Latentic LabsPublished by Latentic Labs with scime atlasscime
The scime Dispatch

ICLR 2026 edition · every accepted paper, counted and compared

ICLR 2026 · 5,358 accepted papersNovember 2026Free · atlas.scime.ai/newspaper

← All editions

Reasoning and Chain of Thought: 1.8% of ICLR 2025, 6.5% this year

ICLR grew 45% to 5,358 accepted papers. Reasoning and chain of thought grew faster still, from 67 to 347.

Reasoning and chain of thought (Large language models) went from 67 papers at ICLR 2025 to 347 this year, from 1.8% to 6.5% of the conference. Inside it, RL with verifiable rewards for reasoning grew fastest (0 → 110 papers). Among them: When More is Less: Understanding Chain-of-Thought Length in LLMs and ReTool: Reinforcement Learning for Strategic Tool Use in LLMs.

Also gaining: Multimodal LLMs (3.6% → 6.0%) and LLM agents (1.6% → 3.7%). Losing ground: Language, safety and interpretability (6.5% → 5.4%) and GNN architectures (2.1% → 1.0%).

Qwen is the model family ICLR 2026 papers build on most: 1,284 of 4,973 papers with result tables (26%), ahead of Llama (17%) and GPT (OpenAI) (11%). The benchmark named most is GSM8K, then ImageNet.

Not everything helped. Of the 58,584 changes the papers measured, 30% made no clear difference or made things worse, as the papers report them. One of 24 on page D:

A paper says it overturns

5,254 papers name an assumption their results go against; 20 of them on page E.

Accepted ICLR papers per year in scime atlas.

AI-assisted. Topics come from the atlas clustering; what each paper tested and measured was extracted by AI from its full text. Nothing here was checked by hand and some lines are wrong; open the paper before you cite it.

ICLR 2026 in numbers

Accepted papers5,358
ICLR 20253,703
Read in full by AI5,295
Findings recorded (change → effect)58,584
Reported as helping66%

Gained most, 2025 → 2026

Reasoning and chain of thought1.8 → 6.5%
Multimodal LLMs3.6 → 6.0%
LLM agents1.6 → 3.7%
RL from feedback0.8 → 2.5%
Robot action and navigation0.7 → 1.4%

Lost most

Language, safety and interpretability6.5 → 5.4%
GNN architectures2.1 → 1.0%
Alignment and preferences2.6 → 1.7%
Pruning and efficient tuning2.2 → 1.4%

Largest areas, 2026

Large language models22.8%
Vision-language and multimodal10.4%
Image generation8.9%
Architectures and efficiency8.7%

Try it on your idea

Describe your own idea in scime atlas to see the closest of 108,967 papers and what similar changes did in them.

Writing it up? scime draft is a free LaTeX editor with AI and the ICLR template.

Free for signed-in readers

Read the rest of this edition

Five more pages: where the 5,358 papers are, the topics that grew and shrank, what 5,129 papers measured and when it did not help, what the papers say they overturn, and the benchmarks, baselines and models everyone uses. Sign in with a free scime account to keep reading.

  • B Where the papers are
  • C What grew, what shrank
  • D What the papers measured
  • E What the papers overturn
  • F Benchmarks, baselines and models

Sign in free

Already signed in? Reload this page. Accounts are free; we only use your email to sign you in and, if you agree, to tell you about new editions.

G

About the data

how this edition was made

Papers. The 5,358 accepted ICLR 2026 papers and 3,703 of 2025 in scime atlas, grouped with all 108,967 atlas papers (ICML, NeurIPS, ICLR, CVPR and AAAI, 2016–2026) into 20 areas and their topics by meaning (an open-source embedding model and clustering; names written with AI).

Full texts. For 5,295 of the 2026 papers an AI model read the full text and recorded the result tables (benchmark, method, base model, number) and each change the paper tested, on what, the metric and whether it helped, plus what the paper says it challenges. These were not checked by hand. Each paper links to its page on scime atlas, which links to the paper itself.

Counting. Shares count each paper once. Citation counts are from OpenAlex and are still low for 2026 papers; papers are ordered by citations, then by how many findings were recorded.

Caveat. Published papers report successes more often than failures. Corrections: contact@latenticlabs.com. Not affiliated with or endorsed by ICLR or any of the conferences.

Get the next scime Dispatch and a monthly field update

One email when a new edition is out (ICLR, CVPR, ICML, NeurIPS, ACL and special editions), plus a short monthly note on what moved in machine learning research.

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy

← All editions · © 2026 Latentic Labs