AI Paper Atlas

Research map / Probabilistic and scientific modeling / Dynamical systems

Text Generation as Continuous Latent Dynamics via Reinforcement Learning

Chen Jia

ICML 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Latent dynamics models

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 KL regularization to the reference LLM (ablation)CT-LLM · AlpacaEval 2length-controlled win rate: 15.4 -> 19.2Improved
add KL regularization to the reference LLM (ablation)CT-LLM · AlpacaEval 2ROUGE-L: 30.5 -> 36.5Improved
add KL regularization to the reference LLM (ablation)CT-LLM · AlpacaEval 2raw win rate: 12.6 -> 16.6Improved
add KL regularizationTransformer Dec-only · AlpacaEval 2length-controlled win rate: 13.1 -> 12.8No clear effect
add KL regularizationTransformer Enc-Dec · AlpacaEval 2length-controlled win rate: 6.2 -> 8.9Improved
replace Transformer Dec-only with continuous latent ODE policy (both RL+KL)CT-LLM · AlpacaEval 2length-controlled win rate: 12.8 -> 19.2Improved
replace Transformer Dec-only with continuous latent ODE policy (both RL+KL)CT-LLM · AlpacaEval 2ROUGE-L: 32.5 -> 36.5Improved
scale CT-LLM from 0.4B to 2.1B parametersCT-LLM · AlpacaEval 2 (model scaling)length-controlled win rate: 3.6 -> 22.0Improved
scale CT-LLM from 0.4B to 2.1B parametersCT-LLM · AlpacaEval 2 (model scaling)perplexity: 9.8 -> 6.8Improved
scale CT-LLM from 1.2B to 2.1B parametersCT-LLM · AlpacaEval 2 (model scaling)length-controlled win rate: 19.2 -> 22.0Improved
use CT-LLM instead of 8B Transformer under same training FLOPsCT-LLM · AlpacaEval 2 (compute-matched, 1.4e20 FLOPs)ROUGE-L: 30.2 -> 36.8Improved
use CT-LLM instead of 1.7B Transformer under same training FLOPsCT-LLM · AlpacaEval 2 (compute-matched, 1.4e20 FLOPs)ROUGE-L: 28.5 -> 36.8Improved

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