Work
seri
2026A cross-platform coding-agent harness, open source, built from scratch.
Chainforge
2026A Go framework for agents. The core has no external dependencies, and it speaks MCP.
GRPO
2026Post-training a 1.7B model with GRPO on GSM8K. Accuracy moved from 66.9% to 87.3%.
Papers
Most Self-Improving LLM-Agent Papers Fail Two or More Evaluation Traps
202639 of 41 papers fail two or more known eval traps.
The Plurality Paradox
2026When disagreement between providers stops working as an alarm.
The Plausibility Trap
2026Past a reviewer's frontier, plausibility can make acceptance favor false claims. A pre-registered test did not find that reversal.