SISL

Check it out! 📄 Title: "Zono-Conformal Prediction: Zonotope-Based Uncertainty Quantification for Regression and Classification Tasks" Paper: jmlr.org/papers/v26/25-… Congratulations to Laura Lützow, Michael Eichelbeck, @aiprof_mykel, and Matthias Althoff!

On synthetic and real-world benchmarks, ZCPs consistently produce tighter prediction sets than classical conformal predictors and interval predictor models, while maintaining comparable coverage.

🆕 New SISL #JMLR paper: When deploying ML models in safety-critical settings (think autonomous vehicles, healthcare, robotics) you don't just need accurate predictions. You need to know how uncertain those predictions are, with formal guarantees.

Your deduplication pipeline was built for small models. At scale, it's broken. New preprint: "Scale Dependent Data Duplication" 1/10

📄 “One Bias After Another: Mechanistic Reward Shaping and Persistent Biases in Language Reward Models” Paper: arxiv.org/abs/2603.03291 Code: github.com/drfein/OneBias… By @DanielFein7*, @MLamparth*, Violet Xiang, @aiprof_mykel, and @nickhaber (* equal contribution).

Sycophancy and model-style sensitivity resist this treatment: their signals are too entangled with genuine reward quality for linear interventions to isolate cleanly, highlighting future research challenges.

🚨 New SISL preprint: State-of-the-art language reward models are still badly biased. Past fixes overcorrect, some can be fixed with simple latent interventions, and some indicate the need for larger efforts.

By Michelle Ho, @mfadhilginting, Isaac R. Ward, @andrzejreinke, @aiprof_mykel, Ali Agha, Shayegan Omidshafiei. Great collaboration between Stanford University and Field AI

Check out the paper: "World Model Failure Classification for Autonomous Inspection" just accepted at ICRA 2026 to be presented in Vienna! Website: …toinspection-classification.github.io Paper: arxiv.org/abs/2602.16182 Code: github.com/autoinspection… Video: youtu.be/FskMku6oI2k

New SISL publication for #ICRA 2026: ⚡Over 90% accuracy in detecting inspection failures before a human observer would notice. We combine world model-based video understanding with conformal prediction to classify outcomes as success, known failure, or OOD anomaly in real time.