Add DIVERSED (AISTATS 2026) under Paradigms > Parallel

[kasakh]

Adds DIVERSED: Relaxed Speculative Decoding via Dynamic Ensemble Verification (AISTATS 2026).

• Paper: https://arxiv.org/abs/2604.07622
• Code: https://github.com/comeusr/diversed
• Project page: https://kasakh.github.io/relaxed-spec-decoding/

TL;DR. Standard speculative decoding rejects many plausible draft tokens because it enforces an exact distributional match with the target. DIVERSED introduces a learned ensemble verifier ν(x) = w·p_target(x) + (1−w)·q_draft(x) with a task- and context-dependent weight. The static ensemble exactly traces the Pareto frontier between acceptance rate and output fidelity (acceptance = 1 − w·TV(q,p)); dynamic per-token weighting then pushes beyond it. A single draft–target pair serves multiple latency–quality SLAs without retraining or extra weights.

Results. On Llama-3.1-8B / Llama-3.2-1B (and generalising to Qwen3 and Gemma-3): draft-token acceptance jumps 21–45% → 70–85%, delivering up to 3× end-to-end latency reduction with task quality preserved (ROUGE-2 9.46 → 12.11; math accuracy held at 67%; code Pass@1 held at 53%). At low temperature, acceptance reaches 89–91%.

Authors. Ziyi Wang, Siva Rajesh Kasa, Ankith M S, Santhosh Kumar Kasa, Jiaru Zou, Sumit Negi, Ruqi Zhang, Nan Jiang, Qifan Song.

Happy to adjust the placement or formatting if it's better suited to a different section — please let me know.