Browse › Natural Language Processing › Language Modelling › C4

Language Modelling archive 2025-07-28

C4 Benchmark (Language Modelling)

9 rows 9 with code listed 3 metrics Dataset page

Syntology Since the archive (Syntology): 3 placed; 27 newer papers judged to report on this board so far.

A language model is a model of natural language. Language models are useful for a variety of tasks, including speech recognition, machine translation, natural language generation (generating more human-like text), optical character recognition, route optimization, handwriting recognition, grammar induction, and information retrieval.

Large language models (LLMs), currently their most advanced form, are predominantly based on transformers trained on larger datasets (frequently using words scraped from the public internet). They have superseded recurrent neural network-based models, which had previously superseded the purely statistical models, such as word n-gram language model.

Source: Wikipedia

The archive carries no text for this table; the description above is the archive's text for the task Language Modelling. archive 2025-07-28

Over time archive 2025-07-28

Direction inferred from the metric name, not from the archive: Perplexity (lower is better). Not inferred (points only, no best-so-far line): TPUv3 Hours, Steps. Points are placed at the row's paper date, not an evaluation date; 9 of 9 rows carry one.

The chart needs JavaScript; the table below carries every value.

Results archive 2025-07-28

Archive rows end at the archive snapshot, 2025-07-28: no result published after that date is in this table. Rank is the archive's row order at that snapshot; not re-ranked here. Metric values are the archive's strings. Column headers sort the table in your browser; each row keeps its archive rank.

Paper Code Ran Syntology Report
1 Primer 12.3517.3K1M – Paper Code 2021 official: no sample here; runs from other or unrecorded repositories · 3 ran (of which 0 constructed an object rather than computing a result; 3 with no instrument failure: 1 honoured, 2 violated, 0 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified report
2 Zeropoint LLM.int8 13B (vector-wise + decomp) 12.45 – Paper Code 2022 official: harvested, nothing ran · 4 ran (of which 0 constructed an object rather than computing a result; 4 with no instrument failure: 0 honoured, 0 violated, 4 with no contract checked; 0 where Syntology's instrument failed) · 1 unverified report
3 T5++ 12.6916.5K1M – Paper Code 2021 official: no sample here; runs from other or unrecorded repositories · 3 ran (of which 0 constructed an object rather than computing a result; 3 with no instrument failure: 1 honoured, 2 violated, 0 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified report
4 Original T5 13.2515.7K1M – Paper Code 2021 official: no sample here; runs from other or unrecorded repositories · 3 ran (of which 0 constructed an object rather than computing a result; 3 with no instrument failure: 1 honoured, 2 violated, 0 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified report
5 LLM.float32 6.7B 13.3 – Paper Code 2022 official: harvested, nothing ran · 4 ran (of which 0 constructed an object rather than computing a result; 4 with no instrument failure: 0 honoured, 0 violated, 4 with no contract checked; 0 where Syntology's instrument failed) · 1 unverified report
6 LLM.float32 2.7B 14.43 – Paper Code 2022 official: harvested, nothing ran · 4 ran (of which 0 constructed an object rather than computing a result; 4 with no instrument failure: 0 honoured, 0 violated, 4 with no contract checked; 0 where Syntology's instrument failed) · 1 unverified report
7 N-Grammer 343M 14.79 – Paper Code 2022 official (archive's flag): 5 ran · 6 ran (of which 0 constructed an object rather than computing a result; 6 with no instrument failure: 0 honoured, 0 violated, 6 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified report
8 N-Grammer 288M 15.01 – Paper Code 2022 official (archive's flag): 5 ran · 6 ran (of which 0 constructed an object rather than computing a result; 6 with no instrument failure: 0 honoured, 0 violated, 6 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified report
9 LLM.float32 1.3B 15.91 – Paper Code 2022 official: harvested, nothing ran · 4 ran (of which 0 constructed an object rather than computing a result; 4 with no instrument failure: 0 honoured, 0 violated, 4 with no contract checked; 0 where Syntology's instrument failed) · 1 unverified report

All 9 rows shown. 9 link to a paper page on this site; 0 are marked as using additional training data in the archive. No GitHub stars are tracked; "Code" is the first repository the archive lists for the row. The archive carries no row tags, review links or community-submitted rows for this table; none are shown. archive 2025-07-28

