All ICLR submissions, 2024 to 2027 · data pulled 26 October 2026

ICLR 2027 Trends

What changed in machine learning between ICLR 2026 and 2027, read from all 42,822 submissions. Start with the eight shifts below; every chapter opens further if you want the detail.

Trends are shares of each year's submissions, since the conference doubled. How this was made

Eight shifts in one year

Each card jumps to its chapter

The conference more than doubled

Submissions and primary areas

ICLR went from 19,814 to 42,822 submissions in one year. Robotics and systems grew fastest and generative models slowest, so every comparison on this page uses shares of the conference.

Submissions per year

Includes withdrawn and desk-rejected papers. 2027 reviews are still running.

Primary areas that moved most

20262027, gained2027, lost
2026 outcomes and 2027 statusaccepted, withdrawn, desk-rejected
  • 2026 for reference: 5,351 accepted (27% of all submissions), 8,349 rejected, 5,206 withdrawn, 908 desk-rejected.
  • 2027 so far: 1,701 desk-rejected (4.0%) and 292 withdrawn. Most withdrawals come after reviews, so this will rise.
  • Placeholders: a handful of 2027 entries have titles such as “xxx” or “111”. They stay in the counts; too few to move any share.
Chapter 1 · Movers

What heated up, what cooled down

Rose: agents, embodied AI and distillation. Fell: reasoning-length work, preference tuning and classic ML.

Change in share of submissions, 2026 to 2027

gained sharelost share
Topics with at least 120 papers in 2027, matched by fixed text patterns in titles and abstracts. ×2 means the topic's share of the conference doubled. Hover a bar for the four-year numbers.

Every tracked topic, with its four-year numbers, is on the trend board →

Chapter 2 · Areas

Area by area, in one line each

Open an area for what is rising, what is cooling, what is new, and papers worth reading.

Areas group the primary areas authors pick at submission. Each brief was written from 110 random 2027 abstracts in the area plus term statistics for the whole area. Sparklines show the area's share of all submissions, 2025 to 2027.

Chapter 3 · New ideas

Sixteen ideas that barely existed in 2026

Pick an idea to see what it is, why it appears now, and the papers that show it best.

Counts come from text patterns and include some unrelated papers; each idea's details say what share of a 65-paper sample was really on topic. Descriptions were written by model readers from those samples, and every cited paper was checked against the data.

Chapter 4 · Crossovers

Which ideas travelled to other areas

A paper is filed under one area, but the ideas in it often come from another. This year LLM agents, world models and on-policy distillation travelled furthest, and LLM efficiency tricks moved into video and robot models.

Where the fastest-spreading ideas went

idea’s share of the area grewline width: 2027 papers
Left, the ideas that grew most outside their main areas; right, the three areas each grew into most. Hover an idea or an area to trace its lines, tap or click to keep it traced, and click a line for papers.

Not more borrowing overall. The share of papers using at least one tracked idea from outside its main areas was 56% in 2026 and 50% in 2027. What changed is which ideas travel.

20262027, rose2027, fell
Ideas whose main area is elsewhere, with at least 15 papers here in 2027, ranked by growth. Click a row for papers.
The full map: every area and every ideadot matrix, ideas grouped by theme

