RESEARCH · RESEARCH · #3965
ARISE: Practical self-evolving memory for enterprise LLM agents (arXiv:2611.23588v1)
This paper introduces ARISE (Agent Refinement through Iterative Self-Evaluation), a memory architecture that lets an agent iteratively attribute its own failures and store micro-corrections which are consolidated into macro rules only if held-out accuracy improves, without updating model weights, tools, or workflow logic. Evaluated on 14 enterprise workflow domains across SOP-Bench and τ^2-bench with five models, ARISE-Teacher (using Claude Opus 4.6 for attribution/consolidation) improves accuracy by 10.9 percentage points on SOP-Bench and 5.8 on τ^2-bench versus baseline, outperforms GEPA by 4.5 pp on τ^2-bench and 1.4 pp on SOP-Bench, and reduces model calls and output tokens by 4.0% and 5.2% respectively at comparable per-task cost (–1.0%); ARISE-Self (the deployment model as attributor) still yields a 10.0 pp gain on SOP-Bench.
KEY POINTS
- This paper introduces ARISE (Agent Refinement through Iterative Self-Evaluation), a memory architecture that lets an agent iteratively attribute its own failures and store micro-corrections which are consolidated into macro rules only if held-out accuracy improves, without updating model weights, tools, or workflow logic.
- Evaluated on 14 enterprise workflow domains across SOP-Bench and τ^2-bench with five models, ARISE-Teacher (using Claude Opus 4.6 for attribution/consolidation) improves accuracy by 10.9 percentage points on SOP-Bench and 5.8 on τ^2-bench versus baseline, outperforms GEPA by 4.5 pp on τ^2-bench and 1.4 pp on SOP-Bench, and reduces model calls and output tokens by 4.0% and 5.2% respectively at comparable per-task cost (–1.0%); ARISE-Self (the deployment model as attributor) still yields a 10.0 pp gain on SOP-Bench.
- ARISE offers a practical mechanism to improve agent accuracy and reduce inference costs by evolving memory without retraining models or changing workflows, which is directly relevant to enterprise LLM deployments.
WHY IT MATTERS
ARISE offers a practical mechanism to improve agent accuracy and reduce inference costs by evolving memory without retraining models or changing workflows, which is directly relevant to enterprise LLM deployments.