acceptodds
Under review as a conference paper at ICLR 2027

Universality and Generalization of Causal Transformers Across Context Lengths

Abstract

Long contexts are central to modern transformer systems, but most expressivity results choose a different network for each fixed sequence length. We study whether one masked transformer can approximate causal token-to-token maps uniformly over sequences of arbitrary length sampling a fixed normalized horizon. To relate sampling resolutions, we model tokens by -Hölder sequences or, more generally, a common modulus of continuity. Our notion of continuity across resolutions characterizes the causal families admitting uniform approximation on these compact input classes by a single transformer with length-independent parameters. The result extends to the infinite-length mean-field limit, where tokens form continuous curves and masked attention becomes a causal time integral. For bounded regression with target maps satisfying a -smooth stability condition defined using regular test functions, quantitative approximation yields a generalization bound: exact empirical risk minimization over suitably sized bounded-weight transformers gives root mean-square prediction error from iid labeled sequences. The bound holds at fixed confidence on the same sampling distribution, with the token dimension and no maximum-length factor. Finally, experiments on physical time series support the Hölder-regular token model at observed scales, with dataset-dependent fitted exponents, whereas text input embeddings provide a contrasting case. Native and dense sampling, shuffled controls, and refinement checks delimit this empirical regularity regime.

Primary area:
learning theory
Keywords:
Causal transformers, universal approximation, long-context modeling, quantitative approximation rates, Hölder regularity, continuous-time attention, positional encoding

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_3r8ROpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper characterizes causal map families that admit uniform approximation across all sampling resolutions by a single transformer with length-independent parameters, under common input regularity. With additional target regularity, it also establishes approximation and generalization upper bounds and finite-bit approximation lower bounds.

Critical strengths
  1. The paper provides a unified treatment of sequence causality, uniform approximation across resolutions with shared parameters, and the continuous-time limit.

  2. Beyond universality, it provides quantitative approximation and generalization upper bounds together with finite-bit approximation lower bounds, examining the complexity required for approximation.

Critical weaknesses

The dependence on token dimension is an important limitation of the quantitative results. Both the approximation and prediction RMSE upper bounds have convergence exponent β/(d+2)\beta/(d+2), which deteriorates with dimension dd for fixed regularity β\beta. This yields weak quantitative guarantees for high-dimensional token representations. Moreover, the finite-bit approximation lower bound has no corresponding dimension penalty in its exponent, leaving unclear whether this deterioration reflects intrinsic difficulty of the target class or losses in the construction and analysis.

Questions for authors

To what extent can the current proofs explain the dimension-dependent gap between the approximation upper and finite-bit lower bounds? In relation to Kovachki, Lanthaler, and Mhaskar (2024), please clarify whether this gap can be traced to losses in the present construction and analysis, or to obstacles in transferring their lower-bound arguments to the present assumptions and complexity measure.

2 comments

Thank you for highlighting the unified theory and the quantitative gap.

Weakness

Both the approximation and prediction RMSE upper bounds have convergence exponent β/(d+2)\beta/(d+2), which deteriorates with dimension dd for fixed regularity β\beta. [...] Moreover, the finite-bit approximation lower bound has no corresponding dimension penalty in its exponent, leaving unclear whether this deterioration reflects intrinsic difficulty of the target class or losses in the construction and analysis.

The factor d+2d+2 comes from our construction; we have not shown it to be intrinsic. Roughly, we resolve distributions in the d+1d+1 token-time coordinates using polynomial moments, recover those moments from attention, and approximate the resulting feature-to-output map by a shallow ReLU network. The number of moments grows with dimension; their unstable recovery requires finer feature precision, adding the extra power in the complexity estimate. The learning bound inherits this exponent from the approximation theorem.

The proof may be pessimistic because it resolves ambient measures and approximates the readout on a full feature cube, rather than exploiting the smaller set of realizable histories. But improving the exponent, especially removing its dimension dependence, appears beyond the present moment-recovery/readout argument: it would require a more stable representation or a quantitatively efficient approximation directly on that history set. We will explain these potential losses without claiming that they are removable or that the upper rate is optimal.

Question

To what extent can the current proofs explain the dimension-dependent gap between the approximation upper and finite-bit lower bounds? In relation to Kovachki, Lanthaler, and Mhaskar (2024), please clarify whether this gap can be traced to losses in the present construction and analysis, or to obstacles in transferring their lower-bound arguments to the present assumptions and complexity measure.

Both issues matter: our upper construction has identifiable costs, while the available lower bounds do not yet match the target class and complexity measure needed to certify those costs. Proposition B.7 already proves logarithmic finite-bit hardness, but uses paths varying in only one token coordinate and therefore does not establish a dimension-dependent exponent.

Kovachki, Lanthaler, and Mhaskar (2024), Theorem 2.10, prove logarithmic lower bounds for Fréchet-smooth functionals on rich compact input sets. Their argument embeds high-dimensional smooth-function problems into the operator class. This supports the relevance of logarithmic rates, but does not directly settle our gap: our smoothness is stability under regular tests of history measures, not Fréchet differentiability; their bounds concern function-value samples or continuous real-valued encodings, whereas ours concerns finite-bit descriptions. A transfer must build a hard subclass satisfying our regularity condition and compare these information models. We will clarify this distinction and identify matching approximation and statistical lower bounds as open problems.

Thank you for clarifying the sources of the dimension dependence and the differences from prior lower bounds. The gap remains a limitation, but the explanation addresses my question and does not change my positive assessment of the unified causal approximation framework. I maintain my weak-accept recommendation.

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

This paper studies the universality and generalization of masked transformer over arbitrary-length sequences sampled from a fixed normalized horizon. The universality result is extended to the infinite-length mean-field limit. The authors also conduct empirical experiments to justify the assumption in the case of time series data.

Critical strengths
  • The paper presents a comprehensive theoretical framework to analyze causal transformers with potentially infinite length. This is the unified treatment for the discrete and continuous versions of transformer, provide a valuable perspective on the expressivity and generalization.
Critical weaknesses
  • The setting is more like sampling from a signal (tokens in sequence are assumed to have time in [0,1][0,1] and tokens close in time or position will close in value). These assumptions seem to support time series data rather than generalizing to text data. The empirical diagnosis section of the papers has also shown this.
Questions for authors
  • Please clarify the advantage of the framework proposed in this work over previous work [1], as it has already investigated transformer universality with arbitrarily many tokens.
  • It seems Theorem 2.9 bounds prediction error under the same training and testing distribution, including the distribution of sequence length. Please clarify if this result can generalize over length, theoretically or empirically, (eg, test sequence longer than train sequence and vice versa), and what additional assumptions need, as length generalization seems to be the motivation for this paper.

Ref:

[1] Furuya et al. Transformers are universal in-context learners. ICLR 2025

2 comments

Thank you for recognizing the unified discrete/continuous theory.

Weakness

These assumptions seem to support time series data rather than generalizing to text data. The empirical diagnosis section of the papers has also shown this.

We agree: the natural setting is regular time series sampled over a fixed horizon, not unrestricted text. We will state this more clearly in the final version, including the abstract, and present text embeddings as a contrasting case.

Question - comparison with prior work

Please clarify the advantage of the framework proposed in this work over previous work [1], as it has already investigated transformer universality with arbitrarily many tokens.

The published proof of causal universality in Furuya et al. (ICLR 2025) is incorrect. A correct proof requires a different treatment of causal histories and their infinite-token limit. In particular, Lemma 10 incorrectly claims that normalized truncation is continuous at temporal atoms. Moreover, the attention functions used there cannot distinguish joint token-time distributions when time is used only for masking. These issues concern the causal proof, not the unmasked result.

We follow the basic ideas of token-time lifting and Stone-Weierstrass, but use continuous paths and temporal-integral attention to construct a compact space of causal histories. Proposition D.5 shows that these histories can be identified continuously and exactly. Together with separation by the positional attention functions, this allows us to apply Stone-Weierstrass. The key difference is the use of positional information in the attention values and residual features.

We also prove a new negative result in Theorem A.2. Even with arbitrary position-dependent attention scores, the score-only positional architecture considered there cannot uniformly approximate the clock map Fn(z)i=i/nF_n(z)_i=i/n over all sequence lengths with error below 1/21/2.

We will explain these differences from Furuya et al. more clearly in the introduction.

