CORTEXA
← Browse
arxivcs.LGcs.AIstat.ML2026-07-13

Are we Merging the Right Models? Impact of Expert Training Duration on Model Merging for LLMs

Nikita Kozodoi, Zainab Afolabi, Jack Butler

Multi-task model merging combines separately trained expert models into a single model that handles all tasks without co-training. Standard practice merges experts at their optimal validation loss. We challenge this convention by systematically studying how training duration of domain experts affects the quality of the merged model. We fine-tune experts on five domains (Math, Code, Instruction Following, Multilingual, and Safety) across three model sizes (Qwen 3.5 0.8B, 2B, and 4B), saving checkpoints from 25% to 500% of the optimal training steps and evaluating five merging methods at each duration. Our findings reveal a striking method-dependent pattern: simple averaging degrades sharply with overfitting, while sparsification-based methods achieve their best performance well past the validation optimum. We formalize this through bias-variance decomposition analysis, drawing a parallel to random forests where averaging benefits from high-variance individual learners. These results suggest that training duration and merging method should be chosen jointly rather than independently.

View free PDFSource page

Related papers

arxivcs.LGcs.AIstat.ML2026-06-29

Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback

Ved Sriraman, Peihan Liu, Daniel Hsu, Adam Block

Imitation Learning is a natural framework for learning in sequential decision-making systems and has emerged as the dominant paradigm through which we understand language model training. A central puzzle is that, while in theory offline IL can be horizon-free and optimal, in prac…

View free PDFSource page
arxivcs.LGcs.AIstat.ML2026-06-29

Pessimism's Paradox: Conservative Offline Training Amplifies Reward Hacking During Online Adaptation in Reasoning Models

Subramanyam Sahoo, Aman Chadha, Vinija Jain, Divya Chaudhary

Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model. We challenge this intuition empir…

View free PDFSource page
arxivcs.LGcs.AIcs.CLstat.ML2026-06-26

What LLMs explain is not what they believe: Evaluating explanation sufficiency under models' own input beliefs

Nhi Nguyen, Shauli Ravfogel, Rajesh Ranganath

Large language models (LLMs) are increasingly deployed in high-stakes domains, where free-text explanations such as chain-of-thought and post-hoc rationales are used to justify model outputs. Yet it remains unclear whether these explanations are sufficient, i.e., if they contain…

View free PDFSource page
arxivcs.AIcs.CLcs.LGstat.APstat.ML2026-07-02

Online Safety Monitoring for LLMs

Mona Schirmer, Metod Jazbec, Alexander Timans, Christian Naesseth, Maja Waldron, Eric Nalisnick

Despite alignment training, LLMs remain prone to generating unsafe outputs at deployment time. Monitoring outputs online and raising an alarm when safety can no longer be assumed is therefore critical. We study a simple real-time monitor that turns a verifier signal from an exter…

View free PDFSource page
arxivcs.AIcs.LGstat.ML2026-07-10

ConfidenceBench: Evaluating Confidence Calibration in Large Language Models

Matthew ffrench-Constant, Daniel Yang, Xinmeng Huang, Sanyam Kapoor

Large language models (LLMs) are increasingly deployed in settings where fluent but incorrect answers can be costly. In these settings, accuracy alone is insufficient: models must also know when they are likely to be wrong. We present ConfidenceBench, a calibration benchmark that…

View free PDFSource page
arxivcs.CLcs.AIcs.LGstat.ML2026-07-05

Telescope: Improving Zero Shot Detection of LLM Generated Content By Measuring Token Repetition Probability

Christopher Nassif, Josh F. Cooper

Distinguishing Large Language Model (LLM) generated text from human writing is a critical and difficult challenge. While LLMs are trained to write like humans, we hypothesize that this training leaves an indelible mark. LLMs develop a particularly strong aversion to token repetit…

View free PDFSource page