Lossless Tensor Compression as Program Synthesis

Chengran Yang; David Lo; Jieke Shi; Junda He; Peixin Zhang; Shidong Pan; Thong Hoang (James); Weifeng Sun; Wenjia Jiang; Xiwei Xu (Sherry)

REVIEW 2 major objections 3 minor 61 references

Lossless Tensor Compression as Program Synthesis

T0 review · 2 major / 3 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Brevis represents each tensor as a tiny reversible program, cutting 2.13 TB of checkpoints to 1.41 TB without losing a bit.

desk verdict Brevis is a solid engineering contribution framing checkpoint compression as DSL synthesis; the gzip baseline is under-configured and the synthesis search adds less than the DSL itself, but the bit-exactness proof and evaluation are clean. read the letter →

arxiv 2608.02162 v1 pith:XG2I4UMO submitted 2026-08-03 cs.SE cs.AIcs.PL

Jieke Shi , Junda He , Wenjia Jiang , Weifeng Sun , Shidong Pan , Zhensu Sun , Chengran Yang , Peixin Zhang

show 6 more authors

Yifan Jia Zhou Yang Thong Hoang (James) Xiwei Xu (Sherry) Zhenchang Xing David Lo

This is my paper · ORCID

classification cs.SEcs.AIcs.PL

keywords losslesscompressionprogramsynthesismodelcheckpointstensordomain-specificlanguageA*searchbit-exactreconstructionfloating-pointfields

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Brevis is a lossless compressor for model checkpoints that treats compression as program synthesis. For each tensor it searches a small domain-specific language for a reversible program whose execution regenerates the tensor's bits exactly, and stores that program as the archive. Because every tensor can get its own program, the method adapts to structures general-purpose compressors miss, such as repeated regions, constant fields, and correlations between adjacent values. On ten public checkpoints totaling 2.13 TB, Brevis produces 1.41 TB of archives, a 33.93% storage reduction, while staying on the throughput Pareto frontier. The significance is that checkpoint archival and transfer become cheaper without sacrificing exact reproducibility.

What carries the argument

The central object is a typed DSL whose seven reversible operators (Lit, Const, Concat, Repeat, Map, Scan, Merge) have inverse rules that decompose a target word stream into child streams, making correctness local and checkable rather than verified by repeated execution. A bounded A* search, guided by a checkpoint-specific production prior learned from a small tensor sample, explores the grammar; candidates are ranked by exact serialized byte size. The decoder validates the typed program and executes it without any search, so decompression is just program evaluation.

What would settle it

Construct a synthetic checkpoint whose tensors have row-wise repeated values separated by strides (e.g., a 2D array where every other row is identical) and run Brevis; if its archive is no smaller than zstd's, the flat-stream assumption fails. Alternatively, shuffle the elements of a real checkpoint while preserving its histogram and show Brevis's saving drops to the level of generic compressors.

Watch

Extended reading notes

Core claim

The paper's central claim is that bit-exact tensor compression is better solved by synthesizing a compact program per tensor than by picking a fixed codec. The DSL's seven reversible operators — literal storage, constants, concatenation, repetition, bijective maps, scans, and field/plane merges — guarantee that any well-formed program reconstructs its target exactly. The synthesis is target-directed: each search hole carries the exact stream it must produce, and an operator is admitted only if its inverse decomposition recomposes that stream. The objective is the complete serialized byte length of the program, not its structure, and a bounded A* search guided by a checkpoint-specific product

Load-bearing premise

The approach assumes that exploitable redundancy in a checkpoint appears as contiguous patterns in the flat, checkpoint-order word stream; if a tensor's structure is only visible through its 2D/ND layout or after regrouping elements, Brevis will miss it and its advantage over generic compressors shrinks.

