Value Realignment

A post-training pass — usually plain supervised fine-tuning on examples written to a chosen viewpoint — that changes which values a model expresses in its answers while leaving its general capabilities intact. The values turn out to sit in a shallow layer: a small amount of targeted training can swap them.

Nous Research demonstrated it on November 6, 2026 with Arete-27B: a supervised fine-tune of Alibaba’s Qwen3.8-27B on about 52.8 million tokens, two epochs, no reinforcement learning and no architectural change. On Nous’s own 803-prompt evaluation, answers adopting the Chinese state’s framing fell from 22% to 0.9% while the fact score rose from 2.32 to 2.98 out of 5 — and general capability barely moved, with its Hermes Bench score at 65.60% against 65.77% for the base model.

The result cuts both ways, and Nous says so itself. The same cheap pass that removes one viewpoint can install another: its own line is that “the values in an open model are one fine-tune away.” The measurement is also the lab’s own — it designed the prompts, the rubric and the judge, an AI model (Grok 4.7) rather than people — and the lab is explicit that Arete reflects Western values, not neutrality. A realigned model is a model with a chosen viewpoint, not a model without one.

The surprising part of Arete is how little training it took. Fifty-two million tokens is a rounding error against pretraining — Nous calls the cost “insignificant” — and the change was applied late and thin: all language-model weights fine-tuned, vision encoder frozen, no reinforcement learning, no preference optimization. The lab reads that as evidence the bias layer is shallow, and its teacher choice is part of the story: it picked Thinking Machines’ Inkling because it needed an open model whose terms permit this kind of tinkering, after finding even Nvidia’s Nemotron carried a “checkered record” on China-related topics.

Do not over-read the headline numbers. The evaluation is Nous’s own instrument: 803 prompts it wrote, required facts it extracted, a rubric it revised three times against a 25-prompt hand-graded pilot, and a single AI judge, Grok 4.7, scoring blind. Secondary coverage adds that most accepted training examples still carry a factual-error tag on their dataset card, and that Israel-related data was absent from the project. And the “thinking on” Arete row is provisional, from two of five runs. The eval is public, so anyone can rerun it — but nobody independent had, at publication.

Treating a realigned model as neutral or debiased. Realignment replaces one set of values with another; the eval only measures distance from the viewpoint it was built to remove.

• blogIntroducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values ↗Nous Research · 2026

  The launch post: a supervised fine-tune of Qwen3.8-27B on 52.8M tokens that cut state-framed answers from 22% to 0.9% on Nous’s own 803-prompt eval, with general capability flat. All figures are the lab’s own; the judge was Grok 4.7.

• reportNous releases Arete-27B with an 803-prompt alignment evaluation ↗AI Primer · 2026

  Independent read of the Arete release: the caveats the launch post does not stress — factual-error tags on most accepted training examples, Israel-related data absent, and dataset population figures that do not map cleanly onto the training run.