Syntology Ran reads “N ran (of which C constructed an object rather than computing a result; K with no instrument failure: H honoured, V violated, P with no contract checked; I where Syntology's instrument failed) · U unverified”: of the code samples Syntology harvested from repositories linked to that row's paper (joined by arXiv id), N executed on a synthesized input and U are unverified (harvested, no recorded run). Of the N, C constructed an object rather than computing a result (a part of N, never taken away from it), K ran with no instrument failure (honoured or violated the contract Syntology drafted, or ran with no contract checked) and I are failures of Syntology's instrument, not of the code (a wrong contract draft, or a contract check that could not run). It counts code from repositories linked to that row's paper, not this result: the row's number was not reproduced and nothing here is a correctness claim. The other cell texts mean no graph line for the row: "linked, not harvested" (the archive links code, Syntology has not harvested it), "no code linked" (no code link in the archive), "not matched" (the row's paper URL matched no paper on this site). 9 rows have a graph line, from 3 distinct papers. Rows whose paper has a sample that ran: 9 (9 with a run with no instrument failure, 0 where every run was a failure of Syntology's instrument). Papers: 3 with a sample that ran: 3 with a run with no instrument failure, 0 where every run was a failure of Syntology's instrument. Whose code ran. When the archive marks a repository official for the row's paper, the cell starts with that repository's state (“official” is the archive's flag, not a verdict on who wrote the code): “official (archive's flag): N ran” counts the samples from it that ran, instrument failures included; “official: harvested, nothing ran”; when it has no sample for this paper and other samples ran, “community repositories only” if every one came from a repository the archive lists for the paper but neither marks official nor records as mentioned in it, or one Syntology harvested outside the archive's links, and “official: no sample here; runs from other or unrecorded repositories” if any came from a repository named in the paper or found in its text by Syntology, or has no repository recorded; “official: harvested for another paper” (no sample from it is linked to this paper) or “official: not harvested”. No state is printed when the archive marks none official. The cell's tooltip names the repositories the samples that ran came from. Of the 9 rows with a graph line, 9 have an archive-official repository: on 2 a sample from it ran (2 with a run with no instrument failure), on 0 the samples that ran came from community repositories only, and on 3 from other or unrecorded repositories. The filters above hide rows and never renumber them; the first two count, by default, only a run with no instrument failure, and offer any run, instrument failures included. Counting each paper once, of 14 samples: 13 ran (of which 0 constructed an object rather than computing a result; 13 with no instrument failure: 1 honoured, 2 violated, 10 with no contract checked; 0 where Syntology's instrument failed) · 1 unverified. Separately, 5 of those 14 are pointer-only (licence): the site points at that code rather than redistributing it, a licence property recorded for ran and unverified samples alike; each cell's tooltip carries the row's own pointer-only count. Syntology's record for this page last changed ; when this build read Syntology's graph is in the build record. Per-sample status is on the paper page; each Ran cell links to that paper's sample list. For agents: get_harvested_code_for_paper(arxiv_id) lists each paper's samples; how to connect.

Since the archive: results placed by Syntology Syntology

Of the 8,888 newer papers whose tables Syntology has read (8,886 checked in full and 2 with a leaderboard still awaiting a verdict), 64 had this board among the leaderboards the extractor considered for their tables (it keeps at most 10 boards per paper and 4 per dataset; the lists of 1,574 of the 8,888 were already full without this board, so if it was crowded out of them it is not counted), and 27 proposed a result; of those 27, 17 gave no cell that passed the extractor's own checks, 7 were rejected by the independent check, 0 await a verdict and 3 were placed. Syntology has checked 8,886 of the 18,637 papers on this site that are newer than the archive (for 3,717 of them no archive leaderboard matched the paper's tables, so there was nothing further to check); 733 were read and have nothing to place (arXiv has no HTML version of the paper, or that version has no tables), 41 could not be read (the extractor's reply could not be parsed), and results from the other 8,977 appear after they are checked.

