What Pretraining and Midtraining Make Learnable from Rewards?
cs.LG, cs.AI, cs.CL
Submitted: 2026-09-29
Updated: 2026-09-29
Comments: 160 pages, 26 figures, 32 tables
Project page: https://yuxinchen2020.github.io/publications/transfo
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
The gist: A reward can identify a correct answer while leaving the computation needed for new inputs undetermined.
Terminology
Abstract
A reward can identify a correct answer while leaving the computation needed for new inputs undetermined. We study how pretraining and midtraining supply the information and computation that make reward adaptation effective. In sequential state computation and contextual memory, we characterize mechanisms that agree on every training reward yet demand different held-out answers. Task-independent source observations resolve this ambiguity. We construct finite sampled Adam paths from specified random initializations through source prediction and reward adaptation in the same parameters, proving how prediction acquires execution or retrieval and rewards learn their task-specific use. Experiments with pretrained Qwen2.5 checkpoints test this division of labor. Across eight worlds, Sequential models trained with correct source and first-operation supervision reach 82.61% success, versus 44.15% for a private-random source control. Memory replay preserves retrieval during reward adaptation, and an independent eight-world confirmation achieves 75.32% task success versus 49.86% after matched alternative-retrieval training. GSM8K and HotpotQA separate accuracy at reward entry, subsequent gain and final performance. Together, these results connect information acquisition, executable computation and reward-guided task learning.
Sources
- RL Post-Training Builds Compositional Reasoning Strategies
- Verify to Amplify: Improving Reasoning via Learned Chain-of-Thought Verification
- DeepCoder: Learning to Write Programs
- Titans: Learning to Memorize at Test Time
- When can transformers reason with abstract symbols?
- Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers
- Training Verifiers to Solve Math Word Problems
- Frustratingly Simple Black-Box Adaptation of Language Models via Logit Bias
- Defeating Prompt Injections by Design
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- What can a Single Attention Layer Learn? A Study Through the Random Features Lens
- On the Statistical Query Complexity of Learning Semiautomata: a Random Walk Approach
- Instruction Following by Principled Boosting Attention of Large Language Models
- IH-Challenge: A Training Dataset to Improve Instruction Hierarchy on Frontier LLMs
- Transformers Provably Learn Sparse XOR with Polylogarithmic Parameters
- When Can Model-Free Reinforcement Learning be Enough for Thinking?
- A Mean-Field Analysis of Multi-Head Self-Attention under Cross-Entropy Training
- Transformers Provably Learn to Internalize Chain-of-Thought
- On the Emergence of Implicit Curriculum in RLVR Learning Dynamics
- Adam: A Method for Stochastic Optimization
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks