Instead, we should permit differentiating based on the context. Ex: synagogues in America are legally allowed to discriminate by religion when hiring rabbis. Work with Michelle Phan, Daniel E. Ho, @sanmikoyejo.bsky.social arxiv.org/abs/2502.01926

Collaboration with a bunch of lovely people I am thankful to be able to work with: @hannawallach.bsky.social , @angelinawang.bsky.social , Olawale Salaudeen, Rishi Bommasani, and @sanmikoyejo.bsky.social. 🤗

Very excited we were able to get this collaboration working -- congrats and big thanks to the co-authors! @rajiinio.bsky.social @hannawallach.bsky.social @mmitchell.bsky.social @angelinawang.bsky.social Olawale Salaudeen, Rishi Bommasani @sanmikoyejo.bsky.social @williamis.bsky.social

3) Institutions and norms are necessary for a long-lasting, rigorous and trusted evaluation regime. In the long run, nobody trusts actors correcting their own homework. Establishing an ecosystem that accounts for expertise and balances incentives is a key marker of robust evaluation in other fields.

which challenged concepts of what temperature is and in turn motivated the development of new thermometers. A similar virtuous cycle is needed to refine AI evaluation concepts and measurement methods.

2) Metrics and evaluation methods need to be refined over time. This iteration is key to any science. Take the example of measuring temperature: it went through many iterations of building new measurement approaches,

Just like the “crashworthiness” of a car indicates aspects of safety in case of an accident, AI evaluation metrics need to link to real-world outcomes.

We identify three key lessons in particular.

1) Meaningful metrics: evaluation metrics must connect to AI system behaviour or impact that is of relevance in the real-world. They can be abstract or simplified -- but they need to correspond to real-world performance or outcomes in a meaningful way.

We pull out key lessons from other fields, such as aerospace, food security, and pharmaceuticals, that have matured from being research disciplines to becoming industries with widely used and trusted products. AI research is going through a similar maturation -- but AI evaluation needs to catch up.

Thanks @gdemelo.bsky.social for the kind words. I blame the students 😃

Can Symbolic Scaffolding and DPO Enhance Solution Quality and Accuracy in Mathematical Problem Solving with LLMs by Shree Reddy, Shubhra Mishra fine-tune the Qwen-2.5-7B-Instruct model with symbolic-enhanced traces, achieving improved performance on GSM8K and MathCAMPS benchmarks.

Heterogeneity of Preference Datasets for Pluralistic AI Alignment by Emily Bunnapradist, Niveditha Iyer, Megan Li, and Nikil Selvam propose objectively quantifying a dataset's diversity and evaluating preference datasets for pluralistic alignment.
n/n

HP-GS: Human-Preference Next Best View Selection for 3D Gaussian Splatting by Matt Strong and Aditya Dutt presents a simple method for guiding the next best view in 3D Gaussian Splatting.
github.com/peasant98/ac...
Video link: youtube.com/watch?v=t3gC....
6/n

Information-Theoretic Measures for LLM Output Evaluation by
Zachary Robertson, Suhana Bedi, and Hansol Lee explore using total variation mutual information to evaluate LLM-based preference learning
github.com/zrobertson46...
5/n

PrefLearn: How Do Advanced Replay Buffers and Online DPO Affect the Performance of RL Tetris with DQNs by Andy Liang, Abhinav Sinha, Jeremy Tian, and Kenny Dao proposes PrefLearn with superior performance and faster convergence
tinyurl.com/preflearn
4/n

Cost and Reward Infused Metric Elicitation by
Chethan Bhateja, Joseph O'Brien, Afnaan Hashmi, and Eva Prakash extend metric elicitation to consider additional factors like monetary cost and latency.
arxiv.org/abs/2501.00696
3/n