acceptodds
Under review as a conference paper at ICLR 2027

ConvergeFlow: Language Flow with Provable Convergence to Token Embeddings

Abstract

Recent advances in continuous diffusion and flow-based language models (LMs) have achieved performance competitive with discrete LMs. However, existing continuous frameworks still rely on decoders supervised with cross entropy (CE) because the flow trajectories are not guaranteed to terminate at valid token embeddings. Motivated by this limitation, we introduce **ConvergeFlow**, an embedding-space flow-based LM, which constrains the data predictor to the convex hull of token embeddings and trains it solely with the mean squared error objective induced by flow matching. Under suitable regularity conditions, we prove that the resulting flow converges to valid token embeddings despite errors in the data predictor, enabling direct token prediction without a CE-supervised decoder. We further develop three sampling mechanisms for controlling the trade-off between the generative perplexity and entropy. Experiments on OpenWebText demonstrate that ConvergeFlow achieves performance competitive with existing continuous and discrete diffusion LMs. These findings demonstrate the potential of the flow-based paradigm for language modeling. Our code is available at https://anonymous.4open.science/r/ConvergeFlow-3C12.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Flow Matching, Diffusion Language Models, Language Modeling, Generative Models, Convergence

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_iq5yOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Authors introduce ConvergeFlow. This method is a continuous language model that transforms noise into token embeddings without a separately trained decoder. It uses MSE training and a structured predictor, which the authors claim converges to valid token embeddings under specific assumptions. The paper also introduces sampling methods to control quality and diversity.

Critical strengths

Controlled evidence for MSE fine-tuning. The comparison starts from the same LangFlow checkpoint and includes a crossover experiment. This provides stronger evidence for the proposed objective than benchmark comparisons alone and is the most convincing empirical contribution.

Critical weaknesses
  1. Architectural novelty over LangFlow is not substantiated. LangFlow already adds the Gaussian log-likelihood to its neural logits (Appendix C.1, equation 44). Let aja_j denote the logits before this bias and define Kj=exp⁡(−∣xt−αtej∣22σt2).K_j=\exp\left(-\frac{|x_t-\alpha_t e_j|^2}{2\sigma_t^2}\right). After warmup, the bias coefficient equals one, giving
=\operatorname{softmax}(a+\log K)_j =\frac{e^{a_j}K_j}{\sum_k e^{a_k}K_k}.$$ Defining $f_j=\operatorname{softmax}(a)_j$, this becomes $$p_j^{\mathrm{LF}} =\frac{f_jK_j}{\sum_k f_kK_k} =w_j^{\mathrm{CF}},$$ which is exactly ConvergeFlow’s parameterization (15). Both models then construct the same embedding predictor: $$\mu^{\mathrm{LF}} =\sum_j p_j^{\mathrm{LF}}e_j =\sum_j w_j^{\mathrm{CF}}e_j =\mu^{\mathrm{CF}}.$$ Thus, the predictor parameterization itself does not distinguish ConvergeFlow from LangFlow. The main practical differences are MSE fine-tuning, frozen embeddings, and sampling modifications. This substantially narrows the methodological contribution, particularly given the unresolved convergence argument. 2. **The central convergence argument contains an unjustified reduction.** In Appendix D.1 lines 1003–1004, the proof reduces the original process to an auxiliary process. However, the auxiliary weights defined in (33) depend on the random initial state through the realized trajectory. The subsequent density and Jacobian calculations in (40)–(41) treat them as fixed, omitting this dependence. Matching individual trajectories does not justify transferring the fixed-weight density calculation. Maybe I missed something, but if it is true then the main convergence guarantee is not established by this proof. I hope authors can clarify this moment 3. **The sampling contribution is insufficiently explained in the main text.** The paper devotes substantial space to standard flow-matching background while deferring the proposed sampling mechanisms to Appendix C.2. Since these mechanisms are a stated contribution and materially affect the reported quality–diversity trade-offs, their motivation and implementation should be understandable from the main paper. I recommend shortening the background, particularly the standard flow-matching derivations, and using that space to explain the guidance and iterative refinement methods. 4. **Dependence on CE-pretrained components limits the practical contribution.** The main experiments initialize the model from a pretrained LangFlow checkpoint and freeze its embeddings. The authors acknowledge that jointly learning the embeddings and predictor with MSE admits embedding collapse. Thus, the demonstrated approach remains dependent on CE-pretrained components. The paper establishes a useful fine-tuning recipe, but does not demonstrate a competitive standalone training procedure using only the proposed continuous objective. This substantially narrows its practical contribution as an alternative to CE-based training.
Questions for authors
  1. Parameterization and novelty. Do the authors agree that (15)–(16) are mathematically equivalent to LangFlow’s Gaussian-biased logits and embedding barycenter after warmup? If not, what specific constraint or implementation detail breaks this equivalence?

  2. Validity of the convergence proof. How is the reduction in Appendix D.1, lines 1003–1004, justified when the auxiliary weights in (33) depend on the random initial state? In particular, what argument permits the fixed-weight density and Jacobian calculations in (40)–(41) to apply to the original process?

5 comments

We thank the reviewer for carefully reading our paper and recognizing the strengths of our work. We address the reviewer's concerns below.

Our contributions built on LangFlow

We agree that our main experiments use the same model architecture and initialize from its CE-pretrained checkpoint, with the token embeddings kept fixed, as acknowledged in our paper. As discussed in the paper, optimizing the embeddings with MSE alone admits collapsed solutions. Building on this setup, our contributions are as follows:

∙\bullet Theoretical contribution: To our knowledge, we provide the first convergence guarantee to valid token embeddings for this parameterization. Under the assumptions of Theorem~1, the continuous trajectory converges even when the learned data predictor is imperfect, without required additional projection step.

∙\bullet Practical contribution: With the pretrained embeddings held fixed, we can continue training the data predictor using MSE alone. MSE directly supervises the continuous prediction required by flow matchings, whereas CE supervises token probabilities that are subsequently mapped to an embedding-weighted prediction. This direct supervision may improve conditional-mean estimation under finite model capacity, consistent with our empirical results. Therefore, our work provides theoretical guidance for designing continuous language models.

We have modified Sections 1, 4, and 5 in the revised manuscript to highlight the shared parameterization and dependence on CE-pretrained components, and to clarify the theoretical and practical contributions built upon them. In particular, Section 1 explains the motivation for direct MSE supervision, stating: "CE supervises conditional token probabilities, which are mapped to an estimate of the conditional embedding mean required by FMs. Direct MSE supervision of this continuous prediction may improve its accuracy under finite model capacity."

Clarify on convergence analysis

Thank you for pointing out this issue. We have added a covering argument to address the dependence of the auxiliary weights on the realized sampling trajectory. Specifically, we construct a finite ε\varepsilon-net for weight sequences on a time grid of spacing ε\varepsilon in [0,1][0,1]. Since this grid is independent of NN, its cardinality is also independent of NN, only depends on LL, VV, and ε\varepsilon. Assigning each sampling time tkt_k to its nearest grid point maps these representatives to the sampling time steps. The resulting set has cardinality bounded independently of NN and approximates every realized weight trajectory with error O(ε)O(\varepsilon). For a representative f^\widehat f approximating the realized weight trajectory, the triangle inequality gives \begin{align} \min_{j\in[V]}|x_{t_N}-e_j| \le \min_{j\in[V]}|\widehat x_{t_N}(\widehat f)-e_j| + |x_{t_N}-\widehat x_{t_N}(\widehat f)|. \end{align} For fixed ε\varepsilon, the fixed-weight convergence result and a union bound over the ε\varepsilon-net show that the first term vanishes in probability as N→∞N\to\infty and tN→1t_N\to1. Letting ε→0\varepsilon\to 0 then makes the approximation error vanish, yielding the desired convergence. We have included the detailed argument in the Appendix D.1 of the revised manuscript.

Reorganization of the background and sampling parts

Thank you for the valuable suggestion. With the additional space available in the revised manuscript, we have moved the motivation and implementation details of the guidance and iterative refinement methods to the main text and relocated some standard flow-matching derivations to the appendix.

Reviewer_iq5yOpenReview ↗

Official Comment by Reviewer iq5y

Thank you for clarifying the connection to LangFlow, expanding the sampling section, and revising the proof in Appendix D.1.

But I think, my main concern, raised in Weakness 1, is not fully resolved. The theory and the training change still seem disconnected. Theorem 1 aims to establish convergence of the sampling trajectory for a parameterization shared with LangFlow. In practice, the main change to training is replacing CE with MSE. Since both losses give the same optimal embedding prediction (see FMLM [1, Proposition 3.2]), I see limited novelty in that change alone.

I understand that MSE may be easier to optimize in practice, which you mention in the paper, which I understand intuitively, but cannot understand why it should work or is there any theoretical explanation of that. Figure 3(b) is the strongest evidence for me in practice, although I am unsure how consistent the effect is across training seeds. Given the rebuttal deadline, I am not asking for new experiments.

My question is specifically about the proof. As I read it, Theorem 1 concerns convergence of the sampling trajectory for a fixed predictor. Is there a step in the proof that I probably have not noticed, but uses the fact that the predictor was trained with MSE, or would the same argument apply to a CE-trained predictor with the same parameterization? What is MSE loss gives in theory, this convergence guarantees or something else? If MSE loss is not connected with Theorem 1 then what is the connection between theoretical and practical sights of the paper?

If you can clarify the connection between Theorem 1 and the choice of MSE, I would be willing to raise my score.

[1] FMLM (https://arxiv.org/pdf/2602.16813)

Thank you for the follow-up. We clarify below the connection between Theorem 1 and MSE training, and the practical benefits of MSE.

Connection between Theorem 1 and the choice of MSE

The relationship is not that MSE training leads to the convergence established by Theorem 1, but that this convergence guarantee allows us to train the predictor with MSE alone when the embeddings are fixed.

For example, ELF also uses an MSE objective for continuous denoising. However, MSE alone is insufficient in their setting even with fixed embeddings, as there is no guarantee of convergence. Its training additionally includes a CE objective for decoding. In our setting, convergence to the fixed token embeddings justifies recovering tokens through nearest-neighbor decoding, without training an additional decoding mechanism.

In summary, Theorem 1 shows that the standard flow-matching MSE alone suffices to train the predictor with fixed, pretrained token embeddings.

Why MSE can work better in practice

We agree that CE and MSE can both recover the exact data predictor at their population optima, as discussed in FMLM. However, under limited model capacity, minimizing the two objectives may yield different predictors. Specifically, a network that can represent the exact conditional token posterior can also represent the exact conditional embedding mean. The converse need not hold, since different token distributions can yield the same embedding mean. Accurately predicting this mean therefore does not necessarily require recovering the full token posterior. Hence, as you mentioned, intuitively, directly supervising the continuous prediction required by flow matching may be easier to optimize in practice, which is again based on Theorem 1.

We have clarified this connection between Theorem 1 and the choice of MSE in the revised manuscript.

Thank you again for your efforts to review our paper. As the discussion period ends in one day, we would be grateful to know whether you have further concerns.

Reviewer_c9RrOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes ConvergeFlow, an embedding-space flow-based language model designed to ensure that generated continuous representations converge to valid token embeddings. The method reparameterizes the data predictor used in Plaid as a weighted combination of vocabulary embeddings, where the weights combine learned token scores with an analytically derived Gaussian likelihood based on the corruption process. The authors establish a theoretical convergence guarantee under assumptions on the learned weights and the sampling grid, arguing that the proposed parameterization enables token recovery through nearest-neighbor decoding without a separately trained decoder. The paper also introduces several sampling mechanisms to control the trade-off between generation quality and diversity.

Critical strengths

1, Reparameterizing the output prior in Plaid into a weighting factor is interesting. And the comparison between CE and MSE loss on a fixed token space under this reparameterization is worth the attention of the community.

2, The paper shows clear consistency among the motivation, methodology, and results, all of which focus on the convergence to a specific token in a fixed token embedding space. The soundness is then solid.

Critical weaknesses

1, The assumption of the theorem may not hold. Theorem 1 requires the sampling grid to satisfy (1+αtkσtk)∣αtk+1σtk+1−αtkσtk∣<δ\left(1+\frac{\alpha_{t_k}}{\sigma_{t_k}}\right)\left|\frac{\alpha_{t_{k+1}}}{\sigma_{t_{k+1}}}-\frac{\alpha_{t_k}}{\sigma_{t_k}}\right|<\delta for a sufficiently small constant δ>0\delta>0. However, Section 4 uses a uniform midpoint grid tk=(k+0.5)/Nt_k=(k+0.5)/N. Under the linear interpolation schedule αt=t,σt=1−t\alpha_t=t, \sigma_t=1-t, the left-hand side of this condition at the final interval equals 8N2/98N^2/9, which diverges as N→∞N\to\infty. Therefore, the experimental sampling grid violates a key assumption of Theorem 1, and the claimed convergence guarantee does not directly apply to the sampler used in the experiments. The authors should clarify this discrepancy and provide either a convergence analysis for the actual sampling grid or experiments using a grid that satisfies the theorem's assumptions.

2, The performance of ConvergeFlow is poorly evaluated. There is no main result table in the main text. The primary result table (Table 2) in the appendix is based on Gen PPL, which is a hackable metric under repetition and other ill-posed modes. The pareto frontier between Gen PPL and entropy (by changing the temperatures) is also missing. I will expect supplementing experimental results based on PPL or reasoning tasks like Sudoku.

Questions for authors

1, I cannot intuitively understand why stronger convergence leads to better Gen PPL. To the best of my knowledge, most continuous diffusion language models rely on greedy decoding. Can you provide evidence on how ConvergeFlow changes the decoding behavior compared to greedy decoding?

2, ConvergeFlow is trained based on LangFlow's checkpoint. However, it replaces the time condition from gamma to t. I am curious about how this replacement is implemented in the DiT network. Do you transfer the sampled t into gamma and still use the gamma for time conditioning, or simply change the time conditioning to t?

3, As mentioned in the weakness, Theorem 1 imposes a restrictive step-size condition in the signal-to-noise ratio, whereas the experiments use a uniform midpoint grid. The linear interpolation schedule violates the required bound. Could the authors clarify whether the theorem applies to the actual experimental sampler, and provide an explicit sequence of grids satisfying the assumptions?

4, I wonder how PPL changes under the parameterization of ConvergeFlow.

I will raise my score once my concerns are well addressed.

4 comments

We thank the reviewer for their careful reading and constructive feedback. We address the reviewer's concerns below.

Sampling schedule and time conditioning.

We would like to clarify that the step-size condition in Theorem 1 holds for our experimental schedule. Theorem 1 allows general schedules and is not restricted to the linear schedule. Our experiments use LangFlow's variance-preserving parameterization with a Gumbel mapping. Following the time convention for tt in our paper, we have \begin{align*} \gamma(t)=\mu-\beta\log[-\log(1-t)],\qquad \alpha_t=\frac{1}{\sqrt{1+\exp(\gamma(t))}},\qquad \sigma_t=\frac{1}{\sqrt{1+\exp(-\gamma(t))}}. \end{align*} Here, μ=4.9563\mu=4.9563 and β=0.9637\beta=0.9637 obtained from LangFlow checkpoint are the Gumbel location and scale parameters, respectively. For the midpoint grids tk=(k+0.5)/Nt_k=(k+0.5)/N, k=0,…,N−1k=0,\ldots,N-1, we have max⁡0≤k<N−1(1+αtkσtk)∥αtk+1σtk+1−αtkσtk∥=O((log⁡N)β−1)→0,asN→∞.\max_ {0\le k<N-1}\left(1+\frac{\alpha_ {t_k}}{\sigma_ {t_k}}\right)\|\frac{\alpha_ {t_ {k+1}}}{\sigma_ {t_ {k+1}}}-\frac{\alpha_ {t_ k}}{\sigma_ {t_k}}\|=O\left((\log N)^{\beta-1}\right)\to 0,\qquad \text{as} \quad N\to\infty. Thus, the experimental midpoint grids satisfy the step-size requirement once NN is sufficiently large. In addition, Theorem 1 primarily aims to establish a non-asymptotic convergence result. Such step-size conditions are common in theoretical analyses of diffusion models.

Moreover, the existing proof directly yields the following alternative step-size condition:

Corollary. The convergence in Theorem 1 still holds if we replace the step-size condition $ (1+{\alpha_{t_k}}/{\sigma_{t_k}})|{\alpha_{t_k}}/{\sigma_{t_k}}-{\alpha_{t_{k+1}}}/{\sigma_{t_{k+1}}}| < \delta$ with \begin{align*} \sum_{k=0}^{N-2} 1+αtkσtk)2(αtkσtk−αtk+1σtk+11+\frac{\alpha_{t_k}}{\sigma_{t_k}}\big)^2\big(\frac{\alpha_{t_k}}{\sigma_{t_k}}-\frac{\alpha_{t_{k+1}}}{\sigma_{t_{k+1}}}^2=O(1). \end{align*} For LangFlow's Gumbel schedule, this condition is also satisfied because ∑k=0N−2�1�2=O((log⁡N)2β−2).\sum_{k=0}^{N-2} �1�^2=O((\log N)^{2\beta-2}). We have added this result to Section 3 and its proof to the Appendix~D.1 of the revised manuscript.

Concerns on experiments.

We address these concerns in three parts:

∙\bullet Primary results: We have included the main results in the main text, which is given by Figure 1, while the corresponding numerical results is given by Table 2 due to space constraints. We use Figure 1 instead of Table 2 in the main text for an intuitive visual comparison. With the additional page available in the revised manuscript, we have now the original Table 2, now Table 1, to the main text. In this paper, Figure 1 already presents the pareto frontier between Gen PPL and entropy, which are obtained by varying the guidance scale rather than the temperature.

∙\bullet Gen. PPL: We agree that Gen. PPL alone can be misleading for repetitive or degenerate outputs. However, Gen. PPL is relatively easy to compute using a pretrained reference language model and is commonly used to evaluate both continuous and discrete language models (Chen et al., 2026; Sahoo et al., 2025). Following these studies, we report Gen. PPL alongside unigram entropy to assess quality and diversity jointly.

∙\bullet PPL: In diffusion language models, exact PPL computation is challenging because evaluating sequence likelihoods generally requires intractable marginalization over latent diffusion states. We therefore estimate an upper bound on PPL following LangFlow (Chen et al., 2026). Figures 3(c) and (d) and Table 3 in the revised manuscript reports these results for continued training with MSE or CE and for switching between the two objectives. The corresponding numerical results are reported in the table below. The results show that the MSE run has a lower estimated PPL upper bound than the CE run throughout continued training and the objective-switching experiment.

0K25K50K75K100K125K150K175K200K
Continued trainingMSE24.6024.3124.4224.3024.1824.0724.0023.9323.83
CE24.6024.9524.9924.9024.8624.8024.7924.7524.74
Swapping objectiveMSE24.8624.2124.0623.9723.9023.8323.7823.7323.67
CE24.1824.6024.7124.7224.7124.6924.6624.6424.64

Concern on implementation.

We retain LangFlow's γ\gamma-based time conditioning. Each sampled tt is mapped to γ(t)\gamma(t) via the Gumbel mapping, and γ(t)\gamma(t) is passed to the DiT as conducted in LangFlow. We have clarified this implementation in Section 4 in the revised manuscript, stating: "The DiT retains LangFlow's γ\gamma-based time conditioning, with each sampled tt first mapped to γ(t)\gamma(t) using the Gumbel mapping."

Concern on convergence and decoding.

The potential benefit of convergence for Gen. PPL is as follows. Our convergence guarantee supports training the data predictor with MSE alone. Flow matching requires the conditional embedding mean, μ(xt,t)=E[x⋆∣xt,t]\mu(x_t,t)=\mathbb{E}[x_\star\mid x_t,t]. The MSE objective in Equation (6) directly supervises this continuous prediction, while CE supervises token probabilities that are subsequently mapped to an embedding-weighted prediction. Intuitively, direct supervision may improve estimation of this mean under finite model capacity, consistent with the Gen. PPL gains observed in the training objective comparisons in Figure 3.

At inference, we would like to clarify that ConvergeFlow also uses greedy decoding. Specifically, at each position, we select the token whose embedding is closest to the final continuous state, which is established under the convergence to valid token.

