Fine Until Fine-Tuned: Repeated Solutions Make Reasoning Fragile
cs.LG
Submitted: 2026-09-27
Updated: 2026-09-27
Code: https://github.com/ely2ba/reasoning-durability
Terminology
Sources
- MathArena: Evaluating LLMs on Uncontaminated Math Competitions
- LoRA Learns Less and Forgets Less
- How Do Large Language Models Acquire Factual Knowledge During Pretraining?
- Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting
- Internal Data Repetition Destroys Language Models
- Weight Ensembling Improves Reasoning in Language Models
- Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs
- Early Data Exposure Improves Robustness to Subsequent Fine-Tuning
- Stream of Search (SoS): Learning to Search in Language
- Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs
- Measuring Mathematical Problem Solving With the MATH Dataset
- Self-Improvement in Language Models: The Sharpening Mechanism
- Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead
- Data Repetition Beats Data Scaling in Long-CoT Supervised Fine-Tuning
- Understanding Catastrophic Forgetting in Language Models via Implicit Inference
- Tulu 3: Pushing Frontiers in Open Language Model Post-Training
- Let's Verify Step by Step
- Why Do Reasoning Models Lose Coverage? The Role of Data and Forks in the Road
- gpt-oss-120b & gpt-oss-20b Model Card
- Unveiling Over-Memorization in Finetuning LLMs for Reasoning Tasks
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: Local-Consistency Optimisation for Bayesian Neural Networks
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks