Pith. sign in

REVIEW 3 major objections 6 minor 115 references

Structured Newton Layer Parallelism cuts sequential nonlinear depth in encrypted Transformer inference, using far fewer bootstraps while accumulating less polynomial error.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 16:18 UTC pith:U7LX5WSE

load-bearing objection Solid FHE-systems application of prior SNLP: real multi-model evidence that layer-parallel structure can cut sequential nonlinear depth and not worsen poly-error accumulation, but the 2.65× bootstrap headline is still a symbolic proxy, not encrypted runs. the 3 major comments →

arxiv 2607.04819 v2 pith:U7LX5WSE submitted 2026-07-06 cs.LG cs.CR

Layer-Parallel Inference Reduces Encrypted Nonlinear Depth in Transformers

Ligong Han , Kai Xu , Hao Wang , Ruijiang Gao , Han Gao , Akash Srivastava This is my paper
classification cs.LG cs.CR
keywords fully homomorphic encryptionTransformer inferencelayer parallelismSNLPpolynomial approximationCKKSbootstrappingerror amplification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Encrypted Transformer inference is expensive because each layer still needs polynomial approximations of nonlinear ops such as softmax, and those stages stack in sequence, burning modulus and forcing frequent bootstraps. This paper argues that Structured Newton Layer Parallelism changes the inter-layer graph rather than the local operators: a prefix of layers runs as usual, then many suffix layers are evaluated in parallel over a small number of Newton-style iterations and corrected with cheap linear updates that add little or no multiplicative depth. In a Chebyshev-polynomial simulation across eight models, that structure lowers error amplification relative to ordinary sequential inference and, on a 0.5B IDN-trained model, cuts symbolic bootstraps from 53 to 20 while keeping encrypted perplexity within about 1.2% of the sequential encrypted baseline. Softmax approximation remains the main local error source; arithmetic noise is negligible. A sympathetic reader cares because bootstrap cost dominates FHE latency, so any method that shortens sequential nonlinear depth without wrecking quality is a practical lever, complementary to better per-block approximations.

Core claim

SNLP reduces layerwise sequential nonlinear depth from L stages to (L-N)+K stages plus linear structured corrections. Under degree-12 Chebyshev simulation, this consistently lowers error amplification versus sequential inference across eight models in four architecture families, and on a 0.5B IDN-trained model yields 2.65 imes fewer symbolic bootstraps (20 vs 53) with encrypted perplexity within 1.2% of sequential encrypted PPL.

What carries the argument

Structured Newton Layer Parallelism (SNLP): solve the residual equation for the layer hidden-state trace by evaluating N suffix blocks in parallel for K iterations, then applying a structured Newton correction (Identity Newton is purely additive with zero FHE multiplicative depth; HC Newton uses a small linear mixing matrix). NFE = (L-N)+K tracks bootstrap count in the symbolic CKKS model.

Load-bearing premise

The Chebyshev-and-symbolic-bootstrap simulation ranks sequential versus layer-parallel inference the same way real encrypted arithmetic would.

What would settle it

Run actual CKKS (or a circuit that includes max-subtraction comparisons and real rsqrt) on the same 0.5B IDN model at degree-12 softmax: if SNLP n24-K4 no longer shows lower error amplification or roughly 2.65× fewer bootstraps than sequential, the claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Encrypted inference can trade a few extra solver iterations for a large cut in sequential nonlinear depth and bootstrap count.
  • SNLP-aware training is required to reach the low-NFE operating region with usable quality.
  • Models with linear residual-stream mixing (mHC-style) are inherently more FHE-friendly because those connections add no nonlinear depth.
  • FHE-optimal SNLP settings can differ from wall-clock-optimal ones: higher K can still be cheaper when cost is dominated by depth.
  • SNLP complements, rather than replaces, better local polynomial approximations of softmax and related ops.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Pairing SNLP with softmax-light or linear-attention blocks could stack the depth savings, since softmax already dominates the simulated error budget.
  • Training that optimizes the depth–quality frontier rather than GPU wall-clock time may unlock more aggressive encrypted operating points.
  • The same residual-equation view of depth may apply to other deep residual stacks under FHE, not only Transformers.
  • Once comparison circuits for stable softmax are counted, absolute bootstrap numbers will rise, but relative ranking should still favor structures whose corrections stay linear.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper studies whether Structured Newton Layer Parallelism (SNLP) can reduce the sequential nonlinear depth of encrypted Transformer inference under FHE. Rather than redesigning per-block operators, SNLP replaces L sequential nonlinear block evaluations with a short sequential prefix plus K parallel suffix iterations and linear structured corrections (IDN/HCN). Using a Chebyshev-polynomial simulation of FHE-unfriendly ops and a symbolic CKKS cost model, the authors define NFE = (L−N)+K as a bootstrap proxy and compare sequential vs SNLP error accumulation across 8 Nanochat models in 4 architecture families. Headline result: on a 0.5B IDN-trained model at degree-12, SNLP n24-K4 cuts symbolic bootstraps 53→20 (2.65×) with encrypted PPL within 1.2% of sequential encrypted PPL and lower amplification (1.36× vs 1.42×); SNLP amplification is lower on all tested models. Ablations attribute most error to softmax approximation and find CKKS arithmetic noise negligible.

