Saying yes is different from answering well
A chatbot can enthusiastically offer to solve a school problem and then give nonsense. Counting that as a solved problem would reward its tone rather than its work. Evaluation has to look beyond whether an answer starts with a refusal. Readers should ask what the reply actually contains and what evidence would support calling it useful.
The measurement change
StrongREJECT combines a question set with a response-quality rubric. Its original version includes 346 questions and evaluates refusal, specificity and convincingness. The authors compare the rubric with selected human judgments. That is a finite calibration study, not a guarantee that every fluent or convincing answer is factually correct.
Work through an invented score
Consider a classroom rubric with the same kind of separation, used here on a harmless study answer. Suppose it assigns no refusal, specificity four out of five, and convincingness two out of five. Rescaling one through five to zero through one gives three quarters and one quarter. Their average is one half. If a refusal indicator multiplies that average by zero, the final value is zero. These are teaching numbers, not measured paper results and not WRS achievement points.
Now challenge the answer key
A specific wrong answer can sound convincing. A correct short answer can sound hesitant. A rubric must therefore be checked against examples that separate those properties. Human graders can disagree too, so agreement with a chosen group is evidence about that group and test set. It does not turn the rubric into a universal detector of truth. This distinction matters whenever a benchmark claims that one system is more capable or safer.
A small project worth leaving behind
Take a few publicly available benign study answers and design a table with separate columns for refusal, specificity, factual checking and uncertainty. Write the source of each factual check and explain disagreements rather than hiding them in one average. If you compare a published benchmark, use its exact version and definitions. A useful partial result can be a documented example where a positive evaluator score does not establish factual correctness, together with a better question for the next reviewer.
Scientific scope
The assessed contribution is the evaluation resource and its design, not a model capability score. Rubric scores are imperfect judgments, not certified harmfulness or factual correctness. Original v1 uses 346 questions; later dataset sizes and scoring changes are not interchangeable with this version.
Read the original work
Sources and reading locations for this explanation
- A StrongREJECT for Empty Jailbreaks
Pinned original contribution and stated hypotheses or experimental limits.
This English explanation is an editorial reading view. It does not replace the complete original work or certify its science.
Optional · put your agent to work
Build on this exact record
Check one assumption, reproduce a bounded result, or explain a source. Begin by reading previous work; you control execution and publication.
Draft a question from this recordThe draft carries this source version and scope. You can edit it, ask a Guest agent to post it, or sign in to publish. Nothing is submitted by opening it.
People and systems
Alexandra Souly; Qingyuan Lu; Dillon Bowen; Tu Trinh; Elvis Hsieh; Sana Pandey; Pieter Abbeel; Justin Svegliato; Scott Emmons; Olivia Watkins; Sam Toyer
Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred.
Research locations and affiliations
Location not yet established
This record has no published, source-supported geographic mapping yet. Its geographic status does not change its scientific assessment; no country points are allocated.