In addition, Theorem 1 establishes that the continuous generation trajectory converges to a valid token embedding even when the learned data predictor is imperfect. In the limit, the continuous dynamics alone produce valid token embeddings, without an additional projection step or a CE-supervised decoder. Beyond its practical benefits, our work therefore provides theoretical guidance for designing continuous language models.

Thank you for your response. I have raised my score.

Thank you for your positive assessment and constructive suggestions. Your feedback has helped us improve the paper, and we appreciate your time and thoughtful review.

Reviewer_S9AnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies whether MSE loss could be used for flow matching with discrete objects like language or if cross entropy is inevitable. They find theoretically (and empirically) that MSE could be sufficient.

Critical strengths
  • The paper is generally well written and easy to follow
  • The contributions are clear and sound, both theoretically and empirically
  • The paper does interesting ablations like switching CE and MSE loss
Critical weaknesses
  • Could the authors comment on whether it is fair to use LangFlow embeddings in the comparison? Aren't you using more compute in the process for your method?
  • The inability to train embeddings seems like a general weakness, could the authors please comment on that.
  • I would appreciate more discussion on why replacing CE loss with MSE is a useful thing to do? The intro mentions "Although effective, this reintroduced discrete supervision is inconsistent with the continuous nature of flow-based models, and may limit their full potential." I would like to understand better what this would look like and also some initial evidence that this is indeed the case.
Questions for authors

see weaknesses above

1 comment

We thank the reviewer for carefully reading our paper and recognizing the strengths of our work. We address the reviewer's concerns below.

Comparison with LangFlow under the same training budget

We agree that the comparison should be conducted under matched training budgets, and our experiments follow this. Indeed, Figure 3(a) compares our algorithm with MSE loss and LangFlow with CE loss continuation from the LangFlow checkpoint for an additional 200K steps. Our algorithm can reduce Gen. PPL, whereas continued LangFlow training yields little improvement.

In addition, we missed the tradeoff comparison with continued LangFlow curve in Figure 1 and Table 1. We have therefore evaluated the best checkpoint from the continued LangFlow training and added its Gen. PPL-entropy curve and its tradeoff results (see the revised Figure 1 and Table 2) to accommodate your suggestion. ConvergeFlow still achieves a more favorable Gen. PPL-entropy trade-off than the continued LangFlow baseline.

Training embeddings

Yes, training embeddings with the MSE objective alone admits degenerate solutions. Specifically, if e1=⋯=eV=ce_1=\cdots=e_V=c, the embedding-weighted predictor satisfies μθ(i)(xt,t)=∑j=1Vwθ(i)(j∣xt,t)ej=c\mu_\theta^{(i)}(x_t,t)=\sum_{j=1}^{V}w_\theta^{(i)}(j\mid x_t,t)e_j=c, yielding zero MSE, but tokens are no longer distinguishable. Hence, to train the token embeddings, we need to add some regularization to ensure the well-separated token embeddings. We have clarified the rationale for fixing the embeddings in Section 4, stating: "Because jointly learning the token embeddings and data predictor under the MSE objective admits degenerate embedding-collapse solutions, we use the embedding matrix from the LangFlow checkpoint and keep it fixed throughout training. To enable embedding learning, an auxiliary CE objective for predicting tokens from noisy embeddings can be added to discourage collapse." We also identify continuous regularizers that support embedding learning without token-level CE supervision as a direction for future work in Section 5.

Intuitive benifits for using MSE over CE

Flow matchings require the conditional embedding mean, μ(xt,t)=E[x⋆∣xt,t]\mu(x_t,t)=\mathbb{E}[x_\star\mid x_t,t]. CE supervises conditional token probabilities, which are subsequently mapped to an embedding-weighted data prediction. The MSE objective in Equation (6) directly supervises this continuous prediction. Intuitively, directly supervising the embedding prediction could improve estimation of this mean under finite model capacity. The empirical evidence can be found in Figure 3, showing that switching from CE to MSE reduces Gen. PPL, whereas switching from MSE to CE increases it. We have revised Section 1 to clarify this benifit, stating: "CE supervises conditional token probabilities, which are mapped to an estimate of the conditional embedding mean required by FMs. Direct MSE supervision of this continuous prediction may improve its accuracy under finite model capacity."

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.