AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Two Heads Are Better than One: Distilling Large Language Model Features into Small Models with Feature Decomposition and Mixture

Tianhao Fu, Xinxin Xu, Weichen Xu, Jue Chen, Ruilong Ren, Bowen Deng, Xinyu Zhao, Jian Cao, Xixin Cao

AAAI 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 1 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
add task decomposition to layer decomposition (ablation)CMM FD(layer) · FUEPnL (1e3): 12.54 -> 17.25Improved
add volatility-type data decomposition (ablation)CMM FD(task) · FUEPnL (1e3): 17.25 -> 22.41Improved
add Hajek-MoE fusion of experts (ablation)CMM FD(type) · FUEPnL (1e3): 22.41 -> 31.39Improved
add Hajek-MoE fusion of experts (ablation)CMM FD(type) · FUMAP (unit): 45 -> 32Improved
distill LLM features with orthogonal decomposition instead of ReviewKDsmall student · FUEPnL (1e3): 10.52 -> 31.39Improved
use CMM instead of X-MoEMoE fusion · FUEPnL (1e3): 19.63 -> 31.39Improved
distill LLM into CMM small modelsLLM-Base · FUEPnL (1e3): 30.47 -> 31.39Improved
distill LLM into CMM small modelsLLM-Base · FUlatency (s): 1.9 -> 0.3Improved
use CMM instead of IMMIMM · RBEPnL (1e3): 16.46 -> 22.69Improved
use CMM instead of IMMIMM · FUEPnL (1e3): 28.1 -> 31.39Improved
use CMM instead of IMMIMM · AGEPnL (1e3): -14.5 -> -8.95Improved
use CMM instead of IMM under extreme marketIMM · simulated flash crash and reversalEPnL (1e3): 7.80 -> 10.50Improved

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