acceptodds
Under review as a conference paper at ICLR 2027

Fork-Ready World Models: Training Distilled World Models to Be Branchable

Abstract

Tree-structured imagination relies on continuations sampled from a shared prefix, but a shared noisy cache can correlate those continuations and reduce the benefit of best-of- selection. We characterize this effect in a linear-Gaussian model and measure it in a B distilled autoregressive video model built on Self-Forcing and Wan2.1. The shipped clean-cache recipe has zero cache-noise variance in our protocol, whereas increasing cache noise reduces the measured variance ratio from to . We introduce Fork-Invariance, a training objective that compares noised-cache rollouts with clean-cache references using a shared sampling tape. Its strongest variance-ratio result () is misleading: full-weight training damages decoded video and reduces sensitivity to the prefix. Capping the objective's gradient norm at of the distillation gradient instead yields a variance ratio of , retains prefix sensitivity, and shows no detected quality degradation under two frozen visual judges. At matched measured variance ratio, the retained best-of- search gain increases from to . These results distinguish cache-noise suppression from preservation of useful conditioning and motivate evaluating branch groups through variance, selection performance, decoded quality, and prefix sensitivity together. The gradient-matched result uses one training run, and judge scores alone do not establish perceptual equivalence.

Primary area:
generative models
Keywords:
world models, tree-structured imagination, copy-on-write KV cache, branch consistency, exchangeability of sampled futures, distillation, autoregressive diffusion, RL in imagination, test-time search, batched serving, model-based RL, determinism and reproducibility

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fNjWOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper identifies and formalizes a previously underappreciated failure mode of copy-on-write (CoW) forking in distilled autoregressive-diffusion world models: when the shared KV cache carries per-token Gaussian noise from the distillation process, forking off that noisy cache produces under-dispersed sibling rollouts. The authors prove a linear-Gaussian under-dispersion law (variance ratio VR=Σ(Σ+σ2WW⊤)−1VR = \Sigma(\Sigma + \sigma^2 WW^\top)^{-1}), confirm the law quantitatively on a 1.4B Self-Forcing / Wan2.1 model across 4 individual prompts, and show that this under-dispersion more than halves best-of-k search gain at k=64, sigma=2. They then introduce Fork-Invariance, a training objective that pairs the noised rollout against a clean-cache reference so the sampling noise cancels in the branch-invariant component. Constraining the FI gradient share to 0.2 of the DMD gradient norm yields VR=0.888, pixel-level generation indistinguishable from base, and restores retained best-of-k search gain from 52% to 84% at matched variance ratio.