Significance. If the relative depth and error-accumulation claims hold under real FHE stacks, this is a useful complementary axis for private Transformer inference: inter-layer graph restructuring rather than only better local polynomial approximations. Strengths include multi-model/multi-architecture evaluation, a clear NFE–bootstrap tracking check (Table 1), per-operation ablations isolating softmax (Table 2), degree and noise sweeps (Tables 3–4), Pareto analysis of NFE vs PPL (Figure 1), and per-layer error traces (Figure 2). The paper is also explicit that SNLP does not remove the need for FHE-friendly block design and that FHE-optimal (N,K) can differ from GPU wall-clock optima. The main contribution is therefore a well-scoped structural hypothesis with reproducible simulation evidence, not a full end-to-end encrypted system.

major comments (3)
  1. §4.1 and Table 1: NFE and the 20 vs 53 bootstrap counts rest on a critical-path multiplicative-depth model (15 usable levels; IDN adds zero mult depth; each SNLP iteration contributes one nonlinear stage for the whole suffix). This is plausible for modulus consumption along a dependency chain, but the manuscript should state more carefully what is and is not counted: total ciphertext multiplications, number of ciphertexts that must be bootstrapped when N suffix states are live, packing/rotation overhead for holding N parallel streams, and whether a practical HE runtime that serializes independent block evaluations would still see bootstrap count track NFE rather than ~(L−N)+K·N. Without that clarification, the headline 2.65× bootstrap reduction can be read as measured FHE cost rather than symbolic critical-path depth.
  2. §4.2 and Limitations: the HE simulation omits comparison circuits for softmax max-subtraction and reciprocal approximation for normalization, and replaces RMSNorm Goldschmidt with controlled residual noise. The relative sequential-vs-SNLP comparison is fair only if these omitted circuits do not change the sequential depth ranking or interact with finite-K SNLP distribution shift. Please either (i) fold a symbolic depth cost for comparison/reciprocal into NFE for both paths and re-report Table 1/5, or (ii) give a short argument/bound that these costs are identical per nonlinear stage and therefore cancel in the NFE ratio, and note any residual risk if SNLP’s parallel suffix changes activation ranges relative to the calibrated Chebyshev intervals.
  3. Table 5 and Figure 1: the strong low-NFE operating points (n24-K4, 2.65×) require SNLP-aware training; baselines are limited to ~1.14× NFE reduction with <1 pp amplification gain. The abstract’s headline numbers are for the IDN-trained 0.5B model. Please make this co-design requirement first-class in the abstract and introduction (one sentence), so readers do not infer that SNLP is a drop-in inference transform for arbitrary pretrained Transformers under FHE.
minor comments (6)
  1. Abstract vs §5: amplification is reported as 1.36× vs 1.42× in the abstract and 1.356× vs 1.419× in Table 5; keep one consistent rounding and state the eval token budget (200k) next to the headline.
  2. Eq. (4): define N and K at first use in the main text before NFE_SNLP; currently N is introduced in prose after the equation in §4.1.
  3. Table 2 caption notes absolute PPL differs from Table 5 due to a smaller eval set; consider reporting the same 200k-token protocol for the ablation or adding a one-row check that relative degradations match at 200k.
  4. Figure 2 uses n24-K1 while the headline config is n24-K4; a short note that K=1 is chosen to isolate single-iteration parallel error (or an extra K=4 panel) would avoid confusion.
  5. Related work: a brief pointer to how NFE relates to multiplicative depth accounting in THE-X / Iron / BOLT / THOR would help FHE readers map the metric to prior circuit-depth tables.
  6. Typos/style: “bootsraps” does not appear, but “FHE-optimal ≠ wallclock-optimal” in §6 is fine; ensure consistent “mHC” vs “mHC-Newton/HCN” naming between §3.2 and Table 5.

