View From Above: A Framework for Evaluating Distribution Shifts in Model Behavior

Tanush Chopra, Michael Li, Jacob Haimes
7/1/2024
cs.CLcs.AIcs.LG

No official code found. Best implementation: view-from-above/VFA (0 stars, MIT, last commit 2024-09-28).

Abstract

When large language models (LLMs) are asked to perform certain tasks, how can we be sure that their learned representations align with reality? We propose a domain-agnostic framework for systematically evaluating distribution shifts in LLMs decision-making processes, where they are given control of mechanisms governed by pre-defined rules. While individual LLM actions may appear consistent with expected behavior, across a large number of trials, statistically significant distribution shifts can emerge. To test this, we construct a well-defined environment with known outcome logic: blackjack. In more than 1,000 trials, we uncover statistically significant evidence suggesting behavioral misalignment in the learned representations of LLM.

View on arXivView PDF

Code Implementations(1)

view-from-above/VFACommunity100%
MIT

Collection of files from ApartResearch hackathons

00Python, Jupyter NotebookJun 29, 20242 years agoMIT

Similar Papers

Recursive Training Loops in LLMs: How training data properties modulate distribution shift in generated data?

85% match2025

Beyond the Black Box: A Statistical Model for LLM Reasoning and Inference

85% match2024

What do Language Model Probabilities Represent? From Distribution Estimation to Response Prediction

85% match2025

Understanding Multimodal LLMs Under Distribution Shifts: An Information-Theoretic Approach

84% match1 repo122025

Bayesian Statistical Modeling with Predictors from LLMs

84% match2024

Do Large Language Models Know What They Are Capable Of?

84% match2025

Cite this paper

@article{chopra2024view,
  title  = {View From Above: A Framework for Evaluating Distribution Shifts in Model Behavior},
  author = {Tanush Chopra and Michael Li and Jacob Haimes},
  year   = {2024},
  eprint = {2407.00948},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2407.00948v3}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2407.00948.svg)](https://codeofpaper.com/paper/2407.00948)

Discussion