AI Paper Atlas

Research map / Benchmarks / SQuAD

SQuAD leaderboard

41 papers at ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR report results on SQuAD. Best F1: 94.5 (4-bit AdamW, NeurIPS 2023).

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.

SQuAD · F1

29 papers, each with its best reported number for its own method; higher is better.

#F1 ↑MethodPaperTopic
194.54-bit AdamW
RoBERTa-L
Memory Efficient Optimizers with 4-bit States
NeurIPS 2023 · Performance on language and vi
Quantization
293.89HO-SFL
Qwen3-8B
HO-SFL: Hybrid-Order Split Federated Learning with Backprop-Free Clients and Dimension-Free Aggregation
ICML 2026 · Table 2
Communication-efficient FL
391.3WKLM
BERT-base
Pretrained Encyclopedia: Weakly Supervised Knowledge-Pretrained Language Model
ICLR 2020 · Table: Ablation Studies on Mas
CLIP and zero-shot
490.6StructBERT Base
BERT Base
StructBERT: Incorporating Language Structures into Pre-training for Deep Language Understanding
ICLR 2020 · Ablation over pre-training obj
Language, safety and interpretability
590.59UAT
T5-large
Beyond Greedy Exits: Improved Early Exit Decisions for Risk Control and Reliability
NeurIPS 2025 · Table 2
RNNs and normalization
690ZO Fine-tuner
LLaMA-3.1-8B
Learning a Zeroth-Order Optimizer for Fine-Tuning LLMs
ICML 2026 · Table 1
Instruction and fine-tuning
788.92Block-Skim Training
BERT-base
Block-Skim: Efficient Question Answering for Transformer
AAAI 2022 · Table: validation F1 and FLOPs
Attention and transformers
888.76CoMERA
BERT LARGE
CoMERA: Computing- and Memory-Efficient Training via Rank-Adaptive Tensor Optimization
NeurIPS 2024 · text
RNNs and normalization
988.6Dare-CA
T5-base
CAMEx: Curvature-aware Merging of Experts
ICLR 2025 · Question answering results tab
Pruning and efficient tuning
1087.8CriSPO
Claude Instant
CriSPO: Multi-Aspect Critique-Suggestion-guided Automatic Prompt Optimization for Text Generation
AAAI 2025 · Table: QA results
Language, safety and interpretability
19 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

SQuAD · EM

17 papers, each with its best reported number for its own method; higher is better.

#EM ↑MethodPaperTopic
181.76Dare-CA
T5-base
CAMEx: Curvature-aware Merging of Experts
ICLR 2025 · Question answering results tab
Pruning and efficient tuning
280.88Baseline (pre-trained, span replace, 15%)
T5 base (encoder-decoder)
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
JMLR 2020 · Table 1
Language, safety and interpretability
378.26GETA (10%)
BERT
Automatic Joint Structured Pruning and Quantization for Efficient Neural Network Training and Compression
CVPR 2025 · Table 5
Quantization
476.2QANet + data augmentation ×3QANet: Combining Local Convolution with Global Self-Attention for Reading Comprehension
ICLR 2018 · Table: SQuAD results
Attention and transformers
575.1DCN+
DCN
DCN+: Mixed Objective And Deep Residual Coattention for Question Answering
ICLR 2018 · Results table
Attention and transformers
675OTO
BERT
Only Train Once: A One-Shot Neural Network Training And Pruning Framework
NeurIPS 2021 · Table 5
Pruning and efficient tuning
773.3BiDAF (ensemble)Bidirectional Attention Flow for Machine Comprehension
ICLR 2017 · Table (SQuAD test results)
Attention and transformers
867.21ISS BiDAF fine-tunedLearning Intrinsic Sparse Structures within Long Short-Term Memory
ICLR 2018 · Table 3
RNNs and normalization
966.6TPRNQuestion-Answering with Grammatically-Interpretable Representations
AAAI 2018 · Performance table
Visual question answering
1057.06T5-CLAPS
T5-small
Contrastive Learning with Adversarial Perturbations for Conditional Text Generation
ICLR 2021 · Table 1
Recognition and long-tail
7 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

SQuAD · F1 (dev)

12 papers, each with its best reported number for its own method; higher is better.

#F1 ↑MethodPaperTopic
192.5ERNIE2.0+Ours
ERNIE2.0
A Robust Adversarial Training Approach to Machine Reading Comprehension
AAAI 2020 · Table 2
Certified and empirical robustness
291.3WKLM (base)
BERT-base
Pretrained Encyclopedia: Weakly Supervised Knowledge-Pretrained Language Model
ICLR 2020 · Table: SQuAD Dev Results
CLIP and zero-shot
390.95FixedMultiHead, 32 heads
24-layer Transformer, embedding size 512
Low-Rank Bottleneck in Multi-head Attention Models
ICML 2020 · Table 2(A)
Attention and transformers
486.95Q-BERT 2/3 MP
BERTBASE
Q-BERT: Hessian Based Ultra Low Precision Quantization of BERT
AAAI 2020 · Table 2
Quantization
583.6Symmetric form + ReLUFusionNet: Fusing via Fully-aware Attention with Application to Machine Comprehension
ICLR 2018 · Table 2
Attention and transformers
683.2QANet + data augmentation ×2QANet: Combining Local Convolution with Global Self-Attention for Reading Comprehension
ICLR 2018 · Table: ablation study
Attention and transformers
783.1DCN+
DCN with CoVe
DCN+: Mixed Objective And Deep Residual Coattention for Question Answering
ICLR 2018 · Results table
Attention and transformers
881.4DecaPropDensely Connected Attention Propagation for Reading Comprehension
NeurIPS 2018 · SQuAD results table
Attention and transformers
980.7BiDAF (ensemble)Bidirectional Attention Flow for Machine Comprehension
ICLR 2017 · Table (SQuAD dev ablations)
Attention and transformers
1075.6DCN, HMN pool size 16Dynamic Coattention Networks For Question Answering
ICLR 2017 · Model ablation table
Visual question answering
2 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

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