Polaris: A Godel Agent Framework for Small Language Models through Experience-Abstracted Policy Repair

Researchers have introduced Polaris, a Gödel agent framework that enables compact language models to repair their own decision-making policies through a structured cycle of failure analysis and minimal code patches, according to a paper posted on arXiv and accepted at ACL 2026 [1][3]. The framework, detailed by Aditya Namdev Kakade, Vivek Srivastava, and Shirish Karande, targets models with as few as 7 billion parameters [1][3]. Unlike response-level self-correction or full parameter tuning, Polaris makes policy-level changes with small, auditable patches that persist and are reused on unseen instances within each benchmark [1][5]. The agent engages in meta-reasoning: it explains its errors, proposes concrete revisions to its own policy, and then updates the policy [2][4]. A core innovation is experience abstraction, which distills failures into compact, reusable strategies that transfer to unseen instances [1][5]. Each cycle follows a defined sequence of reflection, abstraction, and repair, converting execution failures into validated code-level updates while preserving full traceability in the agent’s memory [2][5]. The system controls context growth by limiting the number of failed examples retained for reflection along with reduced tool-call history [5]. The paper, which appears in Findings of the Association for Computational Linguistics: ACL 2026, evaluates Polaris on four benchmarks: MGSM, DROP, GPQA, and LitBench, covering arithmetic reasoning, compositional inference, graduate-level problem solving, and creative writing evaluation [3][4]. A 7-billion-parameter model equipped with Polaris achieved consistent gains over the base policy and competitive baselines without supervision or retraining [1][3]. The submission record shows three versions posted to arXiv between March and June 2026, with file sizes shrinking from 1,342 KB to 788 KB across revisions [1]. The work was processed through the ACL ARR January 2026 review cycle under submission number 9247 and categorized under language models and continual learning [4].