Editorial extensions

If this is right

If Brevis's results generalize, checkpoint archives can shrink by roughly a third with zero information loss, directly cutting storage and transfer costs for model hubs and deployment clusters.
Because decompression is plain program execution, decoding runs at 6.61 GB/s, making the method practical for deployment pipelines, not just cold storage.
The archive preserves the original safetensors layout, so existing loading code and tooling can be adapted without changing on-disk semantics.
The synthesis objective — exact serialized size — gives a principled way to compare candidate programs, and the literal fallback guarantees every tensor remains representable even when no structure is found.
A checkpoint-specific prior lets the search adapt to the dtype and statistics of each model without being retrained on every tensor.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the authors leave implicit is cross-tensor synthesis: learning programs shared across tensors could capture redundancy that per-tensor search cannot see, since many weights repeat across layers and shards.
If the DSL gained dimension-aware operators (e.g., transposes or strided repeats), the method could exploit 2D/ND structure such as attention-head layouts; a testable prediction is that on such structured tensors a strided-repeat operator would yield further gains.
The production prior could plausibly become adaptive or universal, removing the calibration pass; this is an engineering hypothesis not yet demonstrated.
Because the method is agnostic to the byte stream's semantic meaning, it may transfer to other structured binary data beyond model weights, such as scientific array archives, though the evaluation only covers checkpoints.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

What you should know: this paper's real contribution is the typed reversible DSL plus target-directed bounded A* for lossless tensor compression, and the evaluation is large and carefully done. The main soft spot is the gzip baseline: it runs libdeflate at DEFLATE level 1, which is not the standard gzip level and likely understates gzip's compression. The stress-test note is right. The abstract says "up to 30.87% smaller than four general-purpose compressors, including zstd and gzip," but that maximum is over Snappy, and the average margin over gzip could shrink substantially with a fair baseline. This doesn't affect the absolute 33.93% reduction or the comparisons against ZipNN and DFloat11, but it does affect the claimed margins over general-purpose compressors. The fix is a rerun at gzip's default level.

What is actually new: formulating per-tensor lossless compression as program synthesis over a typed reversible DSL, where each tensor becomes a self-contained program that reconstructs it bit-exactly. The DSL operators (merge/field-split, repeat, map/scan) are sensible and cover real structure in floating-point tensors. The target-directed expansion is a nice trick—it only decomposes a stream when the decomposition recomposes it exactly, so correctness is preserved by construction. Proposition 1 is a clean inductive proof under explicit assumptions. The evaluation is unusually thorough: all 60 archives were decoded and verified, the statistical tests are Holm-corrected, and the ablation is honest.

The softer spot, beyond gzip, is that the "program synthesis" part contributes less than the DSL itself. The ablation shows removing A* costs about 3 MB on a 4.98 GB shard, and removing the learned prior costs 389 bytes. That suggests the synthesis search is doing modest work relative to the fixed reversible operators. The authors are transparent about this, but the paper should clarify what exactly the search adds over a greedy operator selection. The flat byte-order assumption—tensors are flattened in checkpoint order and the DSL has no dimension-aware operators—is a real scope limitation, but it's acknowledged as future work, not a hidden flaw.

Bottom line: this is a well-engineered, reproducible result for checkpoint archival and transfer. The bit-exactness guarantee and the scale of the evaluation make it worth a serious referee. Deserves peer review, with the gzip baseline rerun and a tightened claim about the synthesis search's contribution.

Referee Report

2 major / 3 minor

Summary. The paper introduces Brevis, a lossless tensor compressor that represents each tensor as a program in a small reversible DSL. The program is synthesized by a target-guided bounded A* search, using a checkpoint-specific production prior learned from a calibration sample and scoring candidates by complete serialized size. The authors prove compositional bit-exactness (Prop. 1) and evaluate on 10 public checkpoints (2.13 TB), reporting a 33.93% total storage reduction, smaller archives than ZipNN and DFloat11, and bit-exact decoding of all 60 archives. The contribution is framed as a new way to exploit tensor-specific structure without fixed codecs.