Question - generalization across lengths

It seems Theorem 2.9 bounds prediction error under the same training and testing distribution, including the distribution of sequence length. Please clarify if this result can generalize over length, theoretically or empirically, (eg, test sequence longer than train sequence and vice versa), and what additional assumptions need, as length generalization seems to be the motivation for this paper.

Theorem 2.9 is a same-distribution guarantee, including the length distribution; it does not guarantee transfer to lengths excluded from training support. That distribution may have unbounded support, and a finite training sample need not contain every possible test length. This differs from deliberately training only at short lengths and testing outside that range. Our universality theorem supplies one approximant valid at every resolution, but does not show that restricted-resolution training discovers it. We will state this distinction more explicitly and avoid implying an empirical transfer result.

Thank you for the clarifications. I'll keep my score.

Reviewer_tn8JOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether a single causal transformer can approximate a prediction rule across sequences of different lengths without changing its parameters. My understanding is that the inputs are samples of continuous signals over a fixed time interval, with a common regularity assumption, and the prediction rules at different sampling resolutions must approach a common continuous rule. Under these assumptions, the authors show that one attention block followed by a shared MLP is sufficient for uniform approximation across all lengths. They also extend the result to the continuous-time limit and derive logarithmic approximation and learning bounds under additional assumptions on the target. The experiments mainly examine whether time-series data and token embeddings satisfy the input regularity assumption at the scales tested.

I had difficulty understanding how much this says about long-context models in the usual sense. Here, increasing the length means sampling the same time interval more finely, so I am unsure how the theory applies when additional tokens introduce new information or when the model needs to retrieve one particular token. I also found the comparison with earlier expressivity results unclear. Previous work already considers a single model across lengths, including work on formal languages, and I would like to understand more clearly what is added here and how the result differs from Furuya et al. (2025).

Critical strengths
  1. Clean results within the chosen setup. I have doubts about whether the setup fits long-context modeling (see Weakness 1). Still, once one accepts that a longer sequence is a finer sampling of a smooth signal, the results are clean. The paper gives an exact characterization of when one transformer can work across all lengths, with both the sufficient condition (Thm 2.4) and its converse (Cor. D.13), and the same parameters also work in the continuous-time limit (Thm 2.6).
  2. Quantitative results and honest limitations. Within the same setup, the paper goes beyond existence and gives an explicit construction with bounded weights (Thm B.3) and a generalization bound that does not depend on the maximum sequence length (Thm 2.9). The paper is also open about the limits of these results, for example that the rates are slow and that text embeddings do not fit the Hölder model.
Critical weaknesses
  1. Does this setting allow retrieval at all? My main concern is the needle-in-a-haystack question, which is arguably what long contexts are for. One token carries a key fact, and the model must find it later. In this paper, a longer sequence is just a finer sampling of the same time window, so it carries no new information. The smoothness assumption forces each token to be close to its neighbors, so a token that stands out (the needle) is not allowed once nn is large. Information must be spread over a large fraction of the context to be visible. Likewise, the construction keeps attention scores bounded, so every token gets weight roughly 1/n1/n, and the influence of any single token fades as the context grows. Intuitively, the model's resolution on individual tokens gets worse as the context gets longer, which is the opposite of what we want from long-context models. If this is right, I find the contribution questionable. The paper studies a simpler setting in which longer sequences lose information about individual tokens, and the hard part of long-context modeling is exactly keeping that information. Uniform approximation across lengths then seems much easier to achieve, and much less relevant to the long-context transformers that motivate the paper.

  2. The main result looks very close to Furuya, de Hoop & Peyré (ICLR 2025). Their Theorem 2 already shows that one causal transformer with fixed parameters can approximate continuous causal maps on inputs with any number of tokens, even infinitely many. As far as I can tell, the proof idea is also the same. Both papers attach a time stamp to each token, view each prefix as a distribution over (token, time) pairs, assume the tokens vary smoothly in time so that the input class is compact, and then apply Stone–Weierstrass. Their smoothness assumption is in fact the special case ω(r)=Cr\omega(r) = Cr of the one used here. Yet the paper cites this work only briefly, and the introduction suggests the length-independent question is new. The genuine additions (Hölder smoothness, a one-block construction, the converse) seem modest and should be stated clearly as the contribution beyond Furuya et al.

