PaperlayerView exact revision on arXiv

Source-linked AI summary

ALPS: Visual State Encoders for World Models

Yanqing Liu, Letian Zhang, Chengzhi Liu, Zijun Wang, Jieru Mei, Hongliang Fei, Yuyin Zhou, Cihang Xie

arXiv:2610.30120v1cs.CV

TL;DR

World models use visual representations to predict scene changes and plan, but existing encoders typically specialize in only one or two of state information, language alignment, and pixel reconstruction. ALPS combines these capabilities in patch tokens through two-stage pretraining on still images and text. It achieves the lowest aggregate LIBERO state-probe error among ViT-B encoders, matches the 3.5× larger V-JEPA 2-L, and supports downstream control, planning, prediction, and visual inspection.

  • Problem

    Existing encoders typically specialize in only one or two of physical-state information, native language alignment, and pixel reconstruction, although world models use visual representations for prediction and planning.

    Source: §1 · p. 2

  • Method

    ALPS trains a ViT-B/16 from scratch on still images and text with language, reconstruction, and spatial objectives, then refines it through cross-view self-distillation.

    Source: §1 · p. 2

  • Results

    ALPS has the lowest aggregate LIBERO state-probe error among ViT-B encoders, matching the 3.5× larger V-JEPA 2-L while retaining strong language alignment and accurate pixel reconstruction.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Frozen ALPS tokens support robot policies, world-model prediction and planning, and pixel decoding of predicted scenes for inspection.

    Source: §1 · p. 2

    The evaluation focuses on a ViT-B/16 backbone and simulation benchmarks; larger scales and real-robot deployment remain future work.

    Source: §6 · p. 11

Abstract

from arXiv · show

World models turn visual representations into predictions and actions. We argue that their visual foundation should combine physical-state information, native language alignment, and pixel reconstruction, so that a single representation can support prediction, instruction, and inspection. However, existing encoders typically specialize in only one or two of these capabilities; we therefore introduce ALPS, a pretrained visual state encoder that unifies all three in a single set of patch tokens. Using only still images and text, we pretrain a ViT-B/16 from scratch with language, reconstruction, and spatial correspondence objectives, and then refine it via cross-view self-distillation. In a unified evaluation across fifteen public encoders, ALPS achieves the lowest aggregate LIBERO state-probe error among ViT-B models (matching the 3.5x larger V-JEPA 2-L) while maintaining strong language alignment and accurate pixel reconstruction. Downstream, frozen ALPS tokens reach the highest mean policy success on LIBERO (85.5%) when paired with our native text tower, match or exceed leading encoders on world-model planning across PushT, Wall, and Granular, and allow predicted future tokens to be decoded directly into pixels for visual inspection. We hope this work helps establish visual pretraining as a shared foundation for understanding, prediction, and action in world models.

1. Introduction

World models depend on visual tokens that capture object state, align with language, and can be rendered as images, but existing encoders usually combine only one or two of these capabilities. ALPS trains and refines one representation for all three, with competitive state-probe performance and downstream support for control, planning, and inspection.

1. Introduction

  • World-model encoders must represent object locations and states; language-aligned and image-decodable tokens also connect scenes to instructions and allow observed or imagined scenes to be inspected.Many world models operate in the latent space of a frozen pretrained visual encoder.Source: §1 · p. 2
  • ALPS uses a two-stage ViT-B/16 trained from scratch on still images and text, combining language, reconstruction, and spatial objectives before cross-view self-distillation strengthens state information.The refinement predicts a frozen first-stage teacher’s patch tokens at corresponding locations across augmented and clean views.Source: §1 · p. 2
  • ALPS has the lowest aggregate LIBERO state-probe error among ViT-B encoders, matching the 3.5× larger V-JEPA 2-L while retaining strong language alignment and accurate pixel reconstruction.The paper also reports that cross-view refinement reduces state-probe error by 17% while changing image-text retrieval by at most 0.5 points.Source: §1 · p. 2
  • Frozen ALPS tokens support robot policies and world models, including LIBERO policy success with its native text tower, planning across PushT, Wall, and Granular, and pixel decoding of predicted futures.The paper presents a unified evaluation spanning state information, language alignment, reconstruction, control, planning, and prediction.Source: §1 · p. 2

2. Related Work

Prior work develops spatial, language-aligned, reconstruction-based, and predictive visual representations, but these capabilities are often addressed separately. ALPS jointly initializes language, pixel, and correspondence features, then refines them with geometry-conditioned cross-view self-distillation.