Significance. Brevis is a fresh and clean formulation: compression as program synthesis, with a rigorous correctness argument and a reproducible artifact (all archives decoded). The evaluation is extensive in volume (2.13 TB) and the results over ZipNN, though modest (≈0.7%), are consistent across all 10 checkpoints. However, the headline comparison against general-purpose compressors is weakened by the non-standard gzip baseline (DEFLATE level 1) and by imprecise wording in the abstract, so the quantitative claims need revision before the strengths can be fully credited.

major comments (2)

[§4.1, Table 1 footnote, Figure 4] gzip is run with libdeflate 1.19 in gzip mode at DEFLATE level 1, while zstd and LZ4 are run at level 9. This is not the standard gzip configuration (GNU gzip defaults to level 6) and understates gzip's compression. The reported mean saving of 13.61% over gzip (Fig. 4) is therefore not a reliable measure of superiority over standard gzip, and the abstract's claim 'including zstd and gzip' inherits this problem. Please rerun gzip at the default or a matched level (e.g., level 6) and update all affected numbers, or provide a clear rationale for level 1 as the appropriate point on the speed/ratio tradeoff.
[Abstract and §1] The 'up to 30.87%' figure is the mean per-checkpoint saving over Snappy, not the maximum over the four general-purpose compressors. The maximum saving vs Snappy is about 34.28% (Llama-3.1-70B), while savings vs zstd and gzip are 12.94% and 13.61%, respectively. The phrasing 'archives up to 30.87% smaller than those of four general-purpose compressors, including zstd and gzip' is misleading because it suggests the 30.87% applies to zstd/gzip as well. Please report per-baseline figures and use 'on average' for the Snappy comparison.

minor comments (3)

[§3.1, Eq. (3), §5] The DSL flattens every tensor into a word stream in checkpoint order and has no dimension-aware or reordering operators. Thus any structure not contiguous in that linear order (e.g., row-wise patterns separated by strides) is invisible. This is a scope limitation worth stating explicitly in the Limitations section, which currently only mentions independent tensor synthesis.
[§4.4, Table 2] The ablation shows that removing the production prior changes archive size by only 389 bytes on the Llama-3.1-8B shard, and the prior's contribution is only visible in combination with A*. The main text should discuss this quantitatively instead of describing the prior as a key component without qualification.
[§4.1] The gzip level is disclosed only in the Table 1 footnote; please move this configuration detail to the main text of §4.1 so readers see the asymmetry immediately.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline compression result is an externally measured empirical outcome, and the only fitted component (the production prior) is not stored and affects archive size by only 389 bytes in the ablation.

full rationale

The paper's central claim is an empirical measurement: 2.130 TB of public checkpoints compress to 1.407 TB, with bit-exact reconstruction verified by decoding all 60 archives. No 'prediction' is derived from a fitted parameter. The only fitted component is the checkpoint-specific production prior used to guide bounded A* search. This prior is not stored with the archive, does not enter the serialized size L(P), and the ablation (Table 2) shows removing it changes archive size by only 389 bytes on the Llama-3.1-8B shard, so the 33.93% reduction is not forced by the fit. The target-directed decomposition condition (Eq. 8) is a soundness invariant, not a mechanism for smuggling the target into the result; candidate quality is determined by exact serialized size (Eq. 5), with the literal fallback guaranteeing representability. There is no load-bearing self-citation: the cited PHOG/Euphony prior is standard external work, and no uniqueness theorem from the authors is invoked. The DSL's floating-point field merge resembles prior transforms such as ZipNN/TDT, but the paper evaluates Brevis against those baselines rather than renaming them; this is a novelty or fairness discussion, not circularity. The gzip baseline running at DEFLATE level 1 is a legitimate measurement-fairness concern, but it is not a circularity step and does not affect the internal derivation. Overall, the derivation is self-contained and externally validated.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a handful of design choices: the definition of Bits(X), the reversibility of the DSL operators, the invertibility of literal codecs, and the representativeness of the calibration sample. These are not invalid, but they are assumed rather than derived. The learned production prior is a fitted distribution, though it is not stored and its effect is small (389 bytes on the ablated shard).

free parameters (6)

