AI
- ParetoSaddler: Pareto-Guided Optimization of Agent Harnesses for Quality and Cost — Separating candidate promotion from final deployment selection lifts agent-harness quality from 70.86% to 81.35% while cutting cost from $1.17 to $1.01/task on ClawMark, and from 56.22% to 62.89% quality while cutting cost from $0.78 to $0.45/task on GAIA2.
- “Shall We Approve?” Consequence-Aware Approval for Autonomous Coding Agents — Caper, which runs a proposed shell command in a sandbox and shows an approver its observed effects before committing or rolling it back, cuts the harmful-outcome rate to 6.7% (SABER) and 7.8% (the authors’ new YoloFS-Ext, where harm is hidden inside invoked tools) versus at least 14.8% and 35.3% for six existing approval systems.
- SpecPilot: Autonomous Algorithm Discovery and Systems Optimization for Speculative Decoding — An autonomous search loop over speculative-decoding configurations lifts single-request speedup over autoregressive decoding from a stock 4.46x to 10.87x, and batched speedup from 1.94x to 3.93x, with no training or quality loss, across four target models, three drafters, and two GPU platforms.
- Axion 1.0: Open Agent Foundation Models, Simplified — A fully open agent foundation model trained with only two SFT stages plus RL (no mid-training, DPO, or on-policy distillation) scores 52.2% (35B) and 59.0% (122B) on the restricted Humanity’s Last Exam, beating the similarly sized Agents-A1 (47.6%) by 4.6 points.