How much each area uses each idea

share grew ≥25%within ±25%fell ≥20%idea’s main area
Scroll sideways for more ideas. Rows are areas, columns are ideas. A dot’s size is the share of that area’s 2027 submissions mentioning the idea; its colour is the change since 2026 (pale dots have too few papers to say). Click a dot for papers, an area name for what it borrows, an idea name for where it spread.
How crossovers are countedmain areas, rankings, what to distrust
  • Ideas. 81 tracked terms from the trend board (chapter 7): fixed text patterns over title plus abstract. Model names, benchmark names and three generic terms (language model, reasoning, transformer) are left out. A mention is not the same as use; some abstracts mention an idea only in passing.
  • Areas. The primary area authors chose at submission, grouped into the 12 areas of chapter 2. Each paper has exactly one.
  • Main areas. For each idea, the area with the most papers using it in 2026, plus the 2027 leader if different (2027 only for ideas with under 40 papers in 2026). A crossover is an idea used in any other area. The map marks main areas with a ring.
  • The opening chart. Ideas are ranked by papers in non-main areas beyond what those areas’ own growth would give: for each area, 2027 papers minus 2026 papers times the area’s growth. It shows the top eight ideas (or fourteen) and, for each, the three largest gains with at least 20 papers in 2027.
  • What each area borrows. Ideas outside their main areas with at least 15 papers in the area in 2027 and under 5% of the area in 2026, ranked by change in share (adding 3 papers to each year to damp small counts).
  • Ideas that combine. Two ideas in the same abstract. Both had at least 40 papers in 2026, they are not variants of one family (GRPO and RL, say), they share at least 25 papers in 2027, and the pair grew at least as fast as the two ideas would predict on their own. Ranked by change in the share of submissions using both, at most three pairs per idea. Arrows (A → B, A applied to B) were set by a model reader from the example titles.
  • Examples. Papers whose title names the idea come first, then a fixed random order. They illustrate the count; they are not picked for quality.
  • Notes. A model reader wrote the short descriptions from the example titles behind each idea, area and pair (3 to 15 titles each), so they describe typical papers, not every paper. All counts come from the data.
Chapter 5 · Models

The model zoo turned over

Qwen went from 0.74% of submissions in 2025 to 8.7%, overtaking Llama. GPT-5 and Claude replaced GPT-4 as the frontier systems papers name.

Share of submissions naming a model family in the title or abstract. Grey lines are other families; hover for names.

Benchmarks moved with them. Reasoning benchmarks fell (AIME 0.92% → 0.47%, GPQA 0.38% → 0.21%), while agent and code benchmarks rose (SWE-bench 0.19% → 0.62%, tau-bench 0.02% → 0.12%). All named benchmarks are on the trend board →

Chapter 6 · Style

The abstracts sound different this year

The stock words of the GPT-4 era collapsed in one year. A new set, about controls and exact effect sizes, replaced them in every area at once.

2025 house style · illustrative
“This paper proposes a novel framework for multi-task learning, meticulously designed to enhance performance across tasks. This innovative approach leverages a shared encoder to guide each task’s predictions…”Written for this page to show the style; not taken from a submission. The highlighted words are real examples of the vocabulary that fell by 2027.
2027 house style · illustrative
“We first train a shared encoder on all tasks, then freeze it and fit a small head for each task. With matched training budgets, the method lowers held-out error compared with a single shared head in five of six settings…”Written for this page to show the style; not taken from a submission. The highlighted words are real examples of the vocabulary that rose in 2027.
79%was 28%
of abstracts use none of twenty GPT-era stock words (novel, leverage, crucial…)
60%was 21%
use at least one of twenty new stock words (matched, frozen, held-out…)
34%was 23%
report an exact % figure in the abstract

One vocabulary out, another in

20262027, fell2027, rose
Abstracts per 1,000 containing the word. Both panels share one scale. Hover for 2024 and 2025.
How widespread is the switch?every area, and desk rejections
3+ old stock words3+ new stock words
Share of abstracts using at least three words from each list of twenty (lists in the method notes).
  • Every area: heavy-style abstracts range from 3% in learning theory to 12% in interpretability (heavy = four or more new stock words).
  • Desk rejections: heavy style is about twice as common among desk-rejected submissions (10.8%) as among those under review (5.8%).
Titles and abstracts over four yearscolons, question titles, “Towards”, em dashes

Titles

Abstracts

How to read thisa change in tools more than topics
  • Tools, not topics. The change is abrupt, uniform across areas and concentrated in words that carry no topic. That fits a change in writing tools better than a change in research.
  • Old signature out. The 2025 buzzwords (novel, leverage, seamless, showcase) were the known signature of earlier assistants and faded as authors and tools learned to avoid them.
  • New signature in. The new set (matched, frozen, held-out, percentage points, held fixed, pre-registered) reads like a newer assistant’s idea of rigor.
  • Some real substance. 34% of abstracts now give an exact % figure (up from 23%), more name seeds and controls, and a genre of controlled-measurement papers has appeared (chapter 3.5).
  • For readers. Generic hype words no longer separate weak papers from strong ones, and the vocabulary of control does not guarantee the controls were run.
  • Caveat. Text alone cannot show who or what wrote an abstract; these numbers describe the corpus, not any paper.
Chapter 7 · Explore

Check any term yourself

Type up to four phrases, separated by commas, to chart how often submissions used them each year.