A* expansion budget B = 1 (main); 32/256 (ablation)
Chosen to trade compression time against archive size; headline results use B=1 (Section 4.4).
Calibration sample size = ≤4 tensors, ≤6 expansions/tensor, ≤1,048,576 elements/tensor
Deterministic sample for learning production prior; hand-picked caps affect the prior but not stored bytes.
Program size limits = 64 nodes, depth 4, 512 MiB
Bound the synthesis search space; influence which programs are reachable within budget.
Additive smoothing beta = unspecified (β>0)
Smoothing for Eq. 9; value not reported, affects prior and thus search order.
Production prior counts N(r,κ) = learned from calibration sample
The checkpoint-specific distribution in Eq. 9 is fitted to a small sample of the same checkpoint; not stored, but guides search.
Concurrency/worker count = 32 workers; AMD EPYC 9654 (192 cores)
Affects throughput; acknowledged as a practical, non-identical configuration.

assumptions (5)

domain assumption Flattening a tensor into Bits(X) in checkpoint order preserves bit-exactness and does not hide exploitable structure.
Section 3.1 Eq. 3; the DSL contains no transpose/reshape, so only order-contiguous patterns are discoverable.
domain assumption Each DSL operator is reversible for its stored parameters, and Eq. 8 validly guarantees parent reconstruction.
Figure 3 well-formedness rules; Proposition 1 relies on this to prove compositional bit-exactness.
domain assumption All physical literal codecs are exactly invertible and serialization/parsing preserve the typed tree.
Proposition 1 assumptions (i)-(ii); verified by decoding all 60 archives, but still an implementation invariant.
standard math The A* heuristic h(s) is admissible and the byte lower bound LB_L(s) is valid in the relaxed grammar.
Section 3.3; needed so no potentially smaller program is pruned incorrectly.
domain assumption The calibration sample is representative of each checkpoint's tensor population.
Section 4.1: at most 4 tensors, deterministic stratified sample; if unrepresentative, the prior can misorder expansions, though ablation shows small impact.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Lossless Tensor Compression as Program Synthesis." pith.science (2026). https://pith.science/paper/XG2I4UMO

@misc{pith2026260802162,
  author       = {Pith},
  title        = {Pith review of: Lossless Tensor Compression as Program Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XG2I4UMO}},
  note         = {Machine review of arXiv:2608.02162}
}

read the original abstract

Model checkpoints are growing in both number and size, which makes archival, transfer, and deployment increasingly costly. General-purpose compressors can reduce storage requirements but ignore tensor structure, whereas existing tensor-specific compressors rely on fixed and format-specific pipelines. We present Brevis, which formulates lossless tensor compression as program synthesis. We design a typed domain-specific language (DSL) that captures recurring tensor structures, such as repeated regions and floating-point fields, through a set of reversible operators. Given a tensor, Brevis synthesizes a self-contained DSL program that reconstructs it bit-exactly. A checkpoint-specific production prior, learned from a small representative sample of tensors, guides a bounded A* search to synthesize compact programs, which can later be executed directly for bit-exact decompression. On 10 public checkpoints spanning language, audio, and image generation models, Brevis reduces 2.13 TB of checkpoint data to 1.41 TB, a 33.93% storage reduction. It produces archives up to 30.87% smaller than those of four general-purpose compressors, including zstd and gzip, and smaller archives than the tensor-specific compressors ZipNN and DFloat11. Under a practical concurrency configuration, Brevis achieves 3.60 GB/s compression and 6.61 GB/s decompression while preserving every source byte.

Figures

Figures reproduced from arXiv: 2608.02162 by the authors.

**Figure 1.** Growth in public model repositories and the compression tradeoff. (a) Hugging Face repositories (log scale). (b) Storage reduction versus compression throughput. et al. 2025]. They are designed for efficient inference and often require specific numerical formats (e.g., GGUF [Gerganov 2026]) or execution environments, making them unsuitable for archival or exact checkpoint transfer because the original weights can… view at source ↗

**Figure 2.** Compression as program synthesis in Brevis. Each tensor is represented by a short reversible program whose execution regenerates its bits exactly. (1) Read as IEEE 754 fields, 6 FP32 words expose 3 different regularities, and no single codec captures all of them. (2) Target-guided 𝐴 ∗ searches the typed grammar and ranks candidates by the complete serialized size 𝐿(𝑃), which includes the literal payload and codec ta… view at source ↗