Circularity Check

2 steps flagged

Bootstrap counts are definitional from NFE under the symbolic model; FHE error-amplification claims are empirical and non-circular, with only mild self-citation of SNLP.

specific steps
  1. self definitional [§4.1 Eq. (4); Table 1; Abstract]
    "We define NFE(Nonlinear Forward Evaluations) as the number of sequential nonlinear stages in the computation graph: NFE_seq = L, NFE_SNLP = (L−N) + K ... With 15 usable CKKS levels before bootstrapping, the number of bootstraps scales linearly with NFE. In this symbolic model, NFE tracks the bootstrap count within 1–2% across all tested configurations (Table 1)."

    Headline “53 to 20 (2.65×) bootstraps” is NFE reduction (32→12 for n24-K4) times a fixed levels factor. Given the definitions, any SNLP (N,K) with NFE < L reduces “symbolic bootstraps” by construction of the cost model; the number is not an independent FHE measurement.

  2. self citation load bearing [§3.2; Han et al., 2026; §5 cross-model table (m4 IDN Reg.)]
    "Structured Newton Layer Parallelism (Han et al., 2026) treats the hidden-state trace h=(h1,…,hL) as the solution of a nonlinear residual equation… SNLP-aware training adds an auxiliary loss during pretraining that makes the finite-iteration SNLP solve match the sequential trace…"

    Aggressive low-NFE configs (e.g. n24-K4 on m4) and the solver itself rest on the authors’ prior/concurrent SNLP paper and training recipe. This is load-bearing for the method existing, but does not make the measured HE amplification ranking a restatement of that prior work.

full rationale

The paper’s load-bearing scientific claim is comparative error accumulation (amplification = PPL_HE/PPL_exact) under shared Chebyshev approximations for sequential vs SNLP graphs. That ratio is measured, not fitted or defined into existence: both paths use the same per-block polynomials, and the paper isolates HE-induced error from finite-K solver error by using each method’s own exact PPL as denominator. The 2.65× “symbolic bootstraps” headline, by contrast, is almost tautological once NFE ≔ (L−N)+K and bootstraps are stipulated to scale linearly with NFE at 15 usable levels—so choosing n24-K4 yields 12 vs 32 stages and thus 20 vs 53 bootstraps by construction of the cost model, not from encrypted runs. Dependence on Han et al. (2026) for the SNLP solver and SNLP-aware training is real but ordinary method inheritance; it does not force the amplification ranking. No uniqueness theorem, fitted-then-predicted target, or renamed known law appears. Overall circularity is minor and confined to presenting the NFE cost model as bootstrap savings.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The central claim rests on a simulation proxy for FHE, a symbolic multiplicative-depth model, and SNLP surrogates from prior work. Free parameters are mostly engineering choices (degree, N/K, intervals, usable levels) that set the operating point rather than force the qualitative ranking. No new physical entity is postulated; NFE is a derived cost metric. The largest non-standard load is treating the poly-approx simulator plus symbolic bootstrap model as faithful for relative sequential-vs-SNLP conclusions.

free parameters (6)
  • Chebyshev polynomial degree d = 12 (default; sweep {8,10,12,14})
    Default d=12 chosen after degree sweep; controls accuracy–depth tradeoff and headline amplification numbers.
  • SNLP configuration (N, K) = headline n24-K4; others e.g. n8-K4, n20-K4, n24-K8
    Suffix width and solver iterations selected per model; headline uses n24-K4 on 0.5B IDN model.
  • Usable CKKS levels before bootstrap = 15
    Symbolic cost model assumes 15 usable levels; bootstrap counts scale linearly with NFE under this choice.
  • Softmax Chebyshev fitting interval = [-20, 0]
    Interval [-20,0] after max-subtraction, calibrated from validation percentiles; clamping affects approximation error.
  • RMSNorm residual noise scale = 10^{-(n+1)} for n iterations
    Goldschmidt not fully implemented; residual error modeled as multiplicative noise 10^{-(n+1)}.
  • SNLP-aware training regularization strength λ = λ = 1/16–0.5 (reported range)
    IDN/HCN regularization enables aggressive low-NFE configs; baseline models cannot use them well.
axioms (5)
  • domain assumption NFE = (L−N)+K is a faithful proxy for sequential nonlinear FHE stages and bootstrap count when IDN corrections add zero multiplicative depth.
    Introduced in §4.1 and validated only inside the authors’ symbolic CKKS model (Table 1, ratio 0.99–1.02×).
  • domain assumption Shared per-block polynomial approximations make relative sequential-vs-SNLP error accumulation comparable even if absolute FHE costs omit comparison/reciprocal circuits.
    Stated in §4.2 and Limitations; underpins the claim that simulation gaps do not flip the ranking.
  • domain assumption Identity/HC structured Newton surrogates sufficiently propagate depth corrections for finite K after SNLP-aware training.
    Imported from Han et al. 2026 and used throughout experiments; without it low-NFE quality collapses (Figure 1 baseline vs IDN).
  • standard math CKKS approximate arithmetic and bootstrap cost structure as used in prior FHE-ML (Cheon et al.; Agrawal et al.).
    Background §3.1; standard in the subfield though parameter choices are paper-specific.
  • ad hoc to paper Degree-d Chebyshev polynomials on calibrated intervals adequately stand in for encrypted nonlinear ops for error-accumulation study.
    Method §4.2 simulation design; not a full encrypted circuit equivalence.
invented entities (1)
  • NFE (Nonlinear Forward Evaluations) no independent evidence
    purpose: Scalar proxy for sequential nonlinear stages and bootstrap burden under SNLP vs sequential graphs.
    Defined as NFE_seq=L and NFE_SNLP=(L−N)+K; useful metric but not independently measured outside this cost model.

pith-pipeline@v1.1.0-grok45 · 15728 in / 3795 out tokens · 32299 ms · 2026-07-14T16:18:47.181084+00:00 · methodology

0 comments
read the original abstract

Fully homomorphic encryption (FHE) enables computation on encrypted data, but practical encrypted Transformer inference is bottlenecked by the sequential composition of many nonlinear blocks. We study whether Structured Newton Layer Parallelism (SNLP) can make this inter-layer composition more FHE-friendly: each Transformer block still requires polynomial approximations for operations such as softmax and RMSNorm, but SNLP reduces the layerwise sequential nonlinear depth from L stages to a small number of solver iterations plus linear structured corrections. Using a simulation framework based on Chebyshev polynomial approximations, we measure error accumulation under sequential versus SNLP inference across 8 models and 4 architecture families. On a 0.5B IDN-trained model, SNLP reduces symbolic bootstraps from 53 to 20 (2.65x) with only +1.2% perplexity degradation, while lowering error amplification (1.36x vs. 1.42x). Across all tested models, SNLP has lower amplification than sequential inference. Ablations show that softmax approximation dominates the error budget and CKKS arithmetic noise is negligible in our setting, suggesting that SNLP is complementary to block-level FHE-friendly operator design rather than a replacement for it.

Figures

Figures reproduced from arXiv: 2607.04819 by Akash Srivastava, Han Gao, Hao Wang, Kai Xu, Ligong Han, Ruijiang Gao.

Figure 1
Figure 1. Figure 1: NFE vs. PPL under degree-12 HE approximation. Each point is one [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Per-layer relative error ∥hHE − hexact∥/∥hexact∥ under degree-12 polynomial ap￾proximation (0.5B IDN model, N = 24, K = 1). SNLP uses 3.5× fewer bootstraps (15 vs. 53) yet achieves comparable per-layer error to sequential inference. Prefix layers (0–7) are shared. In the parallel layers (8–31, shaded), SNLP error stays at or below sequential despite evaluating all suffix blocks in parallel with only one Ne… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

115 extracted references · 40 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    Simple and effective masked diffusion language models , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    Advances in neural information processing systems , volume=

    Simplified and generalized masked diffusion for discrete data , author=. Advances in neural information processing systems , volume=

