PKU-Wu-Lab/LTE-Qwen3-8B-Base

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

LTE-Qwen3-8B-Base is an 8 billion parameter language model developed by PKU-Wu-Lab, based on the Qwen3 architecture. It utilizes the Learning to reason from Trial and Error (LTE) approach, a Reinforcement Learning with Verifiable Rewards (RLVR) method. This model is designed to improve reasoning capabilities by learning from its own self-generated errors, mitigating exploration stagnation without requiring external expert guidance. It is particularly suited for tasks requiring enhanced exploitation and exploration during training.

Loading preview...

Overview

LTE-Qwen3-8B-Base is an 8 billion parameter language model from PKU-Wu-Lab, implementing the Learning to reason from Trial and Error (LTE) approach. LTE is a Reinforcement Learning with Verifiable Rewards (RLVR) method that enhances a language model's reasoning abilities by leveraging its own past mistakes as learning signals. This technique aims to overcome exploration stagnation in LMs by providing self-generated hints, eliminating the need for external expert supervision.

Key Capabilities

  • Self-Correction: The model learns from errors it generates during training, using these as internal hints for improvement.
  • Autonomous Learning: It operates without requiring external expert guidance or human feedback to refine its reasoning processes.
  • Enhanced Exploration and Exploitation: The LTE approach is designed to improve both the exploration of new solutions and the exploitation of learned knowledge during training, leading to a higher performance ceiling for LMs.

Good For

  • Applications requiring robust reasoning capabilities where models can learn and adapt from their own outputs.
  • Scenarios where external expert feedback is limited or unavailable.
  • Research into self-improving language models and reinforcement learning techniques for reasoning tasks.