Pith. sign in

REVIEW 3 major objections 5 minor 52 references

LaCache removes wasted computation in diffusion LLM inference by caching the intermediate states of unchanged tokens, achieving 1.3x speedup alone and up to 40.2x when combined with other accelerators.

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 · deepseek-v4-flash

2026-08-02 00:56 UTC pith:MD7F23SJ

load-bearing objection LaCache's caching idea is real and the standalone speedup is plausible, but the lossless claim fails on the paper's own pseudocode and the 40.2x headline is unsupported. the 3 major comments →

arxiv 2607.16339 v2 pith:MD7F23SJ submitted 2026-07-16 cs.AI

LaCache: Exact Caching and Precision-Adaptive Inference for Diffusion Large Language Models

Xingru Chen , Zelang Liang , Yongjia Ma , Jiqing Zhan , Shuling Yang , Lian Wen , Kun Zhan This is my paper
classification cs.AI
keywords diffusion large language modelssemi-autoregressive decodinginference accelerationcachingFlashAttentiononline softmaxFP8 quantizationmixed precision
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.

Diffusion-based large language models generate text in semi-autoregressive blocks: within a block, the model denoises the same sequence over many steps, but only the tokens in the current block change while the prefix and masked suffix keep the same token IDs. LaCache exploits this by memoizing the intermediate results of those unchanged tokens—embedding outputs, first-layer pre-attention states after RoPE, and the FlashAttention online-softmax state—so later steps inside the block recompute only the changed tokens. On top of that, it quantizes feed-forward linear layers to FP8 with per-group scales matched to the layer- and step-dependent activation distributions. The paper reports about 1.3x end-to-end speedup from LaCache alone across three model variants, and 3.7–12.3x typical (up to 40.2x on long contexts) when composed with existing step-reduction accelerators, with accuracy effectively unchanged or slightly better.

Core claim

The paper's central claim is that within a semi-autoregressive block, the unchanged tokens' contributions to the forward pass can be cached exactly and reused, so redundant embedding, QKV, RoPE, and first-layer attention work disappears without retraining or output change. The claim rests on three caches: EmbedCache (embedding outputs), RoPECache (first-layer token-wise pre-attention states), and FACache (the running max, normalizer, and unnormalized output accumulator of FlashAttention for query and KV tiles fully outside the current block). Combined with a per-group FP8 quantization of feed-forward layers, LaCache reports 1.3x end-to-end speedup alone and up to 40.2x when layered on existi

What carries the argument

FACache is the load-bearing piece: it snapshots FlashAttention's online-softmax state (row max m, normalizer ℓ, and output accumulator O-tilde) for query blocks that do not overlap the current generation block, and in later steps initializes attention from the snapshot, skipping every KV tile that also lies outside the block. EmbedCache and RoPECache handle the token-wise operators that precede attention; the per-group FP8 quantization (1x128 activation groups, 128x128 weight groups) fused with AddNorm and SwiGLU addresses memory-bandwidth bottlenecks in feed-forward layers.

Load-bearing premise

FlashAttention's online max-and-normalizer state, once cached, can be resumed as if the skipped attention tiles had been processed, with bit-identical results; the paper gives no floating-point error analysis of this resume.

What would settle it

Run a SAR block twice with identical inputs: once with FACache active and once with attention fully recomputed, under the same FP16/BF16 kernels, and compare the output logits bitwise across all tokens and steps. Any nonzero difference for the unchanged-region outputs would refute the 'exact/lossless' characterization, while still permitting approximate speedup.

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

If this is right

  • LaCache removes roughly 10–43% of end-to-end latency on the tested benchmarks (about 3% from caching alone, the rest from FP8).
  • Because it skips recomputation rather than denoising steps, LaCache composes with sampling-level accelerators; combined results reach 3.7–12.3x typical and up to 40.2x on long contexts.
  • Accuracy stays within noise, and the paper reports slightly higher strict-match scores on some reasoning tasks, attributed to longer generated answers under the FP8 path.
  • The scheme is training-free and applies to existing diffusion LLMs without modifying weights, so it can be dropped into current deployments.
  • FACache alone adds a further ~6% speedup on top of an existing sampling accelerator in the paper's large-batch test, indicating the attention cache is useful even when other optimizations shrink the number of steps.

