Review: LLM readers

This page lists each source on how language models read a prompt, with what it found. It first says how the search ran; its six strands then run from position in the context to practice.

Terms are defined in frame.md; numbers live in state-of-the-art.md.

How the search ran

Strand	Academic name	Acronym	Author seed
position	"lost in the middle", positional bias	LiM, NIAH	Liu; Hsieh
format	prompt formatting sensitivity, table serialization	FormatSpread	Sclar
distraction	irrelevant context, distractibility	GSM-IC	Shi
instruction files	repository context files, cursor rules	AGENTS.md	Gloaguen; Treude
writer	self-correction, LLM-generated text detection	IFEval	Huang; Kamoi

Each strand: arXiv search by title and abstract phrase, then by author, then a web search for non-arXiv sources.

Surveys, read first

Mei 2025 frames the context as an information payload to optimise, over about 1,400 papers.
Liu 2025 surveys long-context models, including how they are evaluated and analysed.
Kamoi 2024 audits self-correction research and gives a checklist for fair experiments.
Wu 2025 surveys the detection of LLM-generated text and its open problems.
Xu 2024b classifies knowledge conflicts: between context and memory, inside the context, inside memory.

Strand 1: position in the context

Seminal: Liu 2024 moved the answer document through 10, 20 and 30 documents, and keys through key-value lists, for six models.
The result was a U-shape: best at the start or end, worst in the middle.
Placing the question both before and after the data (query-aware contextualization) fixed key-value retrieval but barely moved document QA.
Hsieh 2024b traced the U-shape to an attention bias toward the start and end, and calibrated it away.
Hsieh 2024a (RULER) showed that plain NIAH is nearly solved, then added multi-hop and aggregation tasks that are not.
Modarressi 2025 (NoLiMa) removed word overlap between question and needle, and long-context accuracy fell sharply.
Veseli 2025 measured length relative to each model's window: LiM fades near the limit, and a recency bias remains.
Hong 2025 tested 18 models: accuracy fell as input grew, and needle position did not matter in their task.
Levy 2024 padded the same reasoning task to longer inputs and saw accuracy fall far below the advertised limit.

Strand 2: format sensitivity

Seminal: Sclar 2024 varied separators, casing and spacing that keep meaning, and measured large accuracy swings.
Format rankings in Sclar 2024 correlated only weakly between models, so one model's best format is not another's.
He 2024 wrote the same prompts as plain text, Markdown, JSON and YAML; the larger model was more robust.
Sui 2024 found table understanding depends on table format, content order and partition marks.
Ahsan 2026 (TABVERSE, preprint) held table content fixed across HTML, Markdown, LaTeX and images.
Tam 2024 showed that asking for JSON or XML output lowers reasoning.
Lee 2026 located that loss in the format instruction itself, and recovered it by reasoning first and formatting second.

Strand 3: distraction, length, order, redundancy

Seminal: Shi 2023 (GSM-IC) added one irrelevant sentence to grade-school maths problems.
Yoran 2024 characterised when retrieval hurts and trained models to ignore irrelevant passages.
Cuconasu 2024 found related-but-useless passages hurt more than random ones.
Chen 2024b permuted premises away from the order a proof uses them.
Xu 2024a (Re2) had the model read the question twice, citing causal attention as the reason it helps.
Leviathan 2025 (preprint) repeated the whole prompt for non-reasoning models of four families.
Jiang 2023 (LLMLingua) compressed prompts heavily with little loss, so much prompt text is redundant to the model.

Strand 4: instruction files for agents

Howard 2024 proposed llms.txt: one required H1, a summary blockquote, link lists, and an "Optional" section an agent may skip.
The AGENTS.md site (Agentic AI Foundation 2026) defines a plain Markdown file for agents; the nearest file in the tree wins, and the chat prompt overrides it.
Neither site cites an evaluation of accuracy.
Jiang 2025 built a taxonomy of cursor rules from 401 repositories: conventions, guidelines, project information, directives, examples.
Zarzycki 2026 censused 22,854 repositories and annotated 341,852 statements in their instruction files.
dos Santos 2026 catalogued six smells in instruction files, including bloat and conflicting instructions.
Treude 2026 named stale instruction files context rot and detected them with an existing documentation checker.
Gloaguen 2026 ran agents with and without instruction files on SWE-bench Lite and on a new benchmark of repositories that ship one.
Lulla 2026 ran agents on 124 pull requests with and without AGENTS.md, measuring time and tokens.
Khatri 2026 ablated instruction files across two agents with equivalence tests.
McMillan 2026 varied file size, rule position, file layout and contradictions in a factorial design.
Zhou 2023a (IFEval) built instructions a program can verify, such as word counts and required keywords.
Jaroslawicz 2025 (IFScale) raised the number of simultaneous instructions to 500.
Zhou 2023b (DocPrompting) retrieved library documentation before code generation.

Strand 5: the LLM as writer

Kobak 2025 tracked excess style words across 15 million biomedical abstracts.
Liang 2024 estimated the share of LLM-modified text in AI conference peer reviews.
Reinhart 2025 compared human and LLM text on Biber's grammatical and rhetorical features.
Sadasivan 2023 broke detectors, watermarks included, with recursive paraphrasing.
Liang 2023 showed detectors flag non-native English writing as machine-made.
Huang 2024 removed oracle stopping labels and found self-correction of reasoning fails without external feedback.
Gou 2024 (CRITIC) and Chen 2024a (self-debugging) gave models tool and test feedback instead.
Olausson 2024 counted the cost of self-repair and found feedback quality is the bottleneck.
Panickssery 2024 showed evaluators recognise and favour their own outputs.
Kumar 2025 (SCoRe) trained self-correction with reinforcement learning.

Strand 6: Practice

Each page below is its owner's claim about its own product. None states a method for its numbers, except where noted.

Anthropic 2026a (company guide): long documents at the top, the query at the end; it reports gains "up to 30 percent" with no method.
Anthropic 2026b (product docs): target under 200 lines per CLAUDE.md; longer files reduce adherence; contradictory rules may be picked arbitrarily.
Anthropic 2026b also says memory files are context, not enforcement, and points to a hook to block an action.
OpenAI 2025 (company cookbook): instructions at both start and end beat one end only; XML delimiters did well and JSON did particularly poorly.
OpenAI 2026a (product docs): AGENTS.md files are layered from root to current directory, with a combined cap of 32 KiB by default.
AGENTS.md 2026 (de facto specification): the page claims use by over 60k open-source projects and lists logos of 23 tools (counted here); stewarded under the Linux Foundation. The use claim is below the quality bar: an example, not evidence.
Howard 2024 (llms.txt): the proposal; its page says only that thousands of sites publish one.
Ahrefs 2026 reports llms.txt on 28% of 137,210 domains using Ahrefs Web Analytics, an upper bound.
Ahrefs 2026 (SEO company, stated method, server logs, 2026-05): 97% of llms.txt files received zero requests; AI retrieval bots made 1.1%. The page's heading says of total requests, but its body says of AI bot requests.
Ahrefs 2026 states that a fetch is not a read, so these shares are ceilings on use.
Vercel 2026 (vendor evaluation, one domain): baseline 53%; a skill by default 53%; a skill with explicit instructions 79%; an 8 KB docs index in AGENTS.md 100%.
In Vercel 2026 the skill was never invoked in 56% of cases; the page gives no model name and no task count.
Chatlatanagulchai 2026 (preprint) describes content only: in 2,303 files from 1,925 repositories, test procedures appear in 75.9%, security in 14.8%.