Research map / Large language models / Language, safety and interpretability

RIZZ: Routing Interactions to Near Zero-Interference Zones for Continual Adaptation of Black-Box Agents

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RIZZ, a verifier-gated, routed external-memory framework that lets a frozen black-box LLM agent continually adapt across task streams via dynamically spawned memory branches and a budget-aware prompt compiler.

Problem

Black-box LLM agents deployed over nonstationary task streams need to adapt without weight access, but single global prompts or memories suffer interference across task families, and rollout-heavy methods are costly.

Approach

Route each input to a dynamically spawned, branch-local memory zone (hierarchical judge labels with embedding snapping and Jaccard de-duplication), compile a bounded prompt from branch-local and global memory, and write back to memory only when a deterministic verifier approves the outcome.

Key results

• On TRACE, RIZZ reaches average reward 0.706 vs 0.637 for Frozen (+6.9 pp) at $12.79 cost.
• On LongMemEval-S, RIZZ reaches 0.484 average reward vs 0.202 for Frozen and 0.322 for AMem.
• On tau-Bench, ReAct + RIZZ reaches overall pass@1 0.400 vs 0.352 for ReAct + Frozen at roughly equal cost ($14.48 vs $14.57).

Surprising findings

• ACE's playbook grew beyond 130K tokens and overflowed Haiku's 200K context window on TRACE and LongMemEval-S, causing hard failures; the authors attribute this to advisory-only budget enforcement.
• On StreamBench, RIZZ underperforms Frozen on the DS-1000 subtask, which fragments into many low-volume code branches.

Limitations

• Gains depend on recurring local structure; sparse or fragmented streams (e.g., DS-1000) leave too little experience per branch.
• Frozen base model limits steering to capabilities the model can already express via prompting and context.
• Persistent memory raises privacy, security, and memory-poisoning risks that the framework does not itself address.

Experimental setup

Datasets: StreamBench TRACE LongMemEval-S tau-Bench

Benchmarks: StreamBench TRACE LongMemEval-S tau-Bench

Metrics: average reward pass@1 cost ($) tokens (M) wall-clock hours

Baselines: Frozen AMem ACE ReAct + Frozen

Builds on: Dynamic Cheatsheet ACE A-Mem Reflexion GEPA

Closest papers by meaning

• AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.88)
• PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.87)
• Revealing Behavioral Plasticity in Large Language Models: A Token-Conditional Perspective (ICML 2026 · similarity 0.87)
• MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents (ICLR 2026 · similarity 0.86)
• InduceKV: Fixed-Footprint Continual Adaptation of Multimodal LLMs via Inducing KV Memories (NeurIPS 2026 · similarity 0.86)
• SkillOpt: Executive Strategy for Self-Evolving Agent Skills (NeurIPS 2026 · similarity 0.86)