Zhejiang University and Peking University Propose SPADER for Multi-Answer QA

Researchers from Zhejiang University and Peking University have published SPADER, a reinforcement learning method for questions that have more than one correct answer. The work has been accepted to the main program of EMNLP 2026. The announcement from Jiqizhixin contains no benchmarks or tables yet, so only the idea and the problem setup can be assessed. The claimed gains in numbers will have to be checked against the paper itself.

The problem: the model names only the obvious

Multi-answer QA requires returning a complete set of answers: all films by a director, all countries that meet a condition, all drugs of a given class. Metrics here are usually computed over the whole set, and recall matters as much as precision. A standard RL pipeline with a correctness reward pushes the model toward a safe strategy. It names the two or three most common answers, collects the reward, and stops searching. Rare answers from the "long tail" stay out of the picture.

The authors put exactly this skew in the title: SPADER stands for Step-wise Peer Advantage with Diversity-Aware Exploration Rewards. The name contains two techniques.

What's inside SPADER

GRPO (Group Relative Policy Optimization) works without a separate critic model. For a single question, it generates a group of answers and computes each one's advantage relative to the others in the group. Judging by the name and the announcement's headline, SPADER makes this calculation step-wise: the comparison among "peers" (peer advantage) covers not only the final outcome but also the steps of reasoning or search. For a long chain, this gives a denser signal. The model can see at which step it found a new answer and at which it was spinning its wheels.

The second component is a reward that accounts for diversity (diversity-aware exploration reward). It rewards finds that differ from those already found, which pushes the policy to explore the tail of the distribution. The announcement does not disclose the exact formula, the way diversity is measured, or the weight of this reward.

What is known so far

• Authors: teams from Zhejiang University and Peking University
• Venue: EMNLP 2026, Main Conference
• Task: question-answering systems with multiple correct answers
• Algorithm: GRPO with step-wise advantage calculation between "neighboring" trajectories
• Reward: a bonus for diversity among the answers found
• Metrics, base models, and the size of the improvement are not stated in the available excerpt

Where this could be useful

Multi-answer tasks come up in enterprise search, in fact-gathering for analytics, in legal and medical reference work, and in RAG systems, where an incomplete list is worse than an honest "I don't know." If SPADER delivers a noticeable gain in recall without losing precision, the method is of interest to teams that fine-tune models with RL on their own data. GRPO has already become a working tool for this kind of fine-tuning, and the step-wise scheme does not require a separate critic.

Before adopting anything, check four things in the paper. First: which datasets and which models produced the result. Second: how the balance between recall and precision shifts, since a diversity reward can easily encourage inventing extra answers. Third: by what factor training costs grow because of the step-wise calculation. Fourth: whether the code is open. Until those details are in hand, SPADER remains an intriguing research idea, not a ready-made recipe.