LLM Cascade with Multi-Objective Optimal Consideration

Kai Zhang, Liqian Peng, Congchao Wang, Alec Go, Xiaozhong LiuSubmission to ICLR 2025 · Rejected

TL;DR

Standard LLM cascading frameworks route difficult queries from lightweight on-device models to cloud servers solely based on performance–cost trade-offs, often ignoring critical real-world constraints like data privacy. This paper introduces a multi-objective cascading framework that jointly evaluates generation quality and secondary objectives like sensitive data leakage. Through a specialized multi-head loss-tuning strategy, the local model learns when to defer queries to the server without exposing private user data.

What Problem Does This Paper Address?

Running large language models (LLMs) on resource-constrained consumer devices (e.g., smartphones) requires deploying smaller, weaker models (like Gemma-2B). To maintain output quality, systems use cascading: an on-device model first attempts the query, and if its deferral module predicts poor quality, it forwards the query to a powerful cloud-hosted server model (such as Gemma-7B).

Existing cascades make routing decisions strictly on cost versus task accuracy. In real-world applications, however, queries frequently contain personally identifiable information (PII) or protected health data. Blindly routing uncertain queries to a remote server risks severe privacy violations. The authors investigate how to design an LLM cascade that handles multi-objective decision-making—balancing downstream accuracy, routing cost, and privacy preservation.

Methods

The authors evaluate both training-free and training-based strategies to teach the local model $\Phi(L)$ cascade logic before handing off decisions to a confidence/logit-based deferral module $D(\cdot)$:

  1. Multi-Objective Prompting (Training-Free): Utilizes zero-shot and few-shot instructional prompts instructing the local model to perform the primary task while simultaneously evaluating confidence and privacy status.
  2. Multi-Objective Instruction Tuning: Fine-tunes the local model on step-by-step instructional prompts with ground-truth task answers and binary privacy annotations.
  3. Multi-Objective Loss Tuning (Novel Approach): Deconstructs the output into distinct tasks using multiple heads $\{h_1, \dots, h_i, h_L\}$ on the local model to separately generate task responses and predict objective criteria (e.g., privacy). The training loss dynamically combines:
  4. Binary cross-entropy for auxiliary objectives (e.g., privacy),
  5. Task prediction loss for the local model,
  6. A server distillation loss gated by a modified Heaviside step function $\alpha(t) = H(\text{logit}_{y^L}, t)$, which selectively pulls knowledge from the server model $\Phi(S)$ only when local logit confidence falls below a set threshold $t$.

The deferral module routes the prompt to the server model only if the local response's logit distribution indicates both low task confidence and the absence of privacy concerns.

          +-----------------------+
          |  User Query (x)       |
          +-----------+-----------+
                      |
                      v
          +-----------------------+
          | Local Model \Phi(L)   |
          | (Multi-Head/Tuned)    |
          +-----------+-----------+
                      |
           [Task Logits & Privacy Logits]
                      |
                      v
          +-----------------------+
          |    Deferral Module    |
          +-----------+-----------+
                     / \
   High Confidence  /   \  Low Confidence
   OR Private Data /     \ AND Safe to Send
                  v       v
         +-----------+  +------------------------+
         | Local     |  | Server Model \Phi(S)   |
         | Answer y^L|  | Generates Answer y^S   |
         +-----------+  +------------------------+

Datasets

The authors evaluated their framework across three core benchmarks:

  • GSM8K: Grade-school math word problems, modified following prior privacy research to embed personal identifiers (e.g., names, account details) to test privacy leakage alongside numerical reasoning.
  • MedQSum: A medical query summarization dataset containing genuine personal health information (PHI) subject to HIPAA regulations.
  • WMT22: A standard machine translation benchmark used as a control to measure cascading performance without explicit privacy constraints.

Privacy token extraction accuracy was verified against PII and HIPAA standards (achieving >99% extraction accuracy and validated with human verification).

Key Insights and Ideas

  • Prompting Fails on Small Models: Zero-shot and few-shot prompting fail to establish reliable multi-objective cascading on small models (e.g., Gemma-2B). The model either treats confidence as a static classification task or routes queries semi-randomly, leaking privacy tokens at high rates.
  • Loss Tuning Separates Decision Signals: Standard instruction tuning helps, but multi-head loss tuning provides a distinctly bimodal logit distribution. Correct local responses and private queries cluster at high logits, making threshold-based deferral exceptionally reliable.
  • The Cascade Can Beat the Server Alone: By combining the strengths of the local model on straightforward queries with selective routing on difficult queries, the cascaded system outperformed the standalone server model on all three benchmarks.

Key Results

  • Task Accuracy & Performance:
  • On GSM8K, the loss-tuned cascade achieved 55.92% accuracy at an 81.2% server call rate, surpassing the fine-tuned Gemma-7B server model's standalone performance (52.85%) by +3.07%.
  • On MedQSum, loss tuning reached 62.95 ROUGE-Sum (+1.73 over server) at a 97.3% call rate.
  • On WMT22, loss tuning achieved 39.69 ROUGE-Sum (+3.18 over server) at an 80.6% call rate.
  • Privacy Detection:
  • Zero-shot prompting failed entirely to detect privacy risks on GSM8K (0% precision and recall).
  • Multi-objective loss tuning achieved 91.79% precision and 87.24% recall on GSM8K privacy detection (and 90.10% precision / 82.99% recall on MedQSum), dramatically suppressing private token leaks compared to prompting.

Limitations and Future Work

  • Objective Scalability: The experimental evaluation was limited primarily to two concurrent objectives (task performance and binary privacy). Handling three or more complex, competing constraints remains unvalidated.
  • Binary Privacy Formulation: Privacy was modeled primarily as a binary classification problem based on token-level PII extraction, which may not capture contextual or implicit privacy nuances.
  • Future Directions: The authors plan to investigate how the number and complexity of simultaneous objectives affect local model capacity, as well as explore memory-based methods to maintain low computational overhead.

Summary updated

Abstract

Large Language Models (LLMs) have demonstrated exceptional capabilities in understanding and generating natural language. However, their high deployment costs often pose a barrier to practical applications, especially. Cascading local and server models offers a promising solution to this challenge. While existing studies on LLM cascades have primarily focused on the performance-cost trade-off, real-world scenarios often involve more complex requirements. This paper introduces a novel LLM Cascade strategy with Multi-Objective Optimization, enabling LLM cascades to consider additional objectives (e.g., privacy) and better align with the specific demands of real-world applications while maintaining their original cascading abilities. Extensive experiments on three benchmarks validate the effectiveness and superiority of our approach.
BibTeX
@misc{zhang2024llm,
  author = {Kai Zhang and Liqian Peng and Congchao Wang and Alec Go and Xiaozhong Liu},
  title = {{LLM Cascade with Multi-Objective Optimal Consideration}},
  year = {2024},
  eprint = {2410.08014},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.08014},
}
Figure