State of the art: who holds each record, and how it was measured

Upper bound: the best result achieved so far. Lower bound: the worst failure shown so far. The two are separate claims; improving one says nothing about the other.

Reading long inputs

Sub-question Upper bound Lower bound Measured how
literal fact in long input most of 17 models near-perfect on plain NIAH (Hsieh 2024a) only half of models claiming 32K tokens stay satisfactory at 32K on harder tasks (Hsieh 2024a) synthetic tasks, accuracy against length
fact without word overlap GPT-4o: 99.3% short, 69.7% at 32K, among the best (Modarressi 2025) 11 of 13 models below half their short-input score at 32K (Modarressi 2025) NoLiMa needles, 13 models
fact in the middle question before and after the data: key-value retrieval near-perfect, from a 45.6% worst case (Liu 2024) GPT-3.5-Turbo with the answer in the middle: 53.8% (20 docs), below its 56.1% closed-book; its 16K version 50.5% (30 docs), below 56.0% (Liu 2024) multi-document QA and key-value retrieval, six models
fixing the middle attention calibration: up to 15 points on RAG tasks (Hsieh 2024b) RAG benchmarks
length alone focused inputs of about 300 tokens beat full ones of about 113k tokens, all model families (Hong 2025) accuracy falls far below the technical limit with padding alone (Levy 2024) 306 LongMemEval prompts; padded QA

Format and order

Sub-question Upper bound Lower bound Measured how
input format GPT-4 more robust to template changes than GPT-3.5 (He 2024) 76 accuracy points between formats, LLaMA-2-13B, few-shot (Sclar 2024); up to 40% for GPT-3.5-turbo on code translation (He 2024) same content, meaning-preserving variants
answer format most recent closed-weight models show little loss (Lee 2026) stricter format constraints, larger reasoning loss (Tam 2024) free-form against JSON, XML, YAML
premise order premises in proof order: best accuracy (Chen 2024b) permuting them: drops of over 30% (Chen 2024b) deductive tasks and R-GSM
repetition question read twice improves reasoning on most of 14 datasets, a few drops on vanilla ChatGPT (Xu 2024a) 112 experiments

Distraction

Sub-question Upper bound Lower bound Measured how
irrelevant sentences self-consistency and an "ignore irrelevant information" instruction recover part of the loss (Shi 2023) large drop on GSM-IC across the prompting methods tested (Shi 2023) grade-school maths with one added sentence
irrelevant passages 1,000 training examples make a model robust (Yoran 2024) retrieval lowers accuracy on some questions (Yoran 2024) five open-domain QA sets
redundancy up to 20x prompt compression with little loss (Jiang 2023) GSM8K, BBH, ShareGPT, arXiv

Instruction files for agents

Sub-question Upper bound Lower bound Measured how
task success developer-written files: +2.4%, not significant (Gloaguen 2026) generated files: -0.5% (SWE-bench Lite, 300 tasks) and -2% (138 instances), not significant (Gloaguen 2026) gold tests, several agents and models
cost median runtime -28.64%, output tokens -16.58% (Lulla 2026) cost +20% and +23% on average with generated files, about +19% on average with developer files (Gloaguen 2026) 124 pull requests; agent traces
effect size bound correctness change bounded to at most 10-15 points (Khatri 2026) 288 runs, equivalence tests
file structure size, rule position, layout, contradictions: no detectable effect (McMillan 2026) pooled OR 0.944 per generated function (about 5.6% lower odds), non-monotonic, found post hoc (McMillan 2026) 1,650 sessions, one trivial target rule
many rules best model 68% at 500 instructions (Jaroslawicz 2025) earlier instructions favoured over later ones (Jaroslawicz 2025) 20 models, keyword-inclusion rules
where files rot checker flags stale code references in 23.0% of 356 repositories, 36% of inspected flags false or ambiguous (Treude 2026) documentation consistency checker

The model as writer

Sub-question Upper bound Lower bound Measured how
self-correction RL-trained: self-correction gain +15.6 points on MATH, +9.1 on HumanEval over the base model (Kumar 2025) prompted, no feedback: accuracy fell for every model on every benchmark tested (Huang 2024) reasoning benchmarks without oracle labels
with test feedback up to +12% with unit tests (Chen 2024a) often modest, sometimes absent, once the cost of repair is counted (Olausson 2024) MBPP and TransCoder (Chen); HumanEval and APPS (Olausson)
prevalence of LLM text at least 13.5% of 2024 PubMed abstracts LLM-processed (Kobak 2025); 6.5-16.9% of AI-conference peer-review text substantially LLM-modified (Liang 2024) excess vocabulary; corpus-level estimation
per-document detection defeated by recursive paraphrasing (Sadasivan 2023); false positives on non-native writing (Liang 2023) stress tests of detectors

Most adopted: the practice record

This record is kept apart from the research record above. Holders differ, and none of these numbers comes with a method.

Sub-question Most adopted practice Measured how Same holder as the research record?
file length under 200 lines (Anthropic 2026b); 32 KiB combined (OpenAI 2026a) none stated no: research tests rule count, not lines (Jaroslawicz 2025)
where instructions go long data first, query last (Anthropic 2026a); both ends (OpenAI 2025) vendor tests, no data shown agrees in direction with Liu 2024
delimiters XML tags (Anthropic 2026a); XML good, JSON poor (OpenAI 2025) vendor tests, no data shown partly: Sclar 2024 finds model-specific results
llms.txt 28% of 137,210 domains using Ahrefs Web Analytics publish it, an upper bound (Ahrefs 2026) server logs, 2026-05 no: 97% of files had zero requests
docs the model lacks an 8 KB index in AGENTS.md: 100% against 53% (Vercel 2026) vendor evaluation, one domain no: Gloaguen 2026 finds no net gain from files