Few-Shot Degradation Is Not What It Seems: Behavioral Evidence, Representation Analysis, and a Random-Text Control Across 12 Models, 2 Tasks, and 2 Architectures

arXiv:2609.15990v1 Announce Type: new Abstract: Few-shot prompting sometimes degrades language models instead of helping them, but why this happens is unknown. We evaluate 12 open-weight models on two Ukrainian tasks news classification and legal case outcome prediction and find that the effect is strongly task-dependent: the same models that gain +24 pp on news show only +3.4 pp on legal text, with two models degrading. To understand why, we look inside the models. Prior work measures how much hidden states shift between zero-shot and few-shot modes, but few-shot prompts are much longer, and that length difference alone moves representations. We propose a simple fix: replace demonstrations with length-matched random text to measure the shift caused by prompt length, then subtract it. The resulting metric content delta isolates how much the model's representations change because of what the demonstrations say, not how long they are. This changes the picture entirely: raw shift does not predict whether few-shot helps or hurts (r = 0.20), but content delta does (rho = +0.65, p = 0.043). Models that restructure representations more from demonstration content benefit more the opposite of the intuitive "distortion" explanation. Masking demonstrations in Llama 3.3 70B confirms the finding causally, recovering accuracy above the zero-shot baseline.

Abstract:Few-shot prompting sometimes degrades language models instead of helping them, but why this happens is unknown. We evaluate 12 open-weight models on two Ukrainian tasks news classification and legal case outcome prediction and find that the effect is strongly task-dependent: the same models that gain +24 pp on news show only +3.4 pp on legal text, with two models degrading. To understand why, we look inside the models. Prior work measures how much hidden states shift between zero-shot and few-shot modes, but few-shot prompts are much longer, and that length difference alone moves representations. We propose a simple fix: replace demonstrations with length-matched random text to measure the shift caused by prompt length, then subtract it. The resulting metric content delta isolates how much the model's representations change because of what the demonstrations say, not how long they are. This changes the picture entirely: raw shift does not predict whether few-shot helps or hurts (r = 0.20), but content delta does (rho = +0.65, p = 0.043). Models that restructure representations more from demonstration content benefit more the opposite of the intuitive "distortion" explanation. Masking demonstrations in Llama 3.3 70B confirms the finding causally, recovering accuracy above the zero-shot baseline.

12 pages, 6 figures, 4 tables. Data: this https URL

Computation and Language (cs.CL); Machine Learning (cs.LG)

arXiv:2609.15990 [cs.CL]

(or arXiv:2609.15990v1 [cs.CL] for this version)

https://doi.org/10.48550/arXiv.2609.15990 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

analiza aiqbench

Few-Shot Degradation Is Not What It Seems: Behavioral Evidence, Representation Analysis, and a Random-Text Control Across 12 Models, 2 Tasks, and 2 Architectures

Subiecte conexe