acceptodds
Under review as a conference paper at ICLR 2027

Learning Skills from Historical Action Trajectories: Action Experience Dictionary for World Action Models

Abstract

World Action Models (WAMs) couple visual dynamics prediction with action generation, yet they do not explicitly support the reuse of action experience across manipulation tasks. Furthermore, existing WAMs struggle to capture underlying cross-task semantic relationships that could guide target action prediction, as redundant background elements interfere with the extraction of key visual information. To address these challenges, we develop a novel Action Experience Dictionary (AED) that encodes historical physical action trajectories into shared action embeddings to support skill reuse and model cross-task relationships. Specifically, we first aggregate historical actions to align with visual observations and retrieve action embeddings from the AED using a pretrained action tokenizer. Subsequently, we visually condition the pooled embeddings through cross-attention and prepend them to noisy action tokens, providing interaction context and action intent for prediction. To encourage action-related motion modeling and reduce reliance on irrelevant background cues, we introduce a motion-aware transition loss that supervises visual feature change prediction over random temporal intervals. Experiments on simulation benchmarks and in real-world cross-embodiment settings demonstrate that our model outperforms the baselines. The anonymous project website is available at [https://anonymous.4open.science/w/supplementary_materials-4586/](https://anonymous.4open.science/w/supplementary_materials-4586/).

Primary area:
applications to robotics, autonomy, planning
Keywords:
World action model

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_ymMgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the experience reuse of world action models. The authors argue that WAMs struggle to capture underlying cross-task semantic relationships that could guide target action prediction. To this end, the authors propose Action Experience Dictionary to encode historical physical action trajectories. To model action-related motion and reduce reliance on irrelevant background cues, the authors propose motion-aware transition loss. Experiments show the effectiveness of the proposed method

Critical strengths
  1. Overall the paper is well-written and relatively easy to understand.
  2. The experiments show the effectiveness of the proposed method.
Critical weaknesses
  1. The proposed method can not be distinguished from prefix tuning, or to be clearer, adding more learnable parameters to the model. The proposed AED is essentially an embedding layer. The dictionary encoder is also essentially more learnable parameters. As the losses are standard loss functions for world action models, no structural prior was introduced for these modules. Therefore there is no evidence that the proposed modules do learn the expected knowledge.

  2. The proposed method rely on historical experiences, while the authors did not explain how to cold-start the model.

  3. FAST tokenizer tokenizes physical trajectories described by rot6d, which hardly contains any semantic meaning such as grasping or carrying. Therefore, if FAST tokenizer outputs an ID, it does not necessarily mean "grasping". The soundness of the proposed method is suspicious.

  4. The proposed AED does not require large-scale pretraining. Therefore, there is no guarantee that it can learn a wide range of semantics used by robot manipulation tasks.

Questions for authors
  1. Please clarify what the AED and dictionary encoder has learnt through experiments. Even qualitative visualization is better than no further explanation.

  2. Please explain how to start the model when the model has no historical experiences.

  3. Does the cooperation of pre-defined FAST tokenizer and AED trained from scratch have concrete meanings? When FAST tokenizer outputs a certain ID, what does this ID mean to AED?

23 comments
AuthorsOpenReview ↗

Response to Reviewer ymMg: Clarification on AED (1/5)

1. AED is not an unconstrained prefix tuning method. a) AED is a shared dictionary of 2,048 learnable entries, and pretrained FAST+ action-token indices retrieve these entries. Thus, occurrences assigned the same index across tasks are tied to the same learned vector: this imposes a specific cross-task parameter-sharing structure. The retrieved action embeddings are then conditioned on historical visual features and compressed into eight prefix tokens; they are not eight independent, freely learned prefix vectors. The dictionary contains approximately 2.10M parameters, but its defining property is the indexing and sharing constraint, not parameter count alone. MT is a separate, explicitly defined objective, not a relabeling of the standard WAM flow-matching loss. In Eq. (7), the model uses action-expert hidden states conditioned on the AED representations to predict visual-feature changes over randomly sampled temporal intervals. This directly couples reusable action representations to their visual consequences and provides action-conditioned transition supervision. It is jointly optimized with the standard flow-matching objective as the total loss in Eq. (10).

AuthorsOpenReview ↗

Response to Reviewer ymMg: Clarification on cold start (2/5)

2. Cold start. Appendix C.1, “Model and Representation Configuration,” states that AED contains 2,048 valid entries plus a separate PAD entry; at cold start, the PAD entry represents the empty action history. Initialization details are given in Sec. 4.1 and Appendix C.1: the video expert is initialized from pretrained Wan2.2 weights (Sec. 4.1; Appendix C.1), the ActionDiT action expert is initialized by linear interpolation of the video-expert parameters (Appendix C.1), and the AED action embeddings are randomly initialized (Sec. 3.2).

AuthorsOpenReview ↗

Response to Reviewer ymMg: Clarification on FAST+ tokenizer (3/5)

3: FAST+ tokenizer and AED. a) A high-level skill can comprise multiple motion patterns, each indexed by a FAST+ ID, an single ID therefore denotes a discretized action pattern. A high level skill such as “grasp” would be represented by a sequence of ID. FAST+ maps each action segment to a discrete ID squence based on its motion pattern; AED uses that ID as an address to retrieve a trainable embedding from its shared dictionary. Thus, when FAST+ outputs ID kk, it tells AED which dictionary vector to retrieve and share with other occurrences of ID kk across tasks. b) In Fig. 5 and Fig.9 we provide the visualization results.

Reviewer_yon1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes an Action Experience Dictionary (AED) for world action models. Historical actions are aggregated, tokenized using FAST+, mapped to shared learnable embeddings, and conditioned on historical visual features. The resulting tokens prefix the action expert. A motion-aware transition objective additionally predicts visual feature differences over sampled temporal intervals. The paper reports results on LIBERO, RoboTwin 2.0, LIBERO-Plus, and two real-robot platforms.

Critical strengths
  • Combining action history with visual context is well motivated: similar motions can correspond to different interaction stages and goals.
  • The evaluation covers several simulation benchmarks and real-robot settings. The LIBERO-Plus improvement over Fast-WAM, from 70.8% to 86.2%, is particularly promising.
  • The appendix provides useful implementation details and comparisons involving continuous-action encoding, tokenization order, visual-memory size, and prefix injection.
Critical weaknesses
  • The central claim of cross-task skill reuse is not sufficiently isolated. Shared FAST+ token IDs demonstrate shared dictionary access, but do not by themselves establish semantic skill transfer. The five-versus-ten-task experiment also changes data availability and potentially optimization exposure. Although an MLP history-encoding baseline is included, it is not evaluated under the same transfer experiment. The paper therefore does not yet distinguish AED-specific transfer from ordinary multitask learning and history conditioning.
  • The mechanism behind the robustness gains remains unclear. Component ablations are primarily cumulative and evaluated on clean LIBERO-10 with one seed. Removing MT changes success from 97.4% to 96.8%, corresponding to three outcomes among 500 trials. Since robustness is the strongest empirical result, component ablations on LIBERO-Plus would be more informative.
Questions for authors
  • Under matched target-task exposure and optimization budgets, does AED obtain a larger benefit from auxiliary tasks than continuous history encoding or a task-specific dictionary control?
  • How do AED, visual conditioning, and MT individually affect LIBERO-Plus performance, particularly camera and noise robustness? Are the smaller clean-benchmark gains stable across training seeds?
5 comments
AuthorsOpenReview ↗

Response to Reviewer yon1: Clarification on cross-task skill reuse (1/2)

1. Cross-task skill reuse Under matched experimental conditions, including the same optimization budget and exposure to the target tasks, we compare FastWAM, history-MLP, and our method in the five- versus ten-task settings. As shown in the table, our method achieves a gain of 3.6, compared with 2.2 for FastWAM and 2.0 for history-MLP. These results suggest that the improvement from AED cannot be attributed to multi-task training alone and AED improve performances with cross task skill reuse.

Method5 tasks10 tasksGain
Fast-WAM93.695.8+2.2
history-MLP93.495.4+2.0
Ours94.097.6+3.6
AuthorsOpenReview ↗

Response to Reviewer yon1: Clarification on ablaton (2/2)

2. LIBERO Plus ablation

VariantCam.Robot.Lang.Light.Back.Noise.Layout.Avg.
Ours (full)43.950.186.695.765.578.966.469.6
-MT37.051.475.986.760.173.161.663.7
-VC37.348.970.281.753.469.959.360.1
-AED36.844.367.681.442.764.556.756.3

On LIBERO-10 task of the LIBERO Plus, the full model achieves 69.6% average success. Removing MT, VC, and AED reduces the average to 63.7%, 60.1%, and 56.3%, respectively (drops of 5.9, 9.5, and 13.3 percentage points). The full model scores higher than each ablated variant in six of seven categories. We has added the ablation study (Tab. 6) in the submission.

Thanks for adding the five-versus-ten-task comparison and the LIBERO-Plus ablations. These are useful additions that strengthen the evidence for transfer and robustness.

Could you also report results across multiple training seeds for the transfer comparison and key ablations, including means and standard deviations? Also, does each row in Table 6 remove only the named component while keeping the other components and training settings unchanged, or are the removals cumulative as in Table 3?

Reviewer_E35xOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the Action Experience Dictionary (AED), which maps historical actions into shared learnable action representations after aggregation and FAST encoding. These representations are then conditioned on historical visual information and injected into the action prediction branch of WAM in the form of prefixes. In addition, the authors propose the motion-aware transition (MT) loss to supervise the prediction of future visual feature changes from action hidden states. The paper is evaluated in simulation and on two real robot platforms, demonstrating overall performance improvements.

Critical strengths

1. This paper is well-written and easy to follow, with a clear structure spanning motivation, problem definition, method design, and experiments. The ablation studies and quantitative results on both simulation and real-robot experiments demonstrate that the proposed model is competitive with state-of-the-art methods. The provided demo videos and project pages are also clearly presented and easy to read.

2. This paper proposes AED, which learns a set of FAST-tokenizer-indexed action pattern embeddings as reusable skills across various manipulation tasks. The idea of employing learnable embeddings as action representations, embedding reusable skills for WAMs, is conceptually sound.

3. The motivation of discussing how similar motions serve different purposes depending on the objects being manipulated and their spatial relationships is insightful and novel.

Critical weaknesses

1. My major concern is whether AED preserves and exploits semantic or functional distinctions within the same skill across different tasks.

The paper motivates its approach by observing that similar motions can serve different purposes depending on the manipulated objects and their spatial relationships, and presents visual conditioning as supplying interaction context and action intent. However, these implementations still seem to lack modelling of differences in semantic intent for the same skills across different tasks.

For example, a similar reach–grasp–lift sequence involving the same layout, including a cup and a bottle, may precede either placing the bottle upright on a shelf or pouring its contents into the cup. While the historical reachreach pattern can be shared, the affordance place of graspgrasp might be quite different.

I consider that AED lacks explicit learning for representing the diversity and ambiguity within each reusable skill. AED relies on learnable soft prompts and the FAST tokenizer indexed update mechanism; this approach may actually blur the semantic distinction between tasks that share the same skill. Moreover, using only VAE features for supplementing the semantic context may be insufficient to achieve the goal, compared with features from visual-linguistic pretrained models.

I would ask the authors to provide additional analysis on whether the learned AED embeddings exhibit sufficient intra-skill diversity and cross-task discriminability. I have already observed in Appendix Figure 9 that some skills, such as grasp and transport, show a multimodal distribution in cross-task similarity. However, I would like to see further evidence, for example, first, the internal variance of the learned AED embeddings, and second, whether reusable skills (embeddings) are indexed by different inherent patches across tasks.

If the learned AED embeddings do not achieve a sufficiently rich cross-task semantic pattern for the same skill, then AED's role may lie more in summarizing identical action patterns into skills, a direction that has in fact already been explored in prior work. What is more, the quality of AED embeddings might be sensitive to the configuration of training data and evaluation tasks. I suggest the authors additionally evaluate this method on sim or real tasks that require unseen skills or novel intents.

2. Limited improvement on simulation benchmarks.

AED reports 98.8% on LIBERO and 92.8% on RoboTwin 2.0, closely matching Faster-WAM, which achieves 99.0% on LIBERO and 92.6% on RoboTwin 2.0, and AHA-WAM, which achieves 92.8% on RoboTwin 2.0. However, its reported inference latency is 518.7 ms, which is substantially higher than the 252.95 ms of Faster-WAM and the 41.37 ms of AHA-WAM. It would be valuable to know whether AED also works for VLA baselines.

The method would be much more convincing if it could be shown to enhance VLA baselines while reducing latency, given that it appears to make only minor modifications to the video expert branch.

Questions for authors

1. Motivation of Motion-Aware Transition Loss.

I am kind of confused about the motivation for Motion-Aware Transition Loss, as the DINOv3 feature already contains tolerance for background noise, so decoupling the background and foreground features seems to be incremental.

What is more, the JEPA feature might focus more on transitions and dynamics. It seems to be more straightforward to replace the VAE feature with JEPA latents as the historical visual latent.

2. Is AED suitable for cross-embodiment? Can AED embeddings learned from trajectories of Spirit AI MOZ1 enhance the performance of the checkpoints of a new embodiment?

3. After all, I am mostly concerned about whether the selected historical visual latent could provide enough semantic or intent priors for aggregating and using AED; the semantic labels, for example, the instructions, might be more useful to this model.

I suggest the authors further explain or show demonstrations of how their method distinguishes or uses the semantic and intent cues of similar motions that serve different purposes to improve the generalization or performance of baseline models.

I will raise my score if most of my major concerns are addressed.

11 comments
AuthorsOpenReview ↗

Response to Reviewer E35x: Clarification on Semantic or functional distinctions (1/5)

  1. Semantic or functional distinctions. AED preserves semantic or functional distinctions between tasks sharing the same skill via visually conditioned action representations, MT Loss and instruction-conditioned prediction.
  • Modeling of differences of semantic intent . a) Visual conditioning (VC) compresses historical visual latents into z^vh\hat{\mathbf{z}}^h_v with fixed size via learnable queries (Eq. 5); the dictionary encoder attends to z^vh\hat{\mathbf{z}}^h_v using retrieved action embeddings as queries (Eq. 6), encoding manipulated objects, spatial relations, and interaction states—so a shared dictionary entry yields task-dependent representations. Text instructions further condition the model via cross-attention, specifying the goal while visual history supplies scene context. b) The MT loss supervises AED-conditioned states to predict patch-level visual-feature differences Δh=F(ot+Δ)−F(ot)\Delta h=F(o_{t+\Delta})-F(o_t) (Eq. 7), where static features partially cancel, emphasizing interaction-relevant changes. On LIBERO-Plus, removing MT drops success from 69.6% to 63.7%, and removing VC to 60.1%, confirming both visual and instruction-based priors encode intent differences.

  • Intra-skill diversity. Across 2,048 L2-normalized dictionary embeddings, total variance is 0.9995 with entropy effective rank 798.05/1,024. Analyzing 60,960 action intervals across 40 tasks, task-dependent differences explain 9.1%–24.9% of variance within six kinematic skill categories (e.g., 9.1% grasp, 17.4% lift, 22.0% transport), showing representations retain large task variation even within the same skill.

  • Cross-task discriminability. On two held-out LIBERO90 tasks with identical layouts—place the front bowl on a plate vs. stack the front bowl on another bowl—the shared grasping stage attends to the gripper and bowl, while placement shows distinct spatial attention (Fig. 7). Token sequences reflect this: at gripper closure, placing retrieves [431, 335, 300, 274] and stacking [639, 300, 274, 309] sharingID∗∗300∗∗and∗∗274∗∗sharing ID **300** and **274**, the placing task retrieves [300, 1049, 295, 293] while the stacking task retrieves [1310, 770, 298, 291]. Shared entries with task-dependent attention and token combinations provide qualitative evidence of cross-task discriminability.

AuthorsOpenReview ↗

Response to Reviewer E35x: Clarification on improvement on simulation benchmarks. (2/5)

  1. Improvement on simulation benchmarks. a) We measured all inference latencies on the same NVIDIA A100 80G GPU. Faster-WAM takes 387.7 ms with 10 step diffusion. On the more challenging LIBERO-Plus benchmark, which evaluates robustness under seven types of perturbations, Faster-WAM achieves 73.6% success and ours achieves 86.2%, an improvement of 12.6%. This indicates AED's stronger robustness in these OOD settings. b) AHA-WAM takes 620 ms without deployment optimization (TensorRT and CUDA Graphs) about 100 ms slower than ours while diffusing 10 steps. After using TensorRT and CUDA Graphs, the latency would be 35.52 ±\pm 4 ms, while AHA-WAM is 41.37 ms. c) We add employ ours method on a π\pi{0.5}, kept the same settings with version from lerobot, π\pi{0.5} with AED achieve 97.2%, while the lerobot version is 96.0%.
AuthorsOpenReview ↗

Response to Reviewer E35x: Clarification on MT loss (3/5)

  1. MT loss. a) The MT loss encourages the action embeddings to highlight the task-relevat regions. In Eq. (7), we randomly sample a starting time tt and a temporal interval Δ\Delta to form two observations, oto_t and ot+Δo_{t+\Delta}, from the same trajectory. A pretrained visual encoder extracts patch-level features from each observation, and their difference, Δh=F(ot+Δ)−F(ot)\Delta h = F(o_{t+\Delta}) - F(o_t), represents visual changes over the interval. Since relatively static scene content is attenuated in this difference, the target emphasizes changes associated with the interaction. The MT loss supervises the AED-conditioned action hidden states to predict Δh\Delta h, thereby linking action representations to their visual consequences as shown in Fig. 6(b) in the submission. Consistent with this motivation, the LIBERO-Plus ablation in Tab. 6 shows that removing MT lowers average success from 69.6% to 63.7%. b) WAMs model visual dynamics through a video-generation backbone and typically use its compatible VAE latent space. Replacing these latents directly with JEPA features may introduce a feature-distribution mismatch with the pretrained backbone, which could account for the observed degradation.
Reviewer_stU1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an Action Experience Dictionary (AED) for incorporating historical action trajectories into World Action Models. Historical actions are aggregated, tokenized with FAST+, mapped to learnable dictionary embeddings, and conditioned on visual history before being prepended to the action sequence. A motion-aware transition loss additionally predicts visual feature changes over sampled temporal intervals. Experiments cover LIBERO, RoboTwin 2.0, LIBERO-Plus, and two real-world robot platforms, with particularly substantial improvements over Fast-WAM on LIBERO-Plus.

Critical strengths

1.The history-conditioning design is sensible and compact. Combining action history with visual context addresses the ambiguity of interpreting similar movements in different scenes. Compressing 32 historical actions into eight prefix tokens also provides a practical way to incorporate this information.

2.The robustness improvements are substantial. On LIBERO-Plus, AED improves over Fast-WAM from 70.8% to 86.2%, with particularly large gains under camera and noise perturbations (Table 2). These results are more compelling than the small differences on nearly saturated standard benchmarks.

3.The evaluation covers a useful range of settings. The paper includes standard manipulation tasks, 50-task RoboTwin evaluation, several perturbation categories, and experiments on two physical robot platforms. This breadth provides encouraging evidence that the approach is useful beyond a single simulated setting.

4.The appendix examines meaningful design alternatives. Comparisons include continuous-action MLP encoding, different aggregation orders, visual encoders, and prefix designs. The MLP comparison in Figure 12 is particularly useful for assessing the dictionary representation. Table 5 also reports only a modest inference overhead relative to Fast-WAM.

Critical weaknesses

1.The central skill-reuse claim is not sufficiently isolated. Figure 6(a) compares AED trained on five versus ten tasks, but does not repeat this comparison for Fast-WAM or the history-MLP baseline. The improvement could therefore reflect ordinary multitask learning rather than an advantage specific to AED’s proposed reuse mechanism.

2.The interpretation of the MT loss exceeds the supporting evidence. Theorem 1 bounds the additive consistency of feature-change predictions; it does not establish background invariance or action relevance. The reported robustness gains are not accompanied by an MT ablation on LIBERO-Plus, where this explanation could be directly tested.

3.The smaller gains lack uncertainty estimates. The LIBERO advantage over the strongest listed baseline is only 0.3 percentage points. Likewise, the MT ablation’s 0.6-point difference represents three successes out of 500 trials. The reported single-seed evaluations do not establish how stable these differences are.

4.The real-world protocol is insufficiently documented. Appendix E omits demonstration counts, evaluation trial counts, and baseline training details. It also does not clarify whether the two robots share learned AED entries or use separately trained models, limiting what can be concluded about cross-embodiment experience reuse.

Questions for authors

1.Cross-task transfer: When repeating the five-versus-ten-task experiment with Fast-WAM and the history-MLP baseline, while controlling optimization budget and target-task exposure, does AED provide a larger transfer benefit?

2.MT mechanism: How does removing MT affect LIBERO-Plus performance, particularly under camera, background, and noise perturbations? Would directly predicting endpoint features provide similar gains?

3.Statistical stability: Can the authors report repeated-seed results and uncertainty estimates for the principal benchmark comparisons and MT ablation, preferably using matched evaluation initial states?

4.Real-world protocol: What are the training-demonstration and evaluation-trial counts for each platform? Were baselines trained on the same data, and were AED entries or other model weights transferred between the two robots?

5 comments
AuthorsOpenReview ↗

Response to Reviewer stU1: Clarification on skill-reuse claim (1/4)

1. Five-versus-ten-task comparison. Under matched experimental conditions, including the same optimization budget and exposure to the target tasks, we compare FastWAM, history-MLP, and our method in the five- versus ten-task settings. As shown in the table, our method achieves a gain of 3.6, compared with 2.2 for FastWAM and 2.0 for history-MLP. AED achieves better improvement (3.6%).

Method5 tasks10 tasksGain
Fast-WAM93.695.8+2.2
history-MLP93.495.4+2.0
Ours94.097.6+3.6
AuthorsOpenReview ↗

Response to Reviewer stU1: Clarification on MT loss (2/4)

2. MT loss and LIBERO-Plus ablation. a) Theorem 1 analyzes the role of random temporal-interval sampling in MT supervision: the sampled objective is an unbiased estimator of the sampling-weighted transition risk, which bounds temporal-composition error. Its theoretical conclusion concerns consistent transition prediction across temporal scales, rather than background invariance or action relevance. We clarify this distinction in the revised discussion.

b) Following Eq. (7), we randomly sample a starting time t and a temporal interval Δ, forming two observations, o_t and o_{t+Δ}, from the same trajectory. A pretrained visual encoder extracts patch-level features from each frame, and their difference, Δh = F(o_{t+Δ}) − F(o_t), represents the visual changes over that interval. This differencing attenuates features from relatively static regions and emphasizes patches changing with the interaction. The MT loss then supervises the AED-conditioned action hidden states to predict (\Delta h), encouraging the model to capture the visual consequences associated with each action.

c) We additionally report a component ablation on the LIBERO-10 task suite under seven LIBERO-Plus perturbation settings, as shown in Tab.~6. Removing MT reduces average success from 69.6% to 63.7%, with decreases from 43.9% to 37.0% under camera perturbations and 65.5% to 60.1% under background perturbations. Removing VC or AED reduces average success to 60.1% and 56.3%, respectively. These results provide empirical evidence for the components’ contributions to robustness, complementing the transition visualizations. c) We add employ ours method on the π0.5\pi_{0.5}, kept the same settings with version from lerobot, π0.5\pi_{0.5} with AED achieve 97.2%, while the lerobot version is 96.0%.

AuthorsOpenReview ↗

Response to Reviewer stU1: Clarification on the smaller gains lack uncertainty estimates (3/4)

3. Ablation study on LIBERO Plus.

VariantCam.Robot.Lang.Light.Back.Noise.Layout.Avg.
Ours (full)43.950.186.695.765.578.966.469.6
-MT37.051.475.986.760.173.161.663.7
-VC37.348.970.281.753.469.959.360.1
-AED36.844.367.681.442.764.556.756.3

On LIBERO-10 task of the LIBERO Plus, the full model achieves 69.6% average success. Removing MT, VC, and AED reduces the average to 63.7%, 60.1%, and 56.3%, respectively (drops of 5.9, 9.5, and 13.3 percentage points). The full model scores higher than each ablated variant in six of seven categories. We have added the ablation study (Tab. 6) in the submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.