Results that newer papers report for their own method, placed here by Syntology. A model pointed at the cell in the paper's own table; the number was read from that cell and checked against this leaderboard's metric, dataset, split and scale; an independent check that saw this leaderboard's other rows and every other leaderboard on the same dataset accepted it. Not reviewed by the paper's authors or by the archive's editors, and not ranked against the archive rows above. Measured before publishing, on held-out archive papers: the extractor was given 883 of them and produced output for 881 (the other 2 had no title and were excluded before any model call, so they are not part of this measurement). A model adjudicator (Claude Opus 5.5), blind to which entries were real, judged 108 of 110 accepted entries correct on those held-out archive papers; its own controls: 69 of 70 archive-confirmed entries judged true, 101 of 103 corrupted copies judged false. It was not measured on the newer papers whose results are placed on this site.

Paper Method (configuration) Perplexity Date Where in the paper Code Syntology Report
HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM... arXiv:2605.29843 4 HARP · Llama 2 13B 6.14 28 May 2026 Table 1, row “4 HARP”Caption: “Table 1: Perplexity (PPL ↓ ) for QuIP# with fixed RHT and Mixed-Radix HARP.…” 9 harvested: 8 ran (of which 2 constructed an object rather than computing a result; 6 with no instrument failure: 3 honoured, 1 violated, 2 with no contract checked; 2 where Syntology's instrument failed) · 1 unverified report
RaBiT: Residual-Aware Binarization Training for Accurate and Efficient LLMs arXiv:2602.05367 RaBiT (Ours) · Llama2-13B 6.95 5 Feb 2026 Table 2, row “RaBiT (Ours)”Caption: “Table 2 : Comparison with […] 2-3-bit Methods on Llama Models. We…” 6 harvested: 5 ran (of which 1 constructed an object rather than computing a result; 1 with no instrument failure: 0 honoured, 0 violated, 1 with no contract checked; 4 where Syntology's instrument failed) · 1 unverified report
Revisiting Adaptive Rounding with Vectorized Reparameterization for LLM Quantization arXiv:2602.02151 4 bits VQRound+GPTQ · LLaMA-13B 7.10 2 Feb 2026 Table 3, row “4 bits VQRound+GPTQ”Caption: “Table 3 : LLaMA family perplexity on Wikitext2 and C4. Lower is better.” 3 harvested: 3 ran (of which 1 constructed an object rather than computing a result; 2 with no instrument failure: 0 honoured, 0 violated, 2 with no contract checked; 1 where Syntology's instrument failed) · 0 unverified report

Syntology 3 entries, one per paper, newest first by month (the arXiv date, else the month in the arXiv id), then by arXiv id. Each value is the cell text as the paper prints it; under "Where in the paper" is the start of the table's caption as the paper prints it (ranking words shown as […]; a caption carrying TeX this page does not render, or quoting a percentage figure, is left to the hover and says so), and hovering there gives the caption start and each value's column header. Not part of the archive and not in the chart above. Drawn from 8,886 of 18,637 newer papers checked so far (733 more had nothing to place; 41 could not be read).

Newer papers whose tables Claude Sonnet 4.5 judged to report on this benchmark (a model's judgement, not a result; newest first by arXiv id)

