Continual-ARC benchmark: top coding agents fail to learn from experience, a fine-tuned 9B model doesn't
CodeByPoonam · x · 2026-11-12
Unify.ai released Continual-ARC, a new benchmark testing whether agents can memorize from experience and generalize. OpenClaw, Hermes Agent, and Prime Agent all failed to meaningfully learn from experience — none memorized anything of value or generalized to new tasks, instead re-reading the repo every session. In contrast, training a small 9B model's weights achieved both memorization and generalization. The result highlights a fundamental gap in current agent architectures around continual learning.
More from coding & agent
- United Airlines to Ban Personal AI Agents From Booking Flights for Customers — Polymarket · 2026-11-12
- Agentic Search explained: a 4-search loop costs about 5x one search — and queries leak to more companies — adnan_hashmi · 2026-11-12
- AWS says thousands of autonomous AI agents wrote every line of AWS Emulator — SumitGup · 2026-11-12
- freeCodeCamp launches free 2-hour course on agentic coding with Codex — adnan_hashmi · 2026-11-12
- A criminally underrated resource comparing tool-calling reliability across models and providers — ricklamers · 2026-11-12
- Organizing 55k data points is easy: enterprise data prep for agents is the hard part — BenBajarin · 2026-11-12