Image-text and unified visual pretraining

  • ALPS first jointly trains unnormalized patch tokens for language alignment, pixel reconstruction, and dense correspondence, then uses cross-view refinement to sharpen state information while preserving language and pixel interfaces.This differs from adapting a frozen backbone or encoding autoencoder latents.Source: §2 · pp. 2–3

Visual representations for robot learning

  • Robot-representation work spans egocentric video, masked autoencoding, language supervision, foundation-model distillation, and neural rendering, while comparative studies report no single dominant representation.ALPS uses general-domain images and text to provide frozen state tokens and a native text tower for language-conditioned manipulation.Source: §2 · p. 3

Representations for world models

  • World models may learn representations with dynamics, use pretrained features, or generate observations; ALPS instead focuses on pretraining frozen visual tokens as the state representation for downstream world models.Prior studies cited in the passage report that off-the-shelf features do not automatically improve sample efficiency and reconstruction fidelity alone is an unreliable proxy for action recovery.Source: §2 · p. 3

3. Method

ALPS exposes frozen patch tokens to probes, policies, world models, and pixel decoders, with a language interface derived from projected and pooled tokens. Its first stage jointly trains language, reconstruction, and spatial objectives; its second stage refines the encoder by predicting corresponding teacher tokens across views.

3. Method

  • Frozen patch tokens serve downstream state probes, policies, world models, and pixel decoders, while a projected, pooled, and normalized image embedding pairs with a text encoder for language alignment.The encoder maps an image to patch tokens; the Stage 1 image encoder, text encoder, and image projection are retained after pretraining.Source: §3 · p. 3, §3.1 · p. 4

3.1. Stage 1: joint pretraining from scratch

  • Stage 1 jointly supervises shared patch tokens with contrastive alignment, captioning, reconstruction, and dense correspondence, connecting tokens to language, preserving visual detail, and supervising local relationships across views.The image encoder is a ViT-B/16; the pretraining reconstruction decoder and dense-correspondence projector are discarded afterward.Source: §3.1 · pp. 3–4

3.2. Stage 2: cross-view self-distillation on single images

  • Stage 2 initializes a student and frozen teacher from Stage 1, then predicts standardized teacher tokens for corresponding patches across augmented and clean crops using geometry queries.Only teacher patches whose centers lie inside the student crop contribute; patches outside the overlap are excluded, and the student crop is resampled when fewer than 16 patches overlap.Source: §3.2 · p. 4
  • After cross-view training on single-image samples, the predictor is discarded and the refined student becomes ALPS, paired with the unchanged Stage 1 text encoder and image projection.The figure’s state-probe illustration shows readout from the last 16 frames or prediction 1 or 3 seconds ahead given actions.Source: §3.2 · p. 5, Figure 3 · p. 5

4. Evaluation protocol

The evaluation freezes encoders and compares them under shared protocols across state, language, reconstruction, control, planning, and prediction. It uses native text interfaces where available and lightweight task-specific probes or adapters for downstream measurements.

4. Evaluation protocol

  • The comparison includes fifteen public encoders, using last-layer patch tokens under common frozen-encoder protocols; the primary comparison uses ViT-B image towers at 224 × 224 resolution.V-JEPA 2 ViT-L is a video encoder that takes the last 16 frames and outputs tokens for the current frame.Source: §4 · p. 5
  • The LIBERO state probe reads the last 16 frames to estimate current object state or predict state 1 or 3 seconds ahead given robot actions; language evaluation uses native text towers for zero-shot accuracy and image-text retrieval.The probe estimates movable-object positions, orientations, and binary predicates against simulator ground truth.Source: §4 · pp. 5–6
  • Table 1 reports ALPS’s lowest state-probe error among ViT-B encoders, alongside language and pixel comparisons; its state metric is LIBERO position error relative to ALPS (Stage 1).The table marks ViT-L references separately and uses n/a when an encoder has no released text tower.Source: Table 1 · p. 6
  • Pixel reconstruction is measured by training a lightweight convolutional adapter from frozen tokens to frozen SDXL autoencoder latents, then evaluating held-out images with ℓ1 error and LPIPS.The adapter uses 19,000 ImageNet validation images for training and 1,000 held-out images for evaluation.Source: §4 · p. 6
  • LIBERO control policies use frozen visual tokens and behavior cloning with 50 demonstrations per task; world-model tasks train identical dynamics predictors over frozen patch tokens for planning and multi-step prediction.Control instructions use either frozen BERT or each encoder’s native text tower; world-model evaluation follows task-specific DINO-WM or LeWorldModel protocols.Source: §4 · p. 6

