Can Reasoning Models Keep Improving? TEMPO Uses EM to Stop Reward Drift

The paper introduces TEMPO, a test‑time training framework inspired by the Expectation‑Maximization algorithm, which alternates policy optimization (M‑step) with Critic calibration (E‑step) to prevent reward‑signal drift, and demonstrates on Qwen3 and OLMO3 models that it continuously improves reasoning performance and maintains output diversity beyond the saturation point of existing TTT methods.

EM algorithmLarge Language ModelsReinforcement Learning

0 likes · 14 min read

Can Reasoning Models Keep Improving? TEMPO Uses EM to Stop Reward Drift

Model Perspective

Mar 1, 2023 · Artificial Intelligence

Mastering the EM Algorithm: Theory, Math, and Python Implementation

This article explains the Expectation‑Maximization (EM) algorithm, detailing its iterative E‑step and M‑step processes, mathematical formulation, and practical Python implementation for estimating parameters of mixed linear regression models, while highlighting convergence considerations and common pitfalls.

EM algorithmPythonexpectation maximization

0 likes · 12 min read

Mastering the EM Algorithm: Theory, Math, and Python Implementation

Model Perspective

Jan 8, 2023 · Artificial Intelligence

Unlock Hidden Patterns: A Deep Dive into Unsupervised Learning Techniques

This article introduces unsupervised learning, covering its motivation, Jensen's inequality, key clustering methods such as EM, k‑means, hierarchical clustering, evaluation metrics, and dimensionality‑reduction techniques like PCA and ICA, providing clear explanations and illustrative diagrams.

EM algorithmICAK-Means

0 likes · 8 min read

Unlock Hidden Patterns: A Deep Dive into Unsupervised Learning Techniques

DataFunTalk

Nov 10, 2021 · Artificial Intelligence

Learnable Index Structures for Large‑Scale Retrieval: Deep Retrieval Model and Training Methods

This article introduces ByteDance's Deep Retrieval (DR) framework, describing its learnable index structure that aligns embedding training with retrieval objectives, detailing the core model, structure‑loss training via EM and online EM algorithms, beam‑search serving, multi‑task learning, and practical insights from Q&A.

Beam SearchEM algorithmRecommendation Systems

0 likes · 11 min read

Learnable Index Structures for Large‑Scale Retrieval: Deep Retrieval Model and Training Methods