AI Paper Atlas

Research map / Architectures and efficiency / Neural architecture search

Composer: A Search Framework for Hybrid Neural Architecture Design

Bilge Acun, Prasoon Sinha, Newsha Ardalani, Sangmin Bae, Alicia Golden, Meghana Madhyastha, Chien-Yu Lin, Fei Sun, Neeraja J. Yadwadkar, Carole-Jean Wu

ICLR 2026 · Architectures and efficiency › Neural architecture search · approach: Bi-level differentiable NAS · 7 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
use Composer-discovered hybrid architecture instead of Llama 3.21B LLM · DCLM validationvalidation loss: 2.80 -> 2.77Improved
use Composite instead of Llama 3.21B LLM · 6 downstream tasksaverage accuracy: 58.69 -> 60.83Improved
use 1:8 Striped Attention (fewer attention layers, no search)1B LLM · DCLM validationvalidation loss: 2.80 -> 2.85Hurt
scale down width during search (ablation)Composer 6-layer search · 1B DCLM validationvalidation loss and search cost: loss -0.02 to -0.04; cost 6.38x lowerImproved
use one-shot or middle-incremental search instead of end-incremental searchComposer Search Engine · 1B DCLM validationvalidation loss: 0.01-0.02 lower lossImproved
stretch rather than stack discovered architecture when search depth >= 16 layers (ablation)Composer Extrapolator · 1B DCLM validationvalidation lossImproved
stretch rather than stack from very small searches (e.g. 2A+4M) (ablation)Composer Extrapolator · 1B DCLM validationvalidation lossHurt
search over more than 16 layers (ablation)Composer Search Engine · 1B DCLM validationvalidation lossHurt
use MAD synthetic dataset instead of Large-Scale DCLM for evaluator (ablation)Composer Evaluator · search costsearch cost: over 8x lower costImproved
use Chinchilla-scaled small DCLM (4M model) for search (ablation)Composer Evaluator · 1B DCLM validationvalidation lossNo clear effect
use 1:2 attention-to-MLP ratio instead of 1:1 (ablation)hybrid LLM · 1B DCLM validationvalidation lossImproved
use Composite instead of Llama 3.21B LLM · Training and inference efficiencythroughput/latency: 1.25x throughput, 1.33x latency, 1.69x KV cacheImproved

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