Paper arXiv date Ran Syntology Outcome
Selective Disclosure Watermarking for Large Language Models arXiv:2607.05353 6 Jul 2026 10 harvested: 5 ran (of which 2 constructed an object rather than computing a result; 3 with no instrument failure: 1 honoured, 0 violated, 2 with no contract checked; 2 where Syntology's instrument failed) · 5 unverified refused by a rule: the table's column header does not name the leaderboard's metric
UniSVQ: 2-bit Unified Scalar-Vector Quantization arXiv:2606.10520 9 Jun 2026 2 harvested: 2 ran (of which 0 constructed an object rather than computing a result; 1 with no instrument failure: 1 honoured, 0 violated, 0 with no contract checked; 1 where Syntology's instrument failed) · 0 unverified refused by a rule: the table's column header does not name the leaderboard's metric
DAG-MoE: From Simple Mixture to Structural Aggregation in Mixture-of-Experts arXiv:2606.01062 31 May 2026 4 harvested: 1 ran (of which 0 constructed an object rather than computing a result; 1 with no instrument failure: 0 honoured, 0 violated, 1 with no contract checked; 0 where Syntology's instrument failed) · 3 unverified rejected by the independent check
Memory-Efficient LLM Training with Dynamic Sparsity: From Stability to Practical Scaling arXiv:2606.00888 30 May 2026 13 harvested: 11 ran (of which 7 constructed an object rather than computing a result; 9 with no instrument failure: 2 honoured, 0 violated, 7 with no contract checked; 2 where Syntology's instrument failed) · 2 unverified refused by a rule: the table reports a different split or setting than the leaderboard
HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM Quantization arXiv:2605.29843 28 May 2026 9 harvested: 8 ran (of which 2 constructed an object rather than computing a result; 6 with no instrument failure: 3 honoured, 1 violated, 2 with no contract checked; 2 where Syntology's instrument failed) · 1 unverified placed
AliMark: Enhancing Robustness of Sentence-Level Watermarking Against Text Paraphrasing arXiv:2605.29434 28 May 2026 3 harvested: 3 ran (of which 0 constructed an object rather than computing a result; 1 with no instrument failure: 1 honoured, 0 violated, 0 with no contract checked; 2 where Syntology's instrument failed) · 0 unverified refused by a rule: the model gave no value to check against the cell
GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs arXiv:2605.23078 21 May 2026 1 harvested: 1 ran (of which 1 constructed an object rather than computing a result; 1 with no instrument failure: 0 honoured, 0 violated, 1 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified; the one sample that ran constructed an object rather than computing a result rejected by the independent check
Prune, Update and Trim: Robust Structured Pruning for Large Language Models arXiv:2605.18331 18 May 2026 5 harvested: 3 ran (of which 0 constructed an object rather than computing a result; 1 with no instrument failure: 1 honoured, 0 violated, 0 with no contract checked; 2 where Syntology's instrument failed) · 2 unverified refused by a rule: the table's column header does not name the leaderboard's metric
Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs arXiv:2605.15491 15 May 2026 22 harvested: 17 ran (of which 0 constructed an object rather than computing a result; 6 with no instrument failure: 3 honoured, 0 violated, 3 with no contract checked; 11 where Syntology's instrument failed) · 5 unverified refused by a rule: the row is an ablation variant, not the method itself
Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts arXiv:2605.12476 12 May 2026 repository linked, no samples harvested refused by a rule: the table's column header does not name the leaderboard's metric
BCJR-QAT: A Differentiable Relaxation of Trellis-Coded Weight Quantization arXiv:2605.10655 11 May 2026 9 harvested: 4 ran (of which 0 constructed an object rather than computing a result; 1 with no instrument failure: 1 honoured, 0 violated, 0 with no contract checked; 3 where Syntology's instrument failed) · 5 unverified refused by a rule: the table's column header does not name the leaderboard's metric
Fitting Is Not Enough: Smoothness in Extremely Quantized LLMs arXiv:2605.08894 9 May 2026 repository linked, no samples harvested refused by rules: the row is an ablation variant, not the method itself; the table's column header does not name the leaderboard's metric
Swift-SVD: Theoretical Optimality Meets Practical Efficiency in Low-Rank LLM Compression arXiv:2604.01609 2 Apr 2026 1 harvested: 1 ran (of which 0 constructed an object rather than computing a result; 0 with no instrument failure: 0 honoured, 0 violated, 0 with no contract checked; 1 where Syntology's instrument failed) · 0 unverified rejected by the independent check
GlowQ: Group-Shared LOw-Rank Approximation for Quantized LLMs arXiv:2603.25385 26 Mar 2026 3 harvested: 3 ran (of which 0 constructed an object rather than computing a result; 0 with no instrument failure: 0 honoured, 0 violated, 0 with no contract checked; 3 where Syntology's instrument failed) · 0 unverified rejected by the independent check
Surgical Repair of Collapsed Attention Heads in ALiBi Transformers arXiv:2603.09616 10 Mar 2026 1 harvested: 1 ran (of which 0 constructed an object rather than computing a result; 1 with no instrument failure: 1 honoured, 0 violated, 0 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified refused by a rule: the table's column header does not name the leaderboard's metric
