Conflicting Supervision Moves Commitment, Not Capability: A 12.29 sigma arrangement effect that is exactly zero under a convention-agnostic score
cs.AI
Submitted: 2026-09-23
Updated: 2026-09-23
Code: https://github.com/wenhui-ml/Capability-ConvergenceHypothesis
Sources
- How Learning Rate Decay Wastes Your Best Data in Curriculum-Based LLM Pretraining
- The Capability Convergence Hypothesis: Capability from Access Structure, Not Scale
- Knowledge Conflicts for LLMs: A Survey
- Why Supervised Fine-Tuning Fails to Learn: A Systematic Study of Incomplete Learning in Large Language Models
- Truth as a Compression Artifact in Language Model Training
- The 'Problem' of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation
- Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead
- Hallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the Answer
- Are We Evaluating Knowledge or Phrasing? Mitigating MCQA Sensitivity with ParaEval
- Random Reshuffling: Simple Analysis with Vast Improvements
- GraB: Finding Provably Better Data Permutations than Random Reshuffling
- Shuffling the Data, Stretching the Step-size: Sharper Bias in constant step-size SGD
- CommonIT: Commonality-Aware Instruction Tuning for Large Language Models via Data Partitions
- Demystifying Data Organization for Enhanced LLM Training
- Optimizer Memory Makes Shuffle Order a First-Order Source of Fine-Tuning Noise
- The Geometry of Sequential Learning: Lie-Bracket Prediction of Transfer Order
- DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining
- DoGE: Domain Reweighting with Generalization Estimation
- Data Mixing Can Induce Phase Transitions in Knowledge Acquisition
- Mid-Training of Large Language Models: A Survey
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Agents' Last Exam
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems