.

  1. "The main paper does not provide comparisons with baseline methods." The main paper compares PAP against four external baselines in Fig. 4 and Sec. 4.2 (Oracle Semantic, DMP, SPiRL, and Endpoint), where the first paragraph of Sec. 4.2 states what each baseline represents (see next comment: external baselines in the main paper). Furthermore, following your and Reviewer ZdaD's feedback, we have now added end-to-end control comparisons against two least-squares baselines and against DMP replay and DMP goal scaling (see our reply end-to-end Invoke comparison against external baselines to Reviewer ZdaD).
  2. "… a linear model trained using least squares with some intuitively selected features." Thank you for suggesting this natural baseline—we implemented and evaluated it in two versions. All four learned PAP constructions achieve lower prediction error than least squares fitted to the same seven pushes across all three tasks (12 of 12 comparisons), and our three nonlinear PAP constructions outperform both per-state least squares and feature-engineered least squares across all three tasks (see comment least squares, and how the model is trained).
  3. "… trained on a couple of new demonstrations … overfitted." To clarify a key architectural point: our response model is trained once on 900 pooled training conditions across all three tasks and then frozen; no weights are ever trained or fine-tuned on new demonstrations. We also tested your overfitting concern directly: under unseen changes to hidden physics, frozen PAP-Endpoint maintains 71–77% success (vs. 72.5% unchanged), and the same frozen, simulation-trained model works on a real robot across two surfaces without fine-tuning (see comment direct tests of overfitting: hidden physics).
  4. Terminology, metrics, and training details. We provide an intuitive, plain-language reading of the Fig. 4 metric in the second comment, and we have drafted a complete glossary, a five-step training description, an explicit contribution list, and a related-work discussion in the last comment.

How our core contribution connects the three steps of the paper. As you perceptively noted, our approach predicts the local interaction law itself (how the object's physical response changes as a function of the candidate intervention), rather than predicting only a single next state. Our paper establishes this paradigm across three connected steps:

  1. Organizing experience (Sec. 4.2, Fig. 4). Two pushes that share the same verb and have tool paths within 0.5 mm still respond differently to new pushes in 47.6% of cases due to hidden physical differences. Organizing experience by the local interaction law (PAP) yields more physically consistent neighborhoods than organizing by verb (Oracle Semantic), trajectory (DMP), learned skill latent (SPiRL), or a single nominal outcome (Endpoint). This is the primary test of our core organizational claim against external baselines.
  2. Predicting responses (Sec. 4.3, Fig. 6). From seven test pushes and a single forward pass, the frozen model predicts physical responses to 18 unseen candidate pushes (Query).
  3. Control and execution (Sec. 4.3–4.5, Table 1). The retained operator selects interventions for new target effects (Invoke) and remains valid when execution timing changes (Substitute) or when hardware is recalibrated (Correct)—both in simulation and across 480 new real-robot trials.

Importantly, the learned network is one concrete construction of a PAP; our formal definition of the primitive interface is model-class agnostic (Sec. 1)

AuthorsOpenReview ↗

External baselines in the main paper, and the Fig. 4 metric in plain words

1. External baselines in the main paper (Fig. 4, Sec. 4.2). The main paper compares PAP against four external baselines—including DMP and SPiRL—in Fig. 4 and Sec. 4.2. As stated in the first paragraph of Sec. 4.2: "DMP (Ijspeert et al., 2013) represents measured tool trajectories, while SPiRL (Pertsch et al., 2021) represents state–action sequences … These comparisons test the physical information preserved by each organization criterion." The section then reports that "response-based neighborhoods have lower independently measured discrepancy than all four comparison organizations, with each paired interval below zero."

Why each baseline was chosen. Each baseline represents a widely used paradigm in robotics for deciding when two interactions belong to "the same reusable skill":

Method in Fig. 4Paradigm it representsCalls two pushes alike when
Oracle Semanticsymbolic action labelsthey share the true verb label
DMPtrajectory primitivestheir tool trajectories have the same shape
SPiRLlearned skill latentstheir state–action sequences map to nearby skill latents
Endpointnominal outcomesone nominal push gives the same measured outcome
PAP (ours)local interaction lawstheir predicted responses agree over a range of pushes

2. Intuition behind the audit test and the Fig. 4 metric. Suppose an organizational method claims that two interactions are alike (neighbors). To test whether they are truly physically compatible, we apply 18 new, independent pushes to both objects and measure how their physical responses differ. In Sec. 4.2, we evaluate 320 query interactions against 640 stored gallery interactions on RollBall across five hidden damping values; each method locks its neighbor rankings before the 18 scoring pushes are collected.

  • Plain-language meaning of the Fig. 4 numbers. The audit discrepancy is the measured difference between a query and its retrieved neighbors across the 18 new pushes, averaged over displacement, rotation, linear velocity, and angular velocity, each divided by its natural physical scale (5 mm, 5°, 0.05 m/s, 1 rad/s). A value of 0.25 corresponds to a 5 mm difference in displacement alone. The "area" simply averages this discrepancy across three neighborhood sizes (the closest 1%, 2%, and 5% of candidates; lower is better). PAP achieves 0.118, reducing discrepancy by 26.7% relative to Oracle Semantic (0.161) and by 19.1% relative to DMP (0.146), with all paired confidence intervals against SPiRL and Endpoint strictly below zero.
  • Two directly interpretable percentages from the same recorded audit.
    • Among pairs of pushes that share the same verb, the same initial context, and tool paths within 0.5 mm, 47.6% are physically distinguishable under new pushes (differing by more than 5 mm or 5° on at least 2 of the 18 test pushes). Neither a semantic label nor a trajectory representation can separate them.
    • When we take pairs that start from the exact same state and differ only in hidden damping, and ask each method for the top 10% of pairs it considers most alike, the fraction of false matches (pairs whose measured responses are in fact physically different) is only 1.6% for PAP, compared to 4.7% for SPiRL, 26.6% for Endpoint, 43.8% for DMP, and 45.3% for Oracle Semantic (random chance: 52%).

Why controlled planar tasks are ideal for this test. Here, using clean, controlled tasks is an intentional design choice: it demonstrates that even in a simple setting without complex multi-stage contact, look-alike interactions conceal substantial physical response differences (47.6%) that standard semantic and trajectory primitives miss.

We will make the role of each external baseline explicit at the beginning of Sec. 4.2 and add the two intuitive percentages above.

AuthorsOpenReview ↗

Least squares, and how the model is trained

Thank you for suggesting this baseline. We implemented and evaluated two least-squares baselines under the exact Fig. 6 protocol. All four PAP constructions achieve lower prediction error than least squares fitted to the same seven pushes across all three tasks (12 of 12 comparisons), and our three nonlinear PAP constructions outperform both per-state least squares and feature-engineered least squares across every task.

Clarification on training: the model is trained once and frozen, never trained on new demonstrations.

  • Each learned PAP construction is trained once on 900 training conditions pooled across the three tasks (300 per task, with no task identity label; Appendix B.5) and then permanently frozen.
  • At test time, for a new state, the robot executes seven test pushes (probes). A single forward pass compresses them into a 32-number summary u (Sec. 3.2), with zero parameter updates.
  • Every evaluation state is generated using a random seed unseen during training.

The two least-squares baselines we evaluated:

  • Least squares on the seven pushes: an affine response model fitted directly via least squares to the seven test pushes at each state. Notably, this serves as the simplest linear instantiation of our PAP interface.
  • Least squares with features: ridge regression fitted on the exact same 900 pooled training conditions as our neural models (16,200 held-out pushes), using intuitively selected features: a constant bias, the candidate push, the context h, the seven measured probe responses, and their cross-products with the candidate push. Regularization was tuned on the development set and locked prior to evaluation.

In the table below, prediction error is the sum of squared errors across the six response components, each divided by its physical unit (5 mm, 5 mm, 5°, 0.05 m/s, 0.05 m/s, 1 rad/s), where 1.0 corresponds to a one-unit error (e.g., 5 mm) in one component. Values are reported as PushT / PokeCube / RollBall.

ModelPrediction error (lower is better)Invoke success, %
Nearest of the seven pushes1.642 / 0.487 / 0.5331.7 / 16.7 / 0.0
Least squares on the seven pushes0.855 / 0.082 / 0.31045.7 / 79.6 / 58.1
Least squares with features0.595 / 0.054 / 0.20339.8 on PushT
PAP-Affine0.618 / 0.060 / 0.21338.8 / 80.2 / 99.6
PAP-Nonlinear0.417 / 0.043 / 0.06362.6 / 85.6 / 86.5
PAP-Endpoint0.376 / 0.046 / 0.06872.5 / 86.7 / 90.3
PAP-ActionSeq0.395 / 0.037 / 0.05472.5 / 87.0 / 100.0
  • Response Prediction (Query). All four learned PAP constructions predict more accurately than least squares on the seven pushes in every task. Least squares with hand-crafted cross-product features closely matches PAP-Affine. Crucially, all three nonlinear PAP constructions outperform both least-squares baselines across all three tasks (reducing prediction error from 0.595 to 0.376 on PushT, and from 0.203 to 0.054 on RollBall).
  • Intervention Selection (Invoke). On PushT and RollBall—where the physical response varies nonlinearly with the push—our nonlinear PAP constructions outperform least squares by 17–27 percentage points on PushT and 28–42 percentage points on RollBall.

We will add both least-squares baselines to Sec. 4.3 and revise Sec. 3.2 to make the one-time pooled training and frozen test-time protocol unmistakably clear.

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

This paper proposes Physical Action Primitives (PAPs), reusable action knowledge objects that retain a local intervention–response operator together with its operating domain and physical conventions. Given a context and candidate intervention, a PAP predicts the physical response, and Invoke uses this operator to select an intervention for a desired effect. The operator can be instantiated by locally affine or nonlinear learned models and is not a physics simulator. Substitute preserves the PAP, request, and selected spatial intervention while replacing the execution realization, such as temporal profiles, and Correct preserves the PAP and request while calibrating the execution mapping from measurements. Experiments show that response-based organization yields more physically compatible neighbors than semantic, trajectory, learned-skill, or nominal-effect organizers, and that retained operators support prediction, invocation, substitution, and correction.

In the matched RollBall organization study, PAP-Affine achieves the lowest independent audit discrepancy (0.11834 vs. 0.16138 for Oracle Semantic and 0.14635 for DMP), while the construction study compares only PAP variants, the execution study shows that temporal substitution preserves performance and that correction recovers success from 0/72 to 59/72 on PokeCube and 72/72 on RollBall but only 1/72 on PushT, and the real-robot study reports Query position MAE of 3.673 mm for PushT and 1.289 mm for PokeCube with Invoke success rates of 10/20 and 12/20 for 5/10 mm requests on PushT and 13/20 and 12/20 on PokeCube.

A key limitation is that the construction, execution, and real-robot experiments compare only PAP variants or conditions, and although the organization study includes external organizers, it lacks an end-to-end Invoke comparison against methods such as DMP or SPiRL.

Critical strengths

Strengths include a clear formalization of PAPs, a carefully designed organization study with locked rankings and independent physical audits, a model-class-agnostic Query–Invoke interface, and a clean separation between retained response knowledge and execution realization, supported by substitution, correction, source-exchange, and real-robot experiments

Critical weaknesses

Although the paper presents a clear and interesting object definition and a convincing organization study, the empirical evidence does not adequately support the broad claim that PAPs are first-class reusable physical operators; in particular, the construction, execution, and real-robot experiments lack end-to-end external baselines, and the operational claims are demonstrated only under narrow conditions.

The real-robot evaluation is limited to small-scale Query and Invoke trials, while Substitute, Correct, source exchange, and the organization audit are simulation-only; since Substitute and Correct concern execution realization and calibration, hardware validation would be important to support the paper’s operational claims.

Questions for authors

The real-robot evaluation is limited to Query and Invoke, while Substitute, Correct, source exchange, and the organization audit are simulation-only. Could the authors clarify whether Substitute and Correct were attempted on hardware, and if not, whether there is a specific reason to expect that the execution-calibration and substitution claims would not hold on a real robot?

7 comments
AuthorsOpenReview ↗

Substitute and Correct on the real robot (480 trials)

Thank you very much for your careful review and for appreciating our formalization of PAPs, the locked-ranking organization study, and the architectural separation between retained response knowledge and execution realization.