3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs arXiv:2603.01376 2 Mar 2026 6 harvested: 4 ran (of which 0 constructed an object rather than computing a result; 2 with no instrument failure: 1 honoured, 0 violated, 1 with no contract checked; 2 where Syntology's instrument failed) · 2 unverified rejected by the independent check
Taming Momentum: Rethinking Optimizer States Through Low-Rank Approximation arXiv:2602.24283 27 Feb 2026 repository linked, no samples harvested rejected by the independent check
RaBiT: Residual-Aware Binarization Training for Accurate and Efficient LLMs arXiv:2602.05367 5 Feb 2026 6 harvested: 5 ran (of which 1 constructed an object rather than computing a result; 1 with no instrument failure: 0 honoured, 0 violated, 1 with no contract checked; 4 where Syntology's instrument failed) · 1 unverified placed
Revisiting Adaptive Rounding with Vectorized Reparameterization for LLM Quantization arXiv:2602.02151 2 Feb 2026 3 harvested: 3 ran (of which 1 constructed an object rather than computing a result; 2 with no instrument failure: 0 honoured, 0 violated, 2 with no contract checked; 1 where Syntology's instrument failed) · 0 unverified placed
WorldCup Sampling for Multi-bit LLM Watermarking arXiv:2602.01752 2 Feb 2026 2 harvested: 1 ran (of which 1 constructed an object rather than computing a result; 1 with no instrument failure: 0 honoured, 0 violated, 1 with no contract checked; 0 where Syntology's instrument failed) · 1 unverified; the one sample that ran constructed an object rather than computing a result refused by a rule: the table's column header does not name the leaderboard's metric
SFMP: Fine-Grained, Hardware-Friendly and Search-Free Mixed-Precision Quantization for... arXiv:2602.01027 1 Feb 2026 12 harvested: 9 ran (of which 0 constructed an object rather than computing a result; 3 with no instrument failure: 0 honoured, 0 violated, 3 with no contract checked; 6 where Syntology's instrument failed) · 3 unverified refused by a rule: the row label the model quoted is not the row's label
FOAM: Blocked State Folding for Memory-Efficient LLM Training arXiv:2512.07112 8 Dec 2025 6 harvested: 5 ran (of which 0 constructed an object rather than computing a result; 1 with no instrument failure: 0 honoured, 0 violated, 1 with no contract checked; 4 where Syntology's instrument failed) · 1 unverified refused by a rule: the number the model quoted is not the number printed in the cell it pointed at
DartQuant: Efficient Rotational Distribution Calibration for LLM Quantization arXiv:2511.04063 6 Nov 2025 16 harvested: 6 ran (of which 0 constructed an object rather than computing a result; 3 with no instrument failure: 2 honoured, 0 violated, 1 with no contract checked; 3 where Syntology's instrument failed) · 10 unverified refused by a rule: one configuration gave two values for one column
MISA: Memory-Efficient LLMs Optimization with Module-wise Importance Sampling arXiv:2511.00056 28 Oct 2025 1 harvested: 1 ran (of which 0 constructed an object rather than computing a result; 1 with no instrument failure: 0 honoured, 0 violated, 1 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified refused by a rule: the table reports a different split or setting than the leaderboard
Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression arXiv:2510.20984 23 Oct 2025 repository linked, no samples harvested rejected by the independent check
PT²-LLM: Post-Training Ternarization for Large Language Models arXiv:2510.03267 27 Sep 2025 24 harvested: 13 ran (of which 0 constructed an object rather than computing a result; 6 with no instrument failure: 0 honoured, 1 violated, 5 with no contract checked; 7 where Syntology's instrument failed) · 11 unverified refused by rules: the row label the model quoted is not the row's label; one configuration gave two values for one column
AMQ: Enabling AutoML for Mixed-precision Weight-Only Quantization of Large Language Models arXiv:2509.12019 15 Sep 2025 6 harvested: 4 ran (of which 0 constructed an object rather than computing a result; 4 with no instrument failure: 1 honoured, 0 violated, 3 with no contract checked; 0 where Syntology's instrument failed) · 2 unverified refused by a rule: one configuration gave two values for one column

Syntology 27 newer papers judged to report on this board: the papers counted above as having proposed a result (3 placed); every one is listed here and in the JSON twin. The order is the arXiv identifier's, newest first, not a rank. The judgement is the model's, read from the paper's own tables; the outcome is what Syntology's checks did with it, and no value from a paper that was not placed is printed. “refused by a rule” names the extractor's rule; “no cell passed the extractor's checks” means the extractor's record holds no cell for this board, passed or refused. The 17 counted above as giving no cell that passed the extractor's own checks are the 17 marked “refused by a rule”. Drawn from 8,886 of 18,637 newer papers checked so far (733 more had nothing to place; 41 could not be read).

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections