What question did this study set out to answer?

The research investigates what fine-tuning and in-context learning can effectively install in language models, particularly concerning behavior and correctness.

June 19, 2026Open Access

Fine-Tuning and In-Context Learning Install Dispositions, Not Correctness: Template-Compatibility, a Truncation Artifact, and When Doubt Pays

View Full Paper

Ask AI

Bookmark

The research investigates what fine-tuning and in-context learning can effectively install in language models, particularly concerning behavior and correctness.
Assessed 70B scale language models with varied instruction templates in fine-tuning experiments.
Evaluated in-context demonstrations without truncation on five substrates.
Analyzed a methodological artifact affecting apparent accuracy gains via generation truncation.
Fine-tuning with a matched instruction template maintained performance (MMLU 82.3%) compared to vanilla models (84.5%), while verbose templates collapsed performance to 8.2%.
In-context learning did not improve accuracy but instilled a disposition for caution, impacting abstention and false-premise rejection.
A short generation cap created an illusory accuracy gain of +10 percentage points by truncating open reasoning before commitment.

Discussion

No takes yet. Share an insight, caveat, or question.

Implication

Randomized trial examines the effects of fine-tuning and in-context learning on model behavior, indicating a shift in dispositions but not correctness.

Cite This Study

Tomas Pødenphant Lund (2026) studied this question.

synapsesocial.com/papers/6a34e06865a5b0777af2f17d https://doi.org/10.5281/zenodo.20562086

View Full Paper

Ask AI

Bookmark

Implication

Randomized trial examines the effects of fine-tuning and in-context learning on model behavior, indicating a shift in dispositions but not correctness.

Fine-Tuning and In-Context Learning Install Dispositions, Not Correctness: Template-Compatibility, a Truncation Artifact, and When Doubt Pays

Key Points

Discussion

Implication

Cite This Study

Fine-Tuning and In-Context Learning Install Dispositions, Not Correctness: Template-Compatibility, a Truncation Artifact, and When Doubt Pays

Key Points

Discussion

Implication

Cite This Study

Also Consider

Also Consider