Type-Safe Is Not Error-Free: Constrained Decision Heads Follow Option Names, Not Their Rubrics
類型安全不等於無誤：約束決策模型偏向選項字面名稱而非綁定規則
While structured outputs guarantee schema compliance, they do not ensure semantic understanding. By swapping option names (e.g., changing 0/1 to no/yes) bound to specific rubrics across 1,200 workflow decisions, researchers observed a massive decision reversal: answer flip rates increased by 70.4% and AUC dropped from 0.94 to 0.23. The model prioritized the semantic polarity of option names over the defined rubric. Interestingly, using neutral or random character strings as option names successfully mitigates this bias without reducing accuracy.
Key points
Type-Safety is Not Logic-Safety
Even with a 0% type-error rate, the semantic correctness of the model's decisions can completely collapse based on naming.
Semantic Polarity Trap
Models heavily prioritize the literal meaning of option names over the detailed rubrics bound to them.
Impact of Read-Out Geometry
Models utilizing mean-pooling over the full option span flipped answers 4.1x less, indicating architecture plays a key role.
Random Names Prevent Bias
Replacing option names with semantically neutral random strings eliminates the bias without sacrificing performance.
Why it matters
As LLM agents and automated pipelines increasingly rely on structured outputs to interface with software, this study exposes a critical vulnerability. Even when code runs without type errors, the decision logic can be silently reversed by simple option naming. This has major safety implications for high-stakes decision systems in finance, law, and healthcare.
Who it affects
- AI Developer
- AI Researcher
- Product Manager
- Enterprise Leader
How to use it
- 1Designing LLM schemas using semantically neutral tags (e.g., A/B or random strings) instead of polar names like yes/no to prevent decision bias.
- 2Evaluating and robust-testing LLM comprehension of complex decision rubrics within RAG and agentic workflows.
Limitations & caveats
- The study focuses on Jev-like and specific open/hosted models; generalizability to other state-of-the-art frontier LLMs remains to be tested.
- While using random strings as option names mitigates the bias, it may reduce schema and prompt readability for developers during debugging.
Related
Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast Diffusion LLMs
Flash-dLLM：為擴散大語言模型打造的 I/O 感知 KV 快取與平行解碼加速框架
Flash-dLLM is a training-free inference acceleration framework that dramatically speeds up Diffusion LLMs (dLLMs) using an IO-aware fused KV cache and a self-contained draft-and-verify decoding strategy.