Critical strengths
  1. Important and well-characterized problem discovery. The paper identifies a subtle but consequential pathology -- shared-cache noise causing under-dispersion that silently undermines best-of-k / MCTS branching -- that prior work on CoW forking (PageAttention, ForkKV) has overlooked. The linear-Gaussian theorem and its quantitative confirmation on a real 1.4B model make this a genuinely useful contribution to the world-model serving literature.
  2. Rigorous experimental methodology. The paper uses a controlled noise injection sweep to isolate the mechanism, cross-validates across 4 prompts and 9 generated frames, and separates signal from nuisance variance via an explicit decomposition. The paired t-test statistics and effect-size reporting (t-values, p-values, 95% CIs) are above the typical bar for ICLR papers.
  3. Practical and actionable repair. The Fork-Invariance objective, combined with the gradient-share cap constraint, is a simple training-time intervention that requires no serving-side changes. The result -- VR restored to 0.888 with pixel-level fidelity indistinguishable from base and best-of-k gain recovered from 52% to 84% -- is directly usable by practitioners deploying forking world models.
  4. Honest failure-mode analysis. The paper explicitly distinguishes two failure modes (mean offset vs. under-dispersion) and shows why naive fine-tuning, serving-side re-keying, and uncapped Fork-Invariance each fail in differen ways. This level of diagnostic clarity is rare and valuable.
Critical weaknesses
  1. Scope is narrow: only one model family and one distillation recipe. All quantitative results are on Self-Forcing / Wan2.1-T2V-1.3B (student: 1.4B) with the shipped few-step sampler and rolling KV cache. The claim that "forks are exactly exchangeable" for clean-cache pipelines is plausible, but the magnitude of the pathology and the effectiveness of Fork-Invariance are not demonstrated on other architectures (e.g., Diffusion Forcing, GameNGen, or image-based world models like DreamerV3).
  2. The gradient-share cap repair feels like a workaround rather than a principled solution. The paper admits that the full-weight Fork-Invariance objective collapses pixels (VR=0.964 but 0-to-10 quality score), and the only configuration that works is the 0.2 FI gradient share constraint. This raises the question: why does the objective as formulated destroy the pixel loss, and is the cap value a universal constant or a per-model tuning knob? Without a deeper mechanistic explanation, the repair risks being seen as a trick.
  3. Long-horizon and downstream-task evaluation is missing. The best-of-k analysis is on held-out frame scoring, not on downstream planning or RL tasks. If forking is used for MCTS-based planning or RL rollouts, does Fork-Invariance improve the downstream policy performance? The paper stops at frame-level fidelity and search gain, which is a narrower contribution than the title "Fork-Ready World Models" implies.
Questions for authors
  1. The under-dispersion law is derived under a linear-Gaussian assumption. How well does the VR fit generalize beyond the 1.4B Self-Forcing / Wan2.1 setting? Have you tested on larger models or different distilled world model architectures (e.g., Diffusion Forcing, GameNGen)?
  2. The Fork-Invariance loss is added on top of the DMD objective, and the gradient-share cap (FI gradient limited to 0.2 of the DMD gradient norm) is the key hyperparameter that makes the repair work. How sensitive are the final VR and pixel-fidelity numbers to the cap value? Is 0.2 chosen on a validation set, and does it transfer across model sizes?
  3. The paper reports that in a real CoW deployment, serving-side path-keying (per-branch fresh noise) moves neither the variance ratio nor the mean offset. What about an idealized path-keying setup that re-derives the committed prefix noise -- would that recover VR without retraining, and what residual offset would remain?
  4. The failure analysis focuses on the shipped distillation recipe. What fraction of the under-dispersion comes from the final denoising steps versus the rolling cache write? Is the effect concentrated in a few timesteps, or is it distributed across the whole trajectory?
1 comment

Thank you for your generous assessment and for your questions on architectural generalization, gradient caps, and long-horizon planning.

Q1 (W1). Generalization to larger models and other world-model architectures. Holding observed histories fixed, we confirmed positive across-write conditional-mean variance (B>0B>0) at a 10×10\times larger Wan scale and across two additional model recipes:

Model & ArchitectureFixed-History Empirical Result
Wan2.1 14B (10×10\times scale)Positive BB on all 8 prefixes at both write timesteps (at t=300t=300, corrected VR=0.9892\text{VR} = 0.9892 in 14B vs. 0.94710.9471 in 1.3B).
Action-Conditioned DFoT-B (Diffusion Forcing)Positive BB on all 8 histories at both noise levels. At noise 0.3040.304, mean BB was 0.009440.00944 under persistent perturbations and 0.003630.00363 under fractional History Guidance.
Causal-rCM (arXiv:2606.25473, noisy context)Positive BB on all 8 prefixes for both future latent frames with noisy prefill. First-frame mean BB was 7.44×10−37.44\times10^{-3} with noisy prefill and zero with clean prefill.

Adapting only DFoT-B's final block and output projection reduced BB by 4.5%4.5\%–4.7%4.7\% (with CIs spanning zero) while retaining 97.0%97.0\%–98.6%98.6\% of the action-swap response.

Q2 (W2). Why the gradient-share cap is needed and hyperparameter sensitivity.

  • Mechanism & pre-training selection. Matching paired means alone can reward insensitivity to the prefix. At the rollout latent, bounding ∥λgF∥≤α∥gD∥\|\lambda g_F\|\le\alpha\|g_D\| ensures ⟨gD,gD+λgF⟩≥(1−α)∥gD∥2\langle g_D,g_D+\lambda g_F\rangle\ge(1-\alpha)\|g_D\|^2 before the generator Jacobian and optimizer. We set α=0.20\alpha=0.20 prior to training at the upper end of a 1010–20%20\% gradient-norm budget; the initial FI-to-DMD norm ratio was ≈8.1\approx 8.1, corresponding to an effective multiplier of ≈0.025\approx 0.025.
  • Post-submission sensitivity sweep. Both caps and the fixed weight substantially reduced BB:
Wan 1.3B Arm (Matched Seed & Budget)Corrected Fixed-Prefix VRReduction in BB vs. Zero FI
Zero FI (Baseline)0.96550.9655Reference
Online Cap α=0.10\alpha = 0.100.99410.994166.9%66.9\%
Online Cap α=0.20\alpha = 0.200.99610.996174.3%74.3\%
Fixed Weight 0.0250.0250.99360.993667.3%67.3\%