5. Results

Across state probing, language alignment, pixel reconstruction, and downstream tasks, ALPS combines strong performance in one frozen representation. Cross-view refinement strengthens state information while retaining ALPS’s other interfaces, with trade-offs among first-stage objectives.

5.1. One encoder for state, language, and pixels

  • ALPS has the lowest aggregate relative position error among compared ViT-B encoders (.829), while matching the 3.5× larger V-JEPA 2-L and retaining strong language alignment and pixel reconstruction.ALPS reaches 93.2/78.9 Recall@1 on Flickr30k and 70.2/51.9 on COCO; its evaluation-decoder ℓ1 error is .120.Source: §5.1 · pp. 6–7

5.2. Effect of cross-view refinement on state and language

  • Stage 2 reduces ALPS’s aggregate state-probe error by 17% while retrieval shifts by at most 0.5 points; SigLIP 2 and CLIP lose substantially more language performance.The same refinement reduces ALPS state-probe error from 1.000 to .829, while SigLIP 2 and CLIP show larger language-metric drops.Source: §5.2 · pp. 7–8

5.3. Control, planning, and prediction on frozen tokens

  • Stage 2 reduces five-step prediction error by 28% on PushT, 17% on Reacher, 13% on Granular, and 11% on Rope.On PushT, an evaluation decoder trained only on real-frame tokens renders five-step open-loop predictions; ALPS has lower LPIPS (.020) than DINOv3 (.024), Stage 1 (.025), and SigLIP 2 (.031).Source: Table 4 · p. 9, §5.3 · p. 9, Figure 6 · p. 10
  • 85.5% mean success is reached by frozen ALPS tokens on LIBERO with the native text tower, while ALPS leads or ties planning on PushT, Wall, and Granular.With BERT instructions, ALPS reaches 83.4% mean success; Stage 2 also reduces five-step prediction error by 11–28% across four tasks.Source: §5.3 · p. 9, Table 4 · p. 9

5.4. Trade-offs between the first-stage objectives

  • Removing reconstruction raises ℓ1 reconstruction error by 72% and 78% in the two variants, while removing reconstruction or captioning lowers state-probe error by 8.5% and 4.3%.The full recipe retains all objectives, trading a small amount of state-probe accuracy for accurate pixel decoding; Stage 2 reduces state-probe error by 14% to 19% across variants.Source: §5.4 · p. 10, Table 5 · p. 10
  • Removing dense correspondence changes aggregate state-probe error by less than 1%, but lowers cross-view patch-matching accuracy by 6.0 points after Stage 1 and leaves a 2.6-point gap after Stage 2.The full recipe reaches 40.8% matching accuracy after Stage 2 versus 38.2% without dense correspondence.Source: §5.4 · pp. 10–11

6. Conclusion

ALPS combines state information, native language alignment, and pixel reconstruction in one frozen visual representation, supporting robot control, world-model prediction and planning, and inspection of predicted scenes. The work is scoped to ViT-B/16 and simulation benchmarks, with larger scales, real-robot deployment, video supervision, and integrated language-guided planning left for future work.

6. Conclusion

  • One frozen ALPS representation supports robot control, world-model prediction and planning, and visual inspection of predicted scenes.The paper presents visual state pretraining as a shared foundation for robot policies and world models.Source: §6 · p. 11
  • ALPS combines competitive state information, native language alignment, and pixel reconstruction after pretraining on still images and text.Cross-view refinement strengthens state representations while preserving the language and pixel interfaces.Source: §6 · p. 11
  • The evaluation focuses on a ViT-B/16 backbone and simulation benchmarks, leaving larger scales and real-robot deployment to future work.The two-stage curriculum balances reconstruction, language, and spatial state; integrating refinement into one stage remains open.Source: §6 · p. 11
  • ALPS is pretrained without video supervision and is less sensitive to fine-grained temporal actions, motivating refinement on video and interaction data.The downstream evaluations also test language-conditioned control and world-model planning separately; a unified language-guided, pixel-verifiable planner is a next step.Source: §6 · p. 11

S1. Training details

Stage 1 trains the image and text encoders with contrastive alignment, captioning, reconstruction, and dense correspondence, while Stage 2 refines spatial representations through cross-view prediction. The appendix specifies the reconstruction and correspondence losses, crop-coordinate query, training schedule, and Stage 2 image-sampling procedure.

S1. Training details

  • Stage 1 trains at 128 × 128 for 12.8 billion samples, then at 224 × 224 for 1.0 billion samples, using different batch sizes, learning rates, and reconstruction weights.The respective batch sizes are 32,768 and 16,384; peak learning rates are 1.0 × 10−3 and 2.6 × 10−5, and wrec is 2 then 1.Source: S1. Training details · p. 16
  • Reconstruction adds Gaussian noise scaled by 0.05 times the per-token RMS of z, then applies a loss combining ℓ1, structural similarity, and image-gradient terms.The three terms have weights 1, 0.2, and 0.1, respectively.Source: S1. Training details · p. 16
  • Dense correspondence uses a 256-dimensional projection and temperature-0.1 InfoNCE loss, balancing same-image location negatives at weight 0.75 against other-image negatives at weight 0.25.At 224 pixels, 64 negatives of each type are sampled, excluding the 3 × 3 neighborhood around the positive location.Source: S1. Training details · p. 16
  • The Stage 2 teacher-patch query combines a learned embedding, fixed sine-cosine positional encoding, and a learned projection of crop-relative coordinates.Patch coordinates are measured in the central-square image frame; p_i is expressed in patch units of student crop A.Source: S1. Training details · p. 16
  • Stage 2 trains for 4,883 steps on 5M sampled images, with a 10−4 predictor learning rate and 10−5 encoder learning rate.Training uses AdamW, gradient clipping, warm-up, cosine decay, and student-crop photometric augmentation.Source: S1. Training details · p. 16
  • Stage 2 samples from curated public image-editing datasets, using edit metadata only for filtering and sampling; crop pairs come from one image and neither edit pairings nor instructions enter the objective.The collection excludes real-video frames, and physical-state records are sampled with probability 0.05.Source: S1. Training details · p. 16

S2. Evaluation details

The evaluation examines visual encoders across LIBERO state probing and policy learning, world-model planning and prediction, token-to-image decoding, predicted-token rollouts, pipeline calibration, stronger-decoder reconstruction, and transfer from static-image pretraining to video. It reports ALPS’s performance across these evaluations, including distinctions among spatial readout, instruction interface, and latent predictability.

State probe

  • On 256×256 re-rendered frames, ALPS has the lowest ViT-B mean relative error (.845), followed by DINOv3 (.862), while V-JEPA 2 ViT-L reaches .856.Higher native resolution reduces position error across all models.Source: S2. Evaluation details · pp. 19–20, Table S6 · p. 20
  • Intermediate-layer probing selects block 6 for SigLIP 2 and PE-Core, yet ALPS’s last block remains lower at .829 versus the best probed test error of .985.Relative error at the selected layers is 1.147 for SigLIP 2 and 1.009 for PE-Core; ALPS reduces error by 28% and 18%, respectively, compared with their best validation layers.Source: S2. Evaluation details · p. 20, Table S7 · p. 20

Policy learning

  • With BERT instructions, ALPS reaches 85.1% on LIBERO-Goal and 86.5% on LIBERO-Spatial, outperforming all five public baselines on both suites.On LIBERO-Object and LIBERO-Long it reaches 92.2% and 69.7%, respectively, performing on par with several leading encoders.Source: S2. Evaluation details · p. 20, Table S8 · p. 21
  • With ALPS’s native text tower, mean policy success across four suites is 85.2% versus 82.1% with BERT, a paired improvement of 3.1 percentage points.The 95% confidence interval for the improvement is [−0.9, +7.1] points, supporting the tower as an instruction interface without external BERT.Source: S2. Evaluation details · pp. 20–21, Table S8 · p. 21