Counts documents whose title or abstract uses the term. Suggestions come from the keywords authors chose, reviewed term by term. Variants are one entry: singular and plural, spelling (fine-tuning, finetuning), abbreviations and full names (SFT and supervised fine-tuning), and generic endings (“reinforcement learning algorithm” counts as reinforcement learning). Type either form.

The trend board: all 112 tracked termssparklines, sorting, exact patterns

Every tracked term with its four-year trajectory, grouped by theme. Hover a row for counts and the exact pattern. The broad terms in the first group are partly distorted by the change in writing style (see the method notes); specific terms are more reliable.

Method

How this was made

OpenReview metadata for every submission since 2024, fixed text patterns, and model readers working from samples.

Data, measures and what to distrust
  • Data. Every ICLR submission for 2024 to 2027 from the OpenReview API, including accepted, rejected, withdrawn and desk-rejected papers: 7,405, 11,672, 19,814 and 42,822. Titles, abstracts, author keywords and primary areas only; no PDFs. 2027 data was pulled on 26 October 2026, during reviews.
  • Shares, not counts. The conference doubled, so every trend is a share of that year’s submissions. “×2” means a topic’s share doubled.
  • Topic patterns. Each tracked topic is a regular expression over title plus abstract (shown on the trend board, chapter 7). Patterns favour recall; the ideas in chapter 3 say how much of a 65-paper sample was on topic (45% to 92%). The steering and mechanistic-interpretability patterns were narrowed to drop other senses of the words (steering a vehicle or the verb “steer”; quantum and electronic circuits).
  • Broad words. “LLM” and “reasoning” are distorted by the style change: 2026 abstracts often opened with a stock sentence about LLMs, 2027 abstracts rarely do. Specific terms (GRPO, KV cache, world model) are more reliable.
  • Style lists. Old: novel, innovative, advancement(s), underscore, crucial, remarkable, comprehensive, leverage, enhance, seamless, intricate, pivotal, paving the way, significant challenge(s), showcase, delve, notably, state-of-the-art, superior performance, nuanced. New: percentage point(s), matched, frozen, held-out, strongest, held fixed, seed(s), ask(s) whether, audit, unchanged, control(s), pre-registered, prespecified, survive(s), keeps, stays, leaves, rather than, reaches, “Together,”. The new list holds the words whose share rose most from 2026 to 2027.
  • Reading. Idea descriptions and area briefs were written by model readers from samples (65 matching abstracts per idea, 110 random abstracts per area), using only what the abstracts say. Paper ids and titles were checked against the data.
  • Links. Example papers link to their OpenReview pages. Reviews are double-blind and still running, so the page names no authors and links no preprints, and desk-rejected or withdrawn papers are counted but not listed.
  • Names. Terms are written “full name (ABBR)”. Tight spots such as chart labels use the abbreviation alone, with the full name on hover.
  • Merged terms. Overlapping tracked topics were merged (RSI into self-improvement, world-action models into world models, looped models into latent reasoning, “agentic” into LLM agents). For the explorer and the term counts in area briefs, variants count as one term: singular and plural, spelling (finetuning, fine-tuning, time series, time-series), abbreviations with their full names (SFT, T2I, KV cache, PDE and others), near-synonyms reviewed term by term (tool calling and tool use), and generic endings (“reinforcement learning algorithm” counts as reinforcement learning). TTS (test-time scaling or text-to-speech), SSL (self- or semi-supervised), SAM (segment anything or sharpness-aware minimization) and MAE (masked autoencoder or mean absolute error) stay unmerged because they are ambiguous.
  • Explorer. Its terms are the keywords authors chose: 2,384 terms used as a keyword at least 8 times across the four years and found in at least 15 titles or abstracts in some year, reviewed term by term by a model reader to merge variants and remove generic entries. Counts are documents whose title or abstract uses the term, and a phrase never spans punctuation. Other single words can be searched but are not suggested.