Across 3 seeds at cap 0.100.10, BB fell by 69.0%69.0\% and both VLM judges cleared the −0.5-0.5 appearance margin on 192192 pairs (with 0.100.10 scoring +0.266+0.266 points higher than 0.200.20 on Qwen3-VL appearance on the matched seed, 95% CI [0.156,0.391][0.156,0.391]). While fixed weight 0.0250.025 also cleared both judges' appearance margins, online cap 0.200.20 achieved 8.39×10−58.39\times10^{-5} lower BB (95% CI [6.55,10.98]×10−5[6.55,10.98]\times10^{-5}) because the fixed arm exceeded the 0.200.20 norm share on 16/2016/20 updates. Across tasks, the 0.200.20 cap bound 240/240240/240 DFoT-B adaptation updates and 0/5,4000/5{,}400 PointMaze updates (our 14B evaluation tested the released checkpoint without fine-tuning).

Q3. Idealized prefix replay vs. FI-gm and serving latency. Independently rewriting the prefix restores population VR=1\text{VR} = 1 and reduces the cache variance contribution to the group mean from BB to B/kB/k (at k=6k=6, residual random cache variance falls to 0.195×10−30.195\times 10^{-3} for Base and 0.047×10−30.047\times 10^{-3} for FI-gm). Because replay still samples from the noisy-write marginal, a squared mean shift per coordinate relative to clean writes remains (6.557×10−36.557\times10^{-3} for zero FI and 10.115×10−310.115\times10^{-3} for FI-gm). Practically, grouped replay recovered 81.3%81.3\%–82.7%82.7\% of full-replay coverage gain in PointMaze while adding only 1.60%1.60\%–1.81%1.81\% end-to-end latency on an A100 in Wan (including video decoding).

Q4. Where does the variation enter along the trajectory? An ANOVA across noise sources shows that the latest prefix write's main effect accounts for 75.5%75.5\% of mean BB in Base (57.9%57.9\% in FI-gm), whereas the final re-noising draw and its interactions account for 17.6%17.6\% and 23.5%23.5\% of within-cache variance AA. Cache-mean variation emerges on the first denoising pass and stabilizes by pass 3 of 4 (reaching 1.028×1.028\times and 0.968×0.968\times its final value in Base and FI-gm).

W3. Downstream planning and multi-block rollouts.

  • PointMaze planning: Across 9696 starts, FI-gm reduced 48-step plan regret by 5.8%5.8\% at training noise and 22.1%22.1\% at double noise (paired cost difference −0.0546-0.0546, 95% CI [−0.0787,−0.0305][-0.0787, -0.0305]). In closed-loop replanning (192192 starts, 4 simulator checks per replan), independent replay reduced capped goal-arrival time by 6.146.14 steps at double noise (95% CI [−11.46,−0.68][-11.46, -0.68]).
  • Multi-block Wan rollouts: In a 3-block test (8 fixed prefixes, one matched seed, cap 0.100.10, initial noisy prefix followed by clean writes), FI-gm reduced third-block BB by 0.619×10−30.619\times 10^{-3} (95% CI for FI-minus-zero [−1.133,−0.246]×10−3[-1.133,-0.246]\times10^{-3}).
Reviewer_SL38OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies branchability in autoregressive diffusion based world models. It studies how shared noise in cache corrupts the future branching trajectories. It studies two branching protocols: Shared-cache branching where cache noise realization is shared among all sibling trajectories and independent cache branching where each trajectory receives an independently generated cache noise.

Critical strengths
  1. Identifies a meaningful statistical issue in world-model branching.
  2. Provides an intuitive mathematical model of the failure mechanism.
  3. Goes beyond synthetic experiments by evaluating a large video generation model.
  4. Evaluates downstream search performance rather than relying exclusively on a variance statistic.
  5. Investigates negative results and training failure modes rather than reporting only favorable outcomes
  6. Gradient-level intervention is a technically interesting engineering contribution
Critical weaknesses
  1. Conflating exchangeability with independence: The authors define exchangeability as requiring the joint distribution of siblings to remain invariant under permutation, with each sibling's marginal distribution matching the model's conditional distribution. However, under the authors' own linear-Gaussian model, the sibling samples appear to satisfy these conditions, even when they share cache noise.
  2. The paper does not sufficiently justify why independently regenerated cache noise defines the correct branching distribution.
  3. The empirical variance-ratio estimator is not generally equivalent to the theoretical variance decomposition.
  4. The proposed gradient-matched FI method has limited replication, and the paper does not establish improvements on a real planning task.
Questions for authors

Answer the weakness for the questions section.

1 comment

Thank you for your constructive review and for highlighting the distinction between statistical sampling properties and downstream planning utility.

W1. Exchangeability vs. Conditional Independence. We agree with your observation and have refined our terminology. With PeP_e denoting the continuation law at a realized cache ee, a shared-cache group has joint law EE[PE⊗k]\mathbb{E}_E[P_E^{\otimes k}], which remains permutation-invariant (exchangeable) after marginalizing over EE. However, a shared draw EE couples siblings given the observed prefix ss, whereas independent replay samples i.i.d. branches from the product marginal (EEPE)⊗k(\mathbb{E}_E P_E)^{\otimes k}.

W2. Why independent cache replay is desirable for search. Sharing yields conditional independence given the realized noisy cache, whereas replay yields conditional independence given the observed prefix. While both share the same single-branch marginal (EEPE\mathbb{E}_E P_E), Jensen's inequality shows that for any success event with cache-dependent probability pEp_E, the two-branch coverage gain from replay is exactly Var⁡(pE)\operatorname{Var}(p_E). We confirmed this directly on 9696 MuJoCo PointMaze starts (evaluating 48-step plans using the exact same individual branches):

Two-Branch Search Protocol (PointMaze, 9696 starts)Coverage Gain in Percentage Points (Paired 95% CI)
Cache Replay at training noise (plan in lowest-cost quartile)+1.91+1.91 pp [1.51,2.35][1.51, 2.35]
Cache Replay at training noise (plan reaches next waypoint)+2.37+2.37 pp [1.91,2.86][1.91, 2.86]
FI-gm vs. zero FI, shared cache at 2×2\times noise (reaches next waypoint)+3.13+3.13 pp [1.26,5.02][1.26, 5.02]

W3. Interaction-aware variance-ratio estimation. You are right that our original denominator double-counted the cache–sampling interaction. We revised the estimator to VR^=A^/(A^+B^)\widehat{\text{VR}} = \widehat{A} / (\widehat{A} + \widehat{B}), where A^\widehat{A} is the mean within-cache sample variance and B^=(nsC^−D^)/(ns−1)\widehat{B} = (n_s\widehat{C} - \widehat{D})/(n_s - 1) isolates the conditional-mean variance (C^\widehat C: variance of cache-column means; D^\widehat D: mean across-cache variance at fixed tapes; nsn_s: number of tapes). The component estimates are unbiased under the crossed design, and their ratio is a plug-in estimate. Across all 400400 possible 3×33\times 3 subgrids per saved grid, the FI-gm reduction remained directionally consistent in 99.2%99.2\% of paired subgrids using B^\widehat{B}, versus 47.6%47.6\% using the unadjusted D^\widehat{D}.

W4. Multi-seed replication and physical plan selection.

  • Wan replication: Across three equal-budget training pairs, FI-gm consistently reduced mean BB by 74.3%74.3\%, 75.6%75.6\%, and 78.5%78.5\%.
  • PointMaze planning: Across three matched model pairs selecting among 4848 candidate plans executed for 4848 MuJoCo steps, FI-gm reduced simulator regret across two cohorts of 9696 starts by 7.3%7.3\% and 5.8%5.8\% at training noise, and by 12.4%12.4\% and 22.1%22.1\% at double noise (with paired 95% CIs excluding zero in both cohorts at double noise).
Reviewer_tE8MOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper is about how sharing noisy KV caches among branches in autoregressive video world models affects continuation diversity and best k selection performance. The authors introduce a linear-Gaussian analysis and conduct experiments on a 1.4B parameter video model to characterize the effects of cache noise. They propose Fork-Invariance, a training objective designed to reduce shared-noise effects while preserving generation quality and sensitivity to the conditioning prefix.

