Clustering 5,000 ICLR 2026 reviews: weak-baseline complaints drop acceptance to ~23%
sarahookr · x · 2026-10-17
The author used adaptionai's Visualize to cluster the harshest reviews of 5,000 ICLR 2026 papers, mapping which complaint types hurt acceptance most.
Key findings:
- Papers criticized for weak baselines or low novelty face the harshest outcome, with acceptance falling to 23%;
- Complaints about compute cost or real-world validity are far more forgiving, leaving acceptance at 36%;
- By volume, "training & task design" is the most common weakness theme (900 papers), followed by novelty & contribution and LLM/benchmark evaluation (630 each) — the latter two have the lowest acceptance rates (22-23%);
- Efficiency/compute-cost and real-world validity themes show the highest acceptance (35%+), with fine-tuning setup complaints also relatively mild (32%).
Takeaway: reviewers are ruthless about unfair comparisons and insufficient novelty, but lenient on efficiency and practicality criticisms.
More from Research
- Dev's take: every new AI benchmark — StarCraft, GUIs, Minecraft, CAD — is secretly a coding benchmark — wateriscoding · 2026-10-17
- AI Beat His 2027 Prediction: Lean Proof of Fields Medal Sphere-Packing Result Arrived by March — TOEwithCurt · 2026-10-17
- Real-Time Hair Rendering: 100 Characters, 100K Strands Each, Just 1.7 MB — TinfoilTricorn · 2026-10-17
- Strand-Aligned 3D Gaussians Enable Strand-Level Hair Reconstruction — TinfoilTricorn · 2026-10-17
- Apple research: a 7B model with 8 retries matches a 32B model using 4x less memory — CatAstro_Piyush · 2026-10-17
- Visualizing the Change-of-Variables Rule and the Jacobian-Inverse Term in Flow Models — CatAstro_Piyush · 2026-10-17