SPADER from Zhejiang and Peking University Tops All Four Multi-Answer QA Benchmarks, Accepted at EMNLP
jiqizhixin · x · 2026-10-26
SPADER, from Zhejiang University and Peking University, was accepted to EMNLP 2026 Main and ranks first in F1 across all four multi-answer QA benchmarks on both base models — up to +86.1% over PPO and GRPO on QAMPARI.
Multi-answer QA requires recovering the full answer set, not the first hit. The paper identifies two RL mismatches: trajectory-level advantage computation makes credit assignment too coarse, and F1 rewards treat head and long-tail entities equally, leaving agents no incentive to keep digging (e.g., searching once, rephrasing, getting the same three entities, and answering). SPADER refines GRPO's trajectory-level advantage to the step level via Step-wise Peer Advantage, pads early-terminated trajectories with zero future reward and masks them, and adds a Diversity-Aware Exploration Reward granting information gain and novelty premiums for new valid entities, jointly optimizing the policy through step-wise GRPO training.
More from Research
- Study: AI reviewing AI misses most reward hacking; best models miss 25%+ of hacks — ronbodkin · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- NLI alone rejects 5/6 new agent memories; a 0.45 similarity gate fixes it all — Alternative_Pin9598 · 2026-10-26
- AI folks, stop saying "biology" when you mean drug development — KevinKaichuang · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26
- Shanghai AI Lab-led team releases 100-hour egocentric wearable dataset for humanoid training — ChuckDBrooks · 2026-10-26