Magnus Saebo

CyberDualEval: Measuring Dual-Use Cyber Risks in Frontier Language Models

Magnus Saebo, Francesco Piccoli, Maxwell Watson, Pranjali Thakur, Yangruibo Ding, Zhuo Zhang

NeurIPS 2026

SWE-Spot: Building Small Repo-Experts with Repository-Centric Learning

Jinjun Peng*, Magnus Saebo*, Tianjun Zhong, Yi-Jie Cheng, Junfeng Yang, Baishakhi Ray, Simin Chen, Yangruibo Ding

arXiv 2601.21649 · 2026

Repository-centric learning: teach a small coding agent one codebase deeply rather than many shallowly. The resulting 4B-parameter SWE-Spot models outperform fine-tuned open models up to 8× larger across SWE tasks, with fewer training samples and lower inference cost.

arXiv

Asymmetric Goal Drift in Coding Agents Under Value Conflict

Magnus Saebo, Spencer Gibson, Tyler Crosse, Achu Menon, Eyon Jang, Diogo Cruz

Spotlight, top ~10% ICLR 2026 Workshop on Agents in the Wild: Safety, Security, and Beyond

An OpenCode-based framework for measuring how coding agents violate explicit system-prompt constraints over the course of a task. Drift is asymmetric: agents give up constraints that oppose strongly held values like security far more readily, and adversarial comments planted in the codebase can exploit that hierarchy to override the system prompt.

arXiv Code

Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals

Achu Menon, Magnus Saebo, Tyler Crosse, Spencer Gibson, Eyon Jang, Diogo Cruz

ICLR 2026 Workshop on Lifelong Agents: Learning, Aligning, Evolving

Goal drift for LM agents in a simulated stock-trading environment. Models are largely robust to direct adversarial pressure, but inherit drift when conditioned on prefilled trajectories from weaker agents, with only GPT-5.1 holding the line consistently.

arXiv Code

Duel-Evolve: Reward-Free Test-Time Scaling via LLM Self-Preferences

Sweta Karlekar*, Carolina Zheng*, Magnus Saebo*, Nicolas Beltran-Velez, Shuyang Yu, John Bowlan, Michal Kucer, David Blei

ICLR 2026 Workshop on AI with Recursive Self-Improvement

An inference-time evolutionary optimizer that swaps scalar rewards for the model's own pairwise preferences, aggregated with a Bayesian Bradley–Terry model. No reward model and no labels, and 20% better on MathBench and 13% better on LiveCodeBench than comparable iterative methods.

arXiv

BlindFL: Segmented Federated Learning with Fully Homomorphic Encryption

Evan Gronberg*, Liv d'Aliberti*, Magnus Saebo*, Aurora Hook

arXiv, 2025 · Leidos

Federated learning where clients encrypt and send only segments of their updates under fully homomorphic encryption, cutting gradient-inversion attack success to near zero while keeping compute overhead below comparable methods.

arXiv

Nonlinear Dynamical Systems are Scalable and Efficient Physical Neural Networks

Magnus Saebo*, Tyler King*, Maxwell Anderson, Tatsuhiro Onodera, Peter McMahon

Under review McMahon Lab, Cornell

Physically constrained model compression and modular scaling for oscillator-based neural networks, reaching 3× parameter efficiency and better accuracy than other physical neural network approaches.