DDOR Cuts AI Overrefusals by 86% Without Touching Model Weights
DDOR borrows a classic software debugging algorithm to automatically find and fix the exact phrases that make LLMs refuse perfectly safe prompts
Takeaways
- What it is: DDOR is an automated framework that finds the exact phrase causing an LLM to refuse a safe prompt, then fixes it.
- Core technique: Applies delta debugging -- a classic software test-minimization algorithm -- to iteratively strip a prompt down to its minimal refusal-triggering fragment (mRTF).
- Test generation: Produces ~1,000 model-specific overrefusal test cases per model, achieving 19.3% higher overrefusal rates than OR-Bench and 5x more cases than naive rewriting.
- Prompt repair: mRTF-targeted rewriting reduces overrefusal by 51.96% on OR-Bench and 86.41% on XSTest while preserving semantic similarity better than full-prompt rewriting.
- Benchmark quality finding: DDOR's dual-model CoT filter flagged 43.29% of OR-Bench-Hard-1K prompts as actually unsafe -- a significant labeling flaw in the existing benchmark.
- Availability: Code is open-sourced at the project repo; paper is under review at ICLR 2026.
Every developer who has shipped an LLM-powered product has hit the same wall: a user asks something completely reasonable, and the model refuses. Ask about "how to kill a Python process" and some models will decline, pattern-matching on the word "kill" rather than understanding context. This is overrefusal -- and it turns out the research community has been borrowing a 20-year-old algorithm from software testing to finally crack it open.
DDOR (Delta Debugging for OverRefusal) is a new automated framework that does two things no prior tool could do together: it tells you exactly which phrase in a prompt is causing an unjustified refusal, and then it fixes it -- all without ever touching the model's weights or seeing inside its safety mechanisms.
The Safety Tax Is Real
While significant research focuses on mitigating harmful content generation, enhanced safety often comes with the side effect of over-refusal, where LLMs may reject innocuous prompts and become less helpful. This is sometimes called the "safety tax" -- the cost in usability paid for every unit of safety gained. The problem is not hypothetical. A comprehensive study measured the over-refusal of 32 popular LLMs across 8 model families, and the results showed a strong correlation: the safer a model is at rejecting truly toxic prompts, the more likely it is to also reject harmless ones.
Existing benchmarks like OR-Bench have tried to quantify this. OR-Bench comprises 80,000 over-refusal prompts across 10 common rejection categories, a subset of around 1,000 hard prompts that are challenging even for state-of-the-art LLMs, and an additional 600 toxic prompts to prevent indiscriminate responses. But these benchmarks have a critical blind spot: a systematic measurement is challenging due to the difficulty of crafting prompts that can elicit the over-refusal behaviors of LLMs. They also offer no explanation for why a refusal happens, and they are static -- they cannot adapt to the quirks of a specific model.
Borrowing from the Bug-Hunt Playbook
The core insight behind DDOR is to treat overrefusal like a software bug. Delta Debugging is a widely used family of algorithms to automatically minimize bug-triggering test inputs. It takes a test input as a list of elements, partitions the list into sets at different granularities, and systematically identifies and deletes partitions that are bug-irrelevant. In classic software testing, you feed it a 10,000-line file that crashes a compiler, and it spits back the 3 lines that actually cause the crash.