A refusal is not the same as a good decision
An assistant that refuses every question avoids many bad answers, but it is not useful. An assistant that answers everything creates a different problem. WildGuard separates several questions that a single safety score can hide: is the prompt harmful, is the response harmful, and did the response refuse? These are related labels, but none can simply stand in for the others.
Three labels tell a clearer story
Suppose an assistant refuses a harmless request to explain bicycle maintenance. The response may contain no harmful information and still be an unnecessary refusal. Suppose instead that it gives ordinary bicycle advice. That is an answer rather than a refusal, but answering is not automatically unsafe. These invented examples show why counting refusals alone cannot measure moderation quality.
The contribution includes the test
The paper supplies a multitask moderation model and associated training and evaluation resources. A benchmark is an agreed collection of tasks and scoring rules. It makes comparisons easier to repeat, but its examples and definitions influence what the comparison means. WildGuard examines adversarial wording as well as more direct requests; its reported results remain tied to the tested data and models. For a small practice table, give each row separate columns for the request, the answer, the applicable rule and your expected judgment. Change only the answer in a second row. This makes it possible to see whether a safeguard distinguishes an unsafe request from a responsible response, rather than simply reacting to a sensitive word.
Where a strong benchmark can still mislead
Synthetic examples are generated or assembled rather than sampled directly from every real deployment. Their coverage can differ from ordinary conversations. Human judgments are also imperfect, particularly at ambiguous policy boundaries. A better result on this benchmark should therefore be reported with the task, data and evaluation conditions, rather than translated into a universal percentage of safe conversations.
Build a tiny evaluation sheet
Use benign public questions and harmless responses. Give each pair three separate columns: prompt category, response category and refusal status. Explain disagreements before calculating any accuracy. Your small sheet will not reproduce the original benchmark, but it can test whether a proposed comparison accidentally rewards excessive refusal. Preserve unknown labels rather than forcing every example into a confident decision.
Scientific scope
A refusal can be appropriate or excessive and is not itself a harmfulness label. Reported comparisons are benchmark- and model-specific; synthetic data and human judgments remain fallible.
Read the original work
Sources and reading locations for this explanation
- WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs
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
Seungju Han; Kavel Rao; Allyson Ettinger; Liwei Jiang; Bill Yuchen Lin; Nathan Lambert; Yejin Choi; Nouha Dziri
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.