This is not a popularity ranking or another compression of eight abstracts. It places the claim people repeat, the result the paper directly reported, and the boundary of that evidence side by side—so later impact is not mistaken for an original experimental result.
PaperBridge Editorial·
How this index treats evidence
Each entry stays with experiments, metrics, and conditions directly reported in the primary paper. A boundary does not dismiss the work; it prevents later product capabilities, historical influence, or marketing language from being projected backward as something the original paper already proved. PaperBridge did not reproduce these experiments, so research and product decisions should still verify the source.
01
2017 · Ashish Vaswani et al.
Attention Is All You Need
Popular claim
The Transformer proved that attention could replace every earlier sequence model.
Direct paper evidence
The paper reported 28.4 BLEU on WMT 2014 English-to-German and 41.8 BLEU on English-to-French, while showing a training design that parallelized more readily than recurrent models.
Evidence boundary
The experiments directly support attention-based sequence transduction for machine translation. They did not establish every later LLM capability or show that recurrence and convolution were useless for every task.
02
2018 · Jacob Devlin et al.
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Popular claim
BERT proved that a pretrained model could simply understand language.
Direct paper evidence
Bidirectional pre-training followed by task fine-tuning produced new results on 11 NLP tasks. The abstract reports 80.5% on GLUE, 86.7% on MultiNLI, and F1 scores of 93.2 and 83.1 on SQuAD v1.1 and v2.0.
Evidence boundary
The evidence came from contemporary English-language benchmarks and supervised fine-tuning. It did not establish general language understanding, generation quality, factual reliability, or multilingual performance.
03
2020 · Tom B. Brown et al.
Language Models are Few-Shot Learners
Popular claim
GPT-3 proved that sufficiently large models no longer need training data or fine-tuning.
Direct paper evidence
A 175-billion-parameter autoregressive model performed many tasks from zero-, one-, and few-shot prompts without updating its weights, with in-context performance generally improving with scale.
Evidence boundary
The paper also documented failures, bias, and inconsistency, and many tasks remained behind task-specific fine-tuned systems. It supports prompt-based adaptation, not the disappearance of data, training, or specialized models.
04
2020 · Patrick Lewis et al.
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Popular claim
RAG prevents hallucinations and guarantees reliable citations.
Direct paper evidence
The original system combined a pretrained generator with dense retrieval over Wikipedia, reported state-of-the-art results on three open-domain QA tasks, and generated more specific, diverse, and factual text than a parametric-only baseline.
Evidence boundary
The experiments show that retrieval can help knowledge-intensive tasks. They do not guarantee correct retrieval, faithful use of evidence, or verifiable citations, and they did not test production access control, stale content, or latency.
05
2020 · Jonathan Ho, Ajay Jain, Pieter Abbeel
Denoising Diffusion Probabilistic Models
Popular claim
The DDPM paper already invented today's text-to-image systems.
Direct paper evidence
The paper turned gradual noising and learned reverse denoising into a high-quality image generator, reporting an Inception Score of 9.46 and FID of 3.17 on unconditional CIFAR-10 and showing 256×256 LSUN samples.
Evidence boundary
It established the quality of unconditional diffusion generation, not the full text conditioning, latent-space diffusion, or production systems used today. Its original sampler also required many denoising steps.
06
2020 · Alexey Dosovitskiy et al.
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Popular claim
ViT proved that convolutional networks were obsolete.
Direct paper evidence
The paper treated image patches as a token sequence and applied a standard Transformer. After large-scale pre-training, ViT matched or exceeded contemporary convolutional networks on several image-classification benchmarks with less training compute.
Evidence boundary
The advantage depended on large-scale pre-training and the reported data regime. It established Transformers as viable vision backbones, not that convolutional inductive bias had no value for smaller data, efficiency, or other vision tasks.
07
2021 · Edward J. Hu et al.
LoRA: Low-Rank Adaptation of Large Language Models
Popular claim
LoRA always matches full fine-tuning at almost no cost.
Direct paper evidence
LoRA froze pretrained weights and learned low-rank updates. For GPT-3 175B, the paper reports up to roughly 10,000 times fewer trainable parameters and about three times lower GPU memory while matching or improving quality on the evaluated models and tasks.
Evidence boundary
Results depend on the low-rank assumption, chosen rank, target layers, task, and data. The paper supports parameter-efficient adaptation in tested settings, not guaranteed equivalence to full fine-tuning everywhere.
08
2022 · Long Ouyang et al.
Training language models to follow instructions with human feedback
Popular claim
RLHF made language models safe, truthful, and aligned with all human values.
Direct paper evidence
InstructGPT used demonstrations, preference rankings, and reinforcement learning. On the paper's prompt distribution and with its labeler population, outputs from the 1.3B model were preferred overall to those from 175B GPT-3, with improvements on several truthfulness, toxicity, and instruction-following measures.
Evidence boundary
Preferences came from a specific labeler group and task distribution, and the metrics did not cover every harm, fact, or value conflict. The result supports an effective instruction-alignment process, not complete safety or universal value alignment.
FAQ
AI paper evidence FAQ
What counts as direct evidence in an AI paper?
Direct evidence is a result reported under stated data, baselines, metrics, and experimental conditions. Later product success, citations, and industry impact matter, but they are not results the original paper necessarily tested.
Why state an evidence boundary separately?
Paper conclusions are often repeated as broader claims. A boundary separates experimental support, author interpretation, and later inference, reducing the chance that one benchmark or historical influence is treated as universal proof.
Does this index replace reading the primary paper?
No. It is a triage tool that shows what to verify. Reproduction, research, and high-stakes decisions still require the paper, appendix, code, and independent follow-up evidence.