Critical strengths
  • Provides empirical evaluation on a large-scale video model across different cache-noise levels, rollout horizons, and configurations.
  • Identifies degenerate solutions of the proposed training objective and highlights the importance of evaluating generation quality and conditioning sensitivity.
Critical weaknesses
  • The central theoretical claim seems to be problematic. The authors argue that sharing a noisy cache violates exchangeability, although their linear-Gaussian formulation produces exchangeable samples by construction. Conditional on the shared noise realization, the continuations are independent and identically distributed, and their joint distribution remains invariant under permutations after marginalization. That is a mischaracterization of the statistical property underlying the main contribution. If the realized cache is part of the model state, sharing it across continuations is consistent with conditional sampling. The authors provide insufficient justification for treating cache noise as a nuisance that must be independently resampled for each branch and thus the existence of the claimed sampling pathology remains unestablished.

  • The claimed equivalence (VR=1\iff\sigma=0) is false when the model is insensitive to the cache, and the matrix formulation of the variance ratio is not generally valid under the stated ordering. The proposed variance decomposition does not account for interactions between cache noise and sampling randomness in nonlinear generators. The Fork Invariance objective only constrains differences between the means of clean cache and noisy cache continuations, providing no guarantee of distributional equivalence or valid joint sampling. The claimed monotonic increase in selection regret with branch count and cache noise is also insufficiently justified since the variance relationship alone does not establish such behavior for arbitrary selection objectives.

  • The main gradient matched result relies on a single training run, comparisons involve unequal supervision budgets, and the reported improvement in selection performance fails to replicate on a second prompt. The quality evaluation uses a limited number of generated videos assessed by frozen vision-language models, without sufficient evidence of perceptual equivalence. The best k evaluation minimizes distance to an estimated conditional mean, which may be unrepresentative of meaningful outcomes in multimodal generative models. These limitations undermine the soundness of the paper’s central conclusions.

Questions for authors

Q1. (a) How exactly is P_\theta(x\mid s) defined, and does the state s include the realized KV cache? (b) If the cache is part of the conditioned state, why should independently resampling its noise for each branch be considered statistically correct?

Q2. Theorem 1 claims that VR=1 if and only if \sigma=0. (a) How does this claim account for the case W=0, where VR=1 for any \sigma? (b) Does the cache-insensitive solution observed in the experiments constitute a counterexample to the theorem?

Q3. (a) What theoretical assumptions establish that selection regret increases monotonically with cache noise and branch count? (b) Does the claimed relationship hold beyond the particular Gaussian model and scoring function considered?

1 comment

Thank you for your rigorous review. Your comments prompted us to clarify exchangeability, extend our selection theory beyond the linear-Gaussian model, and test coverage on plans executed in MuJoCo and on decoded video events.

Q1. Conditioning state and why independent cache replay matters for search. Let ss be the observed prefix and Pe=L(X∣s,E=e)P_e=\mathcal L(X\mid s,E=e) the continuation law at cache write ee. The observed-prefix marginal is Pθ(⋅∣s)=EEPEP_\theta(\cdot\mid s)=\mathbb E_E P_E. Shared and replayed groups have exchangeable laws EE[PE⊗k]\mathbb E_E[P_E^{\otimes k}] and (EEPE)⊗k(\mathbb E_E P_E)^{\otimes k}. If the conditioned state includes the realized cache, sharing correctly samples Pe⊗kP_e^{\otimes k}. However, when search requires independent futures conditioned on the observed prefix ss, replay draws independently from the marginal Pθ(⋅∣s)P_\theta(\cdot\mid s) (keeping the saved prefix tensor fixed while rewriting its cache).

Q2. Theorem 1, nonlinear interactions, and FI scope.

  • Equality condition. The additive trace ratio is tr⁡(Σ)/(tr⁡(Σ)+σ2∥W∥F2)\operatorname{tr}(\Sigma)/(\operatorname{tr}(\Sigma)+\sigma^2\|W\|_F^2). With positive denominator, VR=1\mathrm{VR}=1 iff σW=0\sigma W=0. At σ>0\sigma>0, an exactly cache-insensitive model (W=0W=0)—the degenerate regime diagnosed in §5—is a counterexample to the original unqualified statement. For positive definite TT, the matrix form is T−1/2AT−1/2⪯IT^{-1/2}AT^{-1/2}\preceq I.
  • Nonlinear interactions. Total covariance decomposes as T=A+BT=A+B, with AA the mean within-cache covariance and BB the covariance of the conditional mean. We estimate B^=(nsC^−D^)/(ns−1)\widehat B=(n_s\widehat C-\widehat D)/(n_s-1), where C^\widehat C is the variance of cache-column means and D^\widehat D is the mean across-cache variance at fixed tapes. Across three equal-budget Wan pairs, mean BB fell by 76.1%76.1\% while AA increased.
  • FI scope. FI targets paired conditional means. At a fixed write, i.i.d. noisy-minus-clean differences give E∥Δˉ∥2=∥EΔi∥2+tr⁡Cov⁡(Δi)/k\mathbb E\|\bar\Delta\|^2=\|\mathbb E\Delta_i\|^2+\operatorname{tr}\operatorname{Cov}(\Delta_i)/k. Equal means can coexist with different higher moments: for instance, X=(1+E)ΞX=(1+E)\Xi, with EE uniform on {0,1}\{0,1\} and independent standard normal Ξ\Xi, has B=0B=0 but conditional variances 11 and 44.

Q3. Distribution-free selection guarantees and planning.

  • Beyond Gaussian assumptions. For a fixed event with cache-conditional success probability pEp_E, Jensen's inequality yields Hshared(k)=1−E(1−pE)k≤1−(1−EpE)k=Hreplay(k).H_{\rm shared}(k)=1-\mathbb E(1-p_E)^k\le 1-(1-\mathbb E p_E)^k=H_{\rm replay}(k). The gap at k=2k=2 is Var⁡(pE)\operatorname{Var}(p_E). Applying this inequality at every reward threshold establishes weak first-order stochastic dominance for best-of-kk. The gap can peak at intermediate noise or branch counts: for X=σE+ΞX=\sigma E+\Xi with independent standard normals and success ∣X∣≤1\lvert X\rvert\le1, it vanishes at both noise extremes, and at fixed σ>0\sigma>0 it is zero at k=1k=1, positive at k=2k=2, and vanishes as k→∞k\to\infty.
  • Task-level evidence. We measured next-waypoint coverage and plan regret in MuJoCo, as well as a visible event in decoded Wan videos:
Task & Evaluation MetricEmpirical Result
PointMaze waypoint coverage (replay at training noise)74.07%→76.44%74.07\% \to 76.44\% using the exact same individual branches (95% gain CI [1.91,2.86][1.91, 2.86] pp).
PointMaze single-plan selection (FI-gm at 2×2\times noise)22.1%22.1\% lower simulator regret for model-selected plans (95% cost-diff CI [−0.0787,−0.0305][-0.0787, -0.0305]).
Decoded Wan visible event (puck crossing a line, FI-gm)54.17%→66.15%54.17\% \to 66.15\% 2-branch coverage (8 synthetic prefixes, one matched seed, 95% gain CI [1.04,26.04][1.04, 26.04] pp).

Equal-budget replication, second prompt, and expanded quality checks.

  • Equal budget & second prompt. Three matched Wan pairs each used the same initialization, data, and schedule (100100 critic and 2020 generator updates per arm). In a matched-write-timestep comparison, retained search gain under the latent-mean score rose on both lakeside (0.342→0.6250.342\to0.625) and night intersection (0.324→0.7980.324\to0.798).
  • Video quality & prefix sensitivity. Across 192192 paired clips (three cap-0.100.10 seeds), both VLM appearance 95% CIs ([−0.36,0.13][-0.36,0.13] and [−0.47,−0.02][-0.47,-0.02]) cleared our preset −0.5-0.5 margin, and the DINOv2 mean adjacent-frame ratio was unchanged (−0.017-0.017, CI [−0.045,0.010][-0.045,0.010]). On the matched fixed-weight seed, prefix swaps retained 92.7%92.7\% of the squared latent response (95% CI [82.8%,100.3%][82.8\%,100.3\%]).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.