  3. The rates are too slow to be informative, and the experiments do not test the theory.

    • The rates are extremely slow. The error bound is (log⁡log⁡p/log⁡p)β/(d+2)(\log\log p/\log p)^{\beta/(d+2)} in the number of parameters pp. Even in the easiest case (d=1d = 1, β=1\beta = 1), a trillion parameters (p=1012p = 10^{12}) only guarantees an error of about 0.490.49.
    • We do not know whether the rate is tight. The lower bound has exponent β\beta and the upper bound has β/(d+2)\beta/(d+2). The paper admits that the extra d+2d+2 is "not known to be necessary" (p. 24), but the abstract states the rate without this caveat.
    • The experiments train no model. Section 3 only checks whether real data looks Hölder-smooth, and even this check is not convincing. A signal with a sudden jump still gets a positive Hölder slope after encoding, and for two of four domains the estimated Hölder constant grows by about 1.6×1.6\times under finer sampling (App. H). Saying the experiments "support the Hölder-regular token model" seems too strong.
  4. The comparison with prior expressivity results is unclear to me. I was confused by the abstract's statement that most expressivity results choose a different network for each fixed sequence length. My understanding is that work on formal-language recognition and generation already considers transformers across unbounded lengths, for example Yao et al. (2021) on bounded-depth Dyck languages and Hayakawa and Sato (2025) on Dyck recognition and generation. I realize that the architecture and precision assumptions differ, and these results do not establish the same uniform approximation theorem. Still, do the authors mean specifically existing universal approximation results for continuous sequence-to-sequence maps? It would help to make this distinction explicit and explain what the present guarantee adds relative to the formal-language setting.

Questions for authors
  1. In a needle-in-a-haystack task, if a single token carries information that must be retrieved later, is that input covered by your assumptions? If not, what can the theory say about retrieval, and how does a single token's influence survive as the context grows?
  2. Please state precisely what Theorems 2.4 and 2.6 establish beyond Theorem 2 of Furuya, de Hoop & Peyré (ICLR 2025). Is the delta exactly: general concave modulus, the discrete-family formulation (Def. 2.2), the parallel shallow realization, and the converse (Cor. D.13)? The introduction should say so. This is the question most likely to change my recommendation.
  3. With the token-averaged objective 1∑knk∑k∑i∥⋅∥2\frac{1}{\sum_k n_k} \sum_k \sum_i \|\cdot\|^2 instead of the sequence-averaged loss (20), what bound do you obtain? If a length factor appears, "no maximum-length factor" in the abstract should be qualified.
  4. One small experiment that exercises the theory — train the shallow architecture on Example 2.8 targets with lengths in [16,256][16, 256] and evaluate at lengths up to 4096 — would be far more informative than Section 3 as it stands.
  5. Given the AI Use Statement (p. 10: proofs "developed and audited" with OpenAI Codex), were Lemmas C.1, C.5 and G.2 independently verified line by line by a human author?
3 comments

Thank you for the detailed review. We clarify the scope, the correction of the earlier causal universality argument, and the meaning of the quantitative results below.

Weakness 1

Does this setting allow retrieval at all? [...] The smoothness assumption forces each token to be close to its neighbors, so a token that stands out (the needle) is not allowed once nn is large.

An isolated order-one needle is indeed outside our common-modulus input class; the theorem concerns regular signals sampled increasingly finely, not unrestricted symbolic strings. In fact, with uniformly bounded positional encodings, even the simplest needle-in-a-haystack task, copying the first token, cannot be uniformly approximated over unrestricted input sequences of arbitrary length by a fixed finite-depth, full-prefix causal softmax/ReLU transformer with length-independent parameters and no additional unbounded positional score biases. A simple counterexample is the pair (1,0,…,0)(1,0,\ldots,0) and (−1,0,…,0)(-1,0,\ldots,0): the final outputs approach each other as length grows, whereas the desired outputs remain 22 apart. This is the attention-dilution mechanism: a single token's influence vanishes under bounded-score attention, as in Hahn (2020).