Where Pith is reading between the lines

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

  • The same resume-from-cached-online-softmax idea could apply to any setting where large chunks of the KV stream are invariant across calls: prefix caching in autoregressive LLMs, multi-turn dialogue, or sliding-window attention with stable history.
  • The cache's benefit grows with the ratio of unchanged to changed tokens; short-prompt benchmarks show small gains, so the method pays off most on longer contexts and large batches.
  • The per-group FP8 group size could be chosen adaptively per step by observing activation statistics, trading accuracy for speed more finely than the fixed 1x128 scheme.
  • The appendix's accuracy-speedup curve for caching deeper layers suggests a tunable knob: update the cache every k steps. Automating k from token-stability signals could turn lossless caching into a controllable approximate accelerator.

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 / 5 minor

Summary. LaCache targets semi-autoregressive (SAR) decoding of diffusion LLMs. It caches (i) embedding outputs, (ii) first-layer token-wise pre-attention states after RoPE, and (iii) the online-softmax state of first-layer FlashAttention for tokens outside the current generation block, and it additionally applies per-group FP8 quantization to non-first-layer FFN linears. The paper claims training-free, lossless state memoization and reports roughly 1.3× standalone end-to-end speedups, up to 40.2× when composed with Parallel/DPad, with comparable accuracy on GSM8K, MATH, HumanEval, and MBPP across LLaDA-base, LLaDA-instruct, and LLaDA-1.5.

Significance. The idea of exploiting the fact that only a block changes during SAR denoising is sound and orthogonal to sampling-level accelerators. The cache definitions are parameter-free (quantization scales are computed online from activations, Eq. 8), and the empirical speedups, though modest for caching alone (~3% in Table 3), are consistent across models. If the exactness of FACache can be established, the paper would make a useful engineering contribution. The paper also honestly lists limitations (first-layer-only lossless caching, approximate multi-layer extension, FP8 hardware requirement). However, the current manuscript contains a load-bearing inconsistency in the FACache algorithm that must be resolved before the central exactness claim is credible.

