AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Understanding Dynamic Compute Allocation in Recurrent Transformers

Ibraheem Muhammad Moosa, Suhas Lohit, Ye Wang, Moitreya Chatterjee, Wenpeng Yin

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Efficient language transformers · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace fixed depth 8 with ANIRA-E adaptive depthRecurrent-Llama-3.2 · GSM-Symbolic Mainexact accuracy (%): 46.26 -> 45.70No clear effect
replace fixed depth 8 with ANIRA-O adaptive depthRecurrent-Llama-3.2 · GSM-Symbolic Mainexact accuracy (%): 46.26 -> 43.68Hurt
replace fixed depth 8 with ANIRA-E adaptive depthRecurrent-Llama-3.2 · GSM-Symbolic P1exact accuracy (%): 23.54 -> 24.06No clear effect
replace fixed depth 8 with ANIRA-O adaptive depthRecurrent-Llama-3.2 · GSM-Symbolic P2exact accuracy (%): 5.80 -> 5.08Hurt
use ANIRA-O instead of fixed depth 8: compute reductionRecurrent-Llama-3.2 · GSM-Symbolic Mainmean allocated depth: 8 -> 4.108Improved
use ANIRA-E instead of fixed depth 8: compute reductionRecurrent-Llama-3.2 · GSM-Symbolic Mainmean allocated depth: 8 -> 6.060Improved
harder GSM-Symbolic split (P2 vs Main) with ANIRA-EANIRA-E · GSM-Symbolicmean allocated depth: 6.060 -> 5.982No clear effect
harder GSM-Symbolic split (P2 vs Main) with ANIRA-OANIRA-O · GSM-Symbolicmean allocated depth: 4.108 -> 4.001No clear effect
online halting (ANIRA-O) instead of ANIRA-E allocation ruleANIRA · BREVOfeature explaining compute: DFS traversal depth coefficient: -0.005 -> +0.302Mixed
categorical graph-size effects instead of linear N in the regression (ablation)ANIRA-E · BREVO compute regressionR^2 (%): 68.66 -> 73.37Improved
categorical graph-size effects instead of linear N in the regression (ablation)ANIRA-O · BREVO compute regressionR^2 (%): 70.00 -> 70.88No clear effect
adaptive depth allocation (ANIRA-O) for extrapolation to longer MANO expressionsANIRA-O · MANO extrapolationaccuracyMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

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