AI Paper Atlas

Research map / Benchmarks / PG19

PG19 leaderboard

10 papers at ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR report results on PG19. Best perplexity: 8.6499 (HiP, ICLR 2025).

Numbers are as each paper reports them, read by AI from its results tables and not checked by hand. Papers can differ in setting (training data, model size, extra data); the topic column shows what each paper is about. Open the paper before you compare or quote a number.

PG19 · perplexity

10 papers, each with its best reported number for its own method; lower is better.

#perplexity ↓MethodPaperTopic
18.6499HiP
Llama3.1-8B
A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention
ICLR 2025 · text
Attention and transformers
210.37BST:SH:unstruct-LBlock-State Transformers
NeurIPS 2023 · Table 1
Attention and transformers
311.37Memorizing Transformer, memory 65K
12-layer Transformer
Memorizing Transformers
ICLR 2022 · Table 1
Attention and transformers
413.205Vist
TinyLlama
Vision-centric Token Compression in Large Language Model
NeurIPS 2025 · Table: Long-context language m
Reasoning and chain of thought
514.79Meta-Attention + RoPELearned Meta-Tokens for Language Modeling
ICLR 2026 · Table 3
Attention and transformers
618.4348-block DenseFormer (1×1)
Transformer, 48 blocks
DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging
NeurIPS 2024 · Table, PG-19
Attention and transformers
728.9Perceiver AR (2048)General-purpose, long-context autoregressive modeling with Perceiver AR
ICML 2022 · Table: PG-19 language modeling
Attention and transformers
831.04PMI-TR
Transformer
A Framework for Inference Inspired by Human Memory Mechanisms
ICLR 2024 · PG-19 results table
RNNs and normalization
933.636L Compressive Transformer
TransformerXL
Compressive Transformers for Long-Range Sequence Modelling
ICLR 2020 · Table: Eval. perplexities on P
Attention and transformers
1040Zero massive activations and sink value (h = v = 0)
LLaMA2-7B
Towards Understanding Massive Activations in Attention Sink Mechanism
ICML 2026 · Table 3
Attention and transformers

Not affiliated with any conference, journal or the benchmark's authors. Corrections: contact@latenticlabs.com.