VISPA: Training-Free Pluralistic Alignment via Automatic Value Selection Lands EMNLP 2026 Oral
preslav_nakov · x · 2026-10-25
- VISPA (Waterloo, Melbourne, UIUC, MBZUAI, Macquarie) presented as an oral at EMNLP 2026 in Budapest.
- Addresses a key alignment gap: LLMs in high-stakes domains should reflect diverse perspectives, not averaged preference.
- Existing methods rely on prompt-level tweaks, cover few values, or need manual value selection and heavy compute.
- VISPA is training-free: it automatically selects scenario-relevant values and applies activation steering to generate value-conditioned responses.
- Supports three pluralistic alignment modes — Overton, Steerable, Distributional — with no finetuning needed to add new values.
- Strong results across multiple models and settings, including healthcare.
More from Research
- TensorFold's experimental Living Weights writes new facts into local model weights — GabGarrett · 2026-10-25
- Researcher argues AI tooling should filter junk papers so scientists focus on work worth their time — hugo_larochelle · 2026-10-25
- New agent evals field guide draws fire: only 15 of 313 citations predate 2023 — fredahshi · 2026-10-25
- ArtHOI accepted to TPAMI: turns AI-generated video into supervised-free 4D articulated human-object interaction data — liuziwei7 · 2026-10-25
- Lure-TTS: 600M voice models trained for under $200, running 4x faster than real time — smirkishere · 2026-10-25
- Math arXiv paper counts diverge by field: steady growth 2022-2025, then a shift in 2026 — IgorCarron · 2026-10-25