Lifting the boundedness assumption on positional encodings is an interesting direction for future work: attention dilution then need no longer hold. However, the augmented token-position domain is no longer compact, and a continuous extension of attention to our completed history space is no longer guaranteed. We will explicitly add these architectural limitations, the regular-signal scope of our theorem, and this future direction to the final paper.

Weakness 2

The main result looks very close to Furuya, de Hoop & Peyré (ICLR 2025). [...] The genuine additions (Hölder smoothness, a one-block construction, the converse) seem modest and should be stated clearly as the contribution beyond Furuya et al.

The published proof of causal universality in Furuya et al. (ICLR 2025) is incorrect. A correct proof requires a different treatment of causal histories and their infinite-token limit. In particular, Lemma 10 incorrectly claims that normalized truncation is continuous at temporal atoms. Moreover, the attention functions used there cannot distinguish joint token-time distributions when time is used only for masking. These issues concern the causal proof, not the unmasked result.

We follow the basic ideas of token-time lifting and Stone-Weierstrass, but use continuous paths and temporal-integral attention to construct a compact space of causal histories. Proposition D.5 shows that these histories can be identified continuously and exactly. Together with separation by the positional attention functions, this allows us to apply Stone-Weierstrass. The key difference is the use of positional information in the attention values and residual features.

We also prove a new negative result in Theorem A.2. Even with arbitrary position-dependent attention scores, the score-only positional architecture considered there cannot uniformly approximate the clock map Fn(z)i=i/nF_n(z)_i=i/n over all sequence lengths with error below 1/21/2.

We will explain these differences from Furuya et al. more clearly in the introduction.

Weakness 3 (rates)

The rates are extremely slow. [...] Even in the easiest case (d=1d = 1, β=1\beta = 1), a trillion parameters (p=1012p = 10^{12}) only guarantees an error of about 0.490.49.

We do not know whether the rate is tight. The lower bound has exponent β\beta and the upper bound has β/(d+2)\beta/(d+2).

The rate is indeed logarithmically slow, a familiar phenomenon in functional regression and operator learning; see Meister (2016) and Kovachki, Lanthaler, and Mhaskar (2024). Those works show that such slow rates can reflect genuine infinite-dimensional difficulty, although they do not establish our exponent's optimality; we discuss this distinction further in our response to Reviewer 3r8R.

The bound is primarily theoretical, not a practical model-size prescription: the aim is to gain intuition about mechanisms enabling all-resolution expressivity, understand how target smoothness improves rates, and guide further improvements. Its contributions are an explicit way to measure the smoothness of causal prediction maps through regular token-time test functions, a construction showing how β\beta improves approximation, and a bounded-weight approximation/covering argument yielding learning guarantees independent of maximum length. Matching lower bounds remain open: the upper bound pays for stable feature recovery and readout, whereas our lower construction measures finite-bit complexity and does not capture that dimension dependence. Closing both gaps requires more than adjusting constants. We will emphasize the theoretical interpretation and unresolved sharpness in the abstract and rate discussion.

Weakness 3 (diagnostics)

The experiments train no model. [...] A signal with a sudden jump still gets a positive Hölder slope after encoding, and for two of four domains the estimated Hölder constant grows by about 1.6×1.6\times under finer sampling (App. H). Saying the experiments "support the Hölder-regular token model" seems too strong.

We agree that these diagnostics do not validate an all-resolution Hölder model, particularly for text embeddings. The intended setting is regular time series; language embeddings are a contrasting case. Even for physical signals, a positive finite-scale slope is not a Hölder certificate: patch encoding can obscure a jump, and the refinement results do not establish uniformly bounded Hölder constants. We will state these limitations directly, describe the experiments as finite-scale input diagnostics rather than tests of trained performance, and temper the abstract's empirical claim.

Weakness 4

My understanding is that work on formal-language recognition and generation already considers transformers across unbounded lengths, for example Yao et al. (2021) on bounded-depth Dyck languages and Hayakawa and Sato (2025) on Dyck recognition and generation. [...] Still, do the authors mean specifically existing universal approximation results for continuous sequence-to-sequence maps?

Hayakawa and Sato (2025) predate our work and explicitly study arbitrary-accuracy next-token distribution approximation across context lengths, but for a different target family specified by Dyck generative processes. Their Definition 7 uses total-variation approximation, so the distinction is not simply approximation versus recognition. Yao et al. (2021) establish recognition and threshold-based generation of Dyck languages, not universal approximation of arbitrary continuous causal maps. Our result instead quantifies over every continuously extendable causal family on the common-modulus input classes. We will cite both works explicitly and distinguish task-specific recognition/distribution approximation from this operator-universality statement.

Question 1

In a needle-in-a-haystack task, if a single token carries information that must be retrieved later, is that input covered by your assumptions?

No, not if the token remains separated from its neighbors by an order-one jump as resolution increases. As explained under Weakness 1, the theorem covers regular-signal histories, and single-token retrieval faces a separate attention-dilution obstruction under the bounded-position architecture assumptions stated there. We will clarify this boundary and discuss unbounded positional encodings as future work.

Question 2

Please state precisely what Theorems 2.4 and 2.6 establish beyond Theorem 2 of Furuya, de Hoop & Peyré (ICLR 2025). Is the delta exactly: general concave modulus, the discrete-family formulation (Def. 2.2), the parallel shallow realization, and the converse (Cor. D.13)?

Those are additional features, but the central distinction is that the earlier causal proof is incorrect, as detailed under Weakness 2. The new completed-history model, continuous representation, and separating positional features establish a valid all-resolution causal universality result and its path limit. We will state this distinction explicitly, alongside the quantitative approximation and learning results.

Question 3

With the token-averaged objective 1∑knk∑k∑i∥⋅∥2\frac{1}{\sum_k n_k} \sum_k \sum_i \|\cdot\|^2 instead of the sequence-averaged loss (20), what bound do you obtain? If a length factor appears, "no maximum-length factor" in the abstract should be qualified.

This is an important, nontrivial question: removing 1/nk1/n_k in (20), up to an overall normalization, gives longer sequences more weight. The same leading logarithmic rate extends to unbounded lengths with E[n2]<∞\mathbb E[n^2]<\infty, but the finite-sample estimation term changes. Set m:=E[n]m:=\mathbb E[n], κ:=E[n2]/m\kappa:=\sqrt{\mathbb E[n^2]}/m, and ap:=(log⁡log⁡p/log⁡p)β/(d+2)a_p:=(\log\log p/\log p)^{\beta/(d+2)}. Under the other assumptions of Theorem 2.9, for sufficiently large pp, exact token-weighted ERM followed by clipping satisfies, with probability at least 1−δ1-\delta,

:=\frac{\mathbb E\!\left[\sum_{i=1}^n \|\widehat F_n(Z)_i-F_n^\star(Z)_i\|^2\right]}{m} \leq C_\delta\left[a_p^2+ \kappa\sqrt{\frac{p\log(CpN)}{N}}\right].$$ Here the expectation is over an independent test sequence, conditional on the training data. The moment factor $\kappa$ affects estimation, not approximation bias. The existing parameter cover, symmetrization, and the square-integrable envelope for length-weighted sequence losses give this extension. With $p=\lfloor\sqrt N\rfloor$ and fixed confidence, the polynomial estimation remainder is asymptotically smaller than $a_p^2$, so the RMSE retains the paper's logarithmic rate. We will add the result and proof, and clarify that the original theorem needs no length moments specifically because it averages losses within each sequence. ### Question 4 > One small experiment that exercises the theory — train the shallow architecture on Example 2.8 targets with lengths in $[16, 256]$ and evaluate at lengths up to 4096 — would be far more informative than Section 3 as it stands. Our theory establishes all-resolution approximation, but does not guarantee learning-based transfer to resolutions excluded from the training distribution. As phrased, the experiment tests precisely this additional transfer property, rather than the guarantee of Theorem 2.9. Smoothness alone does not identify a target away from the training-history support. Could the reviewer clarify whether the intended request is an empirical study of such transfer, or a same-distribution training experiment illustrating our learning bound? Both are meaningful, but address different questions. We will make this distinction explicit; we do not have trained results to report. ### Question 5 > Given the AI Use Statement (p. 10: proofs "developed and audited" with OpenAI Codex), were Lemmas C.1, C.5 and G.2 independently verified line by line by a human author? Yes: Lemmas C.1, C.5, and G.2 were independently verified line by line by a human author.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.