major comments (3)
  1. [§3.3.3 and Algorithm 1, lines 9–11, 22–25] The FACache update in Algorithm 1 does not maintain a cache-local online-softmax state. Lines 22–25 update ℓ_cache and O_cache using the global running maximum m_i from the full inner loop, and RowMax_cache is initialized to 0 (line 9) and never updated. Thus, when KV blocks that overlap the current block are interleaved with U-only KV blocks, the cached (m, ℓ, O) is not the online-softmax state over U. Example with scores [0, 10, 5] and only the middle block overlapping C: the correct future weight for the third block is exp(5−10)=exp(−5), while the algorithm produces exp(−15) after reuse (or a different incorrect weighting if RowMax_cache is set to the global max). This is an exact-arithmetic error, not a rounding issue, and it invalidates the 'lossless' claim for FACache as presented. The text in §3.3.3 describes a separate cache state, but the pseudocode contradicts it. Please correc
  2. [Abstract and §5] The claimed 'up to 40.2× end-to-end speedup' is not supported by any table. The fastest ratio in Tables 4, 6, and 7 is roughly 10× over vanilla (e.g., MBPP: 24.07/2.33 = 10.3×). Table 9 reports FACache-only improvements of about 6%. No experiment in the paper, including the appendix, provides a 40.2× result. Please either add the measurement (hardware, model, benchmark, and composition) or remove/qualify the claim in the abstract and conclusion.
  3. [§3.1, §3.4, and §5] The 'lossless' characterization of the full framework is overstated. The FP8 quantization in Eqs. (7)–(8) changes activations and therefore outputs; Fig. 7 and the accuracy differences in Tables 1/4/5 confirm small output changes. Lossless caching is a property of the caching components only (and, as the Limitations section notes, only for the first layer). The conclusion's phrase 'consistently achieves lossless acceleration' should be revised to state that the cache components are output-preserving in exact arithmetic while the mixed-precision part is approximate.
minor comments (5)
  1. [Algorithm 1] Several typos and notation inconsistencies: line 14 reads 'GenStep=F lase'; ℓ_cache and RowMax_cache are sized Br×4d while the running maximum m is Br-dimensional; line 9 initializes RowMax_cache to 0, which is inconsistent with the m=(-∞) initialization in line 7. Please clean up the pseudocode and align shapes.
  2. [Table 4 / Table 3] Some latency values are duplicated across Tables 3 and 4. In Table 4, HumanEval +DPad+LaCache (11.32 s) is slightly slower than +DPad alone (11.29 s), which contradicts the claimed monotonic speedup; this needs an explanation or repeated measurement.
  3. [§4.1] The hardware is only described as 'an designed GPU' and 'another designed GPU'. Please provide exact GPU model, library/kernel versions, measurement counts, and warm-up details for reproducibility.
  4. [Appendix C] The multi-layer cache update interval is a manually chosen hyperparameter that controls a speed/accuracy trade-off. Please report how it was selected and include sensitivity results.
  5. [References] Several references are incomplete ('and 1 others'), and at least one citation is future-dated ('Ma et al., 2026'). Please clean up the bibliography.

Circularity Check

0 steps flagged

No significant circularity: the caching claims are exact-by-construction memoization, FP8 scales are activation-derived, and self-citations are not load-bearing.

full rationale

The paper's central derivation is a memoization/equivalence argument: because token IDs in U are unchanged within a SAR block, deterministic token-wise operators (embedding, norm, QKV, RoPE) produce identical outputs, and these outputs can be cached and reused. This is a direct application of the definitions, not a fitted or predicted quantity. The per-group FP8 quantization scales are computed online from activations (Eq. 8), so no parameter is fit to the reported speedups or accuracies. The speedup numbers are measurements, not derivations from the cache equations. The self-citations (Ma et al., Yang et al., Feng et al., etc.) are related-work or background references and are not load-bearing: no uniqueness theorem, ansatz, or external mathematical claim is imported from the authors' prior work to force the conclusion. One non-circular correctness concern should be noted: §3.3.3 describes FACache as caching the online-softmax state restricted to U, but Algorithm 1 (lines 22–25) updates the cache using the global running max m_i^{(j)}, which has already incorporated C-overlapping KV blocks; the cached state is therefore not the U-only softmax state in general. This is an omitted proof / implementation inconsistency, not a circular reduction of the result to its inputs, so it does not raise the circularity score. The overall claim remains independently testable against vanilla DLLM inference and against external benchmarks, and the paper is not self-referential in a load-bearing way.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The framework is an engineering method; its main dependencies are the SAR invariant, first-layer token-wise operator independence, associativity of the FlashAttention online-softmax state, and an empirical accuracy-preservation assumption for per-group FP8. No physical entities are introduced; the hand-chosen items are quantization group sizes, clip bounds, and an optional multi-layer cache interval.

free parameters (4)
  • FP8 activation quantization group size = 1×128
    Hand-chosen in §3.4; not swept, though Table 1 only compares per-tensor/per-token granularity.
  • FP8 weight quantization group size = 128×128
    Hand-chosen in §3.4; affects accuracy and kernel efficiency but no sensitivity analysis is provided.
  • FP8 format / clipping bounds (Qmin, Qmax) = unspecified
    Equations 7–8 depend on Qmin/Qmax, but the paper never states the FP8 variant (e.g. E4M3/E5M2) or the numeric clip range.
  • Multi-layer cache update interval (Appendix C) = empirically identified, value not reported
    Appendix C shows a speed-accuracy curve and says a balancing configuration was found empirically; not part of the main first-layer method but indicates an additional hand selection if deeper caching is used.
axioms (4)
  • domain assumption Token IDs outside the current SAR block are unchanged across denoising steps.
    Defines U in §3.3.1; true by construction of SAR decoding, and all caches depend on it.
  • standard math First-layer token-wise operators (embedding, norm, QKV projection, RoPE) are per-token and yield identical outputs for identical token IDs.
    Justifies EmbedCache/RoPECache in §3.3.2; standard transformer property.
  • domain assumption FlashAttention online-softmax state (m, ℓ, O) can be split and resumed at arbitrary KV-block boundaries without changing the final output.
    Used in §3.3.3 and Algorithm 1; exact in real arithmetic but floating-point reassociation may break bit-exactness, and the paper does not verify it.
  • domain assumption Per-group FP8 quantization of FFN layers preserves task accuracy to a comparable level.
    Empirically assumed in §3.4; supported only by single-run benchmark numbers without error bars.

pith-pipeline@v1.3.0-alltime-deepseek · 17386 in / 18157 out tokens · 142043 ms · 2026-08-02T00:56:31.509530+00:00 · methodology

0 comments
read the original abstract

Diffusion-based Large Language Models(DLLMs) enable parallel generation via Semi-Autoregressive (SAR) decoding in text generation. However, current methods suffer from severe operator-level redundancy: they recompute the entire sequence during denoising steps, ignoring that the prefix and masked suffix remain invariant within a block. We propose LaCache, a training-free acceleration framework that alleviates this redundancy through lossless caching and mixed precision. Specifically, LaCache employs Lossless State Memoization (LSM) by caching three types of intermediate results: (i) EmbedCache for embedding outputs, (ii) RoPECache for token-wise pre-attention states, and (iii) FACache for the online softmax statistics within FlashAttention. These caches allow the model to skip redundant computation on unchanged tokens without altering the output. To further alleviate memory-bandwidth bottlenecks, LaCache inegrates a per-group FP8 quantization strategy for FFN layers, tailored to step-dependent activation distributions across the diffusion process. Experiments demonstrate that LaCache alone achieves approximately 1.3X end-to-end speedup over vanilla DLLM. When combined with existing acceleration methods, LaCache reaches up to 40.2X end-to-end speedup while maintaining comparable task accuracy.

Figures

Figures reproduced from arXiv: 2607.16339 by Jiqing Zhan, Kun Zhan, Lian Wen, Shuling Yang, Xingru Chen, Yongjia Ma, Zelang Liang.

Figure 1
Figure 1. Figure 1: Left:DLLMs split total generation tokens into sequential blocks and do parallel decoding across multiple steps inside each block. During one block’s generation, the tokens outside of the block stay un￾changed, which results in repetition calculation. Right: the latency of the modules in LLaDA-Instruct is mainly present at the transformer layers and the final sampler. With our method LaCache, the latency of… view at source ↗
Figure 2
Figure 2. Figure 2: The whole procedure of one inference step. After updating the cache in the first step, in the subsequent [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The update and resuse scheme of our FACache. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The visualization of activations before the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The average inference latency of LaCache and previous methods on LLaDA-instruct model among [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of token generation speed in long [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The response examples of the backbone and LaCache(Ours) generated on different datasets. A. LaCache’s [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The accuracy-speedup curve of caching strat [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The response examples of the DPad backbone and LaCache(Ours) generated with the prompt in MBPP [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

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

52 extracted references · 1 canonical work pages

  1. [1]

    arXiv preprint arXiv:2502.09992 , year=

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

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2505.19223 , year=

    LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models , author=. arXiv preprint arXiv:2505.19223 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2508.15487 , year=

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

    Pith/arXiv arXiv

  4. [4]

    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

  5. [5]

    arXiv preprint arXiv:2508.14148 , year=

    DPad: Efficient diffusion language models with suffix dropout , author=. arXiv preprint arXiv:2508.14148 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2601.04205 , year=

    STDD: Spatio-Temporal Dynamics-Driven Token Refinement in Diffusion Language Models , author=. arXiv preprint arXiv:2601.04205 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2511.21759 , year=

    Orchestrating Dual-Boundaries: An Arithmetic Intensity Inspired Acceleration Framework for Diffusion Language Models , author=. arXiv preprint arXiv:2511.21759 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2307.08691 , year=

    Flashattention-2: Faster attention with better parallelism and work partitioning , author=. arXiv preprint arXiv:2307.08691 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2412.19437 , year=

    Deepseek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , year=

    Pith/arXiv arXiv

  10. [10]

    International conference on machine learning , pages=

    Smoothquant: Accurate and efficient post-training quantization for large language models , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  11. [11]

    arXiv preprint arXiv:2410.19313 , year=

    Coat: Compressing optimizer states and activation for memory-efficient fp8 training , author=. arXiv preprint arXiv:2410.19313 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2511.05811 , year=

    MOSS: Efficient and Accurate FP8 LLM Training with Microscaling and Automatic Scaling , author=. arXiv preprint arXiv:2511.05811 , year=

    arXiv

  13. [13]

    arXiv preprint arXiv:2110.14168 , year=

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

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2107.03374 , year=

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

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2103.03874 , year=

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

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2108.07732 , year=

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

    Pith/arXiv arXiv

  17. [17]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  18. [18]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  19. [19]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  20. [20]

    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

  21. [21]

    arXiv preprint arXiv:2506.17298 , volume=

    Mercury: Ultra-fast language models based on diffusion , author=. arXiv preprint arXiv:2506.17298 , volume=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2512.14080 , year=

    SonicMoE: Accelerating MoE with IO and Tile-aware Optimizations , author=. arXiv preprint arXiv:2512.14080 , year=

    arXiv

  23. [23]

    arXiv preprint arXiv:1905.12322 , year=

    A study of BFLOAT16 for deep learning training , author=. arXiv preprint arXiv:1905.12322 , year=

    Pith/arXiv arXiv 1905

  24. [24]

    arXiv preprint arXiv:1710.03740 , year=

    Mixed precision training , author=. arXiv preprint arXiv:1710.03740 , year=

    Pith/arXiv arXiv

  25. [25]

    Tech Rep , year=

    Advances in neural information processing systems 22 , author=. Tech Rep , year=

  26. [26]

    IEEE/ASME Transactions on Mechatronics , year=

    Aligning cyber space with physical world: A comprehensive survey on embodied ai , author=. IEEE/ASME Transactions on Mechatronics , year=

  27. [27]

    arXiv preprint arXiv:2509.23094 , year=

    d ^2 Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching , author=. arXiv preprint arXiv:2509.23094 , year=

    arXiv

  28. [28]

    arXiv preprint arXiv:2506.10848 , year=

    Accelerating Diffusion Large Language Models with SlowFast: The Three Golden Principles , author=. arXiv preprint arXiv:2506.10848 , year=

    arXiv

  29. [29]

    Proceedings of Machine Learning and Systems , volume=

    Xgrammar: Flexible and efficient structured generation engine for large language models , author=. Proceedings of Machine Learning and Systems , volume=

  30. [30]

    arXiv preprint arXiv:2506.19037 , year=

    Plan for Speed--Dilated Scheduling for Masked Diffusion Language Models , author=. arXiv preprint arXiv:2506.19037 , year=

    Pith/arXiv arXiv

  31. [31]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  32. [32]

    Discrete diffusion language modeling by estimating the ratios of the data distribution , author=

  33. [33]

    Advances in neural information processing systems , volume=

    Advances in neural information processing systems , author=. Advances in neural information processing systems , volume=

  34. [34]

    arXiv preprint arXiv:2112.10741 , year=

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models , author=. arXiv preprint arXiv:2112.10741 , year=

    Pith/arXiv arXiv

  35. [35]

    Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Diffusion-nat: Self-prompting discrete diffusion for non-autoregressive text generation , author=. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  36. [36]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv e-prints , pages=

    The llama 3 herd of models , author=. arXiv e-prints , pages=

  38. [38]

    ACM Transactions on Software Engineering and Methodology , year=

    A survey on large language models for code generation , author=. ACM Transactions on Software Engineering and Methodology , year=

  39. [39]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    LogicAsker: Evaluating and improving the logical reasoning ability of large language models , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  40. [40]

    arXiv preprint arXiv:2401.04883 , year=

    Multi-user chat assistant (MUCA): a framework using LLMS to facilitate group conversations , author=. arXiv preprint arXiv:2401.04883 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2410.18514 , year=

    Scaling up masked diffusion models on text , author=. arXiv preprint arXiv:2410.18514 , year=

    Pith/arXiv 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:2503.16522 , year=

    Adams Bashforth Moulton Solver for Inversion and Editing in Rectified Flow , author=. arXiv preprint arXiv:2503.16522 , year=

    arXiv

  44. [44]

    arXiv preprint arXiv:2508.06511 , year=

    DiTalker: A Unified DiT-based Framework for High-Quality and Speaking Styles Controllable Portrait Animation , author=. arXiv preprint arXiv:2508.06511 , year=

    Pith/arXiv arXiv

  45. [45]

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

    Qr-lora: Efficient and disentangled fine-tuning via qr decomposition for customized generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  46. [46]

    ACM Trans

    Ma, Yongjia and Chen, Junlin and Di, Donglin and Xie, Qi and Fan, Lei and Chen, Wei and Zhao, Na and Yang, Xun , title =. ACM Trans. Multimedia Comput. Commun. Appl. , month = jun, articleno =. 2026 , issue_date =. doi:10.1145/3794855 , abstract =

    work page doi:10.1145/3794855 2026

  47. [47]

    Advances in Neural Information Processing Systems , volume=

    Duquant: Distributing outliers via dual transformation makes stronger quantized llms , author=. Advances in Neural Information Processing Systems , volume=

  48. [48]

    RD-NERF: Neural Robust Distilled Feature Fields for Sparse-View Scene Segmentation , year=

    Ma, Yongjia and Dou, Bin and Zhang, Tianyu and Yuan, Zejian , booktitle=. RD-NERF: Neural Robust Distilled Feature Fields for Sparse-View Scene Segmentation , year=

  49. [49]

    2024 , eprint=

    TrAME: Trajectory-Anchored Multi-View Editing for Text-Guided 3D Gaussian Splatting Manipulation , author=. 2024 , eprint=

    2024

  50. [50]

    Proceedings of the 7th ACM International Conference on Multimedia in Asia , articleno =

    Xie, Qi and Ma, Yongjia and Di, Donglin and Gao, Xuehao and Yang, Xun , title =. Proceedings of the 7th ACM International Conference on Multimedia in Asia , articleno =. 2025 , isbn =. doi:10.1145/3743093.3770948 , abstract =

    arXiv 2025

  51. [51]

    TV-3DG: Mastering Text-to-3D Customized Generation with Visual Prompt , year=

    Yang, Jiahui and Di, Donglin and Ma, Baorui and Cui, Jianxun and Yang, Xun and Ma, Yongjia and Sun, Wenzhang and Chen, Wei and Xue, Zhou and Wang, Meng and Liu, Yebin , journal=. TV-3DG: Mastering Text-to-3D Customized Generation with Visual Prompt , year=

  52. [52]

    2026 , eprint=

    CogPortrait: Fine-Grained Eye-Region Control in Portrait Animation via Hierarchical Agent Planning , author=. 2026 , eprint=

    2026