Both Substitute and Correct hold on the real robot, directly validating our core design: because a PAP decouples the retained physical response operator from downstream execution, the same frozen simulation-trained operator remains valid on hardware when execution timing changes or when a miscalibrated arm is recalibrated. While hardware Substitute and Correct were not part of the initial submission, we ran 480 new real-robot trials across two objects and two friction surfaces for this response. Replacing the temporal speed profile (Substitute) preserves task outcomes (13–19 successes of 20, matching the 15–17 reference). A miscalibrated arm fails every time (0 of 160), whereas after Correct estimates the execution scale from physical measurement and rescales the command, success returns to the reference level (16–17 of 20).

Hardware Setup. On a real Franka Emika Panda arm, we evaluated a 40 mm cube weighing 62 g and the T block across two surfaces with distinct friction (wood, μ ≈ 0.42; foam, μ ≈ 0.57). In every trial, the robot is requested to push the object 20 mm forward. Following Sec. 4.5, the push displacement is selected by the frozen PAP model trained entirely in simulation(with zero real-world training or fine-tuning), and the retained PAP is kept unchanged across all conditions. A trial is counted as a success if the measured object displacement is within 2 mm of the 20 mm target (even stricter than the 5 mm criterion used in the paper). Distances and rotations were measured against start and target reference lines drawn on each surface using a ruler and a protractor.

Successes of 20 (mean distance ± SD, mm)Cube, foamCube, woodT block, foamT block, woodTotal
Linear (constant speed)15 (20.3 ± 1.7)17 (18.8 ± 1.2)16 (19.3 ± 1.5)17 (19.7 ± 1.3)65/80
Substitute: Smoothstep13 (19.2 ± 1.8)17 (19.7 ± 1.6)19 (19.2 ± 0.8)15 (20.2 ± 1.6)64/80
Substitute: Front-loaded17 (19.7 ± 1.3)14 (20.3 ± 1.9)16 (20.1 ± 1.6)15 (20.5 ± 1.7)62/80
Mismatch: arm travels 55% of each command (run twice, 40 trials per cell)0 (11.1 ± 1.5)0 (11.0 ± 1.3)0 (10.9 ± 1.4)0 (10.6 ± 1.8)0/160
Correct: same faulty arm, command rescaled17 (19.4 ± 1.3)17 (20.2 ± 1.5)16 (20.2 ± 1.3)16 (19.8 ± 1.6)66/80
  • Substitute holds on hardware. Replacing the Linear constant-speed profile with Smoothstep or Front-loaded (the exact temporal profiles of Table 1 and Appendix E.2), while keeping the PAP and selected spatial displacement fixed, shifts mean object displacement by at most 1.5 mm.
  • Correct holds on hardware. When the arm travels only 55% of every command, the object stops at roughly 11 mm instead of 20 mm (0/160 successes). Estimating the execution scale from physical measurement—without reading the injected value—yields 0.542. Dividing each command by this measured gain (commanding 37 mm instead of 20 mm) while leaving the arm miscalibrated and the PAP untouched restores mean object displacement to 19.4–20.2 mm (66/80 successes).
  • Under the paper's original criterion (within 5 mm position error and 5° rotation): all 240 trials with the unmodified arm and all 80 corrected trials succeed (100%), whereas 0 of the 160 miscalibrated trials succeed. No trial rotated the object by more than 5°.

We will add this hardware table to Sec. 4.5 and include photographs of the physical setup and measurement marks in Appendix F.

AuthorsOpenReview ↗

End-to-end Invoke comparison against external baselines

We ran this end-to-end Invoke comparison under the exact protocol of Fig. 6. The results highlight our core contribution: methods that reuse a trajectory without retaining an intervention–response operator succeed in only 0–17% of calls once the requested effect changes, whereas PAP constructions achieve 39–100% success.

The table below reports Invoke success (%) on the extrapolation requests of Fig. 6 (15, 18, and 21 mm forward, beyond the range of the seven test pushes), using the exact same seeds, source and receiving states, requests, and success criteria as Fig. 6 (540 calls per baseline; 2,160 calls across four seeds per learned PAP construction).

MethodWhat it reusesPushTPokeCubeRollBall
DMP replaythe demonstrated trajectory (12 basis functions per dimension)3.90.00.0
Nearest-probe replaythe closest of the seven test pushes1.716.70.0
DMP goal scalingthe trajectory, scaled by requested / demonstrated displacement6.30.00.0
Least squares on the seven pushesa linear response model fitted to the seven pushes45.779.658.1
PAP-Affinea learned, locally affine response model38.880.299.6
PAP-Nonlineara learned nonlinear response model62.685.686.5
PAP-Endpointa learned response model with a nominal-effect bottleneck72.586.790.3
PAP-ActionSeqa learned response model with action-sequence supervision72.587.0100.0
  • Trajectory reuse fails once the requested effect changes. When requests remain near the demonstration (6–12 mm), DMP replay succeeds in 67.8 / 62.0 / 33.3% of calls. Once the request moves beyond the demonstration, however, DMP replay, nearest-probe replay, and DMP goal scaling drop to 0–17%. Notably, DMP goal scaling fails despite using DMP's built-in spatial generalization mechanism, because physical object displacement is not proportional to tool travel: on PushT, scaling up the stroke overshoots and induces unwanted rotation. Paired across the 20 source blocks, PAP-Endpoint outperforms DMP replay by 68.6, 86.7, and 90.3 percentage points across the three tasks.
  • Retaining a response operator is what enables goal-conditioned reuse. Fitting least squares directly to the seven test pushes is the simplest linear instantiation of our model-agnostic PAP interface, and it immediately lifts success to 45.7–79.6%. Our three learned nonlinear PAP constructions improve further across all three tasks (reaching 62.6–100.0%), while PAP-Affine is slightly below per-state least squares on the more nonlinear PushT task (38.8 vs. 45.7%).

On SPiRL. You named SPiRL alongside DMP, so we address it explicitly. SPiRL (Pertsch et al., 2021) learns a skill prior and then trains a high-level policy by reinforcement learning for each downstream task. Invoke instead answers a new request from seven test pushes, with no training. SPiRL therefore has no counterpart in this setting: giving it reinforcement learning for every request would compare very different interaction budgets, and using its skill decoder without a trained high-level policy would amount to replaying a demonstrated skill, the case covered by the replay rows above. We compare with SPiRL where it applies directly, as a criterion for organizing experience: in Fig. 4, PAP yields lower audit discrepancy than SPiRL, with the paired interval below zero.

We will add this end-to-end comparison table alongside Fig. 6 in Sec. 4.3.

AuthorsOpenReview ↗

The 1/72 on PushT

Thank you for pointing out this contrast in Table 1. We investigated the 1/72 result on PushT and found that it was caused by a protocol artifact during calibration rather than a limitation of Correct. When the two calibration strokes are executed in free space, Correct achieves 21/72 on PushT against a reference of 23/72 (PAP-Affine), and 48/72 against 50/72 (PAP-Endpoint).

Why Correct was 1/72 on PushT in Table 1. In the submitted script, the two calibration strokes were inadvertently executed against the object. On PushT, the heavy T block obstructed the tool: when commanded to move 11.0 and 16.5 mm, the tool moved only 9.0 and 11.5 mm. Consequently, the difference-based scale estimator severely underestimated the arm's gain (median 0.212 vs. the true value of 0.55, with 11 of 24 estimates hitting the lower clip of 0.2), over-amplifying subsequent commands. On PokeCube and RollBall, the lighter objects did not obstruct the tool during calibration (yielding accurate scale estimates of 0.549 and 0.543).

Rerun in free space. Because Correct is defined to calibrate the tool's execution mapping from tool motion measurements (Sec. 3.3), the two calibration strokes belong in free space. With that single fix and everything else held identical, reference performance is restored:

PushT, successes of 72 (mean position error)ReferenceMismatchCorrected (submitted: calibrated against the object)Corrected (calibrated in free space)
PAP-Affine23 (6.4 mm)01 (39.95 mm)21 (6.52 mm)
PAP-Endpoint500148
  • The estimated execution scales become 0.547 / 0.549 / 0.543 across the three tasks.
  • PokeCube and RollBall remain at their reference levels (60/72 and 72/72).
  • As a reproducibility control, rerunning the submitted setting reproduces all 189 published records identically. The code will also be updated via an anonymous GitHub link.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.