  3. [3]

    arXiv preprint arXiv:2503.01103 , year=

    Direct discriminative optimization: Your likelihood-based visual generative model is secretly a gan discriminator , author=. arXiv preprint arXiv:2503.01103 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2507.15897 , year=

    Redi: Rectified discrete flow , author=. arXiv preprint arXiv:2507.15897 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2202.00512 , year=

    Progressive distillation for fast sampling of diffusion models , author=. arXiv preprint arXiv:2202.00512 , year=

    Pith/arXiv arXiv

  6. [6]

    International Conference on Machine Learning , pages=

    Consistency Models , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  7. [7]

    arXiv preprint arXiv:2509.24526 , year=

    Cmt: Mid-training for efficient learning of consistency, mean flow, and flow map models , author=. arXiv preprint arXiv:2509.24526 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2511.23342 , year=

    Flow Straighter and Faster: Efficient One-Step Generative Modeling via MeanFlow on Rectified Trajectories , author=. arXiv preprint arXiv:2511.23342 , year=

    arXiv

  9. [9]

    arXiv preprint arXiv:2505.13447 , year=

    Mean flows for one-step generative modeling , author=. arXiv preprint arXiv:2505.13447 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2406.07507 , volume=

    Flow map matching , author=. arXiv preprint arXiv:2406.07507 , volume=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2506.17298 , year=

    Mercury: Ultra-Fast Language Models Based on Diffusion , author=. arXiv preprint arXiv:2506.17298 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2508.02193 , year=

    Seed diffusion: A large-scale diffusion language model with high-speed inference , author=. arXiv preprint arXiv:2508.02193 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2502.09992 , year=

    Large language diffusion models , author=. arXiv preprint arXiv:2502.09992 , year=

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2509.26488 , year=

    dparallel: Learnable parallel decoding for dllms , author=. arXiv preprint arXiv:2509.26488 , year=

    arXiv

  15. [15]

    arXiv preprint arXiv:2112.07804 , year=

    Tackling the generative learning trilemma with denoising diffusion GANs , author=. arXiv preprint arXiv:2112.07804 , year=

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2510.06303 , year=

    Sdar: A synergistic diffusion-autoregression paradigm for scalable sequence generation , author=. arXiv preprint arXiv:2510.06303 , year=

    arXiv

  17. [17]

    arXiv preprint arXiv:2505.22618 , year=

    Fast-dllm: Training-free acceleration of diffusion llm by enabling kv cache and parallel decoding , author=. arXiv preprint arXiv:2505.22618 , year=

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2310.14189 , year=

    Improved techniques for training consistency models , author=. arXiv preprint arXiv:2310.14189 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2410.21035 , year=

    Beyond autoregression: Fast llms via self-distillation through time , author=. arXiv preprint arXiv:2410.21035 , year=

    Pith/arXiv arXiv

  20. [20]

    ACM computing surveys , volume=

    Diffusion models: A comprehensive survey of methods and applications , author=. ACM computing surveys , volume=. 2023 , publisher=

    2023

  21. [21]

    The Twelfth International Conference on Learning Representations , year=

    Let's verify step by step , author=. The Twelfth International Conference on Learning Representations , year=

  22. [22]

    arXiv preprint arXiv:2412.01152 , year=

    Intellect-1 technical report , author=. arXiv preprint arXiv:2412.01152 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2508.09192 , year=

    Diffusion llms can do faster-than-ar inference via discrete diffusion forcing , author=. arXiv preprint arXiv:2508.09192 , year=

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2103.03874 , year=

    Measuring mathematical problem solving with the math dataset , author=. arXiv preprint arXiv:2103.03874 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2110.14168 , year=

    Training verifiers to solve math word problems , author=. arXiv preprint arXiv:2110.14168 , year=

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2108.07732 , year=

    Program synthesis with large language models , author=. arXiv preprint arXiv:2108.07732 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2107.03374 , year=

    Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2503.09573 , year=

    Block diffusion: Interpolating between autoregressive and diffusion language models , author=. arXiv preprint arXiv:2503.09573 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2410.21357 , year=

    Energy-based diffusion language models for text generation , author=. arXiv preprint arXiv:2410.21357 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2509.06949 , year=

    Revolutionizing reinforcement learning framework for diffusion large language models , author=. arXiv preprint arXiv:2509.06949 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2601.07568 , year=

    d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation , author=. arXiv preprint arXiv:2601.07568 , year=

    arXiv

  32. [32]

    arXiv preprint arXiv:2508.10875 , year=

    A survey on diffusion language models , author=. arXiv preprint arXiv:2508.10875 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2505.15781 , year=

    dkv-cache: The cache for diffusion language models , author=. arXiv preprint arXiv:2505.15781 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2509.26328 , year=

    Fast-dllm v2: Efficient block-diffusion llm , author=. arXiv preprint arXiv:2509.26328 , year=

    arXiv

  35. [35]

    arXiv preprint arXiv:2504.19720 , year=

    Taming the Titans: A Survey of Efficient LLM Inference Serving , author=. arXiv preprint arXiv:2504.19720 , year=

    Pith/arXiv arXiv

  36. [36]

    ACM Computing Surveys , volume=

    Towards efficient generative large language model serving: A survey from algorithms to systems , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    2025

  37. [37]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Llm in a flash: Efficient large language model inference with limited memory , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  38. [38]

    arXiv preprint arXiv:2508.15487 , year=

    Dream 7b: Diffusion large language models , author=. arXiv preprint arXiv:2508.15487 , year=

    Pith/arXiv arXiv

  39. [39]

    The Thirteenth International Conference on Learning Representations , year=

    Simple Guidance Mechanisms for Discrete Diffusion Models , author=. The Thirteenth International Conference on Learning Representations , year=

  40. [40]

    Advances in neural information processing systems , volume=

    Semi-implicit denoising diffusion models (siddms) , author=. Advances in neural information processing systems , volume=

  41. [41]

    arXiv preprint arXiv:2505.21467 , year=

    Accelerating diffusion language model inference via efficient kv caching and guided diffusion , author=. arXiv preprint arXiv:2505.21467 , year=

    arXiv

  42. [42]

    arXiv preprint arXiv:2506.06295 , year=

    dllm-cache: Accelerating diffusion large language models with adaptive caching , author=. arXiv preprint arXiv:2506.06295 , year=

    Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:2511.19269 , year=

    CDLM: Consistency Diffusion Language Models For Faster Sampling , author=. arXiv preprint arXiv:2511.19269 , year=

    arXiv

  44. [44]

    arXiv preprint arXiv:2510.23606 , year=

    Variational Masked Diffusion Models , author=. arXiv preprint arXiv:2510.23606 , year=

    arXiv

  45. [45]

    arXiv preprint arXiv:2511.19473 , year=

    WavefrontDiffusion: Dynamic Decoding Schedule for Improved Reasoning , author=. arXiv preprint arXiv:2511.19473 , year=

    arXiv

  46. [46]

    Transactions of Mathematics and Its Applications , volume=

    Geometric aspects of data-processing of Markov chains , author=. Transactions of Mathematics and Its Applications , volume=. 2024 , publisher=

    2024

  47. [47]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Di [M] o: Distilling masked diffusion models into one-step generator , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  48. [48]

    Communications of the ACM , volume=

    Generative adversarial networks , author=. Communications of the ACM , volume=. 2020 , publisher=

    2020

  49. [49]

    arXiv preprint arXiv:2209.03003 , year=

    Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. arXiv preprint arXiv:2209.03003 , year=

    Pith/arXiv arXiv

  50. [50]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  51. [51]

    arXiv preprint arXiv:2511.19942 , year=

    Differential smoothing mitigates sharpening and improves llm reasoning , author=. arXiv preprint arXiv:2511.19942 , year=

    arXiv

  52. [52]

    arXiv preprint arXiv:2510.08141 , year=

    Arbitrary Entropy Policy Optimization Breaks The Exploration Bottleneck of Reinforcement Learning , author=. arXiv preprint arXiv:2510.08141 , year=

    Pith/arXiv arXiv

  53. [53]

    Entropy-preserving reinforcement learning , author=

  54. [54]

    Thinking Machines Lab: Connectionism , year =

    Kevin Lu and Thinking Machines Lab , title =. Thinking Machines Lab: Connectionism , year =

  55. [55]

    The Twelfth International Conference on Learning Representations , year=

    Parallelizing non-linear sequential models over the sequence length , author=. The Twelfth International Conference on Learning Representations , year=

  56. [56]

    Advances in Neural Information Processing Systems , volume=

    Towards Scalable and Stable Parallelization of Nonlinear RNNs , author=. Advances in Neural Information Processing Systems , volume=

  57. [57]

    Advances in Neural Information Processing Systems , volume=

    Parallelizing MCMC Across the Sequence Length , author=. Advances in Neural Information Processing Systems , volume=

  58. [58]

    2025 , publisher =

    Andrej Karpathy , title =. 2025 , publisher =

    2025

  59. [59]

    arXiv preprint arXiv:2604.21254 , year=

    Hyperloop Transformers , author=. arXiv preprint arXiv:2604.21254 , year=

    Pith/arXiv arXiv

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Deep equilibrium models , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Scaling up test-time compute with latent reasoning: A recurrent depth approach , author=. Advances in Neural Information Processing Systems , volume=

  62. [62]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Value Residual Learning , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  63. [63]

    2024 , url =

    Keller Jordan and Jeremy Bernstein and Brendan Rappazzo and Boza Vlado and You Jiacheng and Franz Cesista and Braden Koszarsky , title =. 2024 , url =

    2024

  64. [64]

    arXiv preprint arXiv:2512.24880 , year=

    mHC: Manifold-Constrained Hyper-Connections , author=. arXiv preprint arXiv:2512.24880 , year=

    Pith/arXiv arXiv

  65. [65]

    The Thirteenth International Conference on Learning Representations , year=

    Hyper-Connections , author=. The Thirteenth International Conference on Learning Representations , year=

  66. [66]

    International Conference on Machine Learning , pages=

    Accelerating Feedforward Computation via Parallel Nonlinear Equation Solving , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  67. [67]

    Synthesis of Parallel Algorithms , pages=

    Prefix sums and their applications , author=. Synthesis of Parallel Algorithms , pages=. 1990 , publisher=

    1990

  68. [68]

    International Conference on Learning Representations , year=

    Parallelizing Linear Recurrent Neural Nets Over Sequence Length , author=. International Conference on Learning Representations , year=

  69. [69]

    First Conference on Language Modeling , year=

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces , author=. First Conference on Language Modeling , year=

  70. [70]

    2026 , archiveprefix =

    Attention Residuals , author =. 2026 , archiveprefix =. 2603.15031 , primaryclass =

    Pith/arXiv arXiv 2026

  71. [71]

    Qwen2.5: A Party of Foundation Models , url =

    Qwen Team , month =. Qwen2.5: A Party of Foundation Models , url =

  72. [72]

    arXiv preprint arXiv:2401.02385 , year=

    TinyLlama: An Open-Source Small Language Model , author=. arXiv preprint arXiv:2401.02385 , year=

    Pith/arXiv arXiv

  73. [73]

    arXiv preprint arXiv:2503.19786 , year=

    Gemma 3 Technical Report , author=. arXiv preprint arXiv:2503.19786 , year=

    Pith/arXiv arXiv

  74. [74]

    arXiv preprint arXiv:2506.20639 , year=

    DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation , author=. arXiv preprint arXiv:2506.20639 , year=

    Pith/arXiv arXiv

  75. [75]

    Advances in Neural Information Processing Systems , volume=

    Parallelizing Linear Transformers with the Delta Rule over Sequence Length , author=. Advances in Neural Information Processing Systems , volume=

  76. [76]

    International Conference on Learning Representations , year=

    Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding , author=. International Conference on Learning Representations , year=

  77. [77]

    Advances in Neural Information Processing Systems , volume=

    Deeppcr: Parallelizing sequential operations in neural networks , author=. Advances in Neural Information Processing Systems , volume=

  78. [78]

    Applied numerical mathematics , volume=

    An estimator for the diagonal of a matrix , author=. Applied numerical mathematics , volume=. 2007 , publisher=

    2007

  79. [79]

    2023 , eprint=

    C-Pack: Packaged Resources To Advance General Chinese Embedding , author=. 2023 , eprint=

    2023

  80. [80]

    A stochastic estimator of the trace of the influence matrix for

    Hutchinson, Michael F , journal=. A stochastic estimator of the trace of the influence matrix for

Showing first 80 references.