RIZZ: Routing Interactions to Near Zero-Interference Zones for Continual Adaptation of Black-Box Agents
06:00 · June 23, 2026 · arXiv cs.AI RSS
Large language models are increasingly deployed as long-lived agents that must adapt across users, tasks, domains, modalities, and feedback regimes without access to model weights. Existing black-box adaptation methods typically optimize a single prompt, maintain an undifferentiated memory, or rely on repeated rollout-heavy search. However, these designs struggle when streams of input are nonstationary, feedback is sparse, and failures from one task family can contaminate behavior on another. We introduce RIZZ (Routing Interactions to Near Zero-interference Zones), a continual adaptation framework for compound language-model systems that learns entirely through verifier-gated memory, routing, and prompt compilation. RIZZ organizes input streams into dynamically spawned memory branches. At inference time, either while online or offline, a context-aware router selects or creates a branch that retrieves branch-local, global, graph-structured, and working-memory context, which is compiled into a bounded prompt together with retrieved task evidence. After the model acts, task verifiers score the output, and only verified interactions can update memory, promote reusable rules, demote harmful rules, or create anti-patterns. This yields a black-box agent that improves through persistent natural-language feedback while explicitly controlling interference. RIZZ targets the regime where adaptation must occur online under context budgets. Finally, we demonstrate the effectiveness of our framework against state-of-the-art baselines on competitive benchmarks.
Summary
RIZZ addresses the challenge of adapting black-box large language models that operate as persistent agents across shifting users, tasks, and domains. Because model weights remain inaccessible, adaptation must occur through external mechanisms rather than parameter updates. Standard approaches that rely on a single optimized prompt or a shared undifferentiated memory often fail when input streams are nonstationary and feedback arrives only after actions are taken, allowing experience from one task family to degrade performance on others.
The framework organizes incoming queries into dynamically created memory branches, each functioning as a localized store of verified examples, distilled procedural rules, and reward statistics. A context-aware router, operating at inference time, either selects an existing branch or spawns a new one by combining output-shape signals, embedding similarity, and label information. Relevant branch-local, global, graph-structured, and working-memory elements are then assembled with task evidence into a bounded prompt that respects context limits.
After the model produces a response, deterministic task verifiers evaluate the outcome. Only interactions that meet verification thresholds are permitted to update memory: successful traces reinforce reusable rules, while failures are retained as anti-patterns or demoted. This gated process enables the system to discover its own task organization without requiring ground-truth labels or a predefined taxonomy, while preventing unchecked growth in token usage.
Evaluations across streaming, sequential, conversational, and tool-use benchmarks show that the routed, verifier-controlled design improves accuracy over frozen baselines at modest additional cost and with substantially lower overhead than heavier memory-intensive alternatives.
Why it matters
This research is highly relevant for Dutch AI researchers and developers building compound LLM systems, as it offers a novel method for continuous agent adaptation using black-box models like commercial APIs. Its focus on controlled, verifier-gated memory aligns well with the EU's push for reliable, transparent, and predictable AI behavior.
