← Engineermaxxing
Cenednors — Conference Observatory

ICLR 2026

The International Conference on Learning Representations. Explore 223 oral papers as a constellation of knowledge — hover to discover, click to dive deep.

Location Rio de Janeiro Dates April 23–27, 2026 Orals 223 Workshops 40 Awards 3 Our Deep Dives 9

01The Constellation

Each dot is an oral paper. Clusters are research territories. Hover to peek, click to explore. Gold stars are award winners, blue rings are papers we’ve turned into interactive lessons.

Oral Paper Award Winner Deep Dive Available

02Award Winners

Two outstanding papers and one honorable mention, selected from 4,068 submissions.

★ Outstanding Paper

Transformers are Inherently Succinct

Bergsträßer, Cotterell & Lin

A tiny transformer can encode patterns that would require an exponentially larger RNN — and verifying even simple properties about it is provably intractable (EXPSPACE-complete).

◆ Read our interactive lesson →
★ Outstanding Paper

LLMs Get Lost In Multi-Turn Conversation

Laban, Hayashi, Zhou & Neville

39% average performance degradation across 15 LLMs when tasks are delivered gradually across turns. Even reasoning models like o3 and DeepSeek-R1 suffer.

◆ Read our interactive lesson →
☆ Honorable Mention

The Polar Express: Optimal Muon

Amsel, Persson, Musco & Gower

Optimal polynomial approximations for polar decomposition — provably best convergence for the Muon optimizer via approximation theory and the Equioscillation Theorem.

◆ Read our interactive lesson →

03Our Deep Dives

We’ve processed these ICLR 2026 papers into interactive Veanors lessons — full derivations, canvas simulations, and quizzes.

Outstanding Paper · Theory

Transformers are Inherently Succinct

Exponential succinctness gaps between transformers and RNNs. Verification is EXPSPACE-complete.

10 chapters9 simulations1774 lines
Outstanding Paper · Evaluation

LLMs Get Lost In Multi-Turn Conversation

Sharded simulation reveals 39% degradation when tasks spread across turns.

10 chapters9 simulations1799 lines
Honorable Mention · Optimization

The Polar Express: Optimal Muon

Optimal polynomial approximations for polar decomposition in the Muon optimizer.

10 chapters8 simulations1690 lines
Oral · Architecture

Mamba-3

Inference-first SSM with exp-trapezoidal discretization, complex states, and MIMO updates.

10 chapters8 simulations2023 lines
Oral · Interpretability

Temporal Sparse Autoencoders

Contrastive loss disentangles semantic from syntactic features in a self-supervised manner.

10 chapters9 simulations1617 lines
Oral · World Models

World-In-World: Closed-Loop World Models

First benchmark for world models in closed-loop agent-environment interactions.

10 chapters7 simulations2004 lines
Oral · Vision

Depth Anything 3

Multi-view depth and ray prediction from any number of input views with a plain ViT.

10 chapters9 simulations1736 lines
Oral · Alignment

Why DPO is a Misspecified Estimator

DPO as weighted KL-projection causes preference reversal. AuxDPO fixes it.

10 chapters8 simulations1640 lines
Oral · Architecture

Energy-Based Transformers

Predict via gradient-descent energy minimization. 35% faster scaling than Transformer++.

10 chapters8 simulations2208 lines

04Workshop Constellation

40 workshops, 2,863 accepted papers. Each cluster is a workshop — hover to see papers, click to open on ICLR or OpenReview.

Workshop (click → ICLR) Workshop Paper Hover for details

05Full Paper Index

All papers organized by research territory. Expand a cluster to browse.