World-model planning

  • As a Rope pipeline check, DINOv2-S/14 with DINO-WM’s original predictor and schedule scores .567 on the 10 default instances and .547 across all 50.DINO-WM reports .41 on the 10 default instances.Source: S2. Evaluation details · p. 23, Table S11 · p. 24

World-model planning and prediction

  • Across PushT, Granular, Rope, and Reacher, ALPS has lower 5-step prediction error than SigLIP 2 and its first-stage encoder; on Wall, encoders are not meaningfully separated.ALPS also outperforms DINOv3 on PushT and matches it on Granular and Rope, while C-RADIOv3 and DINOv3 have lower error on Reacher. The world models use frozen tokens and a shared predictor and CEM planning protocol across five DINO-WM tasks.Source: S2. Evaluation details · pp. 23, 26

Decoded frames

  • For single-frame PushT reconstruction, the first-stage encoder has the lowest ℓ1 error (.0055), while ALPS scores .0061, comparable to public encoders (.0062–.0066).The standalone SDXL autoencoder scores .0044, and predicting the mean training frame scores .0370.Source: S2. Evaluation details · p. 26, Table S14 · p. 27

Rendered predictions

  • At 5 world-model steps, rendered ALPS predictions achieve LPIPS .020, behind C-RADIOv3 (.018) but ahead of DINOv3 (.024), the first-stage encoder (.025), and SigLIP 2 (.031).DINOv2 is omitted from pixel rendering because its world model uses 196×196 resolution and its tokens have cosine similarity .80 to the decoder-training tokens.Source: S2. Evaluation details · pp. 26–27, Table S15 · p. 27

Predicted-token state readout

  • At k=5, the error gap between ALPS and C-RADIOv3 is 0.7 units on true future tokens, 0.7 on one-step predictions, and 0.8 after rollout.Autoregressive rollout adds about 3.0 units of error for both encoders, compared with 5.7 to 8.9 units for DINOv3, SigLIP 2, and the first-stage encoder; the 0.8-unit gap’s confidence interval includes zero.Source: Table S17 · p. 28, S2. Evaluation details · p. 29

Pipeline calibration

  • The reproduced DINOv2-S/14 pipeline achieves 94% success on PushT and 92% on Wall, closely matching DINO-WM’s reported 90% and 96% and released-checkpoint evaluation’s 92% and 94%.The reproduction uses DINO-WM’s original recipe and training budget and evaluates across 50 goals.Source: S2. Evaluation details · p. 29

Reconstruction

  • The reconstruction evaluation trains a separate decoder on frozen tokens, with direct SDXL autoencoder reconstruction serving as an upper-bound reference (.071 ℓ1, .130 LPIPS) and mean-image prediction as a baseline (.460, .816).The decoder maps tokens into a frozen SDXL autoencoder’s latent space; evaluation uses 1,000 held-out ImageNet validation images.Source: S2. Evaluation details · p. 29

Stronger decoding and training configuration

  • With stronger decoding, ALPS’s ℓ1 error falls from .120 to .107 and LPIPS from .372 to .302, overtaking C-RADIOv3 on LPIPS (.311).All four evaluated encoders improve, but the setting simultaneously increases decoder width and depth, training steps, and training-set size, so it tests a stronger decoding regime rather than isolating capacity. The evaluation decoder is trained on frozen tokens and the standard protocol uses 19,000 ImageNet validation images for training and 1,000 held-out images.Source: Table S18 · p. 29, S2. Evaluation details · p. 29

Temporal transfer

  • Stage 2 raises DAVIS-2017 J &F from .456 to .533 while SSv2 remains largely unchanged (56.4 vs. 55.6), sharpening spatial correspondence without supplying fine-grained motion features.DAVIS uses training-free mask propagation, while SSv2 uses an attentive probe; V-JEPA 2’s SSv2 score of 70.3% is an in-domain reference because SSv2 is in its pretraining data. Rescaling its spatial positions raises its DAVIS J &F from .347 to .526.Source: S2. Evaluation details · pp. 29, 34, Table S19 · p. 34

S3. Stage 2: full comparison and image source

Stage 2 refinement improves ALPS policy success most on LIBERO-Spatial while leaving Goal and Long nearly unchanged; refined SigLIP 2 remains below ALPS on Goal and Spatial but exceeds it on Long. Replacing the curated Stage 2 image collection with DataComp images yields weaker state-probe and DAVIS results, alongside a small ImageNet zero-shot difference.

S3. Stage 2: full comparison and image source

  • ALPS Stage 2 boosts LIBERO-Spatial policy success by 9.7 points, while Goal and Long each change by +0.8 points.For SigLIP 2, refinement shifts policy success by at most 2.1 points across the three suites.Source: S3. Stage 2: full comparison and image… · p. 34
  • Refined SigLIP 2 reaches 79.0% on LIBERO-Goal and 82.4% on Spatial, below ALPS at 85.1% and 86.5%; on Long, SigLIP 2 reaches 74.4% versus ALPS at 69.7%.Source: S3. Stage 2: full comparison and image… · p. 34
  • Using DataComp images for Stage 2 yields state-probe error .886 versus .829, DAVIS J &F .523 versus .533, and ImageNet zero-shot accuracy 73.6 versus 73.4.The passage notes DataComp images have substantially lower native resolution, with 39% below 224 pixels on the short side versus 512–1,024 pixels for most curated images.Source: S3. Stage 2: full comparison and image… · p. 34

S4. Details of the first-stage ablation

The ablation tests auxiliary objectives at one-tenth Stage 1 pretraining scale, followed by the full Stage 2 refinement recipe. Removing captioning or reconstruction lowers state-probe error, while removing dense correspondence chiefly degrades cross-view patch matching.

S4. Details of the first-stage ablation

  • Each ablation variant uses 39k steps at 128 × 128 followed by 6.3k steps at 224 × 224, then 4.9k Stage 2 steps with 5M sampled images.The Stage 1 schedule is 10% of the full schedule; variants without captioning train the text tower solely through contrastive alignment.Source: S4. Details of the first-stage ablation · pp. 34–35
  • Removing captioning or reconstruction lowers state-probe error by 4.3% and 8.5% after Stage 1, and by 10.4% and 13.4% relative to the full recipe after Stage 2.The values correspond, respectively, to removing captioning and removing reconstruction.Source: S4. Details of the first-stage ablation · p. 35
  • After Stage 1, text-to-image recall is highest for contrastive-only training (47.6 on COCO; 74.4 on Flickr30k) and lowest when captioning is removed (45.9; 72.5).Stage 2 reduces recall by 0.4–0.8 points on COCO and 1.1–1.5 points on Flickr30k across variants.Source: S4. Details of the first-stage ablation · p. 35
  • Removing dense correspondence changes state-probe error by less than 1%, but reduces cross-view patch-matching accuracy by 6.0 points after Stage 1 and 2.6 points after Stage 2.Accuracy falls from 37.4% to 31.4% after Stage 1 and from 40.8% to 38.2% after Stage 2, averaged across four view-pair settings on 1,000 held-out ImageNet images.Source: S4. Details of the first-stage ablation · p. 35, Table S21 · p. 35

S5. Representation drift under Stage 2

ALPS representation drift under Stage 2 is lower than for SigLIP 2 and CLIP on the reported comparisons. A lower-learning-rate SigLIP 2 variant retains high embedding similarity but still loses retrieval accuracy.

S5. Representation drift under Stage 2

  • ALPS has mean cosine similarities of .88 on raw tokens, .84 on standardized tokens, and .96 on global language embeddings after refinement; its patch-token linear CKA is .83.SigLIP 2 reports .81, .72, and .84 with CKA .74; CLIP reports .77, .71, and .83 with CKA .59.Source: S5. Representation drift under Stage 2 · p. 35
  • With SigLIP 2’s encoder learning rate reduced from 10−5 to 3 × 10−6, state-probe error improves from 1.160 to .959 and DAVIS score from .503 to .553.ImageNet zero-shot accuracy still drops by 1.9 points; image-to-text Recall@1 drops by 2.3 points on Flickr30k and 2.9 on COCO, while text-to-image Recall@1 drops by 1.0 and 0.6 points.Source: S5. Representation drift under Stage 2 · p. 35
  • The lower-learning-rate SigLIP 2 variant has .97 global image-embedding similarity versus ALPS’s .96, yet suffers a larger retrieval-accuracy drop.The authors suggest that unnormalized patch tokens paired with a linear mean-pooled language readout provide a more stable interface under local cross-view refinement.Source: S5. Representation drift under Stage 2 · p. 35

Related papers

Loading 2610.30120v1…