**Figure 3.** Typed operators in the Brevis language. Lit is the universal fallback; all internal operators are reversible for their stored parameters. Lit stores an arbitrary word stream and provides a universal fallback, whereas Const stores one word and its repetition count. Concat joins programs for adjacent regions, and Repeat expands one nonempty child multiple times. Together, these operators capture arbitrary values, c… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Mean archive saving of Brevis over the five complete-corpus baselines. Lines show 95% bootstrap intervals over checkpoints. by the largest checkpoint. We additionally compare with DFloat11 using its only publicly available and independently validated native result on …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 1 canonical work pages

[1]

2025 , publisher =

Ghorbani, Mahdi and Bauer, Emilien and Grosser, Tobias and Shaikhha, Amir , title =. 2025 , publisher =. doi:10.1145/3720506 , journal =

doi:10.1145/3720506 2025
[2]

Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , year =

Lee, Woosuk and Heo, Kihong and Alur, Rajeev and Naik, Mayur , title =. Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , year =
[3]

Proceedings of the 33rd International Conference on Machine Learning (ICML) , year =

Bielik, Pavol and Raychev, Veselin and Vechev, Martin , title =. Proceedings of the 33rd International Conference on Machine Learning (ICML) , year =
[4]

Syntax-Guided Synthesis , booktitle =

Alur, Rajeev and Bod. Syntax-Guided Synthesis , booktitle =. 2013 , pages =

2013
[5]

Proceedings of the 42nd ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , year =

Ellis, Kevin and Wong, Catherine and Nye, Maxwell and Sabl. Proceedings of the 42nd ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , year =
[6]

Proceedings of the ACM on Programming Languages , volume =

Barke, Shraddha and Peleg, Hila and Polikarpova, Nadia , title =. Proceedings of the ACM on Programming Languages , volume =. 2020 , doi =

2020
[7]

Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =

Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , title =. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =. 2019 , doi =

2019
[8]

, title =

Huffman, David A. , title =. Proceedings of the IRE , volume =

Show all 61 references

[9]

arXiv preprint arXiv:1311.2540 , year =

Duda, Jarek , title =. arXiv preprint arXiv:1311.2540 , year =

arXiv
[10]

IEEE Transactions on Information Theory , volume =

Ziv, Jacob and Lempel, Abraham , title =. IEEE Transactions on Information Theory , volume =
[11]

Peter , title =

Deutsch, L. Peter , title =. 1996 , note =. doi:10.17487/RFC1952 , url =

1996 doi
[12]

2021 , note =

Collet, Yann and Kucherawy, Murray , title =. 2021 , note =. doi:10.17487/RFC8878 , url =

2021 doi
[13]

2023 , url=

Elias Frantar and Saleh Ashkboos and Torsten Hoefler and Dan Alistarh , booktitle=. 2023 , url=

2023
[14]

2024 , doi=

Lin, Ji and Tang, Jiaming and Tang, Haotian and Yang, Shang and Xiao, Guangxuan and Han, Song , journal=. 2024 , doi=

2024
[15]

Jiang, Albert Q. and Sablayrolles, Alexandre and Mensch, Arthur and Bamford, Chris and Chaplot, Devendra Singh and de las Casas, Diego and Bressand, Florian and Lengyel, Gianna and Lample, Guillaume and Saulnier, Lucile and Lavaud, L. Mistral 7B , journal =. 2023 , eprint =

2023
[16]

arXiv preprint arXiv:2407.21783 , year =

Grattafiori, Aaron and others , title =. arXiv preprint arXiv:2407.21783 , year =. 2407.21783 , archivePrefix =

arXiv
[17]

2025 , eprint =

Qwen3 Technical Report , journal =. 2025 , eprint =

2025
[18]

Jiang, Albert Q. and Sablayrolles, Alexandre and Roux, Antoine and Mensch, Arthur and Savary, Blanche and Bamford, Chris and Chaplot, Devendra Singh and de las Casas, Diego and Hanna, Emma Bou and Bressand, Florian and Lengyel, Gianna and Bour, Guillaume and Lample, Guillaume ...

2024
[19]

2602.15763 , archivePrefix =

arXiv preprint arXiv:2602.15763 , year =. 2602.15763 , archivePrefix =

arXiv
[20]

2025 IEEE 18th International Conference on Cloud Computing (CLOUD) , year =

Hershcovitch, Moshik and Wood, Andrew and Choshen, Leshem and Girmonsky, Guy and Leibovitz, Roy and Ozeri, Or and Ennmouri, Ilias and Malka, Michal and Chin, Peter and Sundararaman, Swaminathan and Harnik, Danny , title =. 2025 IEEE 18th International Conference on Cloud Compu...

2025
[21]

23rd USENIX Symposium on Networked Systems Design and Implementation (NSDI 26) , year =

Zirui Wang and Tingfeng Lan and Zhaoyuan Su and Juncheng Yang and Yue Cheng , title =. 23rd USENIX Symposium on Networked Systems Design and Implementation (NSDI 26) , year =
[22]

SC16: International Conference for High Performance Computing, Networking, Storage and Analysis , year =

Burtscher, Martin and Mukka, Hari and Yang, Annie and Hesaaraki, Farbod , title =. SC16: International Conference for High Performance Computing, Networking, Storage and Analysis , year =
[23]

2018 Data Compression Conference (DCC) , year =

Claggett, Steven and Azimi, Sahar and Burtscher, Martin , title =. 2018 Data Compression Conference (DCC) , year =

2018
[24]

IEEE Transactions on Visualization and Computer Graphics , volume =

Lindstrom, Peter and Isenburg, Martin , title =. IEEE Transactions on Visualization and Computer Graphics , volume =. 2006 , doi =

2006
[25]

2024 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) , year =

Rodriguez, Andrew and Azami, Noushin and Burtscher, Martin , title =. 2024 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) , year =

2024
[26]

Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1 , year =

Azami, Noushin and Fallin, Alex and Burtscher, Martin , title =. Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1 , year =
[27]

Advances in Neural Information Processing Systems , volume =

Zhang, Tianyi and Hariri, Mohsen and Zhong, Shaochen (Henry) and Chaudhary, Vipin and Sui, Yang and Hu, Xia and Shrivastava, Anshumali , title =. Advances in Neural Information Processing Systems , volume =. 2025 , eprint =

2025
[28]

Collet, Yann and Terrell, Nick and Handte, W. Felix and Rozenblit, Danielle and Zhang, Victor and Zhang, Kevin and Goldschlag, Yaelle and Lee, Jennifer and Gorokhovsky, Elliot and Komornik, Yonatan and Riegel, Daniel and Angelov, Stan and Rotem, Nadav , title =. arXiv preprint...

doi
[29]

International Conference on Learning Representations (ICLR) , year =

Yoran, Ori and Zheng, Kunhao and Gloeckle, Fabian and Gehring, Jonas and Synnaeve, Gabriel and Cohen, Taco , title =. International Conference on Learning Representations (ICLR) , year =
[30]

arXiv preprint arXiv:2506.18062 , year =

Jamalidinan, Samirasadat and Cheshmi, Kazem , title =. arXiv preprint arXiv:2506.18062 , year =. doi:10.48550/arXiv.2506.18062 , url =. 2506.18062 , archivePrefix =

doi arXiv
[31]

arXiv preprint arXiv:2410.20650 , year =

Hao, Yongchang and Cao, Yanshuai and Mou, Lili , title =. arXiv preprint arXiv:2410.20650 , year =. doi:10.48550/arXiv.2410.20650 , url =. 2410.20650 , archivePrefix =

doi arXiv
[32]

arXiv preprint arXiv:2502.00922 , year =

Yubeaton, Patrick and Mahmoud, Tareq and Naga, Shehab and Taheri, Pooria and Xia, Tianhua and George, Arun and Khalil, Yasmein and Zhang, Sai Qian and Joshi, Siddharth and Hegde, Chinmay and Garg, Siddharth , title =. arXiv preprint arXiv:2502.00922 , year =. doi:10.48550/arXi...

doi arXiv
[33]

International Conference on Learning Representations (ICLR) , year =

Yang, Zeyu and Zhang, Tianyi and Xie, Jianwen and Li, Chuan and Xu, Zhaozhuo and Shrivastava, Anshumali , title =. International Conference on Learning Representations (ICLR) , year =. 2510.02676 , archivePrefix =
[34]

Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS) , pages =

Fan, Ruibo and Yu, Xiangrui and Pan, Xinglin and Li, Zeyu and Luo, Weile and Wang, Qiang and Wang, Wei and Chu, Xiaowen , title =. Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS) ,...

2026
[35]

Proceedings of the 53rd Annual IEEE/ACM International Symposium on Computer Architecture , year =

Yang, Jinwu and Wu, Jiaan and Liu, Zedong and Ma, Xinyang and Zhao, Hairui and Gu, Yida and Huang, Yuanhong and Liu, Xingchen and Huang, Wenjing and Wei, Zheng and Xing, Jing and Ma, Yili and Zhang, Qingyi and An, Baoyi and Hu, Zhongzhe and Liu, Shaoteng and Zhu, Xia and Lu, J...

arXiv
[36]

265: Video Codecs are Secretly Tensor Codecs , author=

LLM. 265: Video Codecs are Secretly Tensor Codecs , author=. Proceedings of the 58th IEEE/ACM International Symposium on Microarchitecture , pages=. 2025 , doi =

2025
[37]

Proceedings of the 53rd Annual International Symposium on Computer Architecture (

Tan, Hongshi and Chen, Yao and Alonso, Gustavo and Wong, Weng-Fai and He, Bingsheng , title =. Proceedings of the 53rd Annual International Symposium on Computer Architecture (. 2026 , eprint =

2026
[38]

Proceedings of the 43rd International Conference on Machine Learning (

Yang, Yuchen and Zhao, Yaru and Yang, Pu and Wang, Shaowei and Zhou, Zhi-Hua , title =. Proceedings of the 43rd International Conference on Machine Learning (. 2026 , eprint =

2026
[39]

IEEE Transactions on Computers , volume =

Burtscher, Martin and Ratanaworabhan, Paruj , title =. IEEE Transactions on Computers , volume =. 2009 , doi =

2009
[40]

2021 Data Compression Conference , pages =

Knorr, Fabian and Thoman, Peter and Fahringer, Thomas , title =. 2021 Data Compression Conference , pages =. 2021 , doi =

2021
[41]

A Compression Scheme for Radio Data in High Performance Computing , journal =

Masui, Kiyoshi and Amiri, Mandana and Connor, Liam and Deng, Meiling and Fandino, Mateus and H. A Compression Scheme for Radio Data in High Performance Computing , journal =. 2015 , doi =

2015
[42]

Proceedings of the ACM on Management of Data , volume =

Afroozeh, Azim and Kuff. Proceedings of the ACM on Management of Data , volume =. 2023 , doi =

2023
[43]

and Zwarico, Amy E

Hsu, William H. and Zwarico, Amy E. , title =. Software: Practice and Experience , volume =. 1995 , doi =

1995
[44]

2021 Data Compression Conference , pages =

Ray, Jessica and Brahmakshatriya, Ajay and Wang, Richard and Kamil, Shoaib and Reuther, Albert and Sze, Vivienne and Amarasinghe, Saman , title =. 2021 Data Compression Conference , pages =. 2021 , publisher =

2021
[45]

Genetic Programming and Evolvable Machines , volume =

Kattan, Ahmed and Poli, Riccardo , title =. Genetic Programming and Evolvable Machines , volume =. 2011 , doi =

2011
[46]

2025 IEEE International Symposium on Performance Analysis of Systems and Software , pages =

Azami, Noushin and Burtscher, Martin , title =. 2025 IEEE International Symposium on Performance Analysis of Systems and Software , pages =. 2025 , doi =

2025
[47]

2024 Data Compression Conference , pages =

Burtchell, Brandon Alexander and Burtscher, Martin , title =. 2024 Data Compression Conference , pages =. 2024 , doi =

2024
[48]

Journal of Functional Programming , volume =

Matsuda, Kazutaka and Wang, Meng , title =. Journal of Functional Programming , volume =. 2024 , doi =

2024
[49]

, title =

Srivastava, Saurabh and Gulwani, Sumit and Chaudhuri, Swarat and Foster, Jeffrey S. , title =. Proceedings of the 32nd ACM SIGPLAN Conference on Programming Language Design and Implementation , pages =. 2011 , publisher =

2011
[50]

arXiv preprint arXiv:2211.09676 , year =

Townsend, James and van de Meent, Jan-Willem , title =. arXiv preprint arXiv:2211.09676 , year =. 2211.09676 , archivePrefix =

arXiv
[51]

Proceedings of the 30th International Conference on Machine Learning , series =

Menon, Aditya and Tamuz, Omer and Gulwani, Sumit and Lampson, Butler and Kalai, Adam , title =. Proceedings of the 30th International Conference on Machine Learning , series =. 2013 , publisher =

2013
[52]

ACM Transactions on Programming Languages and Systems , volume =

Shi, Kensen and Bieber, David and Singh, Rishabh , title =. ACM Transactions on Programming Languages and Systems , volume =. 2022 , month = mar, doi =

2022
[53]

Scaling Neural Program Synthesis with Distribution-Based Search , journal =

Fijalkow, Nathana. Scaling Neural Program Synthesis with Distribution-Based Search , journal =. 2022 , doi =

2022
[54]

Proceedings of the 43rd Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages , pages =

Bornholt, James and Torlak, Emina and Grossman, Dan and Ceze, Luis , title =. Proceedings of the 43rd Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages , pages =. 2016 , publisher =

2016
[55]

Proceedings of the VLDB Endowment , volume =

Chen, Xinyu and Tian, Jiannan and Beaver, Ian and Freeman, Cynthia and Yan, Yan and Wang, Jianguo and Tao, Dingwen , title =. Proceedings of the VLDB Endowment , volume =. 2024 , doi =

2024
[56]

, title =

Hishida, Kaisei and Liu, Chunwei and Paparrizos, John and Elmore, Aaron J. , title =. Proceedings of the VLDB Endowment , volume =. 2025 , doi =

2025
[57]

Summer at Hugging Face , year =
[58]

Caleb Fahlgren , title =
[59]

Georgi Gerganov , title =
[60]

Protocol Buffers: Encoding , howpublished =
[61]

2007 , institution=

The gray code , author=. 2007 , institution=

2007

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

2025 , publisher =

Ghorbani, Mahdi and Bauer, Emilien and Grosser, Tobias and Shaikhha, Amir , title =. 2025 , publisher =. doi:10.1145/3720506 , journal =

doi:10.1145/3720506 2025

[2] [2]

Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , year =

Lee, Woosuk and Heo, Kihong and Alur, Rajeev and Naik, Mayur , title =. Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , year =

[3] [3]

Proceedings of the 33rd International Conference on Machine Learning (ICML) , year =

Bielik, Pavol and Raychev, Veselin and Vechev, Martin , title =. Proceedings of the 33rd International Conference on Machine Learning (ICML) , year =

[4] [4]

Syntax-Guided Synthesis , booktitle =

Alur, Rajeev and Bod. Syntax-Guided Synthesis , booktitle =. 2013 , pages =

2013

[5] [5]

Proceedings of the 42nd ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , year =

Ellis, Kevin and Wong, Catherine and Nye, Maxwell and Sabl. Proceedings of the 42nd ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , year =

[6] [6]

Proceedings of the ACM on Programming Languages , volume =

Barke, Shraddha and Peleg, Hila and Polikarpova, Nadia , title =. Proceedings of the ACM on Programming Languages , volume =. 2020 , doi =

2020

[7] [7]

Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =

Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , title =. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =. 2019 , doi =

2019

[8] [8]

, title =

Huffman, David A. , title =. Proceedings of the IRE , volume =