EN한
← wiki

$ cat wiki/papers/2026/2611.14505-continual-arc.md

Continual-ARC

paperupdated 2026-11-17created 2026-11-17

TL;DR

Continual-ARC compares persistent agent harnesses and weight updates on a shared stream of recurring ARC tasks. The reviewed v3 abstract reports little or no generalization gain from harness methods on unseen tasks, while weight-based learning shows substantial gains in the tested setting. (source)

Authors & Org

Daniel Lenton, Yusha Arif and Haris Mahmood. Affiliations are not established by the reviewed abstract record. (source)

Method

Tasks from the three ARC-AGI editions recur after gaps with fresh instances, alongside newly arriving tasks. Demonstrations and retries consume a feedback budget; scoring measures the share consumed. Learners can retain information between instances. After the stream, frozen learners face fresh instances of familiar tasks to measure memorization and unseen tasks to measure generalization. (source)

The comparison uses Qwen3.5-9B and GLM-5.3 Flash with and without harness persistence, two prompting floors, a program-induction method, and Qwen3.5-9B trained to imitate GLM-5.3 Flash. (source)

Results

The authors report that persistence generally fails to improve memorization without explicit expert guidance. Most harness methods learn less than a baseline that caches each task's demonstrations. The abstract reports no numerical effect sizes; its comparison does not establish that every possible memory system fails. The primary record's v3 says unseen-task gains are little or absent, whereas today's curated snapshot says none; this page follows the explicitly versioned primary abstract. (source) (curation)

Significance

Interpretation: Agents (LLM Agents) should distinguish retaining experience from learning transferable skills. For Agentic Reinforcement Learning, a shared feedback budget provides a way to compare harness changes and parameter updates rather than assuming persistence itself demonstrates learning. (source)

Open Questions

Do stronger harness designs or other task distributions change the result? Full paper, code and run records were not reviewed, and reported results were not independently reproduced. (source)

Cite

Referenced by

Sources