An official website of the United States government
This content will become publicly available on September 1, 2026
Value-Spectrum: Quantifying Preferences of Vision-Language Models via Value Decomposition in Social Media Contexts
- Award ID(s):
- 2415226
- PAR ID:
- 10633628
- Publisher / Repository:
- ACL (Association for Computational Linguistics)
- Date Published:
- ISSN:
- 0736-587X
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
null (Ed.)State-of-the-art value predictors either use control-flow context or data context to predict values. Predictors based on control-flow context use branch histories to remember past values, but these predictors require lengthy histories to predict anything other than constant and strided values. Predictors that use data context---also known as Finite Context Method (FCM) predictors---use a history of past values to predict a broader class of values, but such predictors achieve low coverage due to long training times, and they can become complex due to speculative value histories. We observe that the combination of branch and value history provides better predictability than the use of each history separately because it can predict values in control-dependent sequences of values. Furthermore, the combination improves training time by enabling accurate predictions to be made with shorter history, and it simplifies the hardware design by removing the need for speculative value histories. Based on these observations, we propose a new unlimited budget value predictor, Heterogeneous-Context Value Predictor (HCVP), that when hybridized with E-Stride, achieves a geometric mean IPC of 3.88 on the 135 public traces, as compared to 3.81 for the current leader of the Championship Value Prediction.more » « less
-
As humans interact with autonomous agents to perform increasingly complicated, potentially risky tasks, it is important to be able to efficiently evaluate an agent’s performance and correctness. In this paper we formalize and theoretically analyze the problem of efficient value alignment verification: how to efficiently test whether the behavior of another agent is aligned with a human’s values. The goal is to construct a kind of “driver’s test” that a human can give to any agent which will verify value alignment via a minimal number of queries. We study alignment verification problems with both idealized humans that have an explicit reward function as well as problems where they have implicit values. We analyze verification of exact value alignment for rational agents and propose and analyze heuristic and approximate value alignment verification tests in a wide range of gridworlds and a continuous autonomous driving domain. Finally, we prove that there exist sufficient conditions such that we can verify exact and approximate alignment across an infinite set of test environments via a constant- query-complexity alignment test.more » « less
-
null (Ed.)
